text
stringlengths 39
745k
|
|---|
---
abstract: |
Stock market prediction is one of the most attractive research topic since the successful prediction on the market’s future movement leads to significant profit. Traditional short term stock market predictions are usually based on the analysis of historical market data, such as stock prices, moving averages or daily returns. However, financial news also contains useful information on public companies and the market.
Existing methods in finance literature exploit sentiment signal features, which are limited by not considering factors such as events and the news context. We address this issue by leveraging deep neural models to extract rich semantic features from news text. In particular, a Bidirectional-LSTM are used to encode the news text and capture the context information, self attention mechanism are applied to distribute attention on most relative words, news and days. In terms of predicting directional changes in both Standard & Poor’s 500 index and individual companies stock price, we show that this technique is competitive with other state-of-the-art approaches, demonstrating the effectiveness of recent NLP technology advances for computational finance.
address: |
Department of Electrical and Computer Engineering\
Queen’s University, Canada\
Kingston, ON, Canada K7L 2N8
author:
- Huicheng Liu
bibliography:
- 'references.bib'
title: 'Leveraging Financial News for Stock Trend Prediction with Attention-Based Recurrent Neural Network'
---
Recurrent Neural Network,Stock Prediction,Attention Mechanism,S&P 500
Introduction
============
Stock market prediction is the act of trying to determine the future value of a company stock [@malkiel1985random]. Apparently, the successful prediction of a stock’s future price can yield significant profit, making the prediction problem an area of strong appeal for both academic researchers and industry practitioners. However, Stock market prediction is usually considered as one of the most challenging issues among time series predictions due to its noise and volatile features [@wang2012novel]. During the past decades, machine learning models, such as Support Vector Regression (SVR) [@basak2007support] and Support Vector Machines(SVM) [@hearst1998support], have been widely used to predict financial time series and gain high predictive accuracy [@refenes1994stock; @das2012support; @lu2009financial].
How to accurately predict stock movement is still an open question with respect to the economic and social organization of modern society. The well-known efficient-market hypothesis (EMH) [@malkiel2003efficient] suggests that stock prices reflect all currently available information and any price changes based on the newly revealed relevant information. However, due to the implicit correlations between daily events and their effect on the stock prices, finding relevant information that contribute to the change of price for each individual stock are difficult. Besides, the influences of events to stock prices can occur in indirect ways and act in chain reactions, which sets obstacles for precise market prediction. There are three approaches related to the information required to make a prediction. The first approach, technical analysis, is based on the premise that the future behavior of a financial time series is conditioned to its own past. Secondly, fundamental analysis, is based on external information as political and economic factors. A major source of information are text from the internet, these information are taken from unstructured data as news articles, financial reports or even microblogs. Nofsinger et al [@nofsinger2001impact] shows that in some cases, investors tend to buy after positive news resulting in a stress of buying and higher stock prices, they sell after negative news resulting in a decrease of prices. Finally the third approach considers as all relevant information coming from both, financial time series and textual data.
In this work, our goal is to leverage public released financial news and train a model named Attention-based LSTM (At-LSTM) to make prediction on directional changes for both Standard & Poor’s 500 index and individual companies stock price. Our model consists a Recurrent Neural network(RNN) to encode the news text and capture the context information, self attention mechanism is applied to distribute attention on most relative words, news and days . The model input are financial news titles extracted from Reuters and Bloomberg. Our model take advantages from the rapid development of deep neural networks and we show that our model is competitive with other state-of-the-art approaches, demonstrating the effectiveness of recent NLP technology advances for computational finance.
The rest of the paper are organized as follows. We introduce related work in Section 2. Followed by, we present some background and the methodologies for our proposed prediction model in Section 3. Experimental setup and results are demonstrated in Section 4. The paper finally concludes and point out future directions in Section 5.
Related work
============
Stock market prediction is an intriguing time-series learning problem in finance and economics, which has attracted a considerable amount of research. Efforts on predicting stock market have been carried out based on different resources and approaches. For example, one of the most widely studied approach relies on analyzing recent prices and volumes on the market [@schoneburg1990stock; @akgiray1989conditional; @goccken2016integrating; @adebiyi2014comparison; @kim2012simultaneous].
Analyzing stock market using relevant text is complicated but intriguing [@wu2009stock; @xie2013semantic; @ding2014using; @ding2015deep; @ding2016knowledge; @chang2016measuring; @peng2015leverage; @luss2015predicting; @skuza2015sentiment; @sehgal2007sops]. For instance, a model with the name Enalyst was introduced in Lavrenko et al. Their goal is to predict stock intraday price trends by analyzing news articles published in the homepage of YAHOO finance. Mittermayer and Knolmayer implemented several prototypes for predicting the short-term market reaction to news based on text mining techniques. Their model forecast 1-day trend of the five major companies indices. Wu et al. [@wu2009stock] predicted stock trends by selecting a representative set of bursty features (keywords) that have impact on individual stocks. Vivek Sehgal et al. [@sehgal2007sops] introduced a method to predict stock market using sentiment. Similarly, Michał Skuza et al. [@skuza2015sentiment] used sentiment from postings on twitter to predict future stock prices. However, these methods have many limitations including unveiling the rules that may govern the dynamics of the market which makes the prediction model incapable to catch the impact of recent trends.
More recently, neural networks have been leveraged to further improve the accuracy of prediction. In general, neural networks are able to learn dense representations of text, which have been shown to be effective on a wide range of NLP problems, given enough training sample. This is the case in stock market prediction where the prices of stocks are available together with a great collection of relevant text data. This provides a good setting for exploring deep learning-based models for stock price prediction. More specifically, dense representations can represent related sentiment, events, and factual information effectively, which can then be extremely challenging to represent using sparse indicator features.
Advance of deep learning models has inspired increasing efforts on stock market prediction by analyzing stock-related text. For example, Ding et al. [@ding2015deep] showed that deep learning representation of event structures yields better accuracy compared to discrete event features. They further augmented their approach [@ding2016knowledge] to incorporating an outside knowledge graph into the learning process for event embedding. As another example, Chang et al. [@chang2016measuring] used neural networks to directly learn representations for news abstracts, showing that it is effective for predicting the cumulative abnormal returns of public companies. Other work,e.g., [@kim2012simultaneous; @goccken2016integrating; @adebiyi2014comparison], has proposed different models of neural network that improve the prediction accuracy.
Methodology
===========
In this section, we first introduce some relative background on our prediction models. Followed by, we introduce the design of our proposed model to predict the directional movements of Standard & Poor’s 500 index and individual companies stock price using financial news titles. The model is named as Attention-based LSTM (At-LSTM) and shown in Fig \[fig:traditional-convolutional-network\].
Background
----------
### Bag of Words Model and Word Embedding
The Bag-of-Words model is a simplifying representation often used in Natural Language Processing and Information Retrieval [@Bag]. Also known as the vector space model. In this model, a text (such as a sentence or a document) is represented as a bag of its words, disregarding grammar and even word order but keeping multiplicity. The bag-of-words model is commonly used in document classification where the occurrence of each word represents a feature for training a classifier.
Word Embedding is the collective name for a set of language modeling and feature learning techniques in Natural Language Processing (NLP) where words or phrases from the vocabulary are mapped to vectors of real numbers [@Word]. Conceptually it involves a mathematical embedding from a space with one dimension per word to a continuous vector space with much lower dimension. Neural networks can be used to generate this mapping and the most common algorithm used are continuous bag of words(CBOW) and skip-gram algorithm [@mikolov2013efficient].
Word and phrase embeddings, when used as the underlying input representation, have been shown to boost the performance in NLP tasks such as syntactic parsing and sentiment analysis.
### Convolutional Neural Network
As shown in figure \[fig:traditional-convolutional-network\], Convolutional Neural Network (CNN) is a class of deep, feed-forward artificial neural networks that has successfully been applied to analyzing visual imagery and NLP related tasks. Same as neural network, CNN are made up of neurons that have learnable weights and biases. Each neuron receives some inputs, performs a dot product and optionally follows it with a non-linearity. A Convolutional layer is composed by four consecutive operations: Convolution, subsampling(pooling), activation and dropout. The convolution layer can help select features so that CNN requires minimal pre-processing compared to other deep learning models. The core part of the Convolutional neural network is the convolution filter.
Convolution filters leverage two important ideas that can help improve a machine learning system: sparse interaction and parameter sharing. Sparse interaction contrasts with traditional neural networks where each output is interactive with each input. As a result, the output is only interactive with a narrow window of the input. Parameter sharing refers to reusing the filter parameters in the convolution operations, while the element in the weight matrix of traditional neural networks are used only once to calculate the output.
### Long-Short Term Memory
Long Short Term Memory (LSTM) networks are a special kind of RNN, proposed by Hochreiter in 1997 [@hochreiter1997long]. LSTM units are a building unit for layers of a Recurrent Neural Network(RNN) [@LSTM]. A RNN composed of LSTM units are often called an LSTM network. LSTM are explicitly designed to avoid the long-term dependency problem. A common LSTM unit is composed of a cell, an input gate, an output gate and a forget gate. The cell is responsible for “remembering” values over arbitrary time intervals; hence LSTM is capable of learning long-term dependencies.
(ct) [$c_t$]{}; (int1) ; (x1) [$\times$]{}; (ht) [$h_t$]{}; (x2) [$\times$]{}; (int2) ; (x3) [$\times$]{}; (ft) [$f_t$]{}; (it) [$i_t$]{}; (ot) [$o_t$]{};
i/in [int2/x2, x2/ct, ct/int1, int1/x1, x1/ht, it/x2, ct/it, ct/ot, ot/x1, ft/x3]{} (i)–();
(ct) to\[bend right=45\] (ft);
(ct) to\[bend right=30\] (x3); (x3) to\[bend right=30\] (ct);
(fit) ;
(fit.west|-int2) coordinate (aux)–++(180:7mm) node\[left\][$x_t$]{}; (\[yshift=1mm\]aux)–++(135:7mm); (\[yshift=-1mm\]aux)–++(-135:7mm);
(fit.north-|it) coordinate (aux)–++(90:7mm) node\[above\][$x_t$]{}; (\[xshift=1mm\]aux)–++(45:7mm); (\[xshift=-1mm\]aux)–++(135:7mm);
(fit.north-|ot) coordinate (aux)–++(90:7mm) node\[above\][$x_t$]{}; (\[xshift=1mm\]aux)–++(45:7mm); (\[xshift=-1mm\]aux)–++(135:7mm);
(fit.south-|ft) coordinate (aux)–++(-90:7mm) node\[below\][$x_t$]{}; (\[xshift=1mm\]aux)–++(-45:7mm); (\[xshift=-1mm\]aux)–++(-135:7mm);
The expression long-short term refers to the fact that LSTM is a model for the short-term memory which can last for a long period of time. An LSTM is well-suited to classify, process and predict time series given time lags of unknown size and duration between important events. Besides, LSTM shows the promising result in sentence encoding in many NLP applications [@sutskever2014sequence; @gers2002learning]. The LSTM archticture are shown in figure \[fig:traditional LSTM neural network\], the computations of LSTM cells are:
[1.0]{} $$\begin{aligned}
f_t = \sigma(W_f[h_{t-1},x_t] + b_f)\end{aligned}$$ $$\begin{aligned}
i_t = \sigma(W_i[h_{t-1},x_t] + b_i)\end{aligned}$$ $$\begin{aligned}
\tilde{C}_t = tanh(W_C[h_{t-1}, x_t] + b_C)\end{aligned}$$ $$\begin{aligned}
C_t = f_t \otimes C_{t-1} + i_t \otimes \tilde{C}_t\end{aligned}$$ $$\begin{aligned}
o_t = \sigma(W_o[h_{t-1},x_t] + b_o)\end{aligned}$$ $$\begin{aligned}
h_t = o_t \otimes tanh(C_t)\end{aligned}$$
The forget gate is described in Equation (1) and is used to decide whether a value should be reserved. The Input gate in Equation (2) is used to control which values should be updated through a certain time step. Here, a new value $\tilde{C}_t$ is created as per Equation (3) using the $tanh$ function. Next, with Equation (4), the cell state will be updated from $C_{t-1}$ to $C_t$, $f_t$ and $i_t$ are used here to decide whether an information should be discarded. in Equation (5), an output gate is used to filter out redundant values. The final output is calculated by Equation (6). $W$ is the weight matrix and $b$ refers to the bias.
### Optimization Algorithm and Loss Function
In this section, we will briefly introduce the Optimization algorithm and loss function we used in our prediction model.
- ***Adadelta Optimization Algorithm***: Apart from Stochastic Gradient Descent, Adaptive Gradient Algorithm or the famous Adam algorithm [@kingma2014adam], we chose Adadelta [@zeiler2012adadelta] as our optimization algorithm. Adadelta is an optimization algorithm that can be used to update network weights iterative based on training data. Adadelta combines the advantages of two other extensions of stochastic gradient descent, Adaptive Gradient Algorithm and Root Mean Square Propagation Algorithm. The method dynamically adapts over time using only first order information and has minimal computational overhead beyond vanilla Stochastic Gradient Descent [@zeiler2012adadelta]. The method requires no manual tuning of a learning rate and appears robust to noisy gradient information, different model architecture choices, various data modalities and selection of hyperparameters.
- ***Cross Entropy Loss***: In information theory, the cross entropy between two probability distributions $p$ and $q$ over the same underlying set of events measures the average number of bits needed to identify an event drawn from the set [@Cross]. Cross entropy is the most suitable loss function since we want to measure the output from our prediction model with it’s true output. A binary cross entropy loss formula are shown in Equation (7). In which, $J(w)$ refers to the loss, $N$ refers to the number of examples, $y_n$ is the expected output and $\tilde{y}_n$ is true output. We use this formula as our loss function in the model. $$\begin{aligned}
J(w)=-\frac{1}{N}H(p_n,q_n)=-\frac{1}{N}\displaystyle\sum_{i=1}^{N} [y_n*\log\tilde{y}_n+(1-y_n)*\log(1-\tilde{y}_n)]\end{aligned}$$
Model design
------------
In this subsection, we introduce the design of our Attention-based LSTM model (At-LSTM). The design has four stages: Input and Embedding layer, news-level Bi-LSTM and self-attention layer, day-level Bi-LSTM and self-attention layer and the final output and prediction layer. These stages will be described below and shown in figure \[fig:model\].
![\[fig:model\]Model Structure](Model.png){width="100.00000%" height="3.5in"}
### Input and Embedding Layer
***Model Input***: The same world event can be described in different expression ways. This variety makes the representation space for events very sparse. Several approaches [@ding2015deep; @ding2016knowledge] represent the event as tuple $< S; V; O >$ to gain generalization. However, we argue that this representation is oversimplified and might lose lots of valuable information. Instead, we use the entire news title content as our models input and use LSTM-based encoder to encode it to a distributional representation to tackle with sparsity problem.\
***Word and Character Embedding***: For each input news, we remove the punctuation and then use a word embedding layer to calculate the embedded vectors for each word. The embedding layer takes a sequence of sentences as input, this sequence corresponds to a set of titles of news articles. These embedding are unique vectors of continuous values with length $w = (w_1,...,w_l)$ and $w_i\in R^{m}$ for each word in the training corpus, $m$ is the word level embedding dimension.
Existing pre-trained word embedding such as Glove [@pennington2014glove] and Word2Vec [@mikolov2013efficient] typically comes from general domains( Google News or Wikipedia, etc). However, these word embeddings often fail to capture rich domain specific vocabularies. We therefore train our own word embedding with financial domain news text consisting of news articles from Reuters an Bloomberg. The data are further described in Section 4.
In addition, we leverage character composition from Chen et al. [@chen2017recurrent] and concatenate the character level composition with the original word embedding to gain rich represent for each word. The character composition feeds all characters of each word into a Convolutional Neural Network (CNN) with max-pooling [@kim2014convolutional] to obtain representations $c = (c_1,...,c_l)$ and $c_n\in R^{n}$ for each word in the training corpus, $n$ is the character composition dimension. Finally, each word is represented as a concatenation of word-level embedding and character-composition vector $e_i = [w_i;c_i]$. A matrices $e^s\in R^{k*(m+n)}$ can be used to represent a news after the embedding layer, where $k$ is the length of the news.
### News Level Bi-LSTM and Self-Attention Layer
***Bi-LSTM Encoding***: After the embedding layer, we fed the words and their context in the news title into a Bi-LSTM based sentence encoder to perform distributional representation. Bidirectional LSTM (Bi-LSTM) is a variant of LSTM which shows better result then uni-direction LSTM in recent NLP tasks as they can understand context better. A bidirectional LSTM runs a forward and backward LSTM on a sequence starting from the left and the right end, respectively [@chen2017recurrent]. In this case, Bi-LSTM can not only preserve information from the past, but also catch the information from the future. We obtain the hidden vectors($\overrightarrow{h_i}$ and $\overleftarrow{h_i}$ shown in Equation (8) and (9)) from the sentence encoders and concatenate them to $H^t=[\overrightarrow{h_1};\overleftarrow{h_1},....,\overrightarrow{h_m};\overleftarrow{h_m}]$, $H_i^t\in R^{2u}$ represents the $ith$ news title after encoding in date $t$ and $m$ refers to the sequence number.
[1.0]{} $$\begin{aligned}
\overrightarrow{h_n} = \overrightarrow{LSTM}(e^s)\end{aligned}$$ $$\begin{aligned}
\overleftarrow{h_n} = \overleftarrow{LSTM}(e^s)\end{aligned}$$
***Word level Self-attention layer***: Instead of taking the average of the hidden vector after the sentence encoding, we leverage multi-hop self-attention mechanism [@lin2017structured] on top of the Bi-LSTM layer. The attention mechanism takes the whole LSTM hidden states $H^t_i$ as input, and outputs a vector of weights $A$:
$$\begin{aligned}
A = softmax(W_2tanh(W_1H_i^{t\intercal}))\end{aligned}$$
Shown in figure \[fig:attention\], here $W_1$ is a weight matrix with a shape of $W_1\in R^{d_a*2u}$ which u refers to the hidden unit of the news level Bi-LSTM. and $W_2$ is a vector of parameters with size $W_2\in R^{r*d_a}$, $d_a$ and $r$ are hyper parameter that can be set arbitrarily. $H_i^t$ are sized $H_i^t\in R^{n*2u}$, and the annotation vector $A$ will have a size $A\in R^{r*n}$, the $softmax()$ ensures all the computed weights sum up to 1. Then we sum up the LSTM hidden states $H_i^t$ according to the weight provided by $A$ to get a vector representation $N_i^t$ for the input sentence. We can deem Equation (10) as a 2-layer MLP without bias, whose hidden units numbers are $d_a$ and parameters are ${W_1,W_2}$. We compute the $r$ weighted sums by multiplying the annotation matrix $A$ and LSTM hidden states $H_i^t$:
![\[fig:attention\]Self attention mechanism](attention.PNG){width="100.00000%" height="4.0in"}
$$\begin{aligned}
N_i^t = AH_i^t\end{aligned}$$
Eventually, the sentence encoding vector $H_i^t$ then becomes a matrix $N_i^t\in R^{r*2u}$ and we use $N_i^t$ to represent the $ith$ news title in date $t$ after encoding. By using multi-hop attention mechanism, the vector representation usually focuses on a specific component of the sentence, like a special set of related words or phrases. Therefore, it is expected to reflect an aspect, or component of the semantics in a sentence instead of adding attention on a specific word.
Shown in equation (12), we apply another MLP layer on top of our self-attention layer to learn which attention group should be rewarded with the highest assign value. We name this as attention-over-attention, the weight matrix have a shape of $W_3\in R^r$, and the final representation of the sentence encoding vector $N_i^t$ are shaped as $N_i^t\in R^{2u}$. At last, we use $N_i^t$to represent a news title encoded from the input.
$$\begin{aligned}
N_i^t = tanh(W_3N_i^t+b_1)\end{aligned}$$
***News level Self-attention layer***: Not all news contributes equally to predicting the stock trend. Hence, in order to reward the news that offers critical information, we apply the same structure multi-hop self-attention on top of the encoding layer to aggregate the news weighted by an assigned attention value. Specifically:
[1.0]{} $$\begin{aligned}
A = softmax(W_5tanh(W_4N^{t\intercal}))\end{aligned}$$ $$\begin{aligned}
D_t = AN^t\end{aligned}$$ $$\begin{aligned}
D_t = tanh(W_6D_t+b_2)\end{aligned}$$
Here, $N^t=(N_1^t,....,N_m^t)$ and $N^t\in R^{m*2u}$, in which $m$ refers to the number of news in date $t$. Note that the weights $[{W_4,W_5,W_6,b_2}]$ in the news level attention layer are different from the word level attention layer weight $[{W_1,W_2,W_3,b_1}]$. A vector $D_t$ represents the temporal sequence for all the news proposed in the date $t$. The merit of using multi-hop self-attention mechanism is that it learns and assign different groups of attention value to the news encoding. Formally, the first group of attention reward the news that contains positive sentiments to the stock market(“raise”, “growth” or “decrease”, “down” etc). Whereas the second group of attention assign there reward to the news that mentions the major companies in the S&P 500 (“Microsoft”,“Google” instead of a small company outside of the S&P 500).
Obviously, the attention layer can be trained end-to-end and thus gradually learn to assign more attention to the reliable and informative news based on its content.
### Day level Bi-LSTM and self-attention layer
[1.0]{} $$\begin{aligned}
\overrightarrow{h_i} = \overrightarrow{LSTM}(D_t)\end{aligned}$$ $$\begin{aligned}
\overleftarrow{h_i} = \overleftarrow{LSTM}(D_t)\end{aligned}$$
We adopt day level Bi-LSTM to encode the temporal sequence of corpus vectors $D_i$, $t\in [1,N]$. Shown in Equation (16) and (17), We obtain the hidden vectors($\overrightarrow{h_i}$ and $\overleftarrow{h_i}$) from the day-level Bi-LSTM and concatenate them to $H_i=[\overrightarrow{h_1};\overleftarrow{h_1},....,\overrightarrow{h_N};\overleftarrow{h_N}]$, $H_i$ represents a vector that encodes the temporal sequence $D_t$ where $N$ refers to the sequence number. Since the news published at different dates contribute to the stock trend unequally, we adopt self-attention mechanism again to reward the dates that contribute most to the stock trend prediction, Shown in Equation below:
[1.0]{} $$\begin{aligned}
A = softmax(W_8tanh(W_7H_i^\intercal))\end{aligned}$$ $$\begin{aligned}
V = AD\end{aligned}$$ $$\begin{aligned}
V = tanh(W_9V+b_3)\end{aligned}$$
In the formula, $D=(D_1,....,D_t)$ and $D\in R^{N*2v}$ , $V\in R^{2v}$ represents the final vector for all news proposed before the prediction date $t+1$ in a delay window with size $N$, where $v$ is the hidden unit number in the day level Bi-LSTM. Note that the weight matrix $[{W_7,W_8,W_9,b_3}]$ in the day level attention layer are different from the weight matrices mentioned in the previous section.
### Output and Prediction Layer
The last stage of our At-LSTM model is a traditional fully connected layer with softmax as activation function whose output is the probability distribution over labels. In this work, the objective is to forecast the direction of daily price movements of the stock price, this direction are used to create a binary class label where a label \[1,0\] represents that the stock price will increase and label \[0,1\] represents that the stock price will decrease.
Experiments
===========
Experimental Setup
------------------
### Data
We evaluated our model on a data set of financial news collected from Reuters and Bloomberg over the time period from October 2006 to November 2013. This data set was made publicly available by Ding et al. [@ding2014using] and shown in table \[tab:Data set\]. We further collected data from Reuters for 473 companies listed in the Standard & Poor’s 500 over the time period started from November 2013 to march 2018. Meanwhile, the historical stock price data from October 2006 to March 2018 for all individual shares in Standard & Poor’s 500 are collected from Yahoo Finance. The second part of the data are used for individual stock price prediction and shown in table \[tab:individual\]. Due to the page limit, We only show the major companies listed in the S&P 500, this will also be applied in the result section.
[ |p[3cm]{}||p[3cm]{}|p[3cm]{}|p[3cm]{}| ]{}\
Data set & Training & Development & Testing\
Time interval & 20/10/2006-27/06/2012 & 28/06/2012-13/03/2013& 14/03/2013-20/11/2013\
News& 445,262 & 55,658 & 55,658\
[ |p[3cm]{}||p[3cm]{}|p[3cm]{}|p[3cm]{}| ]{}\
Company Symbol & Training news& Development news& Testing news\
GOOG & 5744 & 718 & 718\
AMZN & 3245 & 406 & 405\
CSCO & 2471 & 309 & 308\
MSFT & 4056 & 406 & 405\
AAPL & 9720 & 1215 & 1214\
INTC & 4355 & 545 & 544\
IBM & 2016 & 252 & 252\
AMD & 1224 & 153 & 153\
NVDA & 1440 & 168 & 167\
QCOM & 924 & 116 & 115\
WMT& 2793 & 350 & 349\
T& 1228 & 154 & 153\
Following Ding et al. [@ding2014using] we focus on the news headlines instead of the full content of the news articles for prediction since they found it produced better results. We use news articles on each day and predict if the S&P 500 index closing price movement (increase or decrease) in the day $t+1$ compared with the closing price on day $t$.
### Implementation Details
As mentioned in the previous section, We pre-trained 100 dimentional word embedding [@mikolov2013efficient] with skip-gram algorithm on the data set shown in table \[tab:Data set\], the size of the trained vocabulary is 153,214. In addition, firm names and an UNK token to represent any words out of the vocabulary are added to the vocabulary set, having an initial embedding initialized randomly with Gaussian samples. The word embedding are fine-tuned during model training. The character embedding has 15 dimensions, and CNN filters length are \[1,3,5\] respectively, each of those are 32 dimensions. The news level Bi-LSTM and day-level Bi-LSTM both have 300 hidden units, the day level LSTM window size $N$ has been set to 7. Hyper-parameters $d_a$ and $r$ in the self attention layer are set to 600 and 10, respectively. Mentioned in section 3, we use Adadelta for our optimization algorithm, the initial learning rate has been set to 0.04. Our model are trained for 200 Epoch.
Base Lines and Proposed Model
-----------------------------
In this subsection, we propose a few baselines to compare with our proposed model. For the sake of simplicity, the following notation identifies each model:
- ***SVM***: Luss and d’Aspremont et al. [@luss2015predicting] propose using Bags-of-Words to represent news documents, and constructed the prediction model using Support Vector Machines (SVMs).
- ***Bag-At-LSTM***: At-LSTM without sentence encoder. We take the average of the word embedding inputs instead of using Bi-LSTM to encode the news title.
- ***WEB-At-LSTM***: Same as our proposed model but without the character level composition.
- ***Ab-At-LSTM***: Instead of the news title, we use the news abstract as input for our model, the model structure remains the same.
- ***Doc-At-LSTM***: We further leverage a Hierarchical Attention Networks proposed by Yang et al. [@yang2016hierarchical] to represent the entire news Document, this model adds a sentence level attention layer into our proposed model to differentiate more and less important content when constructing the document representation.
- ***Tech-At-LSTM***: We concatenate seven Technical indicator leveraged from Zhai et al. [@zhai2007combining] shown in figure \[fig:tech\] with the vector $V$ after the day level LSTM layer and fed together into the prediction layer.
![\[fig:tech\]Self attention mechanism](tech.PNG){width="100.00000%" height="3.0in"}
- ***CNN-LSTM***: We use CNN instead of the news level self-attention layer. Note that the word and day level self-attention layer still remains the same. We want to see how well the self-attention layer works compared to CNN which is good at capturing local and semantic information from texts.
- ***E-NN***: Ding et al. [@ding2014using] reported a system that uses structure event tuples input and standard neural network prediction model.
- ***EB-CNN***: Ding et al. [@ding2015deep] proposed a model using event embedding input and Convolutional Neural Network as prediction model.
- ***KGEB-CNN***: Ding et al. [@ding2016knowledge] further incorporated an outside knowledge graph into the learning process for event embedding. The model structure is the same as EB-CNN.
Result and Discussion
---------------------
Shown in table \[tab:Result\], the results on the comparison between the models SVM and the rest of the models indicates that deep neural network model achieves better performance than the SVM model. Comparison between Bag-At-LSTM and At-LSTM demonstrates that sentence encoding with LSTM have slightly better result than Bag-of-words model. Furthermore, WEB-At-LSTM and At-LSTM indicates that the character level composition helps improved the models accuracy. The technical Indicator leveraged from Zhai et al. [@zhai2007combining] doesn’t show any performance improvement to our model. In contrast, it results in a decline on the accuracy and it might be caused by adding noise to the dense representation vector $V$ before the output and prediction layer. The comparison between CNN-LSTM and At-LSTM shows that the news level self-attention layer can help capture more relevant news titles and their temporal features. As Ding et al. concluded in [@ding2014using], the news titles might contain more useful information where as the abstract or article might cause some negative effect to the model. The Ab-At-LSTM and Doc-At-LSTM confirmed this viewpoint since their accuracy are lower than the proposed model that only used the information from the title.
Our proposed model has a 65.53% max accuracy and a average accuracy of 63.06% which is lower than the KGEB-CNN proposed by Ding et al., this is likely due to Knowledge graph event embedding (KGEB) is a more powerful method for model the content in news titles than the sequence embedding shown in this work.
[|p[3cm]{}||p[3cm]{}|p[3cm]{}|]{}\
Model & Average Accuracy & Max Accuracy\
SVM & 56.38% & –\
Bag-At-LSTM & 61.93% & 63.06%\
WEB-At-LSTM & 62.51% & 64.42%\
Ab-At-LSTM & 60.6% & 61.93%\
Doc-At-LSTM & 59.96% & 60.6%\
Tech-At-LSTM & 62.51% & 64.42%\
CNN-LSTM & 61.36% & 63.06%\
E-NN & 58.83% & –\
EB-CNN & 64.21% & –\
KGEB-CNN & **66.93**% & –\
At-LSTM & 63.06% & **65.53**%\
We use the At-LSTM model for individual stock price prediction after confirmed that it out performs other approaches. The results are shown in table \[tab:indiResult\], each stock shown in the table has more than 66% accuracy and the company WALMART has an average accuracy of 70.36% and max accuracy of 72.06%. Apparently, predicting individual stock price leads to higher accuracy than predicting S&P 500 index, and this is mainly due to the news article we used for input. In terms of the individual stock prediction, the news article we used are more relative to it’s corresponding company. In contrast, we used the full corpus as input for S&P 500 index prediction and certainly this adds noise to our model and hence effects the accuracy.
[|p[3cm]{}||p[3cm]{}|p[3cm]{}|]{}\
Company & Average Accuracy & Max Accuracy\
GOOG & 68.75% & 71.25%\
AMZN & 67.32% & 69.46%\
CSCO & 66.82% & 67.62%\
MSFT & 67.92% & 69.89%\
AAPL & 67.52% & 69.42%\
INTC & 67.12% & 67.63%\
IBM & 69.49% & 71.41%\
AMD & 66.12% & 69.10%\
NVDA & 69.35% & 70.51%\
QCOM & 68.53% & 69.70%\
WMT& **70.36**% & **72.06**%\
T& 68.53% & 69.70%\
Conclusion
==========
This paper has been motivated by the successes of Deep learning methods in Natural Language Processing task. we proposed a Attention-based LSTM model(At-LSTM) to predict the directional movements of Standard & Poor’s 500 index and individual companies stock price using financial news titles. Experimental results suggests that our model is promising and competitive with the state-of-the-art model which incorporate knowledge graph into the learning process of event embeddings [@ding2016knowledge].
There are some directions in our future work. While previous work and our result has found that including the body text of the news performs worse than just the headline, there may be useful information to extract from the body text, other directions include looking at predicting price movements at a range of time horizons, in order to gauge empirically how quickly information are absorbed in the market, and relate this to the finance literature on the topic. The financial time series are known by its volatility, in many cases small changes in the series that can be interpreted as noise. Moreover, the elimination of small variations makes the model focus only on news with significant variation on prices which might lead to accuracy increase.
|
---
abstract: 'This paper presents detailed results of neutron imaging of argon bubble flows in a rectangular liquid gallium vessel with and without the application of external horizontal magnetic field. The developed image processing algorithm is presented and its capability to extract physical information from images of low signal-to-noise ratio is demonstrated. Bubble parameters, velocity components, trajectories and relevant statistics were computed and analysed. A simpler version of the code was applied to the output of computational fluid dynamics simulations that reproduced the experiment. This work serves to further validate the neutron radiography as a suitable method for monitoring gas bubble flow in liquid metals, as well as to outline procedures that might help others to extract data from neutron radiography images with a low signal-to-noise ratio resulting from high frame rate acquisitions required to resolve rapid bubble motion.'
author:
- |
Mihails Birjukovs\
Institute of Numerical Modelling\
University of Latvia\
Riga, Latvia, Jelgavas 3, 1004\
`mihails.birjukovs@lu.lv`\
Valters Dzelme\
Institute of Numerical Modelling\
University of Latvia\
Riga, Latvia, Jelgavas 3, 1004\
`valters.dzelme@lu.lv`\
Andris Jakovics\
Institute of Numerical Modelling\
University of Latvia\
Riga, Latvia, Jelgavas 3, 1004\
`andris.jakovics@lu.lv`\
Knud Thomsen\
Research with Neutrons and Muons\
Paul Scherrer Institut\
Villigen, Switzerland, Forschungsstrasse 111, 5232\
`knud.thomsen@psi.ch`\
Pavel Trtik\
Research with Neutrons and Muons\
Paul Scherrer Institut\
Villigen, Switzerland, Forschungsstrasse 111, 5232\
`pavel.trtik@psi.ch`\
title: Dynamic neutron imaging of argon bubble flow in liquid gallium in external magnetic field
---
Introduction
============
Gas bubble flow through liquid metal in presence of static external magnetic field is of great interest, because this type of flow occurs during liquid metal stirring, purification, homogenization and crystallization processes, as well as in liquid metal column reactors [@r1; @r2]. Flow parameters must be tailored to each process, which requires the means for reliable control. To this end, application of static magnetic field has been proposed. Theoretical considerations and experimental evidence indicate that, depending on field orientation, one could achieve bubble jet stabilization and/or alter bubble velocity [@r3; @r4].
Optimization of industrial processes requires both simulations and suitable experimental methods. Neutron radiography is, in the context of high frame rate flow monitoring, a relatively new and promising technique that can probe optically opaque liquid metals directly [@r5] in the sense that, in addition to qualitative observations and image-based velocimetry, it enables the study of bubble shape dynamics and characteristic oscillation frequencies, which are important comparison criteria for verification of computational fluid dynamics (CFD) models. For some higher-Z metals, neutron imaging, being conceptually similar to X-ray imaging, is more appropriate, enabling one to work with thicker (i.e. more representative) samples [@r5; @r6]. This is especially important for suppressing the influence of bubble-wall interactions. Thus, neutron radiography enables one to probe a broader class of multiphase flow systems [@r7; @r8], which motivates further validation of this approach as a reliable technique for verification of simulation results and for multiphase flow analysis.
To date, several noteworthy studies employing neutron radiography have been conducted, but most were focused on single phase flow analysis using tracer particles [@r5; @r6], solidification dynamics [@r6], or performed only qualitative analysis of bubble flow. The latter was rather restricted, mainly due to the challenges stemming from the low quality of obtained images \[6\]. Notable exceptions include the work by Zboray et al. [@r8; @r9; @r10], wherein image processing and phase tracking algorithms were implemented and flow assessed by neutron radiography was given a quantitative treatment. However, quantitative assessment of gas bubbles in the gas/liquid metal flow context is scarce. The objective of the present study was to acquire sequences of high temporal resolution neutron radiographies of argon (Ar) bubble flow in liquid gallium (Ga) and, for the first time, to perform image-based velocimetry and shape analysis for different Ar flow rates, with and without applied horizontal magnetic field.
The experiment {#sec:experiment}
==============
The MHD system
--------------
Experiments were conducted at the thermal neutron imaging beamline *NEUTRA* at the Paul Scherrer Institute PSI [@r11]. The setup (Figure \[fig:setup\]a) consisted of a thin-walled glass container filled with liquid Ga, wherein Ar bubble flow was introduced via a submerged copper tube with a constricted nozzle, which ejected bubbles such that they ascended without wall interactions. The tube was bent so that the outlet was directed horizontally, yielding a useful effect – for a given flow rate, bubble size was constant regardless of nozzle constriction, and was prescribed by Ar/Ga surface tension. The container was connected to a pressurized Ar vessel and flow rate was measured and controlled using a digital mass flow controller (*MKS Instruments 1179B*).
![(a) Photograph of the experimental setup and (b) its simplified representation with highlighted dimensions.[]{data-label="fig:setup"}](SetupCombined.png){width="\textwidth"}
To study the influence of applied horizontal magnetic field, the container was placed between two arrays of neodymium permanent magnets (Figure \[fig:setup\]b). Magnetic field flux density within the container (Figure \[fig:magfield\]) ranged from $60~mT$ to $500~mT$ and was roughly $300~mT$ within the bubble flow region, as determined by simulations and measurements. As shown in Figure \[fig:magfield\]a, the direction of the field was rather homogeneous along container vertical axis and did not exhibit too strong magnitude variations, as seen in Figure \[fig:magfield\]b.
![Simulated magnetic field (a) within the melt container and (b) along its vertical axis.[]{data-label="fig:magfield"}](MagFieldCombined.png){width="\textwidth"}
Neutron imaging
---------------
The experimental setup was imaged at the measuring position $2$ of the *NEUTRA* beamline using a medium spatial resolution set-up (MIDI). After passing through the sample, the attenuated neutron beam was detected by $200 ~ \mu m$ thick ${}^6$*LiF/ZnS* scintillator screen. The distance between the centre of the liquid Ga vessel and the detector was $32 ~ mm$. A sCMOS camera (*ORCA Flash 4.0*) was used to collect the scintillator light output. The utilized detector optics produced images with isotropic pixel size of $55.1 ~ \mu m$, with a field of view (FOV) of $112.8$ x $112.8 ~ mm$. As such, the FOV was large enough to capture the entire trajectory of a bubble once it detached from the inlet.
All images were acquired with a high frame rate – $100$ frames per second (FPS) – to capture the motion of ascending bubbles in detail. Bubble trajectories were recorded for $10$ to $300 ~cm^3/min$ flow rates, with and without applied magnetic field. $30~s$ sequences were acquired for all flow rates, resulting in $3000$ images per measurement. Images were recorded for two types of $95$ x $150 ~mm$ glass containers, $20$- or $30~mm$ liquid Ga thickness.
Image Processing
================
Preprocessing was performed in *ImageJ*: pixels with exceedingly high intensities were removed and $4000$ frame average dark current signal was subtracted from raw images, followed by normalization with respect to the $8000$ image average of the unobstructed beam signal to compensate for neutron beam non-uniformity. Afterwards, the images were imported into *Wolfram Mathematica 12* for post-processing.
Several criteria dictated the *Mathematica* code development: reliable bubble and shape detection, minimization of false positive and detection failure rates, bubble detection throughout the entire FOV, robustness and applicability for the lowest possible signal-to-noise ratio (SNR). Additional goals included high enough shape detection precision to enable phase interface (Ar/Ga and Ga/air) tracking and shape strain rate determination. After testing a large number of different options, the authors established that, thus far, the best approach is as it appears in Figure \[fig:pipeline\], wherein the constructed image processing pipeline is illustrated.
![The image processing pipeline, from raw images to extracted phase boundary shapes.[]{data-label="fig:pipeline"}](PipelineV2.png){width="\textwidth"}
Curvature flow filter (CFF) was chosen for denoising, since bubble shapes are elliptic and because CFF aggressively erodes sharp edges that are associated with observed image artefacts. CFF diffuses pixel brightness $I$ over a virtual time interval $\tau$ according to Eq. \[eq:curvature-flow\], where diffusion rates tangential and normal to edges (localized at $\Delta I = 0$) are regulated by a control function $c(I,k)$:
$$\label{eq:curvature-flow}
\pdv{I}{t} = |\nabla I| \cdot \nabla \left( c(I,k) \cdot \frac{\nabla I}{|\nabla I|} \right); ~~t \in [0,\tau]$$
where $k$ is a control parameter. Equation \[eq:curvature-flow\] was solved over image pixels using the finite difference method (FDM).
After determining the appropriate parameters for all pipeline elements ($c(I,k) = exp(- |\nabla I|^2/k^2)$ was chosen), frames from all measurements were processed. A representative example of input/output is shown in Figure \[fig:detect\]. Bubble shapes were identified (assumed to be Jordan curves) and analysed, and parameters such as projected area, centroid coordinates and semi axes were determined by fitting ellipses into detected shapes. For this, an additional subroutine consisting of short-range Gaussian blurring, Otsu binarization and morphological thinning was implemented.
![(a) A sample pre-processed image with highlighted characteristic features and (b) a post-processed image with derived air/Ga and Ar/Ga interfaces.[]{data-label="fig:detect"}](ProcessedFrame.png){width="\textwidth"}
Finally, logical and statistical filters were applied to resulting data, removing artefacts left over (if any) from preceding processing stages. Generated output and another set of logical filters were used to compute velocities, trace trajectories and derive parameter correlations.
Simulations {#sec:simulations}
===========
The experiment was modelled numerically to verify that observed effects are not artefacts due to imperfections in the setup. Simulations were carried out in an open-source finite volume method (FVM) package *OpenFOAM* using the volume of fluid (VOF) method, according to incompressible Navier-Stokes equations, a continuity equation and a transport equation for Ga volume fraction, with linear blending functions for density and viscosity. Boundary conditions were (see Figure \[fig:setup\]b for geometry): zero flow velocity at vessel walls; zero relative pressure at the top opening to allow for gas flow circulation; constant mass flow rate at the tube inlet.
Magnetic field within the Ga container was computed using an open-source finite element method (FEM) package *Elmer* according to magnetic induction equation in terms of magnetic vector potential, and relevant force density contributions (gravity, Lorentz force, surface tension) were passed to *OpenFOAM* via *Elmer-OpenFOAM* (EOF) coupling library [@r12].
Material physical properties were as follows - Ga density: $6080~kg/m^3$; Ga viscosity: $1.97~mPa \cdot s$; Ga electrical conductivity: $3.70 \cdot 10^6~S/m$; Ga surface tension: $0.72~N/m$; Ar density: $1.784~kg/m^3$; Ar viscosity: $2.30 \cdot 10^{-5}~Pa \cdot s$.
Preliminary analysis and simulations show that the magnetic Reynolds number is $0.01 < Re_m < 0.1$, so current induced by metal flow through magnetic field should be important. However, in order to accelerate the computations significantly and obtain preliminary results, the authors decided to neglect (for now) the magnetic field produced by induced currents. The hydrodynamic Reynolds number near bubbles is within $10^3 < Re < 10^4$ and $Re \sim 1$ elsewhere, so an appropriate turbulence model must be used. The $k$-equation subgrid-scale (SGS) large eddy simulation (LES) model was chosen to avoid artificial bubble trajectory stabilization due to overestimated turbulent viscosity introduced by a more standard $k$-$\omega$ shear stress transport (SST) model. Estimates indicate that the Eötvös number is $2.1 < Eo < 4.1$, which corresponds to a flow regime wherein bubbles are of slightly oscillating elliptic shapes.
The dynamic FVM mesh was comprised of roughly $450K$ elements with $2$x refinement at phase interfaces. Simulations were performed on a computational cluster at the University of Latvia (UL). A simpler shape detection algorithm was applied to CFD simulation results: short range local adaptive binarization, morphological thinning and filling transform, followed by Shen-Castan edge detection and thinning, yielded bubble contours.
Results {#sec:results}
=======
Figure \[fig:trajectories\] clearly shows the effect of applying magnetic field – bubble trajectory spread is considerably reduced. Stabilization is brought about by electric current due to gallium flow through applied horizontal magnetic field, where flow is induced via fluid displacement by ascending bubbles. Induced current then interacts with magnetic field resulting in the Lorentz force that acts to reduce the velocity component perpendicular to the applied field. This, in turn, leads to bubble wake laminarization, preventing tail vortex detachment stabilizing bubble trajectory [@r2]. One can see that simulations and experiments are in very good agreement. The horizontal displacement of individual bubbles is also greatly reduced, as is evident from both Figure \[fig:trajectories\] and Figure \[fig:velocimetry\]a. While the horizontal velocity component is reduced (Figure \[fig:velocimetry\]a), the vertical, conversely, is increased (Figure \[fig:velocimetry\]b).
![Several initial bubble trajectories for a $100 ~cm^3/min$ flow rate derived from simulations (a,d) and experiments (b,e). In cases (a,b) there is no magnetic field, and in (d,e) the field ($ \sim 0.3 ~ T$) is applied. Simulation frame rate matches that of the experiment. Bubble detection points are color coded by order or appearance, dark purple to white. Inlet and free surface are located right beneath and above vertical boundaries of images, respectively. Bubbles are highlighted in experimental images (b,e) by dashed white circles. In (c,f), entire sets of detected bubbles over all frames are shown, without (c) and with (f) horizontal magnetic field. Dashed red lines indicate bubble set envelopes, derived using the statistics-sensitive non-linear iterative peak-clipping (SNIP) algorithm. $\delta x$ and $\Delta x$ in (c,f) stand for mean bubble set envelope thickness and maximum horizontal bubble spread, respectively. Distance scales in (c,f) are identical.[]{data-label="fig:trajectories"}](TrajectoryVersus.png){width="\textwidth"}
![Experimentally determined averaged (a) horizontal and (b) vertical velocity components of ascending bubbles at different elevations above the inlet, for different gas flow rates, without and with ($\sim 0.3 ~ T$) applied magnetic field. Colored bands represent averaged curves plus their local errors.[]{data-label="fig:velocimetry"}](VelocimetryCombined.jpg){width="\textwidth"}
These observations are in qualitative agreement with known experimental results obtained by other research groups for similar system dimensionless parameters, which indicates that there are no major issues in the experiment, simulations or the image processing algorithm [@r2; @r3; @r4]. As expected from preliminary analysis, bubble shapes are slightly oscillating, and these shape perturbations are damped when horizontal magnetic field is applied - however, this requires further quantitative analysis. Previously conducted experiments indicate that slight vertical acceleration is expected at $Eo$ values considered herein – this is observed as well.
Conclusions {#sec:conclusions}
===========
The influence of static and rather homogeneous horizontal external magnetic field on the motion of bubbles in an open pool of liquid Ga was investigated both experimentally (using neutron imaging) and *in silico* (using simulations).
High temporal resolution neutron imaging (up to 100 FPS) of dynamics of ascending Ar bubbles (at different gas flow rates) in liquid Ga was performed at the thermal neutron beamline *NEUTRA*. The intrinsically low SNR neutron radiographs were processed via a uniquely tailored image processing pipeline, which enabled successfully perform bubble velocimetry for the resulting images.
The derived bubble trajectories reveal clear influence of applied magnetic field on bubble velocities, both horizontal and vertical. Significant suppression of horizontal displacement in trajectories is observed for all flow rates, along with reduced trajectory spread, leading to a much more stable flow regime. Vertical acceleration is observed in presence of magnetic field at elevations where the contrary holds without applied field. These observations are in qualitative agreement with known experimental results.
The preliminary CFD simulations also match the experimental results rather well. Statistical data and correlations regarding bubble size, aspect ratio and other parameters have been obtained as well. Thus, we may consider our methodology (the imaging setup and the image processing pipeline) successfully validated.
Acknowledgements
================
This work is based on experiments performed at the Swiss spallation neutron source SINQ, Paul Scherrer Institute. The authors would like to thank Jevgenijs Telicko (UL) and Jan Hovind (PSI) for their invaluable assistance with hardware during experiments, and Robert Zboray (Empa Dübendorf, Switzerland), for productive discussions regarding data post-processing. The authors acknowledge the support due to the ERDF project ”Development of numerical modelling approaches to study complex multiphysical interactions in electromagnetic liquid metal technologies” (No. 5 1.1.1.1/18/A/108). This is a preprint for the original paper: *IOS Press, International Journal of Applied Electromagnetics and Mechanics, M. Birjukovs et al., Argon Bubble Flow in Liquid Gallium in External Magnetic Field*. The paper has been accepted for publication and is pending print with a tracking ID *HES19-16*. A DOI link, volume number and article volume ID/page numbers will be added once the article is published.
[99]{}
S. Pavlovs, A. Jakovics, E. Baake, V. Sushkovs, Gas bubbles and liquid metal flow influenced by uniform external magnetic field, *International Journal of Applied Electromagnetics and Mechanics* **53** (2016), 1-11.
C. Zhang, Liquid Metal Flows Driven by Gas Bubbles in a Static Magnetic Field, *PhD thesis* (2009).
C. Zhang, S. Eckert, G. Gerbeth, Experimental study of single bubble motion in a liquid metal column exposed to a DC magnetic field, *International Journal of Multiphase Flow* **31** (2005), 824–842.
E. Strumpf, Experimental study on rise velocities of single bubbles in liquid metal under the influence of strong horizontal magnetic fields in a flat vessel, *International Journal of Multiphase Flow* **97** (2017), 168-185.
M. Sarma, M. Scepanskis, A. Jakovics, K. Thomsen, R. Nikoluskins, P. Vontobel, T. Beinerts, A. Bojarevics, E. Platacis, Neutron Radiography Visualization of Solid Particles in Stirring Liquid Metal, *Physics Procedia* **69** (2015), 457-463.
E. Baake, T. Fehling, D. Musaeva, T. Steinberg, Neutron radiography for visualization of liquid metal processes: bubbly flow for CO2 free production of Hydrogen and solidification processes in EM field, *IOP Conference Series: Materials Science and Engineering* **228** (2017).
R. Zboray, P. Trtik, 800 fps neutron radiography of air-water two-phase flow, *Methods X* **5** (2018), 96-102.
R. Zboray, P. Trtik, In-depth analysis of high-speed, cold neutron imaging of air-water two- phase flows, *Flow Measurement and Instrumentation* **66** (2019), 182-189.
R. Zboray, V. Dangendorf, I. Mor, B. Bromberger, K. Tittelmeier, Time-resolved Fast Neutron Radiography of Air-water Two-phase Flows, *Physics Procedia* **69** (2014).
R. Zboray, I. Mor, V. Dangendorf, M. Stark, K. Tittelmeier, M. Cortesi, R. Adams, High-frame rate, fast neutron imaging of two-phase flow in a thin rectangular channel, *Applied Radiation and Isotopes* **90** (2014), 122–131.
E. H. Lehmann, P. Vontobel, Properties of the radiography facility NEUTRA at SINQ and its use as European reference facility, *Nondestructive Testing And Evaluation* **16(2)** (2001), 191-202.
J. Vencels, P. Raback, V. Geza, Open-source Elmer FEM and OpenFOAM coupler for electromagnetics and fluid dynamics, *SoftwareX* **9** (2019).
|
---
abstract: 'We consider an American put option under the CEV process. This corresponds to a free boundary problem for a PDE. We show that this free boundary satisfies a nonlinear integral equation, and analyze it in the limit of small $\rho$ = $2r/ \sigma^2$, where $r$ is the interest rate and $\sigma$ is the volatility. We use perturbation methods to find that the free boundary behaves differently for five ranges of time to expiry.'
---
[On a free boundary problem for an ]{}\
[American put option under the CEV process]{}
[Miao Xu and Charles Knessl]{}
[ Department of Mathematics, Statistics and Computer Science]{}\
[ University of Illinois at Chicago]{}\
[ 851 South Morgan Street]{}\
[ Chicago, IL 60607-7045]{}
[e-mail: mxu6@uic.edu, knessl@uic.edu]{}
Introduction
============
The pricing and hedging of options has its origins in the Nobel prize winning work of Black, Scholes, and Merton $\cite{cm}$, who assume that the price of an underlying asset $S(t)$ follows a geometric Brownian motion with constant volatility. The price $C(S,t)$ of a European call option at time $t$ for an asset with price $S$, strike $K$, and expiry $T$ is then readily established, and is presented in terms of the normal distribution function. However, there is sufficient empirical evidence $\cite{dl}$ to suggest that in many cases the assumption of constant volatility does not match well to the observed market data. Rather, evidence points out that the implied volatility, which is obtained by equating the model price of an option to its market price and solving for the unknown volatility parameter, varies with the strike price across a wide range of markets. This phenomenon is known as the volatility smile or frown, depending on the shape of the curve, and is not captured by the Blacks-Scholes model with a constant volatility. As a result, there have been various ideas as to how to modify and extend the basic Black-Scholes framework, to account for this phenomenon. One of these is the constant elasticity of variance (CEV) diffusion model, which was introduced by Cox and Ross $\cite{cr}$ in the context of European options. Unlike Black-Scholes, the CEV model is capable of reproducing the volatility smile.
Other work on European options under a CEV process include Davydov and Linetsky $\cite{dl}$, Hu and Knessl $\cite{hk}$ and Lo, et. al. $\cite{lyh}$. However, there exists little or no analytic work for the valuation of American options under a CEV process. The analysis of these options are more difficult than the corresponding European options in that the American options may be exercised prior to the expiration dates. Mathematically the American options lead to partial differential equations (PDE) with free boundaries, which can only rarely be solved exactly. In this paper, we apply asymptotic analysis to a CEV model to examine the behavior of the free boundary under different scaling regimes for the time to expiry, in the limit of small $\rho$ = $2r / \sigma^2$, where $r$ is the interest rate, and $\sigma$ is the volatility. This limit has a small interest rate and/or large volatility, and is of particular relevance to the financial status of the current economy. We will employ singular perturbation methods, including matched asymptotic expansions. The main result is the derivation of a nonlinear integral equation that is satisfied by the free boundary, from which we shall analyze its asymptotic structure for five different ranges of time. The main results are summarized in section 2 and derived in section section 3.
Asymptotic analysis and singular perturbation methods have been recently employed in the context of both European and American options, and this work includes Knessl $\cite{ck1, ck2}$, Howison $\cite{sh}$, Kuske and Keller $\cite{kk}$, Addison, et al. $\cite{ahk}$, Evans, et al. $\cite{ekk}$, Fouque, et al. $\cite{fps}$, and Widdicks, et al. $\cite{wd}$.
Problem Statement and Summary of Results
========================================
We let $P(S,T_0)$ denote the price of an American put option for an asset with price $S$ at some time $T_0$ prior to expiry $T_F$. We assume that $S$ satisfies the stochastic differential equation $$dS = \mu S \ dt + \sigma \sqrt{S} \ dW_t.$$ where $W_t$ is a standard Brownian motion, $\sigma$ is the volatility of the underlying asset, and $\mu = r$ is the risk-free interest rate. We note that unlike Black-Scholes, this model only guarantees non-negativity of $S$ ($S \geq 0$), so the chance of absorption at 0, i.e., bankruptcy, occurs with positive probability.
Introducing the new variables $$t = \frac{\sigma^2}{2}(T_F - T_0), \ \ \rho = \frac{2r}{\sigma^2},$$ we find that $P$ satisfies the following boundary value problem $$\label{pde1}
P_t = S P_{SS} + \rho SP_S - \rho P; \ t > 0 , \ S > \alpha(t)$$ $$P(S,0) = \max(K-S,0)$$ $$P(\alpha(t), t) = K - \alpha(t)$$ $$P_S(\alpha(t), t) = -1$$ $$\label{bc1}
P(0,t) = 0$$ where $\alpha(t)$ is the free boundary in the new time variable. We also have $P(S,t) = K - S$ for $0 < S < \alpha(t)$, and $\alpha(0) = K$. For $S \leq \alpha(t)$ the option should be exercised, and for $S > \alpha(t)$ it should be held.
We convert $(\ref{pde1})- (\ref{bc1})$ into an integral equation by first making a change in coordinates, letting $$P(S,t) = K - S + \tilde{P}(V,t), \ V = S - \alpha(t)$$ where $V$ $\geq$ 0. Then $\tilde{P}$ satisfies the PDE $$\label{pde2}
\tilde{P}_t - \alpha'(t) \tilde{P}_V = [V + \alpha(t)]\tilde{P}_{VV} + \rho [V + \alpha(t)]\tilde{P}_V - \rho K - \rho \tilde{P}; \ V, t > 0$$ with the initial and boundary conditions $$\tilde{P}(V,0) = V$$ $$\label{bc2}
\tilde{P}(0,t) = \tilde{P}_V (0,t) = 0.$$ We introduce the Laplace transform $$\label{lp1}
Q(\theta, t) = \int_{0}^{\infty} e^{-\theta V} \tilde{P}(V,t) \ dV.$$ Using $(\ref{lp1})$ in $(\ref{pde2})$ and $(\ref{bc2})$ then yields $$\label{pde3}
Q_t + (\theta^2 + \rho \theta) Q_\theta = [\alpha'(t) \theta + (\theta^2 + \rho \theta) \alpha(t) - (2\theta + 2\rho)] Q - \frac{\rho K}{\theta}.$$ with the initial condition $$\label{bc3}
Q(\theta, 0) = \frac{1}{\theta^2}.$$ Using the method of characteristics, it can be shown that the only acceptable solution to $(\ref{pde3})$ is $$\label{pde4}
Q(\theta,t) = \frac{K \rho}{\theta^2} e^{\alpha(t)\theta} \int_{\theta / \rho}^{\infty} \frac{1}{z+1} \exp \left[-\rho z \alpha \left(t+\rho^{-1} \log \left(\frac{\theta+\rho}{\theta}\frac{z}{z+1} \right) \right) \right] \ dz.$$ The next result readily follows.
The option price $P(S,t)$ for the CEV model has the integral representation
$$P(S,t) = K - S + \frac{1}{2 \pi i} \int_{Br} {e^{\theta V} Q(\theta, t)} \ d\theta,$$
where $\Re(\theta) > 0 $ on the Bromwich contour, and $Q(\theta, t)$ is given by $(\ref{pde4})$.
Moreover, after setting $t = 0$ and using $\alpha(0)=K$ and $(\ref{bc3})$ in $(\ref{pde4})$, it follows that $\alpha(t; \rho)$ satisfies the nonlinear integral equation (IE):
$$\label{IE}
\frac{e^{-K \theta }}{K \rho} = \int_{\theta / \rho}^{\infty} \frac{1}{z+1} \exp \left[-\rho z \alpha \left(\rho^{-1} \log \left(\frac{\theta+\rho}{\theta}\frac{z}{z+1} \right) \right) \right] \ dz.$$
In the next section we use asymptotic methods to analyze this IE for five different scales of time $t$, in the limit of small $\rho$. We let $\rho = e^{-\lambda}$ so that $\lambda = - \log \rho \rightarrow \infty$. The final results for the free boundary $\alpha(t; \rho)$ are listed below, and we sketch the derivations in section 3.\
1. $t = \omega / \lambda = O(\lambda^{-1}), \ 0 < \omega < K$: $$\label{eq1}
\alpha(t; \rho) = (\sqrt{\omega} -\sqrt{K})^2 + \frac{\log \lambda}{\lambda}\frac{\omega - \sqrt{K \omega}}{2} + \frac{1}{\lambda} \frac{\sqrt{K \omega} - \omega}{2} \log \left(\frac{4 \pi K^2 \omega}{K - \ \sqrt{K \omega}} \right) + o(\lambda^{-1})$$
2. $t = K / \lambda + O(\lambda^{-2}), \ \lambda^2 t - \lambda K = \lambda( \omega - K) = \Lambda$: $$\label{eq2}
\alpha(t; \rho) = \frac{1}{\lambda^2} \mathcal{F}(\Lambda),$$ where $\mathcal{F}(\cdot)$ satisfies the nonlinear IE $$\label{eq3}
\frac{e^{-K\nu}}{K} = \int_{0}^{\infty} \frac{1}{\xi} \exp \left[-\xi F \left(-\nu K^2 - \frac{1}{\xi} \right) \right] \ d\xi.$$ For $\Lambda \rightarrow \pm \infty$, we have $$\label{eq4}
\mathcal{F}(\Lambda) \sim \frac{\Lambda^2}{4K} + \frac{\Lambda}{4} \log( -\Lambda) - \frac{\Lambda}{4} \log(8 \pi K^3),\ \Lambda \rightarrow -\infty$$ $$\label{eq5}
\mathcal{F}(\Lambda) \sim \Lambda e^{-\gamma} \exp \left[-\frac{1}{K}\exp \left(\frac{\Lambda}{K} \right) \right], \ \Lambda \rightarrow + \infty$$ where $\gamma$ is the Euler constant.
3. $t = \omega / \lambda = O(\lambda^{-1}), \ K < \omega < \infty$: $$\label{eq6}
\alpha(t; \rho) \sim \frac{\omega - K}{\lambda} e^{-\gamma} \exp \left(-\frac{1}{K} \rho^{K / \omega - 1} \right);$$
4. $t = O(1), \ 0 < t < \infty$: $$\label{eq7}
\alpha(t; \rho) \sim te^{-\gamma}\exp \left[-\left(\frac{1}{2} + \frac{1}{\rho K} \right) e^{-K/t} \right];$$
5. $t = v / \rho = O(\rho^{-1}), \ v > 0$: $$\label{eq8}
\alpha(t; \rho) \sim \frac{1}{\rho} e^{-\gamma} \exp \left(\frac{1}{e^{v}-1} \right)(1-e^{-v}) \exp \left(-\frac{1}{\rho K} \right).$$
We note that in four of the five cases the expression for $\alpha(t; \rho)$ is completely explicit, and only in case (ii) must we solve a nonlinear IE, which is somewhat simpler than the one in $(\ref{IE})$. We can easily compute $P(S,t)$ as $t \rightarrow \infty$, which corresponds to the perpetual American option, where the problem reduces to solving an ordinary differential equation. Setting $P(S,\infty) = P^\infty(S)$ and using $\alpha(\infty)$ to denote the limiting value of the free boundary, we obtain from $(\ref{pde1})-(\ref{bc1})$
$$P^\infty(S) = Ke^{\rho \alpha(\infty)} \int_{1}^{\infty} \frac{1}{z^2}e^{-z \rho S} \ dz,$$
where $\alpha(\infty)$ satisfies $$\label{aper}
K \rho \int_{1}^{\infty} \frac{1}{z}e^{-\rho \alpha(\infty) z} \ dz = e^{-\rho \alpha(\infty)}.$$ For $\rho \rightarrow 0$ we have $$\label{fbper}
\alpha(\infty) = \frac{1}{\rho} e^{-\gamma} \exp \left(-\frac{1}{\rho K} \right) [1 + O(\rho)],$$ which is exponentially small.
Analysis
========
Analysis for $t = \omega / \lambda$, $0 < \omega < K $
------------------------------------------------------
We first examine $(\ref{IE})$ on the $t = O(\lambda^{-1})$ scale, for small $\rho$. Recalling that $\lambda = -\log \rho$, we let $$\theta = \lambda \beta, \ z = \frac{\lambda(\beta + x)}{\rho}, \ \alpha(t; \rho) \sim \alpha_0(\omega; \rho)$$ for $\omega = (-\log \rho)t = O(1)$. Then $(\ref{IE})$ can be approximated by $$\label{ie1}
\frac{e^{\lambda}}{K} = \int_{0}^{\infty} \frac{1}{\beta + x} e^{\lambda \Phi(x; \beta, \rho)} [1 + O(e^{-\lambda})] \ dx,$$ where $\Phi(x; \beta, \rho) = K \beta - (\beta + x) \alpha_0(\frac{x}{\beta(x + \beta)}; \rho)$. For large $\lambda$ and fixed $\beta$, we evaluate the right hand side of $(\ref{ie1})$ by an implicit form of the Laplace method, assuming for now that there is a saddle point where $$\label{saddle}
\frac{\partial \Phi}{\partial x} = -\alpha_0 \left(\frac{x}{\beta(x + \beta)} \right) - \frac{1}{x + \beta}\alpha_0' \left(\frac{x}{\beta(x + \beta)} \right) = 0.$$ Let us denote $x = x_*(\beta)$ as the solution to $(\ref{saddle})$. It follows that at $x = x_*$, $\Phi \sim 1$ so that $$\label{spe}
1 = K\beta - (\beta + x_*)\alpha_0 \left(\frac{x_*}{\beta(x_* + \beta)} \right).$$ Now let $\omega = \frac{x_*}{\beta(x_* + \beta)}$. Then from $(\ref{saddle})$ we have $\beta = \frac{\alpha_0'(\omega)}{\omega \alpha_0'(\omega) - \alpha_0(\omega)}$ which we use to eliminate $\beta$ in $(\ref{spe})$ to obtain the ODE
$$[1-\alpha_0'(\omega)][\omega \alpha_0'(\omega) - \alpha_0(\omega)] = K \alpha_0'(\omega)$$
Rewriting this as $$\omega \alpha_0'(\omega) - \alpha_0(\omega) = \frac{K \alpha_0'(\omega)}{1-\alpha_0'(\omega)},$$ we recognize this as the Clairaut equation. The solutions consist of a one-parameter family of lines and the singular solution $$\label{wsol}
\alpha_0(\omega) = (\sqrt{\omega} - \sqrt{K})^2$$ which is the envelope of this family. The linear solutions $\alpha_0(\omega) = \omega C - \frac{KC}{1-C}$ must be rejected, since these lead to $\alpha_0(0) \neq K$. The above analysis applies only for $0 < \omega < K$, since the solution $(\ref{wsol})$ vanishes as $\omega$ approaches $K$. Hence we expect different asymptotics for $\omega \approx K$.
We next analyze some higher order terms in the expansion of $\alpha_0$. We evaluate $(\ref{ie1})$ by using the Laplace method, which gives
$$\label{laplace}
\frac{e^{\lambda}}{K} = \frac{1}{\beta + x_*} \sqrt{\frac{2\pi}{-\lambda \Phi_{xx}(x_*; \beta, \rho)}} e^{\lambda \Phi(x_*; \beta, \rho)}[1 + O(\lambda^{-1})],$$
and expand $\alpha_0$ as $$\label{exp1}
\alpha_0(\omega; \rho) = \alpha_0(\omega) + \frac{\log \lambda}{\lambda} \alpha_1(\omega) + \frac{1}{\lambda} \alpha_2(\omega) + o(\lambda^{-1}).$$ In order to balance the two sides of $(\ref{laplace})$, we need $\alpha_1$ to cancel the $\sqrt{1 / \lambda}$ factor. Hence, $$\label{la1}
\frac{1}{\sqrt{\lambda}}\exp \left[-(\beta + x_*)(\log \lambda) \ \alpha_1 \left(\frac{x_*}{\beta(x_* + \beta)} \right) \right] = 1.$$ Writing $(\ref{la1})$ in terms of $\omega$ we obtain $$\label{wsol2}
\alpha_1(\omega) = \frac{1}{2} (\omega - \sqrt{K\omega}).$$ To find the third order term $\alpha_2$, we balance the $O(1)$ terms in $(\ref{laplace})$, so that $$\frac{1}{K} = \frac{1}{\beta + x_*} \sqrt{\frac{{2\pi}}{-\Phi_{xx}}} \exp \left[-(\beta + x_*) \alpha_2 \left(\frac{x_*}{\beta(x_* + \beta)} \right)\right].$$ It can be shown that $\Phi_{xx}(x_*; \beta, \rho) \sim -\frac{1}{2} K^{\frac{1}{2}} \beta^{\frac{3}{2}}x_*^{-\frac{3}{2}}(x_* + \beta)^{-\frac{3}{2}}$ and then $$\label{wsol3}
\alpha_2(\omega) = \frac{\sqrt{K\omega} - \omega}{2} \log \left(\frac{4\pi K^2 \omega}{K - \sqrt{K\omega}} \right).$$ With $(\ref{wsol})$, $(\ref{exp1})$, $(\ref{wsol2})$, and $(\ref{wsol3})$ we have established $(\ref{eq1})$.
Analysis for $t = \omega / \lambda$, $\omega \approx K $
--------------------------------------------------------
We return to $(\ref{IE})$ and introduce the scaling $$\label{scale1}
\theta = \lambda \beta, \ \beta = \frac{1}{K} + \frac{\nu}{\lambda}, \ \omega = \lambda t = K + \frac{\Lambda}{\lambda}$$ with $$\label{scale2}
\alpha(t; \rho) = \frac{1}{\lambda^2} \mathcal{F}(\Lambda; \rho) = \frac{1}{\lambda^2} \mathcal{F} \left(\lambda^2 \left(t-\frac{K}{\lambda} \right); \rho \right).$$ Then we have $e^{-\theta K} = \rho e^{-K \nu}$. Also by setting $z = (\theta + y) / \rho $ in $(\ref{IE})$ this equation becomes $$\label{ie2}
\frac{1}{K} e^{-K \nu} = \int_{0}^{\infty} \frac{1}{\theta + y + \rho} \exp \left[-(\theta + y) \alpha \left(\frac{1}{\theta} - \frac{1}{\theta + y} + O(\rho); \rho \right) \right] \ dy.$$ With the scaling in $(\ref{scale1})$ and $(\ref{scale2})$ and the fact that $\rho = e^{-\lambda}$ is exponentially small, we obtain $$\begin{aligned}
\lambda^2 \alpha \left(\frac{1}{\theta} - \frac{1}{\theta + y} + O(e^{-\lambda}); \rho \right) & = \lambda^2 \alpha \left(\frac{K}{\lambda} - \frac{1}{\lambda^2} \left(\nu K^2 + \frac{1}{\xi} \right) + o(\lambda^{-2}); \rho \right)\notag \\
& \sim \mathcal{F} \left(-\nu K^2 - \frac{1}{\xi} \right)\end{aligned}$$ where $\mathcal{F}(\Lambda)$ is the leading term in an expansion of $\mathcal{F}(\Lambda; \rho)$. Then scaling $y = \lambda^2 \xi$ in $(\ref{ie2})$ and letting $\rho \rightarrow 0$ $(\lambda \rightarrow \infty)$ we obtain the limiting IE in $(\ref{eq3})$. It does not seem possible to solve $(\ref{eq3})$ explicitly for $\mathcal{F}(\Lambda)$. But we can infer the behavior as $\Lambda \rightarrow -\infty$ by evaluating the integral in $(\ref{eq3})$ by an implicit Laplace type expansion, similarly to what we did in section 3.1. This will verify the asymptotic matching between the $\omega$-scale (for $\omega < K$) and the $\Lambda$-scale, and lead to $(\ref{eq4})$. Now consider the limit $\Lambda \rightarrow +\infty$. For $|\nu| < 0$, we rewrite $(\ref{eq3})$ as
$$\label{inte}
\begin{split}
\frac{e^{K |\nu|}}{K} & = \int_{\lvert \nu \rvert K^2}^{\infty} \frac{1}{\eta} \exp \left[-\frac{1}{\eta} \mathcal{F}(\lvert \nu \rvert K^2 - \eta) \right] \ d\eta \\ & + \int_{0}^{1} \frac{1}{u} \left \{ \exp \left[-\frac{1}{u} \frac{\mathcal{F}(\lvert \nu \rvert K^2 (1-u))}{\lvert \nu \rvert K^2} \right] - \exp \left[-\frac{1}{u} \frac{\mathcal{F}(\lvert \nu \rvert K^2)}{\lvert \nu \rvert K^2} \right] \right \} \\ & +
\int_{\frac{\mathcal{F}(\lvert \nu \rvert K^2)}{\lvert \nu \rvert K^2}}^{\infty} \frac{e^{-v}}{v} \ dv. \hspace{1in}
\end{split}$$
Here we broke up the integral over $(0, \infty)$ into the two ranges $(0, \lvert \nu \rvert K^2)$ and $(\lvert \nu \rvert K^2, \infty)$ and made some elementary substitutions. Now, for $\Lambda \rightarrow - \infty$ we have $\mathcal{F}(\Lambda) \sim \frac{\Lambda^2}{4K}$ so that the first integral in the right hand side of $(\ref{inte})$ will vanish as $\nu \rightarrow -\infty$. If $\mathcal{F}(\Lambda) \rightarrow 0$ as $\Lambda \rightarrow + \infty$ the second integral in $(\ref{inte})$ will also vanish, and the third may be approximated by using $$\label{expint}
\int_{\varepsilon}^{\infty} \frac{e^{-v}}{v} \ dv = - \log \varepsilon - \gamma + O(\varepsilon),\;\; \varepsilon \rightarrow 0^+.$$ Hence $(\ref{inte})$ can be replaced by the asymptotic relation $$\frac{e^{K \lvert \nu \rvert}}{K} \sim -\log \left[\frac{\mathcal{F}(\lvert \nu \rvert K^2)}{\lvert \nu \rvert K^2} \right] - \gamma$$ which upon exponentiation leads to the asymptotic result given in $(\ref{eq5})$, for $\mathcal{F}(\Lambda)$ as $\Lambda \rightarrow \infty$.
Analysis for $t = \omega / \lambda$, $K < \omega < \infty $
-----------------------------------------------------------
In the remaining time ranges, $\alpha(t; \rho)$ will be exponentially small as $\rho = e^{-\lambda} \rightarrow 0$, and our analysis of $(\ref{IE})$ will rely heavily on the asymptotic form in $(\ref{expint})$. We let $z = Z / \rho$ in $(\ref{IE})$ to obtain $$\label{ie3}
\frac{e^{-K \theta}}{K \rho} = \int_{\theta}^{\infty} \frac{1}{Z+\rho} \exp \left[-Z \alpha \left(\rho^{-1} \log \left(\frac{\theta + \rho}{\theta} \frac{Z}{Z + \rho} \right) \right) \right] \ dZ.$$ Now we scale $Z = \lambda z_*$ and $\theta = \lambda \theta_*$, let $\alpha(t; \rho) = \tilde{\alpha}(\lambda t; \rho)$ and note that in sections 3.1 and 3.2 we have already characterized $\tilde{\alpha}(\lambda t; \rho)$ for $\lambda t = \omega < K$ and $\omega \sim K$. We also simplify the argument of $\alpha(\cdot)$ in $(\ref{ie3})$ using $$\alpha \left(\frac{1}{\rho} \left[\log \left(1 + \frac{\rho}{\theta} \right) - \log \left(1 + \frac{\rho}{Z} \right) \right] \right) = \alpha \left(\frac{1}{\theta} - \frac{1}{Z} + O(\rho) \right) = \tilde{\alpha} \left(\frac{1}{\theta_*} - \frac{1}{Z_*} + O(e^{-\lambda} \lambda) \right).$$ When $Z = \theta$ we have $z_* = \theta_*$ and we rewrite the integral in $(\ref{ie3})$ by splitting the range of integration into $z_* \in (\theta_*, \theta_* / [1 - K \theta_*])$ and $z_* \in (\theta_* / [1 -K \theta_*], \infty)$, thus obtaining $$\frac{e^{-K \lambda \theta_*}}{K \rho} \sim \left( \int_{\theta_*}^{\frac{\theta_*}{1-K \theta_*}} + \int_{\frac{\theta_*}{1-K \theta_*}}^{\infty} \right) \left( \frac{1}{z_*} \exp \left[-\lambda z_* \tilde{\alpha} \left( \frac{1}{\theta_*} - \frac{1}{z_*} \right) \right] \right) \ dz_*.$$ In the first range $\tilde{\alpha}(\omega) \sim (\sqrt{K} - \sqrt{\omega})^2$ and the first integral will be $o(1)$ as $\lambda \rightarrow \infty$, since $\theta_*^{-1} - z_*^{-1} \leq K$ when $z_* \leq \theta_* / [1 - K \theta_*]$. In the second integral $\tilde{\alpha}$ will be exponentially small and the main contribution will come from very large values of $z_*$, where roughly $z_* = O(\tilde{\alpha}^{-1})$. Then we write $\tilde{\alpha}(\theta_*^{-1} - z_*^{-1}) \sim \tilde{\alpha}(\theta_*^{-1})$ and using $(\ref{expint})$ we conclude that $$\label{ie4}
\frac{e^{-K \lambda \theta_*}}{K \rho} \sim \int_{\frac{\theta_*}{1 - K \theta_*}}^{\infty} \frac{1}{z_*} \exp \left[-\lambda z_* \tilde{\alpha} \left(\frac{1}{\theta_*} \right) \right] \ dz_* \sim - \log \left[\lambda \tilde{\alpha} \left(\frac{1}{\theta_*} \right) \right] - \gamma - \log \left[\frac{\theta_*}{1 - K \theta_*} \right],$$ with an error that is $o(1)$ as $\lambda \rightarrow \infty$. Then exponentiating $(\ref{ie4})$ and replacing $\theta_*$ by $\omega^{-1}$ we obtain the asymptotic result in $(\ref{eq6})$.
For $\omega \rightarrow K$ we note that $\rho^{\omega / K - 1} = \rho^{-1} e^{-K / t} = \rho^{-1} \exp \left[-\frac{\lambda K}{K + \frac{\Lambda}{\lambda}} \right] \\
= \rho^{-1} \exp \left[-\lambda + \frac{\Lambda}{K} + O(\lambda^{-1}) \right] \sim \exp \left(\frac{\Lambda}{K} \right)$ and $(\omega - K) / \lambda = \Lambda / \lambda^2$, which can be used to verify the asymptotic matching between the $\Lambda$-scale and the $\omega$-scale for $\omega > K$, in the intermediate limit where $\omega \downarrow K$ and $\Lambda \rightarrow \infty$.
Analysis for $t = O(1), 0 < t < \infty $
----------------------------------------
Next we consider times $t = O(1)$. We scale $z = \theta w / \rho$. Since we again expect $\alpha(t; \rho)$ to be very small we assume a “WKB-type” ansatz of the form $$\label{wkb}
\alpha(t; \rho) \sim g(t) \exp \left[-\frac{1}{\rho} f(t) \right].$$ Expanding $\alpha(t; \rho)$ in $(\ref{IE})$ for fixed $\theta$ and $\rho \rightarrow 0$, and noting that $$\frac{1}{\rho} \left[\log \left(1 + \frac{\rho}{\theta} \right) - \log \left(1 + \frac{\rho}{\theta w} \right) \right] = \frac{1}{\theta} - \frac{1}{\theta w} - \frac{1}{2} \frac{\rho}{\theta^2} + O \left(\rho^2, \frac{\rho}{w^2} \right),$$ we have $$\begin{aligned}
-\theta w \alpha \left( \frac{1}{\rho} \left[\log \left(1 + \frac{\rho}{\theta} \right) - \log \left(1 + \frac{\rho}{\theta w} \right) \right] \right) & \sim -\theta w g \left(\frac{1}{\theta} \right) \exp \left[-\frac{1}{\rho} f \left(\frac{1}{\theta} - \frac{1}{\theta w} - \frac{1}{2} \frac{\rho}{\theta^2} \right) \right] \\
& = -\theta w g \left(\frac{1}{\theta} \right) \exp \left[-\frac{1}{\rho} f \left(\frac{1}{\theta} \right) + \frac{1}{2 \theta^2} f' \left(\frac{1}{\theta} \right) + O(\rho) \right]\end{aligned}$$ Here we also used $f(\theta^{-1} - (\theta w)^{-1}) \sim f(\theta^{-1})$, since $w$ will be scaled to be exponentially large. Then setting $\varepsilon = \theta g \left(\frac{1}{\theta} \right) \exp \left[-\frac{1}{\rho} f \left(\frac{1}{\theta} \right) \right] \exp \left[\frac{1}{2 \theta^2} f' \left(\frac{1}{\theta} \right) \right]$, scaling $w = \varepsilon^{-1} u$ and using $(\ref{expint})$, $(\ref{IE})$ asymptotically becomes $$\label{asy1}
\frac{e^{-K \theta }}{K \rho} \sim \frac{1}{\rho} f \left(\frac{1}{\theta}\right) -\gamma - \log \left[\theta g \left(\frac{1}{\theta} \right) \right] - \frac{1}{2\theta^2}f' \left(\frac{1}{\theta} \right) + o(1)$$ From the $O(\rho^{-1})$ terms in $(\ref{asy1})$ we conclude that $f \left(1/ \theta \right) = K^{-1} e^{-K \theta}$ and then the $O(1)$ terms determine $g(\cdot)$ from $$\label{match}
\theta g \left(\frac{1}{\theta} \right) = e^{-\gamma} \exp \left[-\frac{1}{2 \theta^2} f' \left(\frac{1}{\theta} \right) \right].$$ The above along with $(\ref{wkb})$ establishes the asymptotic result in $(\ref{eq7})$. The asymptotic matching between $(\ref{eq6})$ and $(\ref{eq7})$ is immediate, since $\rho^{K/ \omega - 1} = \rho^{-1} e^{-K / t}$, and $(\omega - K) / \lambda \sim \omega / \lambda = t$ as $\omega \rightarrow \infty$.
Analysis for $t = v / \rho = O(\rho^{-1}), v > 0$
--------------------------------------------------
We assume that time to expiry for the option is large, with $t = v / \rho = O(\rho^{-1})$. On this time scale we assume that $$\label{vwkb}
\alpha(t; \rho) \sim \frac{1}{\rho} \exp \left(-\frac{1}{\rho K} \right) A(v),$$ where $A(\cdot)$ will be determined from $(\ref{IE})$. After scaling $\theta = \rho W$, $(\ref{IE})$ becomes $$\label{vie}
\frac{e^{-K \rho W}}{K \rho} \sim \int_{W}^{\infty} \frac{1}{z+1} \exp \left[-ze^{-\frac{1}{\rho K}}A \left(\log \left[\frac{W+1}{W} \frac{z}{z+1} \right] \right) \right] \ dz.$$ The major contribution to the integral in $(\ref{vie})$ will once more come from large values of $z$, so we approximate $$A \left(\log \left[\frac{W + 1}{W} \frac{z}{z + 1} \right] \right) \sim A \left(\log \left[\frac{W + 1}{W} \right] \right),$$ and then applying this to $(\ref{IE})$ along with $(\ref{expint})$ leads to $$\begin{aligned}
\frac{e^{-K \rho W}}{K \rho} & \sim - \log \left\{e^{-\frac{1}{\rho K}} A \left(\log \left[\frac{W + 1}{W} \right] \right) \right\} - \gamma - \log(W + 1) \\
&= \frac{1}{K \rho} - \gamma - \log \left\{(W + 1) A \left(\log \left[\frac{W + 1}{W} \right] \right) \right\} + o(1).\end{aligned}$$ Then after expanding $e^{-K \rho W} = 1 - K \rho W + O(\rho^2)$ we conclude that $$A \left(\log \left[\frac{W + 1}{W} \right] \right) = e^{-\gamma} \frac{1}{W + 1} e^W$$ which determines the function $A(\cdot)$ and establishes $(\ref{eq8})$.
Finally we verify the asymptotic matching between $(\ref{eq7})$ and $(\ref{eq8})$. For $v \rightarrow 0$ we have $(e^v - 1)^{-1} = v^{-1} - \frac{1}{2} + O(v)$ and $1 - e^{-v} \sim v = \rho t$. For $t \rightarrow \infty$ we have $e^{-K / t} = 1 - \frac{K}{t} + O(t^{-2}) = 1 - \frac{K \rho}{v} + O(\rho^2)$ so that $-(\frac{1}{2} + \frac{1}{\rho K}) e^{-K / t} \sim -v^{-1} + \frac{1}{2}$ and the matching follows. As $v \rightarrow \infty$ we have $A(v) \rightarrow e^{-\gamma}$ and thus the expansion in $(\ref{vwkb})$ agrees with the small $\rho$ expansion of $\alpha(\infty; \rho)$, as given in $(\ref{fbper})$.
[99]{}
J.A Addison, S.D. Howison, J.R. King. Ray methods for free boundary problems. [*Quart. Appl. Math.*]{}, 64(1):41-59, 2006.
F. Black and M. Scholes. The pricing of options and corporate liabilities. [*J. Political Econom*]{}., 81:637-659, 1973.
J. Cox and S. Ross. The valuation of options for alternative stochastic processes. [*J. Financial Econom*]{}., 3:145-166, 1976.
D. Davydov and V. Linetsky. The valuation and hedging of barrier and lookback options under the CEV process [*Management Science*]{}, 47:949-965, 2001.
J.D. Evans, R.A. Kuske, and J.B. Keller. American options on assets with dividends near expiry. [*Math. Finance*]{}, 12:219-237, 2002.
J.P. Fouque, G. Papanicolaou, R. Sircar, and K. Solna. Singular perturbations in option pricing. [*SIAM J. Appl. Math.*]{}, 63(5):1648-1665, 2003.
S.D. Howison. Matched asymptotic expansions in financial engineering. [*J. Engrg. Math.*]{}, 53:385-406, 2005.
F. Hu and C. Knessl. Asymptotics of barrier option pricing under the CEV process. [*Appl. Math. Finance*]{}, 17(3):261-300, 2010.
C. Knessl. A note on a moving boundary problem arising in the American put option. [*Stud. Appl. Math.*]{}, 107(2):157-183, 2001.
C. Knessl. Asymptotic analysis of the American call option with dividends. [*European J. Appl. Math.*]{}, 13(6):587-616, 2002.
R.A Kuske and J.B. Keller. Optimal exercise boundary for an American put option. [*Appl. Math. Finance*]{}, 5:107-116, 1998.
C.F. Lo, P.H. Yuen, and C.H. Hui. Pricing barrier options with square root process. [*Int. J. Theor. Appl. Finance*]{}, 5:805-818, 2001.
M. Widdicks, P.W Duck, A.D. Andricopoulos, and D.P. Newton. The Black-Scholes equation revisited: Asymptotic expansions and singular perturbations. [*Math. Finance*]{}, 15(2):373-391, 2005.
|
---
abstract: 'The Baumslag-Solitar group is an example of an HNN extension. Spielberg showed that it has a natural positive cone, and that it is then a quasi-lattice ordered group in the sense of Nica. We give conditions for an HNN extension of a quasi-lattice ordered group $(G,P)$ to be quasi-lattice ordered. In that case, if $(G,P)$ is amenable as a quasi-lattice ordered group, then so is the HNN extension.'
address: 'Department of Mathematics and Statistics, University of Otago, PO Box 56, Dunedin 9054, New Zealand.'
author:
- Astrid an Huef
- Iain Raeburn
- Ilija Tolich
date: '8 March, 2017'
title: 'HNN extensions of quasi-lattice ordered groups and their operator algebras'
---
[^1]
Introduction
============
Since they were introduced by Nica [@Nica1992], quasi-lattice ordered groups and their $C^*$-algebras have generated considerable interest (see, for example, [@LacaRaeburn1996],[@LacaRaeburn2010]). The amenability of quasi-lattice ordered groups has been a deep subject (see, for example, [@CrispLaca2002],[@CrispLaca2007] and [@Li2012]). Quasi-lattice ordered groups are also examples of the more recent LCM semigroups [@BrownloweLarsenStammeier2016], [@Starling2015]. Here we generalise two recent results about the Baumslag-Solitar group.
First, Spielberg proved that the Baumslag-Solitar group is quasi-lattice ordered [@Spielberg2012]. The Baumslag-Solitar group is an example of an HNN extension of ${\mathbb{Z}}$, and hence we wondered if HNN extensions could provide new classes of quasi-lattice ordered groups. Spielberg also showed that a groupoid associated to the Baumslag-Solitar semigroup is amenable [@Spielberg2012 Theorem 3.22].
Second, Clark, an Huef and Raeburn examined the phase-transitions of the Toeplitz algebra of the Baumslag-Solitar group [@HuefClarkRaeburn2015]. As part of their investigation they proved that the Baumslag-Solitar group is amenable as a quasi-lattice ordered group. The standard way to prove amenability, introduced by Laca and Raeburn [@LacaRaeburn1996], is to use a “controlled map”: an order-preserving homomorphism between quasi-lattice ordered groups. They observed that the height map, which counts the number of times the stable letter of the HNN extension appears in a word, is almost a controlled map, and then they adapted the standard proof in [@HuefClarkRaeburn2015 Appendix A] to fit.
Our innovation in this paper is a more general definition of a controlled map. We prove in Theorem \[p-controlled map implies amenable\] that if $(G,P)$ is a quasi-lattice ordered group and there is a controlled map $\mu$ into an amenable group, and if $\ker\mu$ is an amenable quasi-lattice ordered group, then $G$ is amenable. The motivation for Theorem \[p-controlled map implies amenable\] was two-fold. First, if a normal subgroup $N$ of a group $G$ is amenable and $G/N$ is amenable, then $G$ is amenable, and second, Spielberg’s result on amenability of groupoids [@Spielberg2014 Proposition 9.3].
In Theorem \[(G\*,P\*) is left ql\] we give conditions under which an HNN extension of a quasi-lattice ordered group is quasi-lattice ordered. This result allows us to construct many new examples of quasi-lattice ordered groups. Finally, we use Theorem \[p-controlled map implies amenable\] to prove that an HNN extension of an amenable quasi-lattice ordered group is amenable (Theorem \[HNN extensions are amenable as expected\]).
Preliminaries
=============
Let $P$ be a subsemigroup of a discrete group $G$ such that $P\cap P^{-1}=\{e\}$. There is a partial order on $G$ defined by $$x\leq y\Leftrightarrow x^{-1}y\in P\Leftrightarrow y\in xP.$$ The order is left-invariant in the sense that $x\leq y$ implies $zx\leq zy$ for all $z \in G$. A partially ordered group $(G, P)$ is *quasi-lattice ordered* if every finite subset of $G$ with a common upper bound in $P$ has a least common upper bound in $P$ [@Nica1992 Definition 2.1]. By [@CrispLaca2002 Lemma 7], $(G,P)$ is quasi-lattice ordered if and only if: $$\label{Crisp Laca}
\ \hfill
\parbox{0.9\textwidth}{
if $x\in PP^{-1}$, then there exist a pair $\mu, \nu\in P$ with $x=\mu\nu^{-1}$ such that $\gamma, \delta\in P$ and $\gamma\delta^{-1}=\mu\nu^{-1}$ imply $\mu\leq \gamma$ and $\nu\leq \delta$. (The pair $\mu, \nu$ is unique.)}$$
Let $(G, P)$ be a quasi-lattice ordered group, and let $x, y\in G$. If $x$ and $y$ have a common upper bound in $P$, then their least common upper bound in $P$ is denoted $x\vee y$. We write $x\vee y=\infty$ when $x$ and $y$ have no common upper bound in $P$ and $x\vee y<\infty$ when they have a common upper bound. An *isometric representation of $P$* in a $C^*$-algebra $A$ is a map $W:P\to A$ such that $W_e=1$, $W_p$ is an isometry and $W_pW_q=W_{pq}$ for all $p,q\in P$. We say that $W$ is *covariant* if $$\label{covariance condition}
W_pW^*_pW_qW^*_q=\begin{cases}W_{p\vee q}W^*_{p\vee q}&\text{if $p\vee q<\infty$}\\0&\text{otherwise.}
\end{cases}$$ Equivalently, $W$ is covariant if $$W^*_pW_q=\begin{cases}W_{p^{-1}(p\vee q)}W^*_{q^{-1}(p\vee q)}&\text{if $p\vee q<\infty$}\\0&\text{otherwise.}\end{cases}$$ An example of a covariant representation is $T:P\to B(\ell^2(P))$ characterised by $T_p\epsilon_x=\epsilon_{px}$ where $\{\epsilon_x:x\in P\}$ is the orthonormal basis of point masses in $\ell^2(P)$.
In [@Nica1992 §§2.4 and 4.1] Nica examined two $C^*$-algebras associated to $(G, P)$. The reduced $C^*$-algebra $C^*_r(G,P)$ of $(G,P)$ is the $C^*$-subalgebra of $B(\ell^2(P))$ generated by $\{T_p:p\in P\}$. The universal $C^*$-algebra $C^*(G,P)$ of $(G,P)$ is generated by a universal covariant representation $w$; it is universal for covariant representations of $P$ in the following sense: for any covariant representation $W:P\to A$ there exists a unital homomorphism $\pi_W:C^*(G,P)\to A$ such that $\pi_W(w_p)=W_p$. It follows from that $$C^*(G,P)= \overline{{\operatorname{span}}}\{w_pw^*_q:p, q\in P\}.$$
Nica defined $(G,P)$ to be *amenable* if the homomorphism $\pi_T:C^*(G,P)\to C^*_r(G,P)$ is faithful [@Nica1992 §4.2]. He identified an equivalent condition: there exists a conditional expectation $E:C^*(G,P)\to \overline{{\operatorname{span}}}\{w_pw^*_p:p\in P\}$, and $(G,P)$ is amenable if and only if $E$ is faithful (that is, $E(a^*a)=0$ implies $a^*a=0$ for all $a\in C^*(G,P)$). Laca and Raeburn took this second condition as their definition of amenability [@LacaRaeburn1996 Definition 3.4].
Order-preserving maps and amenability
=====================================
A key technique, introduced by Laca and Raeburn in [@LacaRaeburn1996 Proposition 6.6][^2], is the use of an order-preserving homomorphism between two quasi-lattice ordered groups which preserves the least upper bound structure. Crisp and Laca called such a homomorphism a *controlled map* [@CrispLaca2007]. If $(G,P)$ and $(K,Q)$ are quasi-lattice ordered groups, $\mu:G\to K$ is a controlled map and $K$ is an amenable group, then $(G,P)$ is amenable as a quasi-lattice ordered group by [@LacaRaeburn1996 Proposition 6.6]. Motivated by work in [@HuefClarkRaeburn2015 Appendix A] we now give a weaker definition for a controlled map. We then follow the program of [@HuefClarkRaeburn2015] to generalise [@LacaRaeburn1996 Proposition 6.6]. We state this generalisation in Theorem \[p-controlled map implies amenable\] below; its proof will take up the remainder of this section.
\[expanded controlled map definition\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups. Let $\mu:G\to K$ be an order-preserving group homomorphism. For each $k\in Q$, let $\Sigma_k$ be the set of $\sigma\in \mu^{-1}(k)\cap P$ which are minimal in the sense that $$x\in \mu^{-1}(k)\cap P \text{\ and\ }x\leq\sigma\Rightarrow \sigma=x.$$ We say $\mu$ is a *controlled map* if it has the following properties:
1. \[cm-1\] For all $x,y\in G$ such that $x\vee y<\infty$ we have $\mu(x)\vee \mu(y)=\mu(x\vee y)$.
2. \[cm-2\] For all $k\in Q$, $\Sigma_k$ is complete in the following sense: for every $x\in \mu^{-1}(k)\cap P$ there exists $\sigma\in \Sigma_k$ such that $\sigma\leq x$.
3. \[cm-3\] For all $k\in Q$ and $\sigma,\tau\in\Sigma_k$ we have $\sigma\vee\tau<\infty\Rightarrow\sigma=\tau$.
\[p-controlled map implies amenable\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups. Suppose that $\mu:G\to K$ is a controlled map. If $K$ is an amenable group and $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is an amenable quasi-lattice ordered group, then $(G,P)$ is amenable.
We start by showing that the kernel of a controlled map is a quasi-lattice ordered group.
Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups, and suppose that $\mu:G\to K$ is a controlled map. Then $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is a quasi-lattice ordered group.
It is clear that $\mu^{-1}(e)$ is a subgroup of $G$ and that$\mu^{-1}(e)\cap P$ is a unital semigroup. Suppose that $x,y\in \mu^{-1}(e)$ have a common upper bound $z\in \mu^{-1}(e)\cap P$. We know that $z$ is a common upper bound for $x,y$ in $P$, and hence $x\vee y$ exists in $P$ and $x\vee y\leq z$. Now $\mu(x\vee y)=\mu(x)\vee \mu(y)=e$, and hence $x\vee y\in \mu^{-1}(e)\cap P$. Thus $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is a quasi-lattice ordered group.
To prove Theorem \[p-controlled map implies amenable\] we will show that the conditional expectation $$E:C^*(G,P)\to \overline{{\operatorname{span}}}\{w_pw^*_p:p\in P\}$$ is faithful. We will use the amenability of $K$ to construct a faithful conditional expectation $\Psi^\mu:C^*(G,P)\to{\overline{{\operatorname{span}}}}\{w_pw^*_q:\mu(p)=\mu(q)\}$, and then show that $E$ is faithful when restricted to ${\operatorname{range}}\Psi^\mu$. To construct $\Psi^\mu$ we follow the method of [@LacaRaeburn1996 Lemma 6.5] which uses a coaction.
Let $G$ be a discrete group and let $A$ be a unital $C^*$-algebra. Let $$\delta_G:C^*(G)\to C^*(G)\otimes_{\min} C^*(G)$$ be the comultiplication of $G$ which is characterised by $\delta_G(u_g)=u_g\otimes u_g$ for $g\in G$. A *coaction* of $G$ on $A$ is a unital homomorphism $\delta:A\to A\otimes_{\min} C^*(G)$ such that $$(\delta\otimes {{\operatorname{id}}})\circ\delta=({{\operatorname{id}}}\otimes \delta_G)\circ \delta.$$ We say that $\delta$ is *nondegenerate* if $\delta(A)(1\otimes C^*(G))=A\otimes_{\min} C^*(G)$.
\[faithful coaction exists\] Let $(G,P)$ be a quasi-lattice ordered group. Suppose that there exists a group $K$ and a homomorphism $\mu:G\to K$. Then there exists an injective coaction $$\delta_\mu:C^*(G,P)\to C^*(G,P)\otimes_{\min} C^*(K)$$ characterised by $\delta_\mu(w_p)=w_p\otimes u_{\mu(p)}$ for all $p\in P$.
Let $W:P\to C^*(G,P)\otimes_{\min}C^*(K)$ be characterised by $W_p=w_p\otimes u_{\mu(p)}$. We will show that $W$ is a covariant representation, and then take $\delta_\mu:=\pi_W$. Unitaries are isometries and hence $W_p$ is isometric for all $p\in P$. Observe that $W_e=w_e\otimes u_{\mu(e)}=1\otimes 1$, and $$W_pW_q=w_pw_q\otimes u_{\mu(p)}u_{\mu(q)}=w_{pq}\otimes u_{\mu(pq)}=W_{pq}\text{ for all $p,q\in P$.}$$ Thus $W$ is an isometric representation. To prove $W$ is covariant, we fix $x,y\in P$ and compute: $$\begin{aligned}
W_xW^*_xW_yW^*_y&=w_xw^*_xw_yw^*_y\otimes u_{\mu(x)}u^*_{\mu(x)}u_{\mu(y)}u^*_{\mu(y)}\\
&=\begin{cases}w_{x\vee y}w^*_{x\vee y}\otimes 1&\text{if $x\vee y<\infty$}\\0\otimes 1&\text{otherwise}\end{cases}\\
&=\begin{cases}w_{x\vee y}w^*_{x\vee y}\otimes w_{\mu(x\vee y)}w^*_{\mu(x\vee y)}&\text{if $x\vee y<\infty$}\\0&\text{otherwise}\end{cases}\\
&=W_{x\vee y}W^*_{x\vee y}.
\end{aligned}$$ Thus $W$ is a covariant representation of $P$. By the universal property of $C^*(G,P)$, there exists a homomorphism $\delta_\mu:=\pi_W$, which has the desired properties. Since $W_e=1\otimes 1$ it follows that $\delta_\mu$ is unital.
To prove the comultiplication identity, we compute on generators: for $p,q\in P$ we have $$\begin{aligned}
((\delta_\mu\otimes{{\operatorname{id}}})\circ \delta_\mu)(w_pw^*_q)&=(\delta_\mu\otimes{{\operatorname{id}}})(w_pw^*_q\otimes u_{\mu(pq^{-1})})\\
&=\delta_\mu(w_pw^*_q)\otimes {{\operatorname{id}}}(u_{\mu(pq^{-1})})\\
&=w_pw^*_q\otimes u_{\mu(pq^{-1})}\otimes u_{\mu(pq^{-1})}\\
&=w_pw^*_q\otimes \delta_K(u_{\mu(pq^{-1})})\\
&={{\operatorname{id}}}\otimes\delta_K(w_pw^*_q\otimes u_{\mu(pq^{-1})})\\
&=(({{\operatorname{id}}}\otimes\delta_K)\circ \delta_\mu)(w_pw^*_q).
\end{aligned}$$ Hence $(\delta_\mu\otimes{{\operatorname{id}}})\circ\delta_\mu=({{\operatorname{id}}}\otimes\delta_K)\circ \delta_\mu$. Thus $\delta_\mu$ is a coaction.
To show that $\delta_\mu$ is injective, let $\pi:C^*(G,P)\to B(H)$ be a faithful representation. We will show that $\pi$ can be written as a composition of $\delta_\mu$ and another representation. Let $\epsilon:C^*(K)\to{\mathbb{C}}$ be the trivial representation on ${\mathbb{C}}$ such that $\epsilon(u_k)=1$ for all $k\in K$. By the properties of the minimal tensor product (see [@RaeburnWilliams1998 Proposition B.13]) there exists a homomorphism $$\pi\otimes \epsilon:C^*(G,P)\otimes_{\min} C^*(G)\to B(H)\otimes{\mathbb{C}}=B(H).$$ Since $$(\pi\otimes \epsilon)\circ \delta_\mu(w_p)=(\pi\otimes \epsilon)(w_p\otimes u_{\mu(p)})=\pi(w_p),$$ we have $\pi=(\pi\otimes \epsilon)\circ \delta_\mu$. Now suppose that $\delta_\mu(a)=0$ for some $a\in C^*(G,P)$. Then $0=(\pi\otimes \epsilon)\circ \delta_\mu(a)=\pi(a)$. Since $\pi$ is faithful, $a=0$. Hence $\delta_\mu$ is injective.
To prove that $\delta_\mu$ is a nondegenerate coaction we must show that $$\delta_\mu(C^*(G,P))(1\otimes C^*(K))=C^*(G,P)\otimes_{\min} C^*(K).$$ It suffices to show that we can get the spanning elements $w_pw^*_q\otimes u_k$, and this is easy: $$\delta_\mu(w_pw^*_q)(1\otimes u_{\mu(qp^{-1}) k})=w_pw^*_q\otimes u_{\mu(pq^{-1})}(1\otimes u_{\mu(qp^{-1}) k})=w_pw^*_q\otimes u_k.$$ Thus $\delta_\mu$ is nondegenerate.
Let $\lambda$ be the left-regular representation of a discrete group $K$. There is a trace $\tau$ on $C^*(K)$ characterised by $$ \tau(u_k)= (\lambda_k\epsilon_e\mid \epsilon_e)=\begin{cases}1 &\text{if $k=e$}\\ 0&\text{otherwise}. \end{cases}$$ It is well-known that if $K$ is an amenable group, then $\tau$ is faithful.
\[conditional expectation construction\] Let $(G,P)$ be a quasi-lattice ordered group. Suppose that there exist a group $K$ and a homomorphism $\mu:G\to K$. Let $$\delta_\mu:C^*(G,P)\to C^*(G,P)\otimes_{\min} C^*(K)$$ be the coaction of Lemma \[faithful coaction exists\]. Then $$\Psi^\mu:=({{\operatorname{id}}}\otimes \tau)\circ\delta_\mu$$ is a conditional expectation of $C^*(G,P)$ with range ${\overline{{\operatorname{span}}}}\{w_pw_q^*: \mu(p)=\mu(q)\}$. If $K$ is an amenable group, then $\Psi^\mu$ is faithful.
Since ${{\operatorname{id}}}\otimes \tau$ and $\delta_\mu$ are linear and norm decreasing, so is $\Psi^\mu$. Since $\Psi^\mu(w_e)=1$ the norm of $\Psi^\mu$ is $1$. We have $$\label{expectation construction eqn}
\Psi^\mu(w_pw^*_q)=\begin{cases}w_pw^*_q&\text{if $\mu(p)=\mu(q)$}\\0&\text{otherwise,}\end{cases}$$ and hence $\Psi^\mu\circ\Psi^\mu=\Psi^\mu$. Thus $\Psi^\mu$ is a conditional expectation by [@Tomiyama1957].
From we see that $\overline{{\operatorname{span}}}\{w_pw^*_q:\mu(q)=\mu(p)\}\subseteq {\operatorname{range}}\Psi^\mu$. To show the reverse inclusion, fix $b\in {\operatorname{range}}\Psi^\mu$, say $b=\Psi^\mu(a)$ for some $a\in C^*(G,P)$. Also fix $\epsilon>0$. There exists a finite subset $F\subseteq P\times P$ such that $\|a-\sum_{(p,q)\in F}\lambda_{p,q}w_{p}w^*_{q}\|<\epsilon$. Since $\Psi^\mu$ is linear and norm-decreasing, $$\begin{aligned}
\epsilon&>\Big\|a-\sum_{(p,q)\in F}\lambda_{p,q}w_{p}w^*_{q}\Big\|
\geq \Big\|\Psi^\mu\Big(a-\sum_{(p,q)\in F}\lambda_{p,q}w_{p}w^*_{q}\Big)\Big\|\\
&
=\Big\|\Psi^\mu(a)-\Psi^\mu\Big(\sum_{(p,q)\in F}\lambda_{p,q}w_{p}w^*_{q}\Big)\Big\| =\Big\|b-\sum_{(p,q)\in F\text{, }\mu(p)=\mu(q)}\lambda_{p,q}w_{p}w^*_{q}\Big\|.\end{aligned}$$ Thus $b\in \overline{{\operatorname{span}}}\{w_pw^*_q:\mu(q)=\mu(p)\}$, and ${\operatorname{range}}\Psi^\mu=\overline{{\operatorname{span}}}\{w_pw^*_q:\mu(q)=\mu(p)\}$.
Now suppose that $K$ is amenable. To see that $\Psi^\mu$ is faithful, we follow the proof of [@LacaRaeburn1996 Lemma 6.5]. Let $a\in C^*(G,P)$ and suppose that $\Psi^\mu(a^*a)=0$. Let $f$ be an arbitrary state on $C^*(G,P)$. Then $$\begin{aligned}
0&=f(\Psi^\mu(a^*a))
=f\circ({{\operatorname{id}}}\otimes \tau)\circ \delta_\mu(a^*a)\\
&=(f\otimes \tau)\circ \delta_\mu(a^*a)
=\tau\circ(f\otimes {{\operatorname{id}}})\circ\delta_\mu(a^*a).
\end{aligned}$$ Since $K$ is amenable, $\tau$ is faithful. Hence $(f\otimes {{\operatorname{id}}})\circ\delta(a^*a)=0$. This implies that for all states $f$ on $C^*(G,P)$ and states $g$ on $C^*(K)$, $$g\circ (f\otimes {{\operatorname{id}}})\circ\delta_\mu(a^*a)=(f\otimes g)\circ\delta_\mu(a^*a)=0.$$
To see that $\delta_\mu(a^*a)=0$, let $\pi_1:C^*(G,P)\to H_1$ and $\pi_2:C^*(K)\to H_2$ be faithful representations. Then $\pi_1\otimes\pi_2$ is a faithful representation of $C^*(G,P)\otimes_{\min}C^*(K)$ on $B(H_1\otimes H_2)$ by [@RaeburnWilliams1998 Corollary B.11]. Fix unit vectors $h\in H_1$, $k\in H_2$. There exists a state $f_h\otimes f_k$ on $C^*(G,P)\otimes_{\min}C^*(K)$ defined by $$f_h\otimes f_k(x)=(\pi_1\otimes\pi_2(x)(h\otimes k)\mid h\otimes k).$$ Since $(f\otimes g)\circ\delta_\mu(a^*a)=0$ for all states $f$ of $C^*(G, P)$ and $g$ of $C^*(K)$, we have $$\begin{aligned}
0&=f_h\otimes f_k(\delta_\mu(a^*a))\\
&=(\pi_1\otimes\pi_2(\delta_\mu(a^*a))(h\otimes k)\mid h\otimes k)\\
&=(\pi_1\otimes\pi_2(\delta_\mu(a))(h\otimes k)\mid \pi_1\otimes\pi_2(\delta_\mu(a)) h\otimes k)\\
&=\|\pi_1\otimes\pi_2(\delta_\mu(a))(h\otimes k)\|^2.\end{aligned}$$ Hence $\pi_1\otimes\pi_2(\delta_\mu(a^*a))=0$. Since $\pi_1\otimes \pi_2$ is faithful, $\delta_\mu(a^*a))=0$. But $\delta_\mu$ is injective, and hence $a=0$, and $\Psi^\mu$ is faithful.
Next we investigate the structure of $${\operatorname{range}}\Psi^\mu={\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}.$$ In the statement of the next lemma, note that we can view $C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ as a $C^*$-subalgebra of $C^*(G,P)$.
Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups, and suppose that $\mu:G\to K$ a controlled map. Let $k\in Q$, and let $F$ be a finite subset of $\Sigma_k$. Let $$B_{k,F}:={\operatorname{span}}\{w_\sigma Dw^*_\tau:\sigma,\tau\in F,D\in C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)\}.$$ Then $B_{k,F}$ is a closed $C^*$-subalgebra of ${\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}$.
To see that $B_{k,F}$ is contained in ${\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}$, it suffices to consider $D$ of the form $w_\alpha w_\beta^*$ where $\mu(\alpha)=\mu(\beta)=e$. Then $w_\sigma Dw^*_\tau=w_{\sigma\alpha}w_{\tau\beta}^*$ and $\mu(\sigma\alpha)=\mu(\sigma)=k=\mu(\tau)=\mu(\tau\beta)$. Thus $B_{k,F}$ is a subalgebra of ${\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}$.
We will prove the lemma by showing that $B_{k,F}$ is isomorphic to $$M_{F}({\mathbb{C}})\otimes C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big).$$ By Definition \[expanded controlled map definition\], the elements of $F$ have no common upper bound unless they are equal. So $$(w_{\sigma}w^*_{\tau})(w_{\gamma}w^*_{\delta})=w_{\sigma\tau^{-1}(\tau\vee\gamma)}w^*_{\delta\gamma^{-1}(\tau\vee\gamma)}=\begin{cases}w_{\sigma}w^*_{\delta}&\text{if $\tau=\gamma$}\\0&\text{otherwise.}\end{cases}$$ Thus $\{w_{\sigma}w^*_{\tau}:\sigma,\tau\in F\}$ is a set of matrix units in the $C^*$-algebra $\overline{B_{k,F}}$. This gives a homomorphism $\theta:M_{F}({\mathbb{C}})\to \overline{B_{k,F}}$ which maps the matrix units $\{E_{\sigma,\tau}:\sigma,\tau\in F\}$ in $M_{F}({\mathbb{C}})$ to $\{w_{\sigma}w^*_{\tau}:\sigma,\tau\in F\}\subset {B_{k,F}}$. It is easy to check that the formula $$\psi(D)=\sum_{\gamma\in F}w_{\gamma}Dw^*_{\gamma}$$ gives a homomorphism $\psi:C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)\to B_{k,F}$. We have $$\begin{aligned}
\theta(E_{\sigma,\tau})\psi(D)&=w_{\sigma}w^*_{\tau}\sum_{\gamma\in F}w_{\gamma}Dw^*_{\gamma}\\
&=w_{\sigma}w^*_{\tau}w_{\tau}Dw^*_{\tau}\quad\text{($w^*_{\tau}w_{\gamma}=0$ unless $\tau=\gamma$)}\\
&=w_{\sigma}Dw^*_{\tau}\\
&=(\sum_{\gamma\in F}w_{\gamma}Dw^*_{\gamma})w_{\sigma}w^*_{\tau}\\
&=\psi(D)\theta(E_{\sigma,\tau}).
\end{aligned}$$ Each $A\in M_{F}({\mathbb{C}})$ is a linear combination of the $E_{\sigma,\tau}$, and hence $\psi(D)\theta(A)=\theta(A)\psi(D)$ for all $A\in M_{F}({\mathbb{C}})$ and $D\in C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$. Since the ranges of $\theta$ and $\psi$ commute, the universal property of the maximal tensor product gives a homomorphism $\theta\otimes_{\max} \psi$ of $M_{F}({\mathbb{C}})\otimes_{\max}C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ into $\overline{B_{k,F}}$.
By [@RaeburnWilliams1998 Theorem B.18] $$\begin{aligned}
M_{F}({\mathbb{C}})\otimes&_{\max}C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)\\
&={\operatorname{span}}\{E_{\sigma,\tau}\otimes D: \sigma,\tau\in F \text{\ and\ } D\in C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big) \},\end{aligned}$$ with no closure. So the range of $\theta\otimes_{\max} \psi$ is spanned by $\theta(E_{\sigma,\tau})\psi(D)=w_{\sigma}Dw^*_{\tau}$ and hence is $B_{k,F}$. Thus $B_{k,F}$ is a closed $C^*$-subalgebra of ${\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}$.
Let $\{\epsilon_x:x\in P\}$ be the usual basis for $\ell^2(P)$. Let $T$ be the covariant representation of $(G,P)$ on $\ell^2(P)$ such that $T_p\epsilon_x=\epsilon_{px}$, and let $\pi_T$ be the corresponding homomorphism of $C^*(G,P)$ onto $C^*_r(G,P)$ such that $\pi_T(w_p)=T_p$. For $k\in Q$ we consider the subspaces $$H_k:=\{\epsilon_{\gamma z}:\gamma\in\Sigma_k,z\in \mu^{-1}(e)\cap P\}.$$
\[isometric on H\_k\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups, and suppose that $\mu:G\to K$ is a controlled map. Let $k\in Q$. Then
1. $H_k$ is invariant for $\pi_T|_{B_{k,F}}$;
2. \[isometric on H\_k2\] if $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is amenable, then $\pi_T|_{B_{k,F}}$ is isometric on $H_k$.
For (1), let $\sigma,\tau\in F$ and let $x,y\in\mu^{-1}(e)$ and let $\epsilon_{\gamma z}\in H_k$. Then $w_{\sigma}w_{x}w^*_yw^*_{\tau}$ is a spanning element of $B_{k,F}$. Since $\mu(\tau)=k=\mu(\gamma)$ we have $$\pi_T(w_{\sigma}w_{x}w^*_yw^*_{\tau})\epsilon_{\gamma z}=
\begin{cases}\epsilon_{\sigma xy^{-1} z}&\text{if $\gamma=\tau$ and $y\leq z$}\\0&\text{otherwise.}\end{cases}$$ If $\pi_T(w_{\sigma}w_{x}w^*_yw^*_{\tau})\epsilon_{\gamma z}=0$ we are done. Otherwise, to see that $\epsilon_{\sigma xy^{-1} z}$ is back in $H_k$, suppose that $y\leq z$. Then $y^{-1}z\in P$. Since $\sigma x\in \mu^{-1}(k)\cap P$ we have $\epsilon_{(\sigma x)(y^{-1} z)}\in H_k$. It follows that $H_k$ is invariant for $\pi_T|_{B_{k,F}}$.
For (2) suppose that $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is amenable. We will show that $\pi_T|_{B_{k,F}}$ is faithful on $H_k$. Take $B=\sum_{\sigma,\tau\in F} w_{\sigma}D_{\sigma,\tau}w^*_{\tau}\in B_{k,F}$ and suppose that $\pi_T(B)|_{H_k}=0$. Fix $\gamma,\delta\in F$. Then $$T^*_{\gamma}\pi_T(B)T_{\delta}=\pi_T(w^*_{\gamma})\pi_T(B)\pi_T(w_{\delta})=\pi_T(D_{\gamma,\delta})$$ Since $T_{\delta}$ is an injection from $H_{e}$ to $H_k$ and $\pi_T(B)|_{H_k}=0$, it follows that $\pi_T(B)T_{\delta}|_{H_e}=0$. Thus $\pi_T(D_{\gamma,\delta})|_{H_e}=0$.
But the restriction $$(\pi_T|_{C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)})|_{H_e}$$ is the Toeplitz representation of $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$, and hence is faithful by amenability. Thus $D_{\gamma,\delta}=0$. Repeating the argument finitely many times shows that all the $D_{\sigma,\tau}=0$ and hence that $B=0$. Thus $\pi_T|_{B_{k,F}}$ is faithful on $H_k$, and therefore is isometric.
\[faithful on Bk\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups, and suppose that $\mu:G\to K$ is a controlled map. Let $B_k=\overline{{\operatorname{span}}}\{w_pw^*_q:\mu(p)=\mu(q)=k\}$. Let ${\mathcal{F}}$ be the set of all finite sets $F\subseteq \Sigma_k$. Then $B_k=\overline{\cup_{F\in{\mathcal{F}}}B_{k,F}}$. If $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is amenable, then $\pi_T|_{B_{k}}$ is isometric on $H_k$.
Observe that ${\mathcal{F}}$ is a directed set partially ordered by inclusion with $E,F\in {\mathcal{F}}$ majorismajorizeded by $E\cup F$. If $E\subseteq F$, then $B_{k,E}\subseteq B_{k,F}$. Thus $\{B_{k,F}:F\in {\mathcal{F}}\}$ is an inductive system with limit $\overline{\cup_{F\in{\mathcal{F}}}B_{k,F}}$.
For each $F\in{\mathcal{F}}$ we have $B_{k,F}\subseteq B_k$, and $B_k$ is closed. Therefore $\overline{\cup_{F\in{\mathcal{F}}}B_{k,F}}\subseteq B_k$. To prove the reverse inclusion it suffices to show that the spanning elements of $B_k$ are in $B_{k,F}$ for some $F$. Fix $p,q\in P$ such that $\mu(p)=\mu(q)=k$ and consider $w_pw^*_q$. By Definition \[expanded controlled map definition\], the set $\Sigma_k$ of minimal elements is complete, and there exists $\sigma,\tau\in\Sigma_k$ such that $\sigma\leq p$ and $\tau \leq q$. Hence there exists $x,y\in P$ such that $p=\sigma x$ and $q=\tau y$. Thus $w_pw^*_q=w_{\sigma x}w^*_{\tau y}=w_{\sigma}(w_xw^*_y)w^*_{\tau}$ and $w_xw^*_y\in C^*\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$. Since $\{\sigma,\tau\}\in{\mathcal{F}}$ we have $w_pw^*_q\in B_{k,\{\sigma,\tau\}}$. Thus $B_k\subseteq \overline{\cup_{F\in{\mathcal{F}}}B_{k,F}}$, and equality follows.
Finally, suppose that $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is amenable. Then $\pi_T|_{B_{k,F}}$ is isometric on $H_k$ for all $F\in{\mathcal{F}}$ by Lemma \[isometric on H\_k\]. Since $\pi_T$ is isometric on every $B_{k,F}$, its extension to the closure is also isometric.
Let ${\mathcal{I}}$ be the set of all finite sets $I\subset Q$ that are closed under $\vee$ in the sense that $s,t\in I$ and $s\vee t<\infty$ implies that $s\vee t\in I$.
\[core is limit of CI\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups, and suppose that $\mu:G\to K$ is a controlled map. For each $I\in{\mathcal{I}}$ let $$C_I=\overline{\operatorname{span}}\{w_pw^*_q:\mu(p)=\mu(q)\in I\}.$$ Then $C_I$ is a $C^*$-subalgebra of ${\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}$, $C_I={\operatorname{span}}\{B_k:k\in I\}$ and ${\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}=\overline{\cup_{I\in{\mathcal{I}}}C_I}$.
Fix $I\in{\mathcal{I}}$. To see that $C_I$ is a $C^*$-subalgebra, it suffices to show that ${\operatorname{span}}\{w_pw^*_q:\mu(p)=\mu(q)\in I\}$ is a $*$-subalgebra. It’s clearly closed under taking adjoints. Let $p,q,r,s\in P$ such that $\mu(p)=\mu(q)\in I$ and $\mu(r)=\mu(s)\in I$. Then $$w_pw^*_qw_rw^*_s=\begin{cases}w_{pq^{-1}(q\vee r)}w^*_{sr^{-1}(q\vee r)}&\text{if $q\vee r<\infty$}\\0&\text{otherwise.}\end{cases}$$ If $w_pw^*_qw_rw^*_s=0$ we are done. So suppose that $w_pw^*_qw_rw^*_s\neq 0$. Then $q\vee r<\infty$. Since $\mu$ is a controlled map and $\mu(p)=\mu(q)$, by Definition \[expanded controlled map definition\], $$\mu(pq^{-1}(q\vee r))=\mu(q\vee r)=\mu(q)\vee \mu(r).$$ Similarly, $\mu(sr^{-1}(q\vee r))=\mu(q)\vee \mu(r)$. Since $I$ is closed under $\vee$ we have $\mu(q)\vee \mu(r)\in I$, and hence $w_pw^*_qw_rw^*_s\in {\operatorname{span}}\{w_pw^*_q:\mu(p)=\mu(q)\in I\}$. It follows that $C_I$ is a $C^*$-subalgebra.
For each $k\in I$, we have $B_k\subseteq C_I$, and so ${\operatorname{span}}\{B_k:k\in I\}\subseteq C_I$. To show the reverse inclusion observe that for $w_pw^*_q\in C_I$ we have $w_pw^*_q\in B_{\mu(p)}$. Since the finite span of closed subalgebras is closed, $\overline{\operatorname{span}}\{w_pw^*_q:\mu(p)=\mu(q)\in I\}\subseteq{\operatorname{span}}\{B_k:k\in I\}$. Thus $C_I={\operatorname{span}}\{B_k:k\in I\}$.
\[piT is faithful on core\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups, and suppose that $\mu:G\to K$ is a controlled map. If $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is amenable, then $\pi_T$ is faithful on $\overline{\operatorname{span}}\{w_pw^*_q:\mu(p)=\mu(q)\}$.
By Lemma \[core is limit of CI\], $\overline{\operatorname{span}}\{w_pw^*_q:\mu(p)=\mu(q)\}=\overline{\cup_{I\in{\mathcal{I}}}C_I}$. Thus it suffices to show that $\pi_T$ is isometric on each $C_I$. Fix $I\in {\mathcal{I}}$. Suppose that $\pi_T(R)=0$ for some $R\in C_I$. Then there exist $R_k\in B_k$ such that $R=\sum_{k\in I}R_k$ and then $\sum_{k\in I}\pi_T(R_k)=0$.
We claim that if $k\not\leq j$, then $\pi_T(B_k)|_{H_j}=0$ (it then follows that $\pi_T(R_k)|_{H_j}=0$). To prove the claim, it suffices to show that $\pi_T(w_q^*)\epsilon_{\gamma z}=0$ for all $q\in \mu^{-1}(k)\cap P$ and $\epsilon_{\gamma z}\in H_k$. We have $$\pi_T(w_q^*)\epsilon_{\gamma z}=T^*_q\epsilon_{\gamma z}
=\begin{cases}
\epsilon_{q^{-1}\gamma z} &\text{if $q\leq \gamma z$}\\
0&\text{otherwise.}
\end{cases}$$ But $q\leq \gamma z$ implies $k=\mu(q)\leq \mu(\gamma z)=\mu(\gamma)=j$. So $k\not\leq j$ implies $\pi_T(w_q^*)\epsilon_{\gamma z}=0$. Hence $\pi_T(B_k)|_{H_j}=0$ if $k\not\leq j$ as claimed.
Let $l$ be a minimal element of $I$ in the sense that $x\leq l$ implies $x=l$. Then for $k\in I$, we have $k\not\leq l$ unless $k=l$. Now $$0=\sum_{k\in I}\pi_T(R_k)|_{H_l}=\pi_T(R_l)|_{H_l}.$$ Since $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is amenable, $\pi_T|_{B_l}$ is isometric on $H_l$ by Lemma \[faithful on Bk\]. Thus $R_l=0$.
Let $l_2$ be a minimal element of $I\backslash\{l\}$. Then we can repeat the above argument to get $R_{l_2}=0$. Since $I$ is finite, we can continue to conclude that $R=0$.
We can now prove Theorem \[p-controlled map implies amenable\]
Suppose that $K$ is an amenable group. To see $(G,P)$ is amenable, we will show that the conditional expectation $$E:C^*(G,P)\to \overline{{\operatorname{span}}}\{w_pw^*_p:p\in P\}$$ is faithful. Let $\Psi^\mu$ be the conditional expectation of Lemma \[conditional expectation construction\]. We have $$\begin{aligned}
E(\Psi^{\mu}(w_pw^*_q))&=\begin{cases}\Psi(w_pw^*_q)&\text{if $\mu(p)=\mu(q)$}\\0&\text{otherwise}\end{cases}\\
&=\begin{cases}w_pw^*_p&\text{if $p=q$}\\0&\text{otherwise}\end{cases}\\
&=\Psi(w_pw^*_q),\end{aligned}$$ and hence $E=E\circ\Psi^{\mu}$.
Since $K$ is an amenable group, $\Psi^\mu$ is faithful by Lemma \[conditional expectation construction\]. Let $P_z\in B(\ell^2(P))$ be the orthogonal projection onto ${\operatorname{span}}\{\epsilon_z\}$. It is straightforward to show that the diagonal map $\Delta:B(\ell^2(P))\to B(\ell^2(P))$ given by $$\Delta(T)=\sum_{z\in P}P_zTP_z$$ is a conditional expectation such that $\Delta\circ \pi_T=\pi_T\circ E$ and is faithful.
Now suppose that $R\in C^*(G,P)$ and $E(R^*R)=0$. Then $E(\Psi^{\mu}(R^*R))=0$ and so $\pi_T\circ E(\Psi^{\mu}(R^*R))=0$. This gives $\Delta\circ \pi_T(\Psi^\mu(R^*R))=0$. Since $\Delta$ is faithful, it follows that $\pi_T(\Psi^\mu(R^*R))=0$. Since $\big(\mu^{-1}(e),\mu^{-1}(e)\cap P\big)$ is amenable, Lemma \[piT is faithful on core\] implies that $\pi_T$ is faithful on ${\overline{{\operatorname{span}}}}\{w_pw_q^*:\mu(p)=\mu(q)\}={\operatorname{range}}\Psi^\mu$. Thus $\Psi^\mu(R^*R)=0$, and then $R=0$ since $\Psi^{\mu}$ is faithful. Hence $E$ is faithful and $(G,P)$ is amenable.
Quasi-lattice ordered HNN extensions
====================================
Let $G$ be a group, let $A$ and $B$ be subgroups of $G$, and let $\phi:A\to B$ be an isomorphism. The group with presentation $${G^*=\langle G,t\mid t^{-1}at=\phi(a),a\in A\rangle}$$ is the *HNN extension of G* with respect to $A,B$ and $\phi$. For every HNN extension $G^*$ the *height map* is the homomorphism $\theta:G^*\to {\mathbb{Z}}$ such that $\theta(g)=0$ for all $g\in G$ and $\theta(t)=1$.
Let $c,d\in {\mathbb{Z}}$. The Baumslag-Solitar group $${\operatorname{BS}}(c,d)=\langle x,t\mid t^{-1}x^dt=x^c\rangle =\langle x,t\mid tx^c=x^dt\rangle$$ is an HNN extension of ${\mathbb{Z}}$ with respect to $A=d{\mathbb{Z}}$, $B=c{\mathbb{Z}}$ and $\phi:A\to B$ given by $\phi(dn)=cn$ for all $n\in {\mathbb{Z}}$. Then ${\mathbb{Z}}^*$ satisfies the relation $t^{-1}dt=\phi(d)=c$. Let ${\operatorname{BS}}(c,d)^+$ be the subsemigroup of ${\operatorname{BS}}(c,d)$ generated by $x$ and $t$. Spielberg showed in [@Spielberg2012 Theorem 2.11] that $({\operatorname{BS}}(c,d), {\operatorname{BS}}(c,d)^+)$ is quasi-lattice ordered for all $c,d> 0$; he also proved in [@Spielberg2012 Lemma 2.12] that $({\operatorname{BS}}(c,-d),{\operatorname{BS}}(c,-d)^+)$ is not quasi-lattice ordered unless $c=1$.
To work with an HNN extension we use a normal form for its elements from [@LyndonSchupp1977 Theorem 2.1]. We choose $X$ to be a complete set of left coset representatives for $G/A$ such that $xA\neq x'A$ for $x\neq x'\in X$. Similarly, choose a complete set $Y$ of left coset representatives for $G/B$. Then a (right) *normal form relative to $X$ and $Y$* of $g\in G$ is a product $$g=g_0t^{\epsilon_1}g_1t^{\epsilon_2}\ldots g_{n-1}t^{\epsilon_n}g_n$$ where:
1. $g_n$ is an arbitrary element of $G$.
2. If $\epsilon_i=1$, then $g_{i-1}$ is an element of $X$
3. If $\epsilon_i=-1$, then $g_{i-1}$ is an element of $Y$.
By [@LyndonSchupp1977 Theorem 2.1], for every choice of complete left coset representatives $X$ and $Y$, each $g\in G^*$ has a unique normal form.
Our goal is to generalise the properties of the Baumslag-Solitar group with $c, d>0$ to construct quasi-lattice ordered HNN extensions of other quasi-lattice ordered groups.
Let $(G,P)$ be a quasi-lattice ordered group. Let $G^*$ be the HNN extension of $G$ with respect to subgroups $A$ and $B$ with an isomorphism $\phi:A\to B$. Let $P^*$ be the subsemigroup of $G^*$ generated by $P$ and $t$. In general, $(G^*,P^*)$ is not a quasi-lattice ordered group. For example, if $c>1$, then $({\operatorname{BS}}(c,-d),{\operatorname{BS}}(c,-d)^+)$ is not quasi-lattice ordered by [@Spielberg2012 Lemma 2.12]. We need some conditions on our subgroups $A$ and $B$ and on the isomorphism $\phi$ which ensure that $(G^*,P^*)$ is quasi-lattice ordered.
There are two reasons why $({\operatorname{BS}}(c,d),{\operatorname{BS}}(c,d)^+)$ is easy to work with. The first is that there are natural choices of coset representatives: $\{0,\ldots,d-1\}$ for $A=d{\mathbb{Z}}$ and $\{0,\ldots,c-1\}$ for $B=c{\mathbb{Z}}$. The second is that the subgroup isomorphism $\phi:md\mapsto mc$ takes positive elements to positive elements. In particular, using this choice of representatives, every element $\omega\in{\operatorname{BS}}(c,d)^+$ has a unique normal form $$\omega=x^{m_0}tx^{m_1}t\ldots x^{m_{n-1}}tx^{m_n}$$ where $0\leq m_i<d$ for $i<n$ and $m_n\in {\mathbb{N}}$. This choice of coset representatives is associated to the division algorithm on ${\mathbb{N}}$: for every $n\in {\mathbb{N}}$ we can uniquely write $n=md+r$ for some $m\in {\mathbb{N}}$ and $0\leq r\leq d-1$.
In general, for $G^*$ we would like a natural choice of coset representatives for $G/A$ and $G/B$ so that every element of $P^*$ has a unique normal form that is a sequence of elements in $P$ and $t$.
\[(G\*,P\*) is left ql\] Let $(G,P)$ be a quasi-lattice ordered group with subgroups $A$ and $B$. Suppose that:
1. There is an isomorphism $\phi:A\to B$ such that $\phi(A\cap P)=B\cap P$;
2. Every left coset $gA\in G/A$ such that $gA\cap P\neq \emptyset$ has a minimal coset representative $p\in P$: $q\in gA\cap P \Rightarrow p\leq q$;
3. For every $x,y\in B$, $x\vee y<\infty\Rightarrow x\vee y\in B$.
Let $G^*=\langle G,t\mid t^{-1}at=\phi(a),a\in A\rangle$ be the HNN extension of $G$ and let $P^*$ be the subsemigroup of $G^*$ generated by $\{P,t\}$. Then $(G^*,P^*)$ is quasi-lattice ordered.
Before we can prove Theorem \[(G\*,P\*) is left ql\], we need to prove two lemmas. The first shows that elements of $P^*$ are guaranteed to have normal forms made up of elements of $P$ and $t$ if and only if condition (2) of Theorem \[(G\*,P\*) is left ql\] holds. The second is a technical lemma which we will use several times in Theorem \[(G\*,P\*) is left ql\] and in later proofs.
\[right lower bound gives unique normal form in P\] Suppose that $(G,P)$ is a quasi-lattice ordered group with subgroups $A$ and $B$. Suppose that $\phi:A\to B$ is a group isomorphism such that $\phi(A\cap P)=B\cap P$. Let $G^*=\langle G,t\mid t^{-1}at=\phi(a),a\in A\rangle$ be the corresponding HNN extension of $G$ and let $P^*$ be the subsemigroup of $G^*$ generated by $P\cup\{t\}$. Let $$L_A:=\{p\in P: q\in pA\cap P\Rightarrow p\leq q\}.$$ The following two statements are equivalent:
1. Every left coset $gA\in G/A$ such that $gA\cap P\neq \emptyset$ has a minimal coset representative $p\in P$;
2. There exists a complete set $X$ of left coset representatives such that $L_A\subseteq X$ and every $\alpha\in P^*$ has normal form $$\label{normal form}
\alpha=p_0tp_1t\ldots p_{n-1}tp_n \text{\ where $p_i\in L_A$ for all $0\leq i< n$, $p_n\in P$.}$$
Assume (1). Choose a complete set $X$ of coset representatives which contains $L_A$. Let $\alpha\in P^*$. If $\theta(\alpha)=0$ then $\alpha\in P$, and $\alpha$ has form trivially.
We proceed by induction on $\theta(\alpha)\geq 1$. Suppose that $\theta(\alpha)=1$. We may write $\alpha=q_0tq_1$ for some $q_0,q_1\in P$. Then $q_0A\cap P\neq \emptyset$, and there exists $p_0\in L_A$ such that $p_0A=q_0A$ and $p_0\leq q_0$. Thus$p_0^{-1}q_0\in P\cap A$. Hence $q_0 =p_0a$ for some $a\in A\cap P$. Thus $\alpha$ has normal form $$\alpha=p_0atq_1=p_0t\phi(a)q_1.$$ Since $\phi(A\cap P)=B\cap P$ we have $\phi(a)\in P$ and so $X$ satisfies (2).
Suppose that all $\alpha$ with $1\leq \theta(\alpha)\leq k$ have normal form . Consider $\alpha$ with $\theta(\alpha)=k+1$. We write $$\alpha=q_0tq_1t\ldots tq_ktq_{k+1}.$$ By assumption, we can write the first $2k+2$ terms of $\alpha$ in normal form $$p_0tp_1t\ldots p_{k-1}tr_k$$ where $p_i\in L_A$ for $0\leq i<k$ and $r_k\in P$. There exists $p_k\in L_A$ such that $r_kA=p_kA$ and $p_k\leq r_k$. As above, we can write $r_k=p_ka$ for some $a\in A\cap P$. Then $$\begin{aligned}
\alpha&=p_0tp_1t\ldots tr_ktq_{k+1}\\
&=p_0tp_1t\ldots tp_katq_{k+1}\\
&=p_0tp_1t\ldots tp_kt\phi(a)q_{k+1}.
\end{aligned}$$ We set $p_{k+1}=\phi(a)q_{k+1}$, which is in $P$ because $\phi(a)$ is. Then $\alpha=p_0tp_1t\ldots tp_ktp_{k+1}$ has form . By induction, every $\alpha$ has normal form . This implies (2).
For (2) $\Rightarrow$ (1) we argue by contradiction: we will assume (2) holds but (1) doesn’t. Let $X$ be a set of coset representatives satisfying (2), and suppose that there exists some coset $gA$ such that $gA\cap P\neq\emptyset$ which has no minimal coset representative. Let $p\in X$ be the coset representative of $gA$. Since $p$ is not minimal, there exists $q\in gA\cap P$ with $p\not\leq q$. Thus $p^{-1}q\not\in P$. Consider $qt\in P^*$ in normal form: $$\label{qt in normal form}
qt=pp^{-1}qt=pt\phi(p^{-1}q).$$ Since $\phi(A\cap P)=B\cap P$ we have $\phi(p^{-1}q)\not \in P$. So is not the normal form , a contradiction. Thus $(2)\Rightarrow (1)$.
\[PP-1-closed is vee closed\] Let $(G,P)$ be a quasi-lattice ordered group and let $B$ be a subgroup of $G$. Suppose that for every $x,y\in B$, $x\vee y<\infty\Rightarrow x\vee y\in B$. Then for all $x\in B\cap PP^{-1}$, there exist $\mu,\nu\in B\cap P$ such that $x=\mu\nu^{-1}$ and for all $p,q\in P$ with $pq^{-1}=x$ we have $\mu\leq p$ and $\nu\leq q$.
The lemma says that if $x\in PP^{-1}\cap B$, then the minimal elements of must also be contained in $B$. In particular, if $\phi(A\cap P)=B\cap P$, then $\phi(x)\in PP^{-1}$.
Fix $x\in B\cap PP^{-1}$. Say $x=st^{-1}$ with $s,t\in P$. Then $x^{-1}s\in P$ and $x\leq s$. Also $e\leq s$, and so $x\vee e<\infty$. Since $e,x\in B$ we get $x\vee e\in B$. Let $\mu=x\vee e$ and $\nu=x^{-1}(x\vee e)$. Then $\mu\nu^{-1}=x\vee e(x^{-1}(x\vee e))^{-1}=x$.
Fix $p,q\in P$ such that $x=pq^{-1}$. Then $x^{-1}p=(pq^{-1})^{-1}p=q$, and so $x\leq p$. Therefore $\mu=x\vee e\leq p$. Now $\mu^{-1}p\in P$, and then $\nu^{-1}q=\mu^{-1}p\in P$ gives $\nu\leq q$.
We can now prove Theorem \[(G\*,P\*) is left ql\]. Its proof is based on [@Spielberg2012 Theorem 2.11], and our presentation is helped by Emily Irwin’s treatment of [@Spielberg2012 Theorem 2.11] in her University of Otago Honours thesis.
Fix $x\in P^*P^{*-1}$. We shall prove that there exist $\mu,\nu\in P^*$ with $x=\mu\nu^{-1}$ such that whenever $\gamma\delta^{-1}=x$ we have $\mu\leq \gamma$ and $\nu\leq \delta$ (see ).
Choose $\alpha,\beta\in P^*$ such that $x=\alpha\beta^{-1}$. Choose a complete set $X$ of left coset representatives of $G/A$ that contains $$L_A:=\{p\in P: q\in pA\cap P\Rightarrow p\leq q\}.$$ By Lemma \[right lower bound gives unique normal form in P\] we can write $\alpha$ and $\beta$ in unique normal form: $$\alpha=p_0tp_1t\ldots tp_mtr \text{ where $p_i\in L_A$ and $r\in P$};$$ $$\beta=q_0tq_1t\ldots tq_nts \text{ where $q_i\in L_A$ and $s\in P$}.$$ Now $x=\alpha\beta^{-1}$ is equal to $$\alpha\beta^{-1}=p_0tp_1t\ldots tp_mtrs^{-1}t^{-1}q_n^{-1}t^{-1}\ldots t^{-1}q_0^{-1}.$$ First we look for initial cancellations in the middle of $\alpha\beta^{-1}$: if $rs^{-1}\in B$, then we can replace $trs^{-1}t^{-1}$ with $\phi^{-1}(rs^{-1})$. By assumption (3), Lemma \[PP-1-closed is vee closed\] applies and there exist $b_1,b_2\in P\cap B$ such that $rs^{-1}=b_1b_2^{-1}$. Then $$\phi^{-1}(rs^{-1})=\phi^{-1}(b_1b_2^{-1})=\phi^{-1}(b_1)\phi^{-1}(b_2)^{-1}.$$ Since $\phi(A\cap P)=B\cap P$ we have $\phi^{-1}(b_1)\phi^{-1}(b_2)^{-1}\in PP^{-1}$. Then $$\label{alpha beta inverse}
x=\alpha\beta^{-1}=p_0tp_1t\ldots tp_m\phi^{-1}(b_1)\phi^{-1}(b_2)^{-1}q_n^{-1}t^{-1}\ldots t^{-1}q_0^{-1}.$$ We can repeat this process until there are no more cancellations available in the middle, and so we assume this is the case for the expression . This gives the following cases:
1. there are no $t$ and no more $t^{-1}$,
2. there are no more $t^{-1}$,
3. there are no more $t$,
4. there are $t$ and $t^{-1}$, and then the term with $t$ to the left and $t^{-1}$ to its right is not in $B$.
In each case, we will write down our candidates for $\mu$ and $\nu$ and prove that they are the required minimums.
\(1) Suppose that after initial cancellations, there are no more $t$ and no more $t^{-1}$. Then $\alpha\beta^{-1}=p_0q_0^{-1}$ is already in normal form. By there exist $\sigma,\tau\in P$ such that $p_0q_0^{-1}=\sigma\tau^{-1}$ and for all $c,d\in P$ such that $cd^{-1}=\sigma\tau^{-1}$ we have $\sigma\leq c$ and $\tau\leq d$. So we write $x=\sigma\tau^{-1}$ and choose as our candidates $\mu=\sigma$ and $\nu=\tau$.
Let $\gamma,\delta\in P^*$ such that $x=\gamma\delta^{-1}$. Let $\theta$ be the height map. Then $\theta(x)=0$ and hence $\theta(\gamma)=\theta(\delta)$. We will prove that $\mu\leq \gamma$ and $\nu\leq \delta$ by induction on $\theta(\gamma)$.
For $\theta(\gamma)=0$ we have $\gamma,\delta\in P$ , and then $\mu=\sigma\leq\gamma$ and $\nu=\tau\leq \delta$. Let $k\geq 0$ and suppose that for all $\gamma,\delta\in P^*$ such that $\theta(\gamma)=\theta(\delta)=k$ and $x=\gamma\delta^{-1}$ we have $\mu\leq \gamma$ and $\nu\leq \delta$. Now consider $\gamma,\delta\in P^*$ such that $x=\gamma\delta^{-1}$ and $\theta(\gamma)=\theta(\delta)=k+1$. We write $\gamma=m_0t\ldots m_ktm_{k+1}$ and $\delta=n_0t\ldots n_ktn_{k+1}$ in normal form where $m_i,n_i\in L_A$ for $0\leq i\leq k$ and $m_{k+1}, n_{k+1}\in P$. Next we reduce $x=\gamma\delta^{-1}$ towards normal form. We have $$x=\gamma\delta^{-1}=m_0t\ldots m_ktm_{k+1}n_{k+1}^{-1}t^{-1}n_k\ldots tn_0^{-1}.$$ Since $x$ has a unique normal form with no $t$ or $t^{-1}$, there must be some cancellation. Since the $m_i, n_i\in L_A$ for $0\leq i\leq k$, the cancellation must occur across $m_{k+1}n_{k+1}^{-1}$. So $m_{k+1}n_{k+1}^{-1}\in B$ and $tm_{k+1}n_{k+1}^{-1}t^{-1}=\phi^{-1}(m_{k+1}n_{k+1}^{-1})$, and $$\begin{aligned}
x&=\gamma\delta^{-1}=m_0t\ldots tm_k(tm_{k+1}n_{k+1}^{-1}t^{-1})n_kt^{-1}\ldots tn_0^{-1}\\
&=m_0t\ldots tm_k\phi^{-1}(m_{k+1}n_{k+1}^{-1})n_kt^{-1}\ldots tn_0^{-1}
\end{aligned}$$ By assumption (3), Lemma \[PP-1-closed is vee closed\] applies and there exists $b_m,b_n\in B\cap P$ such that $m_{k+1}n_{k+1}^{-1}=b_mb_n^{-1}$ and $b_m\leq m_{k+1}$ and $b_n\leq n_{k+1}$. Then $$x=m_0t\ldots tm_k\phi^{-1}(b_m)\phi^{-1}(b_n^{-1})n_kt^{-1}\ldots tn_0^{-1}.$$ Since $\phi(A\cap P)=B\cap P$ we have $m_{k+1}\phi^{-1}(b_m),n_{k+1}\phi^{-1}(b_n)\in P$. But now we have $\gamma'=m_0t\ldots tm_k\phi^{-1}(b_m)$ and $\delta'=n_0t\ldots tn_k\phi^{-1}(b_n)$ such that $\gamma'(\delta')^{-1}=x$ and $\theta(\gamma)=\theta(\delta')=k$. By our induction hypothesis we have $\mu\leq \gamma'$ and $\nu\leq \delta'$.
To show that $\mu\leq \gamma$ we compute: $$\begin{aligned}
\gamma&=m_0t\ldots tm_ktm_{k+1}\\
&=m_0t\ldots tm_ktb_mb_m^{-1}m_{k+1}\\
&=m_0t\ldots tm_k\phi^{-1}(b_m)tb_m^{-1}m_{k+1}\quad\text{(replacing $tb_m$ with $\phi^{-1}(b_m)t$)}\\
&=\gamma'tb_m^{-1}m_{k+1}.
\end{aligned}$$ Since $b_m^{-1}m_{k+1}\in P$ we have $tb_m^{-1}m_{k+1}\in P^*$. Now $(\gamma')^{-1}\gamma\in P^*$ and $\gamma'\leq \gamma$. Since $\mu\leq \gamma'$ this gives $\mu\leq \gamma$. Similarly, $\delta=\delta'tb_n^{-1}n_{k+1}$ and so $\nu\leq \delta$. By induction, for all $\gamma,\delta\in P^*$ such that $x=\gamma\delta^{-1}$ we have $\mu\leq\gamma$ and $\nu\leq \delta$.
\(2) Suppose that after the initial cancellations there are no more $t^{-1}$ left. Then we have $x$ in normal form: $$x=\alpha\beta^{-1}=p_0tp_1t\ldots tp_{m-n-1}trs^{-1}$$ where $m\geq n$. By there exist $\sigma,\tau\in P$ such that $rs^{-1}=\sigma\tau^{-1}$ and $\sigma\leq r$ and $\tau\leq s$. So $$\label{eq-x}
x=p_0tp_1t\ldots tp_{m-n-1}t\sigma\tau^{-1}.$$ Our candidates are $$\mu=p_0tp_1t\ldots tp_{m-n-1}t\sigma\quad\text{and}\quad \nu=\tau.$$ Fix $\gamma,\delta\in P^*$ such that $x=\gamma\delta^{-1}$. Say $\gamma=m_0tm_1\ldots tm_i$ and $\delta= n_0t n_1\ldots n_j$ in normal form. Then $$\gamma\delta^{-1}=m_0tm_1\ldots tm_in_j^{-1}t^{-1}\ldots n_1t^{-1}n_0.$$ Since $\theta(\gamma\delta^{-1}) = \theta(\mu \nu^{-1})=m-n$ we get $$i=\theta(\gamma)=\theta(\delta)+m-n$$ and hence $i\geq m-n$. It follows from the uniqueness of normal form that there exists $\gamma'\in P^*$ such that $$\gamma\delta^{-1}=p_0tp_1t\ldots tp_{m-n-1}t\gamma'\delta^{-1}.$$ Thus $p_0tp_1t\ldots tp_{m-n-1}t\leq\gamma$. From we have $(p_0tp_1t\ldots tp_{m-n-1}t)^{-1}x=\sigma\tau^{-1}$. Since $ (p_0tp_1t\ldots tp_{m-n-1}t)^{-1}\gamma,\delta\in P^*$ and $(p_0tp_1t\ldots tp_{m-n-1}t)^{-1}\gamma\delta=\sigma\tau^{-1}$, we can apply (1) above with $\mu'=\sigma$ and $\nu'=\tau$ to see that $\sigma\leq(p_0tp_1t\ldots tp_{m-n-1}t)^{-1}\gamma$ and $\tau\leq \delta$. Hence $\mu= p_0tp_1t\ldots tp_{m-n-1}t\sigma \leq \gamma$ and $\nu\leq \delta$ as required.
\(3) Suppose that after the cancellations, there are no more $t$ left. Then $$x=rs^{-1}t^{-1}q_{n-m-1}^{-1}t^{-1}\ldots t^{-1}q_0^{-1}$$ for some $r,s\in P$. Consider $$x^{-1}=q_0tq_1t\ldots tq_{n-m-1}tsr^{-1}.$$ By there exist $\sigma,\tau\in P$ such that $rs^{-1}=\sigma\tau^{-1}$ and $\sigma\leq r$ and $\tau\leq s$. Let $$\mu=\sigma \quad\text{and}\quad\nu=q_0tq_1t\ldots tq_{n-m-1}t\tau.$$ By (2) they have the property that $x^{-1}=\nu\mu^{-1}$ and for all $\gamma,\delta\in P^*$ such that $x^{-1}=\delta\gamma^{-1}$ we have $\mu\leq\gamma$ and $\nu\leq\delta$. Taking inverses, $x=\mu\nu^{-1}$ and for all $\gamma,\delta\in P^*$ such that $x=\gamma\delta^{-1}$ we have $\mu\leq\gamma$ and $\nu\leq\delta$.
\(4) Suppose that after the initial cancellations there are both $t$ and $t^{-1}$ left. Then the term with $t$ to the left and $t^{-1}$ to its right is not in $B$. There exist $k\leq m$ and $l\leq n$ such that $$x=p_0tp_1t\ldots tp_ktrs^{-1}t^{-1}q_l^{-1}t^{-1}\ldots t^{-1}q_0^{-1}.$$ By there exist $\sigma,\tau\in P$ such that $rs^{-1}=\sigma\tau^{-1}$ and $\sigma\leq r$ and $\tau\leq s$. Our candidate for $\mu,\nu$ are $$\mu=p_0tp_1t\ldots tp_kt\sigma\quad\text{and}\quad \nu=q_0tq_1t\ldots tq_lt\tau.$$ Fix $\gamma,\delta\in P^*$ such that $x=\gamma\delta^{-1}$. By the argument used in (2), there exists $\gamma'\in P^*$ such that $\gamma\delta^{-1}=p_0tp_1t\ldots tp_{j}t\gamma'\delta^{-1}$. Hence $\gamma'=(p_0tp_1t\ldots tp_{j}t)^{-1}\gamma\in P^*$ and $p_0tp_1t\ldots tp_{j}t\leq \gamma$.
Consider $\gamma'\delta^{-1}=\sigma\tau^{-1}t^{-1}q_k^{-1}t^{-1}\ldots t^{-1}q_0^{-1}$. Here $\gamma',\delta\in P^*$ and there are no $t$ in $\gamma'\delta^{-1}$ after cancellation. Applying (3) with $\mu'=\sigma$ and $\nu'=\nu$ to get $\mu'=\sigma \leq \gamma'$ and $\nu'=\nu \leq \delta$. Then $\mu=p_0tp_1t\ldots tp_{j}t\sigma\leq p_0tp_1t\ldots tp_{j}t\gamma' =\gamma$.
Theorem \[(G\*,P\*) is left ql\] gives new examples of quasi-lattice ordered HNN extensions.
We can show that the Baumslag-Solitar group $({\operatorname{BS}}(c,d),{\operatorname{BS}}(c,d)^+)$ with $c,d>0$ is quasi-lattice ordered using Theorem \[(G\*,P\*) is left ql\]. Since $({\mathbb{Z}},{\mathbb{N}})$ is totally ordered it is quasi-lattice ordered. Let $A=\{dm:m\in {\mathbb{Z}}\}$ and $B=\{cm:m\in {\mathbb{Z}}\}$. Every element $n\in {\mathbb{N}}$ has a unique decomposition $n=r+md$ where $m\in {\mathbb{N}}$ and $r\in \{0,1,\ldots, d-1\}$. The remainder $r$ is a choice of coset representative $n+A=r+A$. For all $n'\in n+A\cap {\mathbb{N}}$ we have $n'=r+md+kd$ where $k\in {\mathbb{Z}}$ and $k\leq m$. Thus $r\leq n'$. Hence every coset of ${\mathbb{Z}}/A$ has nontrivial intersection with ${\mathbb{N}}$, and has a minimal coset representative in ${\mathbb{N}}$. Since $B$ is totally ordered it is closed under taking least upper bounds. Define $\phi:A\to B$ by $\phi(dm)=cm$. Then $\phi(A\cap{\mathbb{N}})=B\cap {\mathbb{N}}$. So Theorem \[(G\*,P\*) is left ql\] applies and gives that $({\mathbb{Z}}^*,{\mathbb{N}}^*)$ is quasi-lattice ordered.
\[example Z\^2\] We can generalise the previous example to $({\mathbb{Z}}^2,{\mathbb{N}}^2)$, which is quasi-lattice ordered by [@Nica1992 Example 2.3(2)]. Fix $a,b,c,d\in {\mathbb{N}}$. Then $A=\{(am,bn):m,n\in {\mathbb{Z}}\}$ and $B=\{(cm,dn):m,n\in {\mathbb{Z}}\}$ are subgroups of ${\mathbb{N}}^2$. Let $\phi:A\to B$ be defined by $\phi((am,bn))=(cm,dn)$. This $\phi$ satisfies $\phi(A\cap {\mathbb{N}}^2)=B\cap {\mathbb{N}}^2$. For all $(m,n)\in {\mathbb{N}}^2$, the division algorithm on ${\mathbb{N}}$ gives a unique decomposition $$(m,n)=(r_1,r_2)+(ja_1,ka_2) \text{\ for $j,k\in {\mathbb{N}}$ and $r_1\in \{0,\ldots a_1-1\}$, $r_2\in \{0,\ldots,a_2-1\}$.}$$ Thus $(r_1,r_2)$ is a minimal left coset representative of $(m,n)+A$. For all $(m,n),(p,q)\in B$, we have $$(m,n)\vee(p,q)=(\max\{m,p,0\},\max\{n,q,0\})$$ and hence $(m,n)\vee(p,q)\in B$. So $B$ is closed under $\vee$. By Theorem \[(G\*,P\*) is left ql\], $({\mathbb{Z}}^{2*},{\mathbb{N}}^{2*})$ is a quasi-lattice ordered group with presentation $${\mathbb{Z}}^{2*}=\langle{\mathbb{Z}}^2\cup\{t\}\mid (am,bn)t=t(cm,dn)\rangle.$$ It is straightforward to extend this construction to $({\mathbb{Z}}^n,{\mathbb{N}}^n)$.
\[example free group\] Consider the free group ${\mathbb{F}}_2$ on $2$ generators $\{a,b\}$ and let ${\mathbb{F}}_2^+$ be the subsemigroup generated by $e$, $a$ and $b$. The pair $({\mathbb{F}}_2, {\mathbb{F}}_2^+)$ is quasi-lattice ordered by [@Nica1992 Example 2.3(4)]. Let $A=\{ a^n:n\in{\mathbb{Z}}\}$, $B=\{ b^n:n\in{\mathbb{Z}}\}$ and $\phi:A\to B$ defined by $\phi(a^n)=b^n$. Every $x\in F_2^+$ can be written as a product of $y\in {\mathbb{F}}_2^+$ which does not end in $a$ followed by $a^n$ for some $n\geq 0$. Then $y\in xA$. Every $z\in yA\cap {\mathbb{F}}_2^+$ begins with the word $y$ which is in ${\mathbb{F}}_2^+$. It follows that $y\leq z$. Thus $A$ has minimal left coset representatives in ${\mathbb{F}}_2^+$. Since $B$ is totally ordered, it is trivially closed under $\vee$. It follows from Theorem \[(G\*,P\*) is left ql\] that $({\mathbb{F}}_2^*,{\mathbb{F}}_2^{+*})$ is a quasi-lattice ordered group with presentation $${\mathbb{F}}_2^*=\langle \{a,b,t\}\mid at=tb\rangle.$$
\[example free group 2\] Building on $({\mathbb{F}}_2, {\mathbb{F}}_2^+)$ again, fix $s,u\in {\mathbb{N}}$, and let $A=\{ a^{ms}:m\in {\mathbb{Z}}\}$, $B=\{b^{mu}:m\in{\mathbb{Z}}\}$ and $\phi:A\to B$ be $\phi( a^{ms}) = b^{mu}$. Then $B$ is totally ordered and hence is closed under $\vee$. To see that $A$ has minimal coset representatives, we observe that every $x\in {\mathbb{F}}_2^+$ is a product of a $y\in {\mathbb{F}}_2^+$ that does not end in $a$ followed by $a^n$ for some $n\in {\mathbb{N}}$. We write $n=r+js$ for some $j\in {\mathbb{N}}$ and $r\in \{0,1,\ldots,s-1\}$. We choose $ya^r$ as our coset representative. Then for all $z\in ya^rA\cap {\mathbb{F}}_2^+$ we have $ya^r\leq z$. It follows from Theorem \[(G\*,P\*) is left ql\] that $({\mathbb{F}}_2^*,{\mathbb{F}}_2^{+*})$ is a quasi-lattice ordered group with presentation $${\mathbb{F}}_2^*=\langle \{a,b,t\}\mid a^st=tb^u\rangle.$$ Taking $u=s=1$ gives Example \[example free group\].
Replacing $B$ by $B'=\{a^{mu}:m\in{\mathbb{Z}}\}$ gives a quasi-lattice ordered group $({\mathbb{F}}_2^*,{\mathbb{F}}_2^{+*})$ with presentation $${\mathbb{F}}_2^*=\langle \{a,b,t\}\mid a^st=ta^u\rangle.$$
In the next two examples we show that it is easy to find subgroups which do not have minimal left coset representatives.
Consider the group $$G={\mathbb{Z}}(\sqrt{2})=\{m+n\sqrt{2}:m,n\in {\mathbb{Z}}\}$$ with subsemigroup ${\mathbb{Z}}(\sqrt{2})^+={\mathbb{Z}}(\sqrt{2})\cap [0,\infty)$. Let $A={\mathbb{Z}}(2\sqrt{2})$. We claim there are no minimal coset representatives for $A$. Suppose, aiming for a contradiction, that there exists some coset representative $p\in {\mathbb{Z}}(\sqrt{2})^+$ such that $$q\in [p+{\mathbb{Z}}(2\sqrt{2})]\cap {\mathbb{Z}}(\sqrt{2})^+\Rightarrow p\leq q.$$ Recall that ${\mathbb{Z}}(2\sqrt{2})$ is dense in ${\mathbb{R}}$.[^3] Thus there exists some $a\in {\mathbb{Z}}(2\sqrt{2})\cap (0,p)$ hence $a\leq p$. Thus $p-a\in [p+{\mathbb{Z}}(2\sqrt{2})]\cap {\mathbb{Z}}(\sqrt{2})^+$. But $p-a\leq p$, giving a contradiction.
Consider $({\mathbb{Z}}^2,{\mathbb{N}}^2)$, and let $A$ be the subgroup generated by $\{(1,2),(2,1)\}$. Consider the coset $$(1,0)+A=(0,1)+A.$$ Since $(1,0)$ and $(0,1)$ have no nonzero lower bound, there can be no choice of minimal coset representative.
Amenability of $(G^*, P^*)$
===========================
In this section we prove the following theorem.
\[HNN extensions are amenable as expected\] Let $(G,P)$ be a quasi-lattice ordered group with subgroups $A$ and $B$. Suppose that $\phi:A\to B$ is an isomorphism which satisfies the hypotheses of Theorem \[(G\*,P\*) is left ql\]. Let $(G^*,P^*)$ be the corresponding HNN extension. If $(G,P)$ is amenable, then $(G^*,P^*)$ is amenable.
To prove the theorem we will show that the height map $\theta: G^*\to{\mathbb{Z}}$ is a controlled map, that $(\theta^{-1}(e), \theta^{-1}(e)\cap P^*)$ is amenable, and then apply Theorem \[p-controlled map implies amenable\]. To prove that $(\theta^{-1}(e), \theta^{-1}(e)\cap P^*)$ is amenable, we start by investigating order-preserving isomorphisms between the semigroups of quasi-lattice ordered groups.
\[isomorphism preserves least upper bounds\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups. Suppose that there is a semigroup isomorphism $\phi:P\to Q$. Then $\phi$ is order-preserving. In particular, for $x,y\in P$, $x\vee y<\infty$ if and only if $\phi(x)\vee\phi(y)<\infty$. If $x\vee y<\infty$ then $\phi(x\vee y)=\phi(x)\vee\phi(y)$.
To see that $\phi$ is order-preserving, let $x,y\in P$ and $x\leq y$. Then $y\in xP$, and $\phi(y)\in \phi(x)\phi(P)=\phi(x)Q$. Thus $\phi(x)\leq\phi(y)$, and $\phi$ is order-preserving.
To show that $\phi$ preserves the least upper bound structure, first suppose that $x,y\in P$ such that $x\vee y<\infty$. Since $\phi$ is order-preserving it follows that $\phi(x),\phi(y)\leq \phi(x\vee y)$. Thus $\phi(x),\phi(y)$ have a common upper bound in $Q$. Hence $\phi(x)\vee\phi(y)$ exists and $\phi(x)\vee\phi(y)\leq \phi(x\vee y)$. Second, suppose that $\phi(x)\vee\phi(y)<\infty$ for some $x, y\in P$. Since $\phi^{-1}:Q\to P$ is a semigroup isomorphism it is order-preserving. Thus $\phi^{-1}(\phi(x)\vee\phi(y))$ is an upper bound for $x=\phi^{-1}(\phi(x))$ and $y=\phi^{-1}(\phi(y))$. Hence $x\vee y$ exists and $x\vee y\leq \phi^{-1}(\phi(x)\vee\phi(y))$. Thus $$\phi(x\vee y)\leq\phi\circ\phi^{-1}(\phi(x)\vee\phi(y))=\phi(x)\vee\phi(y).$$ Hence $x\vee y<\infty$ if and only if $\phi(x)\vee\phi(y)<\infty$, and $\phi(x\vee y)=\phi(x)\vee\phi(y)$.
\[semigroup isomorphism preserves universal and amenability\] Let $(G,P)$ and $(K,Q)$ be quasi-lattice ordered groups. Let $\{v_p:p\in P\}$ and $\{w_q:q\in Q\}$ be the generating elements of $C^*(G,P)$ and $C^*(K,Q)$, respectively. Suppose that there is a semigroup isomorphism $\phi:P\to Q$.
1. There exists an isomorphism $\pi_{\phi}:C^*(G,P)\to C^*(K,Q)$ such that $\pi_\phi(v_p)=w_{\phi(p)}$.
2. $(G,P)$ is amenable if and only if $(K,Q)$ is amenable.
\(1) We will show that $T:P\to C^*(K,Q)$ defined by $T_p=w_{\phi(p)}$ is a covariant representation of $P$, and then take $\pi_\phi:=\pi_T$. Fix $p,q\in P$. Since $\phi$ is a semigroup isomorphism we have $$T_pT_q=w_{\phi(p)}w_{\phi(q)}=w_{\phi(pq)}=T_{pq}$$ and $T_e=w_{\phi(e_G)}=w_{e_K}=1$. Hence $T$ is an isometric representation. We have $$\begin{aligned}
T_pT^*_pT_qT^*_q&=w_{\phi(p)}w^*_{\phi(p)}w_{\phi(q)}w^*_{\phi(q)}\\
&=\begin{cases}w_{\phi(p)\vee\phi(q)}w^*_{\phi(p)\vee\phi(q)}&\text{if $\phi(p)\vee\phi(q)<\infty$}\\0&\text{otherwise}\end{cases}\\
\intertext{which, using Lemma~\ref{isomorphism preserves least upper bounds}, is equivalent to}
&=\begin{cases}w_{\phi(p\vee q)}w^*_{\phi(p\vee q)}&\text{if $p\vee q<\infty$}\\0&\text{otherwise}\end{cases}\\
&=\begin{cases}T_{p\vee q}T^*_{p\vee q}&\text{if $p\vee q<\infty$}\\0&\text{otherwise.}\end{cases}\end{aligned}$$ Hence $T$ is covariant. By the universal property of $C^*(G,P)$ there exists a homomorphism $\pi_\phi:C^*(G,P)\to C^*(K,Q)$ such that $\pi_\phi(v_p)=T_p=w_{\phi(p)}$.
Since $\phi^{-1}:Q\to P$ is an isomorphism the argument above gives a homomorphism $\pi_{\phi^{-1}}:C^*(K,Q)\to C^*(G,P)$ such that $\pi_{\phi^{-1}}(w_q)=v_{\phi^{-1}(q)}$. In particular, $$\pi_{\phi^{-1}}(\pi_\phi(v_p))=\pi_{\phi^{-1}}(w_{\phi(p)})=v_{\phi^{-1}(\phi(p))}=v_p$$ and $\pi_{\phi}(\pi_{\phi^{-1}}(w_q))=w_q$ It follows that $\pi_\phi$ is an isomorphism from $C^*(G,P)$ to $C^*(K,Q)$.
\(2) By symmetry it suffices to show that if $(K,Q)$ is amenable then $(G,P)$ is amenable. Let $E_Q$ and $E_P$ be the conditional expectations on $C^*(K,Q)$ and $C^*(G,P)$, respectively. To see $E_P=\pi_{\phi}^{-1}\circ E_Q\circ \pi_{\phi}$ we compute on spanning elements: $$\begin{aligned}
\pi_{\phi}^{-1}\circ E_Q\circ \pi_{\phi}(v_pv^*_q)&=\pi_{\phi}^{-1}\circ E_Q(w_{\phi(p)}w^*_{\phi(q)})\\
&=\begin{cases}\pi_{\phi}^{-1}(w_{\phi(p)}w^*_{\phi(q)})&\text{if $\phi(p)=\phi(q)$}\\0&\text{otherwise}\end{cases}\\
&=\begin{cases}\pi_{\phi}^{-1}(w_{\phi(p)}w^*_{\phi(q)})&\text{if $p=q$}\\0&\text{otherwise}\end{cases}\\
&=\begin{cases}v_pv^*_q&\text{if $p=q$}\\0&\text{otherwise}\end{cases}\\
&=E_P(v_pv^*_q).
\end{aligned}$$ Suppose that $(K,Q)$ is amenable. Then $E_Q$ is faithful. Suppose that $a\geq 0$ and $E_P(a)=0$. Then $\pi_\phi(a)\geq 0$, and $$0=E_P(a)=\pi_{\phi}^{-1}\circ E_Q\circ \pi_{\phi}(a) \Rightarrow 0=E_Q\circ \pi_{\phi}(a) \Rightarrow 0=\pi_{\phi}(a)$$ because $E_Q$ is faithful. Since $\pi_{\phi}$ is faithful, $a=0$. Now $E_P$ is faithful, and hence $(G, P)$ is amenable.
Next we need some lemmas which will be used to show that the height map $\theta$ is a controlled map. In particular we need to identify the minimal elements of Definition \[expanded controlled map definition\]. If $x\in P^*$ has normal form $$x=p_0tp_1t\ldots p_{n-1}tp_n$$ we call $p_0tp_1t\ldots p_{n-1}t$ the *stem* of $x$ and write $${\operatorname{stem}}(x)=p_0tp_1t\ldots p_{n-1}t.$$ The set of stems is our candidate for the minimal elements.
\[p,q have cub in P\* iff they have cub in P\] Let $(G,P)$ be a quasi-lattice ordered group with subgroups $A$ and $B$. Suppose that $\phi:A\to B$ is an isomorphism which satisfies the hypotheses of Theorem \[(G\*,P\*) is left ql\]. Let $p,q\in P$. Then $p$ and $q$ have a common upper bound in $P^*$ if and only if $p$ and $q$ have a common upper bound in $P$.
First suppose that $p$ and $q$ have a common upper bound $r\in P$. Then $r\in P^*$ and so $r$ is a common upper bound for $p$ and $q$ in $P^*$.
Second, suppose that $p$ and $q$ have a common upper bound $x\in P^*$. If $\theta(x)=0$, then $x\in P$ and we are done. Suppose, aiming for a contradiction, that $\theta(x)=k$ for some $k\geq 1$, and that $p,q$ have no common upper bound $y$ with $\theta(y)<k$.
Observe that $p^{-1}x,q^{-1}x\in P^*$, and that $\theta(p^{-1}x)=\theta(x)=\theta(q^{-1}x)=k$. We write $p^{-1}x$ and $q^{-1}x$ in their normal forms: $$p^{-1}x=p_0tp_1t\ldots p_{k-1}tp_k\quad\text{and}\quad q^{-1}x=q_0tq_1t\ldots q_{k-1}tq_k,$$ where $p_i,q_i\in L_A$ for $i<k$ and $p_k,q_k\in P$. Consider $$p^{-1}q=p^{-1}x(q^{-1}x)^{-1}=p_0tp_1t\ldots p_{k-1}tp_kq_k^{-1}t^{-1}q_{k-1}^{-1}\ldots t^{-1}q_0^{-1}.$$ Since $p^{-1}x(q^{-1}x)^{-1}=p^{-1}q$ and $p^{-1}q$ is in normal form we must have some cancellation. Since the first $k$ terms are already in normal form, $p_kq_k^{-1}\in B$. By Lemma \[PP-1-closed is vee closed\], there exist $b_1,b_2\in B\cap P$ such that $b_1\leq p_k$, $b_2\leq q_k$ and $b_1b_2^{-1}=p_kq_k^{-1}$. Then $$\begin{aligned}
p^{-1}q&= p^{-1}x(q^{-1}x)^{-1}=p_0t\ldots p_{k-1}t(p_kq_k^{-1})t^{-1}q_{k-1}^{-1}\ldots t^{-1}q_0^{-1}\\
&=p_0t\ldots p_{k-1}t(b_1b_2^{-1})t^{-1}q_{k-1}^{-1}\ldots t^{-1}q_0^{-1} \\
&=p_0t\ldots p_{k-1}\phi^{-1}(b_1)tt^{-1}\phi^{-1}(b_2)^{-1}q_{k-1}^{-1}\ldots t^{-1}q_0^{-1} \\
&=p_0t\ldots p_{k-1}\phi^{-1}(b_1)\phi^{-1}(b_2)^{-1}q_{k-1}^{-1}\ldots t^{-1}q_0^{-1}.
\end{aligned}$$ Rearranging gives $$p(p_0t\ldots tp_{k-1}\phi^{-1}(b_1))=q(q_0tq_1t\ldots tq_{k-1}\phi^{-1}(b_2))\in P^*.$$ Therefore $y=p(p_0tp_1t\ldots p_{k-1}\phi^{-1}(b_1))$ is a common upper bound for $p$ and $q$ in $P^*$ and $\theta(y)=k-1$, giving us the contradiction we sought. Therefore $p$ and $q$ have a common upper bound $y$ with $\theta(y)=0$, and hence they have a common upper bound in $P$.
The statement of Lemma \[stem preserves least upper bounds\] is adapted from [@HuefClarkRaeburn2015 Lemma 3.4].
\[stem preserves least upper bounds\] Let $x,y\in P^*$ such that $x\vee y<\infty$. Write $$x={\operatorname{stem}}(x)p \text{ and } y={\operatorname{stem}}(y)q \text{\ where $p,q\in P.$}$$
1. If $\theta(x)=\theta(y)$, then ${\operatorname{stem}}(x)={\operatorname{stem}}(y)$ and $p\vee q<\infty$. In particular, $x\vee y={\operatorname{stem}}(x)(p\vee q)$ and $\theta(x\vee y)=\theta(x)=\theta(y)$.
2. If $\theta(x)<\theta(y)$, then there exists $r\in P$ such that $x\vee y=yr$ and $\theta(x\vee y)=\theta(y)$.
In particular, $\theta(x \vee y)=\max\{\theta(x), \theta(y)\}$.
\(1) Suppose that $\theta(x)=\theta(y)$. We know that $x\leq x\vee y$ and $y\leq x\vee y$. Thus, by the uniqueness of normal forms, ${\operatorname{stem}}(x)={\operatorname{stem}}(y)$. Now by left invariance of the partial order we see that $$p={\operatorname{stem}}(x)^{-1}x\leq{\operatorname{stem}}(x)^{-1}(x\vee y)\quad\text{and}\quad
q={\operatorname{stem}}(x)^{-1}y\leq{\operatorname{stem}}(x)^{-1}(x\vee y).$$ Therefore $p$ and $q$ have a common left upper bound in $P^*$ and hence, by Lemma \[p,q have cub in P\* iff they have cub in P\], they have a common left upper bound in $P$ and $p\vee q$ exists in $P$. By left invariance $x\vee y={\operatorname{stem}}(x)p\vee q$. Further, $\theta(x\vee y)=\theta(x)=\theta(y)$.
\(2) Suppose that $\theta(x)<\theta(y)$. Since $x\leq x\vee y$ we have $x^{-1}(x\vee y)\in P^*$. We can write $x^{-1}(x\vee y)=\tau \gamma u$ for some $u\in P$, $\tau\in \Sigma_{\theta(y)-\theta(x)}$ and $\gamma\in\Sigma_{\theta(x\vee y)-\theta(y)}$. Then $x\vee y=x\tau\gamma u$.
Now we have $x\tau\leq x\vee y$ and $\theta(x\tau)=\theta(x)+(\theta(y)-\theta(x))=\theta(y)$. Write $x\tau={\operatorname{stem}}(x\tau) w$ for some $w\in P$. Therefore $x\tau\vee y<\infty$ and $\theta(x\tau)=\theta(y)$ so we can apply (1) to see that ${\operatorname{stem}}(x\tau)={\operatorname{stem}}(y)$ and $x\tau\vee y={\operatorname{stem}}(y)(q\vee w)$.
Now $x\vee y\leq{\operatorname{stem}}(y)(q\vee w)$. Therefore there exists some $r\in P$ such that $x\vee y={\operatorname{stem}}(y)qr=yr$. Then $\theta(x\vee y)=\theta(y)$.
By (1) and (2) we see that $$\theta(x\vee y)=\begin{cases}\theta(x)&\text{if $\theta(x)=\theta(y)$}\\\theta(y)&\text{if $\theta(x)<\theta(y)$.}\end{cases}$$ Thus $\theta(x \vee y)=\max\{\theta(x), \theta(y)\}$.
We will use Theorem \[p-controlled map implies amenable\]; to do so we need to show that the height map $\theta:(G^*,P^*)\to ({\mathbb{Z}}, {\mathbb{N}})$ is a controlled map in the sense of Definition \[expanded controlled map definition\], and that $(\theta^{-1}(e),\theta^{-1}(e)\cap P^*)$ is amenable.
To see that $\theta$ is order-preserving, let $x, y\in P^*$ such that $x\leq y$. Then $x^{-1}y\in P^*$ and $\theta(x^{-1}y)\geq 0$. So $
0\leq \theta(x^{-1}y)=-\theta(x)+\theta(y)
$ and hence $\theta(x)\leq \theta(y)$. By Lemma \[stem preserves least upper bounds\], if $x\vee y<\infty$, then $\theta(x\vee y)=\max\{\theta(x),\theta(y)\}=\theta(x)\vee\theta(y)$.
For every $k\in {\mathbb{N}}$, $\Sigma_k$ is complete: if $x\in \theta^{-1}(k)\cap P^*$, then ${\operatorname{stem}}(x)\in \Sigma_k$ and $x={\operatorname{stem}}(x)p$ for some $p\in P$. Hence ${\operatorname{stem}}(x)\leq x$. By the uniqueness of normal forms, if $\sigma,\tau\in \Sigma_k$ and $\sigma\vee \tau<\infty$ then $\sigma=\tau$. Therefore $\theta$ is a controlled map into the amenable group ${\mathbb{Z}}$.
Suppose that $(G,P)$ is amenable. Then $\theta^{-1}(0)\cap P^*$ is the set of elements of $P^*$ with height $0$, and hence they all have normal form $p_0$ for some $p_0\in P$. Thus $\theta^{-1}(0)\cap P^*$ is isomorphic to $P$. Since $(G,P)$ is amenable, so is $(\theta^{-1}(0),\theta^{-1}(0)\cap P^*)$ by Lemma \[semigroup isomorphism preserves universal and amenability\]. Since $({\mathbb{Z}},{\mathbb{N}})$ is amenable, it now follows from Theorem \[p-controlled map implies amenable\] that $(G^*,P^*)$ is an amenable quasi-lattice ordered group.
Since $({\mathbb{Z}}^2,{\mathbb{N}}^2)$ and $({\mathbb{F}}_2,{\mathbb{F}}_2^+)$ are amenable quasi-lattice ordered groups [@Nica1992 §5.1], Theorem \[HNN extensions are amenable as expected\] implies that the HNN extensions $({\mathbb{Z}}^{2*},{\mathbb{N}}^{2*})$ and $({\mathbb{F}}_2^{*},{\mathbb{F}}_2^{+*})$ in Examples \[example Z\^2\]-\[example free group 2\] are amenable quasi-lattice ordered groups.
[99]{}
N. Brownlowe, N.S. Larsen, and N. Stammeier, *On ${C}^*$-algebras associated to right [LCM]{} semigroups*, Trans. Amer. Math. Soc. **369** (2017), 31–68.
L.O. Clark, A. an Huef, and I. Raeburn, *Phase transitions of the [T]{}oeplitz algebras of [B]{}aumslag-[S]{}olitar semigroups*, Indiana Univ. Math. J. **65** (2016), 2137–2173.
J. Crisp and M. Laca, *On the [T]{}oeplitz algebras of right-angled and finite-type [A]{}rtin groups*, J. Austral. Math. Soc. **72** (2002), 223–245.
J. Crisp and M. Laca, *Boundary quotients and ideals of [T]{}oeplitz ${C}^*$-algebras of [A]{}rtin groups*, J. Funct. Anal. **242** (2007), 127–156.
M. Laca and I. Raeburn, *Semigroup crossed products and the [T]{}oeplitz algebras of nonabelian groups*, J. Funct. Anal. **139** (1996), 415–440.
M. Laca and I. Raeburn, *Phase transition on the [T]{}oeplitz algebra of the affine semigroup over the natural numbers*, Adv. Math. **225** (2010), 643–688.
X. Li, *Semigroup ${C}^*$-algebras and amenability of semigroups*, J. Funct. Anal. **262** (2012), 4302–4340.
R.C. Lyndon and P.E. Schupp, Combinatorial Group Theory, Springer-Verlag, 1977.
A. Nica, *${C}^*$-algebras generated by isometries and [W]{}iener-[H]{}opf operators*, J. Operator Theory **27** (1992), 17–52.
I. Raeburn and D.P. Williams, Morita Equivalence and Continuous-Trace ${C}^*$-Algebras, Mathematical Surveys and Monographs, vol. 60, Amer. Math. Soc., 1998.
J. Spielberg, *${C}^*$-algebras for categories of paths associated to the [B]{}aumslag-[S]{}olitar groups*, J. Lond. Math. Soc. **86** (2012), 728–754.
J. Spielberg, *Groupoids and $C^*$-algebras for categories of paths*, Trans. Amer. Math. Soc. **366** (2014), 5771–5819.
C. Starling, *Boundary quotients of ${C}^*$-algebras of right [LCM]{} semigroups*, J. Funct. Anal. **268** (2015), 3326–3356.
J. Tomiyama, *On the projection of norm 1 on $W^*$-algebras*, Proc. Japan Acad. **33** (1957), 608–612.
[^1]: This research was supported by the Marsden Fund of the Royal Society of New Zealand and by a University of Otago Publishing Bursary
[^2]: There is an error in the statement of [@LacaRaeburn1996 Proposition 6.6]: the final line should read “if ${\mathcal{G}}$ is amenable then $(G,P)$ is amenable".
[^3]: To see denseness observe that $0<(-2+2\sqrt{2})<1$ and $(-2+2\sqrt{2})^n\in {\mathbb{Z}}(\sqrt{2})$ for all $n\in {\mathbb{N}}$. Thus for every open interval $(u,v)$ there exists $n$ such that $(-2+2\sqrt{2})^n<v-u$. Hence there exists $k\in {\mathbb{N}}$ such that $k(-2+2\sqrt{2})^n\in (u,v)$.
|
---
abstract: 'Electric scooters are becoming immensely popular across the world as a means of reliable transportation around many cities. As these e-scooters rely on batteries, it is important to understand how many of these e-scooters have enough battery life to transport riders and when these e-scooters might require a battery replacement. To this end, we develop the first stochastic model to capture the battery life dynamics of e-scooters of a large scooter network. In our model, we assume that e-scooter batteries are removable and replaced by agents called **swappers**. Thus, to gain some insight about the large scale dynamics of the system, we prove a mean field limit theorem and a functional central limit theorem for the fraction of e-scooters that lie in a particular interval of battery life. Exploiting the mean field limit and the functional central limit theorems, we develop an algorithm for determining the number of **swappers** that are needed to guarantee levels of probabilistic performance of the system. Finally, we show through a stochastic simulation and real data that our stochastic model captures the relevant dynamics.'
author:
- |
Jamol Pender\
School of Operations Research and Information Engineering\
Cornell University\
228 Rhodes Hall, Ithaca, NY 14853\
jjp274@cornell.edu\
- |
Shuang Tao\
School of Operations Research and Information Engineering\
Cornell University\
293 Rhodes Hall, Ithaca, NY 14853\
st754@cornell.edu\
- |
Anders Wikum\
School of Operations Research and Information Engineering\
Cornell University\
206 Rhodes Hall, Ithaca, NY 14853\
aew236@cornell.edu\
bibliography:
- 'scooters.bib'
title: A Stochastic Model for Electric Scooter Systems
---
Introduction
============
It’s a bird. It’s a plane. Nah, it’s a scooter! Electric scooter (e-scooter) companies are growing in popularity across the United States looking to take advantage of the ride-sharing and micro-mobility economy by providing an alternative to cars and bicycles. E-scooters were first introduced in September of 2017 in Santa Monica @Hall2018 when the micro-mobility company Bird Rides Inc. placed thousands of scooters all around the city. Bird’s scooters were immediately popular with commuters, since they are convenient and are a low cost alternative to cars. These e-scooters can reach speeds of up to 25 miles per hour., but actually speeds vary by municipality. The e-scooter technology has become a prevalent form of transportation to provide a feasible solution to the last mile problem in the transportation literature.
These web-based e-scooters are controlled by rental networks and are easily operated by smartphones. Customers are able to use e-scooters by downloading mobile applications to their smartphones. The mobile applications (apps) then show customers an image of the nearest e-scooters to their GPS location and this information will direct customers to the nearest available e-scooter. A picture of the mobile application is given in Figure \[Bird\_App\]. Using the information about available e-scooters, a customer will take an e-scooter and after completing their ride, customers can leave their e-scooter anywhere outside restricted zones, as indicated in the mobile application on their phone.
Companies operating e-scooters are rapidly expanding operations in the United States. For example, another company called Lime has a fleet of e-scooters that are available in more than 60 US cities. They also have an international presence in over 10 cities as well. Today, several major companies, including Bird and Lime, offer dockless e-scooter services, and several other companies, including the ride-sharing companies Uber and Lyft, have recently entered the market as the demand continues to grow. Recent financial analyses show that Lime is valued over \$1 billion and its rival Bird is valued at more than \$2 billion. This growth will continue to accelerate as the demand grows for these e-scooters in many of the largest cities around the world.
While e-scooter transportation should reduce emissions, and automobile congestion in local areas, e-scooters are not without accidents, see for example @allem2019electric [@kobayashi2019merging; @carville2018]. Although they are a convenient and affordable solution to transportation gaps, they are operated by batteries and need to be charged. The battery charging operations for these e-scooters consists of private individuals who go around and collect scooters to charge them. Bird scooter collectors are called “Bird Hunters” and Lime scooter collectors are called “Lime Juicers”. These collectors can make significant profits if they charge these scooters, see for example @goshtasb2018proposing. Generally, there is a flat payment for charging the e-scooter and the payment will increase depending on the difficulty of locating the scooter. Typically the “Bird Hunters” will take the scooters home and after charging them overnight, need to drop them off as groups of three at dedicated assigned points called “Nests” @bordes2019impacts. However, this particular way of charging e-scooters is not without incident. There have been several situations where juicers and hunters quarrel over the ability to charge the e-scooters @goshtasb2018proposing. There are also several places where juicers and hunters “own” a specific territory to charge the e-scooters. It would of great interest to eliminate this territorial behavior over charging the e-scooters.
To solve some of these issues, several e-scooter production companies are developing e-scooters with easily removable batteries. In this situation, instead of having “juicers” and “hunters” that need to charge the batteries, batteries are swapped by employees of the scooter company. We call these agents who replaced the dead batteries “swappers”. Many large scale e-scooter systems such as Bird and Lime already give their users the ability to see real-time availability and “battery life” through a smartphone app and web API. A picture of the Bird app showing that $43\%$ battery life for an e-scooter is given in Figure \[Bird\_App\]. This information helps users make better decisions such as where to pick up e-scooters with enough battery life to get to their destination. It is also helpful for “swappers” who will replace the batteries when the scooter’s battery life is below a predetermined threshold.
However, in order to implement the “swapping” for a scooter company, we need to understand the battery life dynamics of a large scale e-scooter system. Understanding these dynamics will enable us to determine how many of these swappers are necessary to achieve the ideal performance of the e-scooter system. To this end, in this paper, we develop the first stochastic model that analyzes this “swapping” process along with the battery dynamics of e-scooters. Our goal in this work is to understand the dynamics of a removable e-scooter system and what proportion of the e-scooters have a particular fraction of battery life. In particular, we focus on understanding how many scooters are actually available to customers at a specific time when they want to travel. Our analysis yields new insights for staffing removable battery e-scooter systems that will be used in the future.
Main Contributions of Paper
---------------------------
In this section, we describe the contributions of our work in this paper.
- We construct the first stochastic e-scooter model using empirical processes, which measure the battery life dynamics. Since our model is difficult to analyze for a large number of scooters, we propose to analyze an empirical process that describes the proportion of scooters that have a certain fraction of battery life remaining. Our model is informed by real data collected from the JUMP API in Washington D.C.
- We prove a mean field limit and a central limit theorem for our stochastic e-scooter empirical process, showing that the mean field limit and the variance of the empirical process can be described by a system of $\frac{K^2+3K}{2}$ differential equations where $K$ is the number equally sized intervals of battery life.
- We develop a novel algorithm based on our limit theorems for staffing the number of swappers that are needed to ensure that the proportion of scooters that have a small amount of battery is smaller than a given threshold.
Organization of Paper
---------------------
The remainder of this paper is organized as follows. Section \[Data\] describes how we obtained and analyzed our e-scooter data. It also includes insights on how the data was used to inform our model and model parameters. In Section \[Sec\_Bike\_Model\], we introduce two stochastic e-scooter models. The first model assume that battery usage is instantaneous , while the second model assumes that battery usage occurs according to an exponential distribution. In Section, \[Mean\_Field\_Limit\], we prove the mean field limit of our stochastic e-scooter model showing that the mean field limit is a system of $K$ differential equations. In Section \[Central\_Limit\], we prove a functional central limit theorem for the empirical process. We also show the variance of the diffusion limit can be approximated by a system of ordinary differential equations that are coupled to the mean field limit. In Section \[Staffing\], we use the mean field and central limit theorems to construct a staffing policy for the number of “swappers” need to satisfy probabilistic performance constraints. In Section \[Numerics\], we show that our results are indeed valid by comparing them to a stochastic simulation of the e-scooter system. In Section \[Conclusion\], we conclude and give directions for future work. Finally, additional proofs and theorems for our second model are given in the Appendix or Section \[Appendix\].
Preliminaries of Weak Convergence
---------------------------------
Following @ko2018strong, we assume that all random variables in this paper are defined on a common probability space $(\Omega, \mathcal{F}, \mathbb{P})$. Moreover, for all positive integers $k$, we let $\mathcal{D}([0 , \infty), \mathbb{R}^k)$ be the space of right continuous functions with left limits (RCLL) in $\mathbb{R}^k$ that have a time domain in $[0, \infty)$. As is usual, we endow the space $\mathcal{D}([0 , \infty), \mathbb{R}^k)$ with the usual Skorokhod $J_1$ topology, and let $M^k$ be defined as the Borel $\sigma$-algebra associated with the $J_1$ topology. We also assume that all stochastic processes are measurable functions from our common probability space $(\Omega, \mathcal{F}, \mathbb{P})$ into $(\mathcal{D}([0 , \infty), \mathbb{R}^k), M^k)$. Thus, if $\{\zeta\}^\infty_{n=1}$ is a sequence of stochastic processes, then the notation $\zeta^n \rightarrow \zeta$ implies that the probability measures that are induced by the $\zeta^n$’s on the space $(\mathcal{D}([0 , \infty), \mathbb{R}^k), M^k)$ converge weakly to the probability measure on the space $(\mathcal{D}([0 , \infty), \mathbb{R}^k), M^k)$ induced by $\zeta$. For any $x \in (\mathcal{D}([0 , \infty), \mathbb{R}^k), M^k)$ and any $T > 0$, we define $$||x||_T \equiv \sup_{0 \leq t \leq T} \ \max_{i = 1,2,...,k} |x_i(t)|$$ and note that $\zeta^n$ converges almost surely to a continuous limit process $\zeta$ in the $J_1$ topology if and only if $$||\zeta^n - \zeta||_T \to 0 \quad a.s.$$ for every $T > 0$.
Insights from Electric Scooter Data {#Data}
===================================
In this section, we describe some of the e-scooter trip data that we collected for this paper. This data is used to inform our stochastic models in the subsequent sections. With the real data, we can understand how e-scooter battery levels change when riders take trips, the rate of arrival to use e-scooters, and how long riders use the e-scooters in terms of time duration and distance. Below we describe how we collected raw geographic bike-share data, reconstructed likely trips, and filtered some data points that did not make sense from a rider perspective.
The Data Collection Process
---------------------------
General Bikeshare Feed Specification (GBFS) is an industry standard for sharing bike-share data that has been adopted by virtually every bike-sharing company, thanks in no small part to it being a requirement for operation in many US cities. GBFS is designed to provide a real-time snapshot of a city’s fleet, which includes vehicle locations and battery levels for bikes that are not in active use. This information is also collected without keeping records of trips or personal information. In addition to providing more detailed monthly reports to the District Department of Transportation, maintaining a public API with GBFS data is a condition of operating in the Washington D.C. metro area.
![Raw GBFS Data[]{data-label="raw_data"}](./Raw_Jump_Data_Screenshot.png)
The dataset analyzed in this paper is based on GBFS data scraped from APIs maintained by JUMP for the D.C. metro area. The data was pulled from the JUMP API once per minute from the dates 01/01/2020 to 03/01/2020 during peak hours (6:00 – 23:59) to coincide with vehicle location updates. After adding time stamps, the scraped data has the form shown in Figure \[raw\_data\]. Note that the data has a vehicle type and some of them say “bike”. This is because JUMP operates both e-bikes and e-scooters and the data is collected for both. Since our analysis is centered around e-scooters, we removed the data for the e-bikes.
---------------- --------------------------------------------------------------------------
$bike\_id$ Unique identification String for each vehicle in fleet.
$is\_disabled$ Boolean, 1 if vehicle is outside of approved geo-fences and 0 otherwise.
$is\_reserved$ Boolean, 1 if vehicle is reserved through the JUMP app, 0 otherwise.
$battery$ Integer between 0 and 100 representing percent battery remaining.
$company$ Company that operates the vehicle.
$type$ Vehicle type, one of ‘bike’ or ‘scooter’.
$lat/lon$ GPS location of vehicle at the given timestamp.
$time$ Date/Time when API call is executed.
$epoch$ Time API call is executed, in seconds since 01/01/1970.
---------------- --------------------------------------------------------------------------
: GBFS Data Fields
\[table:nonlin\]
#### Reconstructing E-Scooter Trips from Raw Data
We are ultimately interested in trip data for the purpose of informing the model parameters for our stochastic models. Though GBFS data explicitly excludes trip records, we were able to reconstruct trips by observing the times and locations at which bikes disappeared from and reentered the GBFS dataset. Using this information, our goal was to determine if a trip actually occurred or was it something else like rebalancing or strange movements of the e-scooter.
More explicitly, we compute the haversine distance between the start and end GPS coordinates. The start location is defined to be where the e-scooter first disappears from the data extraction from the API and the end location is where the e-scooter reappears in the data again. To mitigate noise from potential rebalancing, any disappearance with a distance less than 50 meters was removed from the dataset. Each of the remaining disappearances were tentatively designated as trips, with corresponding start and end times, locations, and battery levels. The distance of the trip was determined by the haversine distance between its start and end GPS coordinates, which is likely an underestimate of the true distance traveled on the e-scooter. Thus, when measuring the e-scooter drain rates, we are definitely overestimating this quantity in our analyses.
![Cleaned Trip Data[]{data-label="trip_data"}](./Cleaned_Trip_Data_Screenshot.png)
Our final data filtering steps included removing trips corresponding to overnight disappearances, trips with an average velocity greater than the maximum theoretical speed of JUMP e-bikes (25 mph), and trips corresponding to recharging in which the e-scooter battery life increased. After the filtering process was complete, we were left with a dataset of 71,518 likely trips with format shown in Figure \[trip\_data\] .
Insights Gained from Data
-------------------------
#### Distance Traveled
In this section, we use the data we collected to understand how long riders travel on the e-scooters. On the right of Figure \[Scooter\_Duration\_cdf\], we plot a histogram of the distances traveled by riders on the top plot. Since most of the data has less than 5 kilometers distance, we restricted the dataset to be less 5 kilometers. On the bottom right of Figure \[Scooter\_Duration\_cdf\], we plot the cumulative density function (cdf) of the distance data. This cdf plot allows us to understand the quantiles or percentiles of the data more clearly than the histogram plot. In this context, we observe that the median distance or 50% quantile is roughly equal to 800 meters or close to half a mile. Moreover, we observe that about 80% of riders are traveling less than 1.5 kilometers.
#### Time Duration of Riders
On the left of Figure \[Scooter\_Duration\_cdf\], we plot a histogram of the time spent by riders with the e-scooters on the top plot. Since most of the data was less than one hour, we restricted the dataset to be less than one hour or 60 minutes. On the bottom left of Figure \[Scooter\_Duration\_cdf\], we plot the cumulative density function (cdf) of the time duration data. This cdf plot allows us to understand the quantiles or percentiles of the data more clearly than the histogram plot. In this context, we observe that the median duration is roughly 8 minutes and about 80% of riders are traveling less than 15 minutes on an e-scooter.
![Scooter Rental Duration Distribution (Left). Scooter Rental Distance Distribution (Right). []{data-label="Scooter_Duration_cdf"}](./histogram_cdf_duration.png "fig:") ![Scooter Rental Duration Distribution (Left). Scooter Rental Distance Distribution (Right). []{data-label="Scooter_Duration_cdf"}](./histogram_cdf_distance.png "fig:")
#### Inter-arrival Times of Riders
On the left of Figure \[Scooter\_interarrival\_cdf\], we plot a histogram of the inter-arrival times of riders to e-scooters in the network. This histogram provides information about how many riders we should expect to arrive to the system during a time period. We should mention that the arrival rate should depend on time, however, we ignore this time dependence when looking at the inter-arrival times here. Since most of the data for the inter-arrival times was less than 20 minutes, we restricted the dataset to be less than 20 minutes. On the bottom left of Figure \[Scooter\_interarrival\_cdf\], we plot the cumulative density function (cdf) of the inter-arrival data. We observe that the median duration is roughly equal to 1 minute, however, this information is a bit misleading because of how the data is collected. Since the scooter API is updated only once per minute, it is impossible to observe an inter-arrival time less than one minute. Moreover, we observe that about 95% of the inter-arrival times are less than three minutes in length. On the right of Figure \[Scooter\_interarrival\_cdf\], we plot the arrival rate as a function of time average over the days of the data set. It is clear that the arrival rate is non-stationary and varies over the time of day. The two hump pattern (one in the morning and one in the afternoon) is also observed in this data and is common in ride-sharing data.
![Scooter Rental Inter-Arrival Time Distribution (Left). Scooter Rental Arrival Rate Throughout the Day (Right). []{data-label="Scooter_interarrival_cdf"}](./histogram_cdf_interarrival.png "fig:") ![Scooter Rental Inter-Arrival Time Distribution (Left). Scooter Rental Arrival Rate Throughout the Day (Right). []{data-label="Scooter_interarrival_cdf"}](./arrival_rate.png "fig:")
![Scooter Trip Duration vs. Change in Battery Life (Left). Scooter Trip Distance vs. Change in Battery Life (Right). []{data-label="Scooter_Duration_regression"}](./reg_duration.png "fig:") ![Scooter Trip Duration vs. Change in Battery Life (Left). Scooter Trip Distance vs. Change in Battery Life (Right). []{data-label="Scooter_Duration_regression"}](./reg_distance.png "fig:")
#### Estimating Battery Life of E-Scooters
Another informative measurement from a data perspective is the battery usage dynamics of riders. On the left of Figure \[Scooter\_Duration\_regression\], we show a scatterplot of trip time duration and the decrease of battery life. The plot measures for each trip, how long the customer used the e-scooter and what was the subsequent drain in the battery life of the e-scooter. On the right of Figure \[Scooter\_Duration\_regression\], we show a scatterplot of trip distance and the decrease of battery life. This plot measures for each trip, how far in meters the customer drove the e-scooter and what was the subsequent drain in the battery life. We should emphasize that the e-scooter was not tracked during the entire time of usage and only the starting and ending GPS locations were used to compute the haversine distance between them. In both plots of Figure \[Scooter\_Duration\_regression\], we observe that the relationship between time or distance with battery life is negative. We use regression analysis to explore these relationships in a formal way. Table \[table:reg\_result\] summarizes the coefficients of different regression methods used to understand the relationship between % battery change and distance/duration.
feature norm intercept slope
--------- ------ ----------- -------
: Regression coefficients for different methods[]{data-label="table:reg_result"}
We observe that in both plots of Figure \[Scooter\_Duration\_regression\] that performing the regression without an intercept increases the absolute value of the negative slopes. Moreover, we see that $L_1$ regression generally yields more negative slopes than their $L_2$ counterparts. This is especially true in the duration vs. battery plot on the left of Figure \[Scooter\_Duration\_regression\]. Finally, we observe that the trip distance seems to be a better estimate of real trips versus the time duration of trips. This is consistent with the estimates of battery life that are reported from the Jump scooter company.
![Scooter Rental % Battery Change Distribution. []{data-label="Scooter_Battery_cdf"}](./histogram_cdf_battery.png)
#### Battery Usage Per Trip
Finally, on the top of Figure \[Scooter\_Battery\_cdf\], we plot a histogram of the battery used by customer. Since most of the data was less than 60% of the total battery life, we restricted the dataset to be less than 60 %. We observe that the histogram for battery usage almost looks exponentially distributed, but probably is more of a gamma distribution. On the bottom of Figure \[Scooter\_Battery\_cdf\], we plot the cdf of the battery usage data. We observe that the median battery usage is about 6% of battery life and about 80% of riders are using less than 10% of battery life of the scooter on each trip. From the data, we also observe that the minimum starting battery is about 21%. Thus, the average scooter can do about 13 trips before its battery needs to be swapped out for a new one.
Stochastic Models and Limit Theorems {#Sec_Bike_Model}
====================================
In this section, we propose two new Markovian queueing models for the empirical process of e-scooters battery life. In our first model, we assume the battery usage time is instantaneous. However, in our second model, we assume the battery usage time is an exponentially distributed random variable. Although one can analyze each e-scooter individually, this is a high dimensional stochastic process and given the large scale of these e-scooter systems, we wish not to take on this intractable endeavor. As a result, we resort to using an empirical process perspective for the e-scooter system. The empirical process perspective reduces the dimension from the number of e-scooters, which is large, to the number of intervals of battery life one would like to keep track of. For example, in Washington D.C, each scooter company is allowed to operate at most 2500 scooters. However, the number of intervals of battery life for the scooters is at most 100. This represents an 25 fold reduction in dimension. However, it may not be necessary to even keep that much granularity for the purposes of this work. We suggest a value of $K = 10$ to give battery intervals of $10\%$. This would yield a 250 fold reduction in dimension.
There is a large literature in the space of bike sharing and the sharing economy, see for example @hampshire2012analysis [@nair2013large; @schuijbroek2017inventory; @faghih2017empirical; @singla2015incentivizing; @jian2016simulation; @freund2020data]. Despite there being much research on bike sharing networks there is much less literature on electric scooters and their impact on transportation networks in large cities. Our goal in this work is to add to the growing literature in the sharing economy, but specifically for e-scooters. Our approach leverages new data resources for the scooters and uses the data to inform the structure of the stochastic models we will build in the sequel. Our new stochastic models leverage techniques from empirical process theory and weak convergence of martingales.
Empirical processes are not new and have been used in a variety of contexts in queueing theory, see for example @graham1997stochastic [@graham2000kinetic; @graham2000chaoticity; @graham2005functional; @li2014mean; @li2016mean; @mitzenmacher2016analyzing; @ying2016approximation; @iyer2011mean; @iyer2014mean; @yang2016mean; @yang2018mean; @yang2019information]. One of the first papers to consider empirical processes in ride-sharing is @mohamed2012mean, where the authors model bike sharing networks as a network of finite capacity single server queues. Using empirical processes, they prove a mean field limit theorem for the number stations that have $k$ bikes. In this context, the dimensionality is reduced from the total number of bikes in the network to roughly the size of the largest station. In large metropolitan cities like New York City and Washington D.C., this reduction is huge and useful. Recently, @tao2017stochastic, prove a central limit theorem for the same bike sharing model, showing that the central limit theorem is quite good at describing the fluctuations of the stochastic bike-sharing network process. Moreover, recent work by @fricker2016incentives [@bortolussi2016mean; @li2016bike; @li2016queueing; @li2017fluid; @el2018using] has also generalized the mean field limit theorems of @mohamed2012mean to the setting of non-stationary bike sharing systems and for Markovian arrival processes (MAPs) for the arrival and service distributions. More recently, @graef2019fractional extend the mean field model from ordinary differential equations to fractional ordinary differential equations. Generally, the mean field limit theorems provide rigorous support for using ordinary differential equations for describing the mean dynamics of the empirical measure. However, @graef2019fractional shows that using fractional ordinary differential equations might be more appropriate as they provide more flexibility than their non-fractional counterparts. Before we get specific about the models that we will describe in the sequel, we give some of the common notation that we will use throughout the remainder of the paper below in Table \[table:nonlin\].
----------- -----------------------------------------------------------------------------------------------
$i$ index for e-scooter
$N$ Number of e-scooters
$N^*$ Number of swappers
$K$ Battery life bucket size
$K_U$ Battery threshold for riding
$\lambda$ Arrival rate of recharger to a e-scooter
$\mu$ Arrival rate of customer to a e-scooter
$p_{ij}$ Probability of battery end up in bucket $j/K$ in a ride with starting battery in bucket $i/K$
$B_i(t)$ Battery life of e-scooter $i$ at time $t$
$Y^N(t)$ Empirical process of e-scooters battery life at time $t$
----------- -----------------------------------------------------------------------------------------------
: Summary of Notation
\[table:nonlin\]
Model 1: Instantaneous Battery Usage {#model_1}
------------------------------------
Here we describe our first model for modeling the battery dynamics of an e-scooter network. We consider an empirical process of the battery life among all e-scooters in the system. The goal here is to model the distribution of battery life as Markov process and study the asymptotic behavior of the system as the number of e-scooters grows towards infinity i.e, $N \to \infty$. More specifically, we analyze a mean field and central limit theorem for the empirical process to help understand how different parameters can affect the system’s performance.
### Modeling Assumptions
#### Customer Arrival (battery usage):
We assume that customers arrive to the system following a Poisson process with rate $\mu N$ (uniform on geographical location). Only e-scooters with battery life above a certain threshold $K_U/K$ can be picked up and used by the customer. For simplicity of the model, we assume that after the customer picks up the e-scooter, the battery life changes immediately according to a probability matrix $P=(p_{ij})_{ij}$. Each element in the matrix $P=(p_{ij})_{ij}$ represents an e-scooter moving from the $i^{th}$ interval of battery life to the $j^{th}$ interval. Using the Jump scooter data we collected, we find the following empirical probability matrix $\hat{P}$ when setting $K=5$, which is equal to
$$\hat{P}=\begin{blockarray}{cccccc}
[0,20\%] & [20\%,40\%] & [40\%,60\%] & [60\%,80\%] & [80\%,100\%] \\
\begin{block}{(ccccc)c}
0 & 0 & 0 & 0 & 0 & [0,20\%] \\
0.097 & 0.903 & 0 & 0 & 0 & [20\%,40\%] \\
0.003& 0.345& 0.652 & 0 &0& [40\%,60\%] \\
0.0008 & 0.022 & 0.329 & 0.648 & 0 & [60\%,80\%]\\
0.00 & 0.004 & 0.021 & 0.446 & 0.529 & [80\%,100\%] \\
\end{block}
\end{blockarray}$$ Here the first row of $\hat{P}$ is zero because the minimum starting battery life is around 21% from the Jump scooters data.
#### Swapper Arrival:
We assume that swappers arrive to the system following a Poisson process with rate $\lambda N^*$. The probability of a e-scooter with battery life in bucket $k/K$ getting recharged is based on a choice model $\frac{Y_k^N(t)g_k}{\sum_{i=0}^{K-1}Y_i^N(t)g_i}$, where $\{g_i\}_{i=0}^{K-1}>0$ is a decreasing sequence on $i$. For simplicity of the model, we assume that after the swappers picks up the e-scooters, the battery life jumps to full immediately (i.e. neglecting swapping time).
#### Remark:
Note that instead of only recharge e-scooters with low battery, here we use a choice model for recharging that gives more weight to e-scooters with low battery life. With the choice model, there is a positive probability to recharge a e-scooter in bucket $\left[\frac{K-1}{K},1\right]$. Without the choice model, we cannot guarantee the Lipschitz property of the drift function for the limiting mean field equations. The Lipschitz property is crucial for proving the mean field and central limit results in this work. However, one can set up the choice model $\{g_i\}_{i=0}^{K-1}$ so that the probability of recharging a e-scooter with high battery is low (In fact we only need $\min_{i}\{g_i\} > 0$).
### Markov Jump Process
Now that we have described the dynamics of the model, we are now free to construct our empirical process model. To this end, we define the empirical process $Y_k^N(t)$ as the proportion of e-scooters with remaining battery life between $[\frac{k}{K},\frac{k+1}{K})$. Thus, we can write $Y_k^N(t)$ as the following equation $$Y_k^N(t)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}\left\{\frac{k}{K}\leq B_i(t) < \frac{k+1}{K}\right\}, \quad k=0,\cdots, K-1$$ where $N$ is the total number of e-scooters in the system, and $B_i(t)$ is the battery life of the $i^{th}$ e-scooter at time $t$. Moreover, we also assume that battery drainage of a single ride follows a discrete probability distribution, i.e. $$P(\text{battery end up in } j/K \text{ but started with battery in } i/K)=p_{ij}, \quad j\leq i=0,\cdots, K-1.$$ If we condition on $Y_k^{N}(t)=y_k$, the transition rates of $y$ are specified as follows.
#### Swapping Batteries:
When there is a swapper arriving to the system to swap an e-scooter’s battery with battery life in the interval $[\frac{k}{K},\frac{k+1}{K}]$, the proportion of e-scooters with battery life in bucket $[\frac{k}{K},\frac{k+1}{K}]$ goes down by $1/N$, the proportion of e-scooters with battery life in bucket $[\frac{K-1}{K},1]$ goes up by $1/N$, and the transition rate $Q^N$ is $$\begin{aligned}
Q^{N}\left(y,y+\frac{1}{N}(\mathbf{1}_{K-1}-\mathbf{1}_{k}) \right) &=& \lambda N^* \frac{y_kg_k}{\sum_{i=0}^{K-1}y_ig_i} .\end{aligned}$$
#### Riding a Scooter:
When there is a customer riding an e-scooter with battery life in the interval $[\frac{k}{K},\frac{k+1}{K}]$ where $k\geq K_U$, the proportion of e-scooters with battery life in the interval$[\frac{k}{K},\frac{k+1}{K}]$ moves down by $1/N$ with probability $p_{kj}$. In addition, the proportion of e-scooters with battery life in the interval $[\frac{j}{K},\frac{j+1}{K}]$ moves up by $1/N$, and the transition rate $Q^N$ is $$\begin{aligned}
Q^{N} \left(y,y+\frac{1}{N}(\mathbf{1}_{j}-\mathbf{1}_{k}) \right) &=&
\mu N p_{kj} y_k \mathbf{1}\{k\geq K_U\} \mathbf{1}\{j\leq k\}.\end{aligned}$$
With the described transitions, one can show that $Y^N(t)$ is a Markovian jump process with the above transition rates. In Figure \[model1\_diagram\], we illustrate the transitions between states in the model proposed above.
Despite some complexity of our model, it is not completely realistic as we make the battery usage transitions instantaneous. What follows in the sequel is a generalization of our first model where the battery life transitions have an exponential distribution.
Model 2: Exponentially Distributed Battery Usage Time {#model_2}
-----------------------------------------------------
In this subsection, we propose a different model for the empirical process of battery life where battery usage time is considered to be exponentially distributed with rate $\mu_U$. We use the same notation from Section \[model\_1\] and introduce a new variable $R^N(t)$ as the number of e-scooters in use (riding by customers) at time $t$. Table \[table:nonlin2\] summarizes the additional notation we need for this new model.
----------- ---------------------------------------------
$1/\mu_U$ Mean trip duration
$R^N(t)$ Number of e-scooters in use at time $t$
$X^N(t)$ Proportion of e-scooters in use at time $t$
----------- ---------------------------------------------
: Summary of Additional Model 2 Notation
\[table:nonlin2\]
### Modeling Assumptions
Now we describe the following modeling assumptions we make for this new model.
#### Customer Arrival (battery usage):
We assume that customers arrive to the system following a Poisson process with rate $\mu (N-R^N(t))$ (uniform on geographical location). Only e-scooters with battery life above a certain threshold $K_U/K$ can be picked up and used by the customer. After the customer picks up the e-scooter, they will ride the e-scooter for a time that is exponentially distributed with rate $\mu_U$, and after this the battery life changes immediately according to a probability matrix $P=(p_{ij})_{ij}$.
#### Swapper Arrival:
We assume that swappers arrive to the system following a Poisson process with rate $\lambda N^*\left(1-\frac{R^N(t)}{N}\right)$, which is proportional to the number of e-scooters available at the time. The probability of a e-scooter with battery life in bucket $k/K$ getting swapped is based on a choice model $\frac{Y_k^N(t)g_k}{\sum_{i=0}^{K-1}Y_i^N(t)g_i}$, where $\{g_i\}_{i=0}^{K-1}>0$ is a decreasing sequence on $i$. For simplicity of the model, we assume that after the swapper picks up the e-scooter, the battery life jumps to full immediately (i.e. neglecting the swapping time).
### Markov Jump Process
Now that we have described the dynamics of the model, we are now free to construct our empirical process model. To this end, we define the fraction of e-scooters in use as $$X^N(t)=\frac{R^N(t)}{N}$$
By conditioning on $(X^N(t),Y^{N}(t))=(x,y)$, the transition rates of $(x,y)$ are specified as follows:
#### Battery Swapping:
When there is a swapper arriving to the system to swap the battery of an e-scooter with battery life in the interval $[\frac{k}{K},\frac{k+1}{K}]$, the proportion of e-scooters with battery life in the interval $[\frac{k}{K},\frac{k+1}{K}]$ goes down by $1/N$, the proportion of e-scooters with battery life in the interval$[\frac{K-1}{K},1]$ goes up by $1/N$, and the transition rate $Q^N$ is $$\begin{aligned}
Q^{N}\left((x,y),\left(x,y+\frac{1}{N}(\mathbf{1}_{K-1}-\mathbf{1}_{k})\right) \right) &=& \lambda N^* (1-x)\frac{y_kg_k}{\sum_{i=0}^{K-1}y_ig_i}\end{aligned}$$
#### Customer Arrival:
When there is a customer arriving to the system to pick up a e-scooter with battery life in the interval $[\frac{k}{K},\frac{k+1}{K}]$ where $k\geq K_U$,, the proportion of e-scooters in use goes up by $1/N$, and the transition rate $Q^N$ is $$\begin{aligned}
Q^{N}\left((x,y),\left(x+\frac{1}{N},y\right) \right) &=& \mu N(1-x)\sum_{k=K_U}^{K-1}y_k\end{aligned}$$
#### Battery Usage:
When there is a customer riding a e-scooter with battery life in the interval $[\frac{k}{K},\frac{k+1}{K}]$ where $k\geq K_U$, the proportion of e-scooters with battery life in the interval $[\frac{k}{K},\frac{k+1}{K}]$ goes down by $1/N$, with probability $p_{kj}$ the proportion of e-scooters with battery life in the interval $[\frac{j}{K},\frac{j+1}{K}]$ goes up by $1/N$, and the proportion of e-scooters in use goes down by $1/N$, and the transition rate $Q^N$ is $$\begin{aligned}
Q^{N} \left((x,y),\left(x-\frac{1}{N},y+\frac{1}{N}(\mathbf{1}_{j}-\mathbf{1}_{k})\right) \right) &=&
\mu_U Nx p_{kj} y_k \mathbf{1}\{k\geq K_U\} \mathbf{1}\{j\leq k\}.\end{aligned}$$ We the above transition rates, we have that $(X^N(t),Y^N(t))$ is a Markov jump process. In Figure \[model2\_diagram\], we illustrate the transitions between states in the model proposed. Note that in order to make the illustration easier to understand, we break down $X^N(t)$ into each battery interval. Since we assume uniform arrival to scooters with battery life in all levels, we do not need to track the proportion of e-scooters in use in each battery bucket for the model to be Markovian. However, it will be needed when arrival rate depends on the battery life of e-scooters, in which case we can extend the state space to $$X_{u, k}^N(t)=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}\left\{\frac{k}{K}\leq B_i(t) < \frac{k+1}{K}, u_i(t)=u\right\}, \quad k=0,\cdots, K-1, u=0,1$$ where $u_i(t)=1$ denotes scooter $i$ is in use at time $t$ and $u_i(t)=0$ indicates when the $i^{th}$ e-scooter is idle at time $t$.
Now that we have two models for the dynamic behavior of e-scooter systems, we want to understand some important behavior of the system. Since the system is quite large and is not easy to analyze directly, we resort to using asymptotic analysis. Thus, in the sequel we will prove mean field and central limit theorems for describe the mean and variance dynamics of the e-scooter system.
Mean Field Limit of Empirical Processes {#Mean_Field_Limit}
=======================================
In this section, we prove the mean field limit for both of our empirical process of e-scooters battery life models. A mean field limit describes the large system dynamics of the e-scooters battery life and usage over time. Deriving the mean field limit allows us to gain insight about the average system behavior when the number of e-scooters is very large. Thus, we avoid the need to study an $N$-dimensional continuous time Markov chain and compute its steady state distribution in this high dimensional setting, which is quite intractable. We first state the mean field limit result for the stochastic model described in Section \[model\_1\], the empirical process of e-scooter battery life with instantaneous battery usage.
\[fluid\_limit\] Let $|. |$ denote the Euclidean norm in $\mathbb{R}^{K}$. Suppose that $\lim_{N\rightarrow \infty}\frac{N^*}{N}=\gamma$, and $Y^{N}(0)\xrightarrow{p} y(0)$, then we have for $\forall \epsilon>0$ $$\lim_{N\rightarrow \infty}P\left(\sup_{t\leq t_0}|Y^N(t)-y(t)|>\epsilon\right)=0,$$ where $y(t)$ is the unique solution to the following differential equation starting at $y(0)$, $$\label{diff_eqn}
{\raisebox{-0.4pt}{$\stackrel{\bullet}{y}$}}=f(y)$$
where $f:[0,1]^{K}\rightarrow \mathbb{R}^{K}$ is a vector field that satisfies $$\begin{aligned}
\label{eqn:b}
f(y)&=&\sum_{k=0}^{K-1}\left[ \left( \frac{\lambda \gamma g_k}{\sum_{i=0}^{K-1}y_ig_i} \right)(\mathbf{1}_{K-1}-\mathbf{1}_{k})+\sum_{j=0}^{k}\mu p_{kj} (\mathbf{1}_{j}-\mathbf{1}_{k})\mathbf{1}_{k\geq K_U}\right] y_k\nonumber \\\end{aligned}$$ or componentwise for $0\leq j \leq K-1$, $$\begin{aligned}
f_j(y)&=&\underbrace{\sum_{k=\max(K_U,j)}^{K-1}\mu p_{k,j}y_k}_{\text{battery usage from scooters in $k$-th bucket}} +\underbrace{\lambda \gamma \mathbf{1}\{j=K-1\}}_{\text{battery recharge to full}}-\underbrace{\mu y_j \mathbf{1}\{j\geq K_U\}}_{\text{battery usage from scooters in $j$-th bucket}}\nonumber\\
& &-\underbrace{\frac{\lambda \gamma y_jg_j}{\sum_{i=0}^{K-1}y_ig_i}}_{\text{battery recharge to scooters in $j$-th bucket}} .\end{aligned}$$
The full proof is provided in the Appendix (Section \[Appendix\]).
Now we state the the mean field limit result for the stochastic model described in Section \[model\_2\], the empirical process of e-scooter battery life with exponentially distributed battery usage time.
\[fluid\_limit\_model2\] Let $|. |$ denote the Euclidean norm in $\mathbb{R}^{K+1}$. Suppose that $\lim_{N\rightarrow \infty}\frac{N^*}{N}=\gamma$, and $(X^N(0),Y^{N}(0))\xrightarrow{p} (x(0),y(0))$, then we have for $\forall \epsilon>0$ $$\lim_{N\rightarrow \infty}P\left(\sup_{t\leq t_0}|(X^N(t), Y^N(t))-(x(t),y(t))|>\epsilon\right)=0$$ where $(x(t),y(t))$ is the unique solution to the following differential equation starting at $(x(0), y(0))$ $$\begin{aligned}
\label{diff_eqn}
{\raisebox{-0.4pt}{$\stackrel{\bullet}{x}$}}&=&f_x(x, y),\\
{\raisebox{-0.4pt}{$\stackrel{\bullet}{y}$}}&=&f_y(x, y),\end{aligned}$$
where $f=(f_x,f_y):[0,1]\times [0,1]^{K}\rightarrow \mathbb{R}\times \mathbb{R}^{K}$ is a vector field that satisfies $$\begin{aligned}
f_x(x,y)=\underbrace{\mu(1-x)\sum_{k=K_U}^{K-1}y_k}_{\text{customers picking up scooters}} -\underbrace{\mu_U x \sum_{k=K_U}^{K-1}y_k}_{\text{customers drop off scooters}},\end{aligned}$$
$$\begin{aligned}
f_y(x,y)&=&\sum_{k=0}^{K-1}\left[ \left( \frac{\lambda \gamma (1-x) g_k}{\sum_{i=0}^{K-1}y_ig_i} \right)(\mathbf{1}_{K-1}-\mathbf{1}_{k})+\sum_{j=0}^{k}\mu_U x p_{kj} (\mathbf{1}_{j}-\mathbf{1}_{k})\mathbf{1}_{k\geq K_U}\right] y_k,\nonumber \\\end{aligned}$$
or componentwise for $0\leq j \leq K-1$, $$\begin{aligned}
f_y(x,y)(j)&=&\underbrace{\sum_{k=\max(K_U,j)}^{K-1}\mu _U x p_{k,j}y_k}_{\text{battery usage from scooters in $k$-th bucket}} +\underbrace{\lambda \gamma(1-x) \mathbf{1}\{j=K-1\}}_{\text{battery recharge to full}}\nonumber\\
& &-\underbrace{\mu_U x y_j \mathbf{1}\{j\geq K_U\}}_{\text{battery usage from scooters in $j$-th bucket}}-\underbrace{\frac{\lambda \gamma(1-x) y_jg_j}{\sum_{i=0}^{K-1}y_ig_i}}_{\text{battery recharge to scooters in $j$-th bucket}}.\end{aligned}$$
The proof ideas for Theorem \[fluid\_limit\_model2\] follow easily from the proof of Theorem \[fluid\_limit\] so we do not prove them in this paper.
Steady State Analysis
---------------------
By the existence and uniqueness of the fluid limit, $y(t)$ has a unique steady state $\bar{y}$ which satisfies
$$\sum_{k=\max(K_U,j)}^{K-1}\mu p_{k,j}\bar{y}_k +\lambda \gamma \mathbf{1}\{j=K-1\} =\mu \bar{y}_j \mathbf{1}\{j\geq K_U\}+\frac{\lambda \gamma \bar{y}_jg_j}{\sum_{i=0}^{K-1}\bar{y}_ig_i}$$ Specifically, $$\sum_{k=K_U}^{K-1}\mu p_{k,j}\bar{y}_k =\frac{\lambda \gamma \bar{y}_jg_j}{\sum_{i=0}^{K-1}\bar{y}_ig_i}, \quad
0\leq j<K_U$$ $$\sum_{k=j}^{K-1}\mu p_{k,j}\bar{y}_k =\mu \bar{y}_j+\frac{\lambda \gamma \bar{y}_jg_j}{\sum_{i=0}^{K-1}\bar{y}_ig_i}, \quad K_U\leq j<K-1$$ $$\mu p_{K-1,K-1}\bar{y}_{K-1} +\lambda \gamma =\mu \bar{y}_{K-1}+\frac{\lambda \gamma \bar{y}_{K-1}g_{K-1}}{\sum_{i=0}^{K-1}\bar{y}_ig_i}, \quad j=K-1$$
Both mean field limits provide insights to the stochastic model, by providing ordinary differential equations that describe the mean proportion of scooters in a particular interval of battery life. In the first model, the dimension is reduced from $N$ scooters to $K$ intervals where $K$ is generally much lower than $N$. In the second model, the dimension is only increased to $K+1$, which is still much smaller than $N$, the number of e-scooters. We will describe in the sequel how these mean field limits can be used to construct staffing algorithms for agents who will swap out the batteries when they are low. However, the mean field limits only describe the mean dynamics of the stochastic models and say nothing about the stochastic fluctuations around the mean field limits. In the next section, we prove central limit theorems, centering around the mean field limits for our stochastic models. The central limit theorems will provide some rigorous support for confidence intervals around the mean field limits.
Central Limit Theorem of Empirical Process {#Central_Limit}
==========================================
In this section, we derive the diffusion limit of our stochastic empirical process of scooters battery life model. Diffusion limits are critical for obtaining a deep understanding of the sample path behavior of stochastic processes around their mean. One reason is that diffusion limits describe the fluctuations around the mean or mean field limit and can help understand the variance or the asymptotic distribution of the stochastic process being analyzed. We define our diffusion scaled e-scooters sharing model by subtracting the mean field limit from the scaled stochastic process and rescaling it by $\sqrt{N}$. Thus, we obtain the following expression for the diffusion scaled scooter battery life empirical process $$D^{N}(t)=\sqrt{N}(Y^{N}(t)-y(t)).$$ Now we state the functional central limit theorem for the empirical process described in Section \[model\_1\].
\[difftheorem\] Consider $D^{N}(t)$ in $\mathbb{D}(\mathbb{R}_{+},\mathbb{R}^{K})$ with the Skorokhod $J_{1}$ topology, and suppose that $\limsup_{N\rightarrow \infty}\sqrt{N}\left(\frac{N^*}{N}-\gamma\right)< \infty$. If $D^{N}(0)$ converges in distribution to $D(0)$, then $D^{N}(t)$ converges to the unique Ornstein Uhlenbeck (OU) process solving $D(t)=D(0)+\int_{0}^{t}f'(y(s))D(s) ds+M(t)$ in distribution, where $f'(y)$ is specified as follows, $$\begin{aligned}
\label{b'}
\frac{\partial f_j(y)}{\partial y_k}&=&\mu p_{k,j}\mathbf{1}\{k\geq \max\{j,K_U\}\}-\mu \mathbf{1}\{k=j\geq K_U\}\nonumber\\
& &+\frac{\lambda \gamma g_kg_jy_j}{(\sum_{i=0}^{K-1}y_ig_i)^2}-\frac{\lambda \gamma g_j}{\sum_{i=0}^{K-1}y_ig_i}\mathbf{1}\{j= k\},\end{aligned}$$ and $M(t) = (M_{0}(t),\cdots,M_{K-1}(t))\in \mathbb{R}^{K}$ is a real continuous centered Gaussian martingale, with Doob-Meyer brackets given by $$\begin{aligned}
\boldlangle M_{k}(t),M_{j}(t)\boldrangle
&=&\begin{cases}
\int_{0}^{t}\left[\sum_{i=\max(K_U,k+1)}^{K-1}\mu p_{i,k}y_i(s) +\left(\mu(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}\right.\right.\nonumber\\
+\left.\left.\frac{\lambda \gamma g_k}{\sum_{i=0}^{K-1}y_i(s)g_i} \right)y_k(s)\right]ds, & k=j<K-1 \nonumber\\
\int_{0}^{t}\left[\mu(1-p_{K-1,K-1})y_{K-1}(s)+\lambda \gamma\left(1- \frac{g_{K-1}y_{K-1}(s)}{\sum_{i=0}^{K-1}y_i(s)g_i}\right)\right]ds, & k=j=K-1\\
-\int_{0}^{t} \left[\mu p_{k,j}y_k(s)+\frac{\lambda \gamma g_j}{\sum_{i=0}^{K-1}y_i(s)g_i}y_j(s)\mathbf{1}\{k=K-1\}\right]ds, & j<k, k\geq K_U\nonumber\\
0. & \text{otherwise}
\end{cases}\end{aligned}$$ Define $\mathcal{A}(t)=f'(y(t))$, $\mathcal{B}(t)=\left(\frac{d}{dt}\boldlangle M_i(t),M_j(t)\boldrangle \right)_{ij}$, then the covariance matrix $\Sigma (t)=\mathrm{Cov}[D(t),D(t)]$ satisfies $$\frac{d\Sigma(t)}{dt}=\Sigma(t)\mathcal{A}(t)^\top+\mathcal{A}(t)\Sigma(t)+\mathcal{B}(t).$$ Moreover, componentwise, for $i=0,1,\cdots, K-2$, $$\begin{aligned}
\frac{d\Sigma_{ii}(t)}{dt} &=&2\sum_{k=\max\{i,K_U\}}^{K-1}\Sigma_{ik}\mu p_{k,i}-2\left( \frac{\lambda \gamma g_i}{\sum_{i=0}^{K-1}y_ig_i}+\mu \mathbf{1}\{i\geq K_U\}\right)\Sigma_{ii} +2\sum_{k=0}^{K-1}\Sigma_{ik}\frac{\lambda \gamma g_kg_iy_i}{(\sum_{i=0}^{K-1}y_ig_i)^2}\nonumber\\
& &+\sum_{k=\max(i+1,K_U)}^{K-1}\mu p_{k,i}y_k +\left(\mu(1-p_{i,i})\mathbf{1}\{i\geq K_U\}+\frac{\lambda \gamma g_i}{\sum_{i=0}^{K-1}y_ig_i} \right)y_i,\end{aligned}$$ and for $i=K-1$, $$\begin{aligned}
\frac{d\Sigma_{K-1,K-1}(t)}{dt}
&=&-2\left( \frac{\lambda \gamma g_{K-1}}{\sum_{i=0}^{K-1}y_ig_i}+\mu \right)\Sigma_{ii}+2\sum_{k=0}^{K-1}\Sigma_{ik}\frac{\lambda \gamma g_kg_{K-1}y_{K-1}}{(\sum_{i=0}^{K-1}y_ig_i)^2}+\mu(1-p_{K-1,K-1})y_{K-1}\nonumber\\
& &+\lambda \gamma\left(1- \frac{g_{K-1}y_{K-1}}{\sum_{i=0}^{K-1}y_ig_i}\right) ,\end{aligned}$$ for $0\leq i<j\leq K-1$, $$\begin{aligned}
\frac{d\Sigma_{ij}(t)}{dt}
&=&\sum_{k=\max\{j,K_U\}}^{K-1}\Sigma_{ik}\mu p_{k,j}-\left( \frac{\lambda \gamma g_j}{\sum_{i=0}^{K-1}y_ig_i}+\mu \mathbf{1}\{j\geq K_U\}\right)\Sigma_{ij}+\sum_{k=0}^{K-1}\Sigma_{ik}\frac{\lambda \gamma g_kg_jy_j}{(\sum_{i=0}^{K-1}y_ig_i)^2}\nonumber\\
& &\sum_{k=\max\{i,K_U\}}^{K-1}\Sigma_{jk}\mu p_{k,i}-\left( \frac{\lambda \gamma g_i}{\sum_{i=0}^{K-1}y_ig_i}+\mu \mathbf{1}\{i\geq K_U\}\right)\Sigma_{ij}+\sum_{k=0}^{K-1}\Sigma_{jk}\frac{\lambda \gamma g_kg_iy_i}{(\sum_{i=0}^{K-1}y_ig_i)^2}\nonumber\\
& &-\mu p_{j,i}y_j\mathbf{1}\{j\geq K_U\}-\frac{\lambda \gamma g_i}{\sum_{i=0}^{K-1}y_ig_i}y_i\mathbf{1}\{j=K-1\}.\end{aligned}$$
In order to prove Theorem \[difftheorem\], we need to prove the following four results listed below step by step.
- (Lemma \[martingale\_brackets\])$\sqrt{N}M^{N}(t)$ is a family of martingales independent of $D^{N}(0)$ with Doob-Meyer brackets given by $$\begin{aligned}
& &\boldlangle \sqrt{N}M^{N}_k(t),\sqrt{N}M^{N}_j(t)\boldrangle\nonumber\\
&=&\begin{cases}
\int_{0}^{t}\left[\sum_{i=\max(K_U,k+1)}^{K-1}\mu p_{i,k}Y^N_i(s) +\left(\mu(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}\right.\right. & \nonumber\\
+\left.\left.\lambda N^*\frac{ g_k}{N\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_k(s)\right]ds, & k=j<K-1 \nonumber\\
\int_{0}^{t}\left[\mu(1-p_{K-1,K-1})Y_{K-1}^N(s)+\lambda \frac{N^*}{N}\left(1- \frac{g_{K-1}Y_{K-1}^N(s)}{\sum_{i=0}^{K-1}Y_i^N(s)g_i}\right)\right]ds, & k=j=K-1\\
-\int_{0}^{t} \left[\mu p_{k,j}Y^N_k(s)+\lambda N^*\frac{ g_j}{N\sum_{i=0}^{K-1}Y_i^N(s)g_i}Y_j^N(s)\mathbf{1}\{k=K-1\}\right]ds, & j<k, k\geq K_U\nonumber\\
0. & \text{otherwise}
\end{cases}\end{aligned}$$
- (Lemma \[L2bound\]) For any $T\geq 0$, $$\limsup_{N\rightarrow \infty}\mathbb{E}(|D^{N}(0)|^2)<\infty \Rightarrow \limsup_{N\rightarrow \infty}\mathbb{E}(\sup_{0\leq t\leq T}|D^{N}(t)|^2)<\infty.$$
- (Lemma \[tightness\]) If $(D^{N}(0))_{N=1}^{\infty}$ is tight then $(D^{N})_{N=1}^{\infty}$ is tight and its limit points are continuous.
- If $D^{N}(0)$ converges to $D(0)$ in distribution, then $D^{N}(t)$ converges to the unique OU process solving $D(t)=D(0)+\int_{0}^{t}f'(y(s))D(s) ds+M(t)$ in distribution.
We provide the proofs of Lemma \[martingale\_brackets\], Lemma \[L2bound\] and Lemma \[tightness\] in the Appendix (Section \[Appendix\]). For step 4), by Theorem 4.1 in Chapter 7 of @Ethier2009, it suffices to prove the following conditions hold
1. $$\label{condition_1}
\lim_{N\rightarrow \infty}\mathbb{E}\left[\sup_{t\leq T}|D^N(t)-D^N(t-)|^2\right]=0,$$
2. $$\label{condition_2}
\lim_{N\rightarrow \infty}\mathbb{E}\left[\sup_{t\leq T}\left|\int_{0}^{t}F^N(Y^N(s))ds-\int_{0}^{t-}F^N(Y^N(s))ds\right|^2\right]=0,$$
3. for $0\leq k,j\leq K-1$, $$\label{condition_3}
\lim_{N\rightarrow \infty}\mathbb{E}\left[\sup_{t\leq T}\left|\boldlangle \sqrt{N}M^{N}_k(t),\sqrt{N}M^{N}_j(t)\boldrangle-\boldlangle \sqrt{N}M^{N}_k(t-),\sqrt{N}M^{N}_j(t-)\boldrangle\right|^2\right]=0,$$
4. for $0\leq k,j\leq K-1$, $$\label{condition_4}
\sup_{t\leq T}\left|\boldlangle \sqrt{N}M^{N}_k(t),\sqrt{N}M^{N}_j(t)\boldrangle-\boldlangle M_k(t),M_j(t)\boldrangle\right|\xrightarrow{p} 0,$$
5. $$\label{condition_5}
\sup_{t\leq T}\left|\int_{0}^{t}\left\{\sqrt{N}[f(Y^N(s))-f(y(s))]-f'(y(s))D^{N}(s)\right\}ds\right|\xrightarrow{p} 0.$$
Condition (\[condition\_1\]) is easy to show by the fact that $D^N(t)$ has jump size of $1/\sqrt{N}$. Condition (\[condition\_2\]) follows by the fact that $F^N(y)$ is a Lipschitz function of $y$ and that condition (\[condition\_1\]) holds. By Lemma \[martingale\_brackets\] and the fact that $Y^N(t)$ has jump size of $1/N$, it is also easy to show that condition (\[condition\_3\]) holds. For condition (\[condition\_4\]), it follows from Proposition \[driftbound\] and Lemma \[martingale\_brackets\] (see for example proof of Equation (\[martingale\_brackets\_convergence\]) for details).
Finally, to show condition (\[condition\_5\]), by Equation (\[b’\]) we know that $f(y(t))$ is continuously differentiable with respect to $y(t)$. By the mean value theorem, for every $0\leq s\leq t$ there exists a vector $Z^{N}(s)$ in between $Y^{N}(s)$ and $y(s)$ such that $$f(Y^{N}(s))-f(y(s))=f'(Z^{N}(s))(Y^{N}(s)-y(s)).$$ Therefore, $$\int_{0}^{t}\left\{\sqrt{N}[f(Y^{N}(s))-f(y(s))]-f'(y(s))D^{N}(s)\right\}ds=\int_{0}^{t}[f'(Z^{N}(s))-f'(y(s))]D^{N}(s)ds.$$ We know that $$\lim_{N\rightarrow \infty}\sup_{t\leq T}|f'(Z^{N}(s))-f'(y(s))|=0\quad \text{in probability}$$ by the mean field limit convergence (Theorem \[fluid\_limit\]) and the uniform continuity of $f'$. By applying Chebyshev’s inequality we have that $D^{N}(s)$ is bounded in probability. Then, by Lemma 5.6 in @ko2018strong we have that $$\sup_{t\leq T}\left|\int_{0}^{t}\left\{\sqrt{N}[f(Y^N(s))-f(y(s))]-f'(y(s))D^{N}(s)\right\}ds\right|\xrightarrow{p} 0.$$
Like in the mean field case, we now state the functional central limit theorem for the empirical process described in Section \[model\_2\], where we consider battery usage time to be exponentially distributed.
\[difftheorem\_model2\] Define $$D^{N}(t)=\sqrt{N}((X^N(t),Y^N(t))-(x(t),y(t))).$$ Consider $D^{N}(t)$ in $\mathbb{D}(\mathbb{R}_{+},\mathbb{R}^{K+1})$ with the Skorokhod $J_{1}$ topology, and suppose that $$\limsup_{N\rightarrow \infty}\sqrt{N}\left(\frac{N^*}{N}-\gamma\right)< \infty.$$ Then if $D^{N}(0)$ converges in distribution to $D(0)$, then $D^{N}(t)$ converges to the unique OU process solving $D(t)=D(0)+\int_{0}^{t}f'(x(s), y(s))D(s) ds+M(t)$ in distribution, where $f'(x,y)$ is specified as follows, $$\begin{aligned}
\frac{\partial f_x(x,y)}{\partial x}&=&-(\mu+\mu_U)\sum_{k=K_U}^{K-1}y_k,\nonumber\\
\frac{\partial f_x(x,y)}{\partial y_k}&=&(\mu(1-x)-\mu_U x)\mathbf{1}\{k\geq K_U\}\nonumber\\
\frac{\partial f_y(x,y)(j)}{\partial x}&=&\sum_{k=\max(K_U,j)}^{K-1}\mu_Up_{kj}y_k-\lambda\gamma\left(\mathbf{1}\{j=K-1\}-\frac{g_jy_j}{\sum_{i=0}^{K-1}g_iy_i}\right)-\mu_U y_j \mathbf{1}\{j\geq K_U\},\nonumber\\
\frac{\partial f_y(x,y)(j)}{\partial y_k}&=&\mu_U x p_{k,j}\mathbf{1}\{k\geq \max\{j,K_U\}\}-\mu_U x \mathbf{1}\{k=j\geq K_U\},\nonumber\\
& & +\frac{\lambda \gamma (1-x) g_kg_jy_j}{(\sum_{i=0}^{K-1}y_ig_i)^2}-\frac{\lambda \gamma (1-x)g_j}{\sum_{i=0}^{K-1}y_ig_i}\mathbf{1}\{j= k\},\nonumber\\\end{aligned}$$ and $M(t)=(M_{x}(t), M_{y,0}(t),\cdots,M_{y,K-1}(t))\in \mathbb{R}^{K+1}$ is a real continuous centered Gaussian martingale, with Doob-Meyer brackets given by $$\begin{aligned}
\boldlangle M_{x}(t) \boldrangle & = & \int_{0}^{t}\left[(\mu(1-x(s))+\mu_U x(s))\sum_{k=K_U}^{K-1}y_k(s)\right]ds,\\
\boldlangle M_{x}(t), M_{y,j}(t)\boldrangle & = & \
-\int_{0}^{t}\left[\mu_U x(s)\sum_{k=\max\{j,K_U\}}^{K-1}p_{k,j}y_k(s)\right]ds,\end{aligned}$$ $$\begin{aligned}
& &\boldlangle M_{y,k}(t),M_{y,j}(t)\boldrangle \nonumber\\
&=&\begin{cases}
\int_{0}^{t}\left[\sum_{i=\max(K_U,k+1)}^{K-1}\mu_U x p_{i,k}y_i(s) +\left(\mu_Ux(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}\right.\right.\nonumber\\
+\left.\left.\frac{\lambda \gamma(1-x) g_k}{\sum_{i=0}^{K-1}y_i(s)g_i} \right)y_k(s)\right]ds, & k=j<K-1 \nonumber\\
\int_{0}^{t}\left[\mu_Ux(1-p_{K-1,K-1})y_{K-1}(s)+\lambda \gamma(1-x)\left(1- \frac{g_{K-1}y_{K-1}(s)}{\sum_{i=0}^{K-1}y_i(s)g_i}\right)\right]ds, & k=j=K-1\\
-\int_{0}^{t} \left[\mu_U x p_{k,j}y_k(s)+\frac{\lambda \gamma(1-x) g_j}{\sum_{i=0}^{K-1}y_i(s)g_i}y_j(s)\mathbf{1}\{k=K-1\}\right]ds, & j<k, k\geq K_U\nonumber\\
0. & \text{otherwise}
\end{cases}\nonumber\\\end{aligned}$$ Define $\mathcal{A}(t)=f'(x(t),y(t))$, $\mathcal{B}(t)=\left(\frac{d}{dt}\boldlangle M_i(t),M_j(t)\boldrangle \right)_{ij}$, then the covariance matrix $\Sigma (t)=\mathrm{Cov}[D(t),D(t)]$ satisfies $$\frac{d\Sigma(t)}{dt}=\Sigma(t)\mathcal{A}(t)^\top+\mathcal{A}(t)\Sigma(t)+\mathcal{B}(t).$$
The proof ideas for Theorem \[difftheorem\_model2\] follow easily from the proof of Theorem \[difftheorem\] so we do not prove them in this paper.
Insights to Staffing Swappers {#Staffing}
=============================
In Sections \[Mean\_Field\_Limit\] and \[Central\_Limit\], we prove mean field and central limit theorems for the empirical process of battery life. In this section, we show how to use these limits for providing insights for staffing the number of swappers to keep the number of e-scooters with low battery life below a pre-specified threshold. Our limit theorems are useful because when the scale of the system $N$ is large enough, the empirical process representing battery life can be approximated by a normal distribution, which is extremely convenient from a computational perspective. As a result, we can approximate the tail probability of the empirical measure process by the following expression $$\begin{aligned}
P(Y_k^N(t)>x)\approx P(y_k(t)+\sqrt{\Sigma_{kk}(t)/N}\cdot Z>x)=1-\Phi\left(\frac{x-y_k(t)}{\sqrt{\Sigma_{kk}(t)/N}}\right). \end{aligned}$$ where $Z\sim N(0,1)$ and $\Phi$ is the cdf of standard normal distribution.
Similarly, if we want to consider several terms of the empirical process $(Y_{\sigma(i)})_{i}$ where $\sigma$ is a permutation on $\{0,1,\cdots,K-1\}$. We have $$\begin{aligned}
P\left(\sum_{i=1}^{m}Y_{\sigma(i)}^N(t)>x\right)&\approx & P\left(\sum_{i=1}^{m}y_{\sigma(i)}(t)+\sqrt{\frac{1}{N}\left(\sum_{i=1}^{m}\Sigma_{\sigma(i)\sigma(i)}(t)+2\sum_{i<j}^{m}\Sigma_{\sigma(i)\sigma(j)}(t)\right)}\cdot Z>x\right)\nonumber\\
&=&1-\Phi\left(\frac{x-\sum_{i=1}^{m}y_{\sigma(i)}(t)}{\sqrt{\frac{1}{N}\left(\sum_{i=1}^{m}\Sigma_{\sigma(i)\sigma(i)}(t)+2\sum_{i<j}^{m}\Sigma_{\sigma(i)\sigma(j)}(t))\right)}}\right)\end{aligned}$$
An important question the above analysis can help answer is how many swappers are needed to keep the proportion of e-scooters with low battery life lower than a threshold, i.e. $P(Y_0>x)<\epsilon$ for given $(x,\epsilon)$. Specifically, we construct the following algorithm for finding the solution numerically.
\[algo\] Given $x>0, \epsilon>0$, we have the following steps of finding the number of swappers needed to satisfy $P(Y_0^N>x)\leq \epsilon$.
1. Initialize $\gamma=1$.
2. Evaluate $$f(\gamma)=\bar{y}_0(\gamma)+\sqrt{\frac{\bar{\Sigma}_{00}(\gamma)}{N}}\Phi^{-1}(1-\epsilon)-x$$ where $\bar{y}_0(\gamma),\bar{\Sigma}_{00}(\gamma)$ are the limiting mean and variance of $Y^N_0(t)$ at equilibrium (which are computed from the mean field and diffusion limits).
If $f(\gamma)>0$, set $\gamma \leftarrow 2\gamma$ (double the value of $\gamma$) and repeat step 2 until $f(\gamma)<0$. Denote the final value of $\gamma$ as $\gamma_{\max}$.
3. Apply bisection method on interval $[0,\gamma_{\max}]$ to find the root $\gamma^*$ to $f(\gamma)$. Then $\gamma^*$ is the optimal number of swapper per e-scooter needed.
The main idea of the algorithm is to use the mean field and central limit theorems to construct quantiles for each interval of battery life. More specifically, we invert the quantiles to find the number of swappers to achieve the probabilistic performance given by system operator. We will demonstrate the usefulness of this algorithm in the next section, which is devoted to numerical examples.
Numerical Examples and Simulation {#Numerics}
=================================
In this section, we use numerical examples and simulation results to provide better insights to the behavior of the e-scooters system. The examples validate our theoretical results in Sections \[Mean\_Field\_Limit\] and \[Central\_Limit\] by showing how accurate the mean field and diffusion limits are for approximating the mean and variance of the empirical process. We also illustrate how to use our results for staffing the number of swappers. The following simulation results are computed with the following parameter settings:
- The number of e-scooters $N=100$,
- The number of swappers $N^*=50$,
- Arrival rate of customers $\lambda=1$,
- Arrival rate of swappers $\mu=1$,
- Battery life bucket size $K=5$,
- Battery threshold for riding $K_U=1$,
- Battery usage probability $p_{ij}=\frac{1}{i+1}\mathbf{1}\{j\leq i\}$.
We initialize the battery life of scooters at time 0 to be uniform in each interval of battery life, i.e. $Y^N(0)=[0.2,0.2,0.2,0.2,0.2]$.
In Figure \[Sim\_1\], we simulate the e-scooter network according to the model dynamics where $K=5$ (intervals of battery life). It is important to note that we have only simulated one sample path in this picture and this is not an average of sample paths. Thus, we find that the mean field limit captures the sample path behavior of the empirical process dynamics for all of the proportions.
In Figure \[Sim\_2\], we average the dynamics over 100 sample paths. When compared to Figure \[Sim\_1\], we observe that the averaged dynamics are closer to the mean field limit trajectories, which is to be expected. We also find that the mean field limit equations capture the averaged sample path behavior of the empirical process dynamics for all of the proportions.
In Figure \[Sim\_3\], we average the variance dynamics for each proportion over 100 sample paths. We observe that the variance dynamics are a bit more stochastic than the mean field limit simulations. However, we find that the variance dynamics are well approximated by the variance of the central limit theorem for the e-scooter process and for all of the proportions.
To give an example on the implementation of Algorithm \[algo\] in Section \[Staffing\], we set $(x,\epsilon)=(10\%,10\%)$, i.e. find the minimum value of $\gamma$ such that the probability that proportion of scooters with low battery life (<20%) is more than 10% is no greater than 10%. The value of $\gamma$ found through the algorithm is 0.527.
In Figure \[hist\_y0\], we validate our algorithm by running 500 simulations using the optimal number of swappers found through Algorithm \[algo\] ($N=100$, $N^*=53$), and plot the distribution of $Y_0^N$ at equilibrium from the 500 simulations. Then we plot the 50%, 80%, 90% and 95% sample quantiles of $Y_0^N$ at equilibrium (pink solid lines) and compare them with the quantiles estimated from the algorithm (red solid lines), and we can see that the approximation using the algorithm is very close to reality. We also plot the normal distribution curve to see how good the approximation is to the distribution of $Y^N_0$. We find that the central limit approximations capture the quantile behavior of the e-scooter system quite well.
$(x,\epsilon)$ 0.01 0.05 0.1 0.15 0.2 0.25 0.3
---------------- ------- ------- ------- ------- ------- ------- -------
0.05 0.828 0.734 0.695 0.672 0.641 0.625 0.609
0.1 0.598 0.551 0.527 0.508 0.496 0.484 0.473
0.15 0.504 0.467 0.447 0.434 0.424 0.414 0.406
0.2 0.443 0.412 0.395 0.383 0.375 0.366 0.359
0.25 0.398 0.370 0.354 0.344 0.336 0.328 0.322
0.3 0.361 0.334 0.320 0.311 0.303 0.296 0.290
: Optimal $\gamma$ for different values of $(x,\epsilon)$ ($\lambda=\mu=1$)[]{data-label="table:gammas"}
Table \[table:gammas\] summarizes the optimal value of $\gamma$ (number of swappers per e-scooter, for different values of $(x,\epsilon)$. In addition, Figure \[gamma\_surface\] provides a surface plot of $\gamma$ over different values of $(x,\epsilon)$ that are given in Table \[table:gammas\] .
Conclusion {#Conclusion}
==========
In this paper, we construct two stochastic models for modeling the battery life dynamics for e-scooters in a large network. In full generality, the model is intractable when wants to keep track of each scooter’s dynamics individually because of the high dimension. However, we propose to use empirical processes to capture the essential dynamics of battery life in e-scooter systems. Empirical processes describe the proportion of e-scooters that have battery life in a particular interval. To this end, we prove a mean field limit and a functional central limit theorem for our e-scooter network. We show that the mean and the variance of the empirical process can be approximated by a system of $\frac{K^2 + 3K}{2}$ differential equations where $K$ is the number battery life intervals we want to keep. We use the mean and variance to also construct a numerical algorithm to compute the number of **swappers** needed to ensure that the fraction of scooters whose battery life is below a pre-determined threshold.
There are many directions for future work. As Figure \[Scooter\_Duration\_cdf\] shows, the trip durations are not exponential and are closer to a lognormal distribution or gamma distribution. An extension to general arrival and service distributions would aid in showing how the non-exponential distributions affect the dynamics of the empirical process. Recent work by @li2017nonlinear [@ko2017diffusion; @li2016mean; @pender2017approximations] provides a Poisson process representation of Markovian arrival processes. Thus, it might be useful to leverage this representation in future work where the arrivals and service distributions are non-renewal processes.
Although not explicitly studied in this work, it would be interesting to explore the impact of non-stationary arrival rates and service rates. This would undoubtedly change the underlying dynamics, however, we should mention that our analysis is easily generalizable to this setting. Moreover, our numerical algorithm for determining the number of **swappers** does not depend on the stationarity of our model and would easily generalize to the non-stationary setting. One important feature that is important to know in the non-stationary rate context is the size of the amplitude and the frequency of the mean field limit when the arrival rate is periodic. One way to analyze the amplitude and the frequency is to use Lindstedt’s method like in @novitzky2019nonlinear.
Lastly, it is also interesting to consider a spatial model of arrivals to the e-scooter network. In this case, we would consider customers arriving to the system via a spatial Poisson process and riders would choose among the nearest scooters with enough battery life to make their trip. This spatial process can model the real choices that riders make and would model the real spatial dynamics of e-scooter networks. We intend to pursue these extensions in future work.
Appendix {#Appendix}
========
Proof of Mean Field Limit Results {#proof-of-mean-field-limit-results .unnumbered}
---------------------------------
Our proof exploits Doob’s inequality for martingales and Gronwall’s lemma, and we use Proposition \[bound\], Proposition \[Lipschitz\], and Proposition \[drift\] in the proof, which are stated after the proof of Theorem \[fluid\_limit\].
Since $Y^{N}(t)$ is a semi-martingale, we have the following decomposition of $Y^N(t)$ , $$\label{semiY}
Y^{N}(t)=\underbrace{Y^{N}(0)}_{\text{initial condition}}+\underbrace{M^{N}(t)}_{\text{martingale}}+\int_{0}^{t}\underbrace{F^N(Y^{N}(s))}_{\text{drift term}}ds$$ where $Y^{N}(0)$ is the initial condition and $M^{N}(t)$ is a family of martingales. Moreover, $\int_{0}^{t}F^N(Y^{N}(s))ds$ is the integral of the drift term where the drift term is given by $F^N: [0,1]^{K}\rightarrow \mathbb{R}^{K}$ or $$\begin{aligned}
\label{F^N}
F^N(y)&=&\sum_{x\neq y}(x-y)Q^N(y,x)\\
&=&\sum_{k=0}^{K-1}\left[ \left( \frac{\lambda N^*}{N} \frac{g_k}{\sum_{i=0}^{K-1}y_ig_i} \right)(\mathbf{1}_{K-1}-\mathbf{1}_{k})+\sum_{j=0}^{k}\mu p_{kj} (\mathbf{1}_{j}-\mathbf{1}_{k})\mathbf{1}_{k\geq K_U}\right] y_k\nonumber \end{aligned}$$
We want to compare the empirical measure $Y^N(t)$ with the mean field limit $y(t)$ defined by $$y(t)=y(0)+\int_{0}^{t}f(y(s))ds.$$
Let $|\cdot|$ denote the Euclidean norm in $\mathbb{R}^{K}$, then $$\begin{aligned}
\left|Y^N(t)-y(t)\right|&=&\left|Y^{N}(0)+M^{N}(t)+\int_{0}^{t}F^N
(Y^{N}(s))ds-y(0)-\int_{0}^{t}f(y(s))ds\right|\nonumber \\
& =& \left|Y^{N}(0)-y(0)+M^{N}(s)+\int_{0}^{t}\left(F^N
(Y^{N}(s))-f(Y^{N}(s))\right)ds \right.\nonumber\\
& & \left.+\int_{0}^{t}(f(Y^{N}(s))-f(y(s)))ds\right|.\nonumber \\\end{aligned}$$ Now define the random function $f^{N}(t)=\sup_{s\leq t}\left|Y^{N}(s)-y(s)\right|$, we have $$\begin{aligned}
f^{N}(t) &\leq& |Y^{N}(0)-y(0)|+\sup_{s\leq t}|M^{N}(s)|+\int_0^t|F^N(Y^{N}(s))-f(Y^{N}(s))|ds \nonumber \\
&&+\int_{0}^{t}|f(Y^{N}(s))-f(y(s))|ds.\end{aligned}$$ By Proposition \[Lipschitz\], $f(y)$ is Lipschitz with respect to Euclidean norm. Let $L$ be the Lipschitz constant of $f(y)$, then. $$\begin{aligned}
f^{N}(t)&\leq& |Y^{N}(0)-y(0)|+\sup_{s\leq t}|M^{N}(s)|+\int_0^t|F^N(Y^{N}(s))-f(Y^{N}(s))|ds \nonumber \\
&&+\int_{0}^{t}|f(Y^{N}(s))-f(y(s))|ds \nonumber \\
&\leq& |Y^{N}(0)-y(0)|+\sup_{s\leq t}|M^{N}(s)|+\int_0^t| F^N(Y^{N}(s))-f(Y^{N}(s))|ds \nonumber \\
&&+L\int_{0}^{t}|Y^{N}(s)-y(s)|ds \nonumber \\
&\leq& |Y^{N}(0)-y(0)|+\sup_{s\leq t}|M^{N}(s)|+\int_0^t| F^N(Y^{N}(s))-f(Y^{N}(s))|ds \nonumber \\
&&+L\int_{0}^{t}f^{N}(s)ds.\end{aligned}$$
By Gronwall’s lemma (See @10.2307/1967124), $$f^{N}(t) \leq \left(|Y^{N}(0)-y(0)|+\sup_{s\leq t}|M^{N}(s)|+\int_0^t| F^N(Y^{N}(s))-f(Y^{N}(s))|ds\right)e^{Lt}.$$
Now to bound $f^{N}(t)$ term by term, we define function $\alpha: [0,1]^{K}\rightarrow \mathbb{R}^{K}$ as $$\begin{aligned}
\alpha_k(y)&=&\sum_{x\neq y}|x-y|^2Q^N(y,x)(k)\nonumber\\
&=&\begin{cases}
\frac{1}{N}\sum_{i=\max(K_U,k+1)}^{K-1}\mu p_{i,k}y_i +\frac{1}{N}\left(\mu(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}+\frac{\lambda N^* g_k}{N\sum_{i=0}^{K-1}y_ig_i} \right)y_k, & k<K-1\nonumber\\
\frac{1}{N}\mu(1-p_{K-1,K-1})y_{K-1}+\lambda \frac{N^*}{N^2}\left(1- \frac{g_{K-1}y_{K-1}}{\sum_{i=0}^{K-1}y_ig_i}\right), & k=K-1
\end{cases}\end{aligned}$$
and consider the following four sets $$\begin{aligned}
\Omega_0 &=& \{|Y^{N}(0)-y(0)|\leq \delta \}, \\
\Omega_1 &=& \left\{\int_0^{t_0}|F^N(Y^{N}(s))-f(Y^{N}(s))|ds\leq \delta \right\}, \\
\Omega_2 &=& \left\{\int_0^{t_0}\alpha(Y^{N}(t))dt \leq A(N)t_0 \right\}, \\
\Omega_3 &=& \left\{\sup_{t\leq t_0}|M_t^{N}|\leq \delta \right\} ,\end{aligned}$$ where $\delta=\epsilon e^{-Lt_0}/3$. Here the set $\Omega_{1}$ is to bound the initial condition, the set $\Omega_{2}$ is to bound the drift term $F^N$ and the limit of drift term $b$, and the sets $\Omega_{2},\Omega_{3}$ are to bound the martingale $M^{N}(t)$.
Therefore on the event $\Omega_0\cap \Omega_1\cap \Omega_3$, $$\label{e}
f^{N}(t_0)\leq 3\delta e^{Lt_0}=\epsilon.$$
Since $\lim_{N\rightarrow \infty}\frac{N^*}{N}=\gamma$, we can choose large enough $N$ such that $$\frac{N^*}{N}\leq 2\gamma.$$ Thus, we have $$\begin{aligned}
\alpha_k(y)&=&\sum_{x\neq y}|x-y|^2Q(y,x)(k)\nonumber\\
&\leq & \frac{1}{N}\sum_{i=\max(K_U,k+1)}^{K-1}\mu p_{i,k}y_i +\frac{1}{N}\left(\mu(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}+2\frac{\lambda \gamma g_k}{\sum_{i=0}^{K-1}y_ig_i} \right)y_k\nonumber \\
&\leq &\frac{1}{N}\left(\mu K +\mu +2\lambda \gamma \right)\nonumber \\
&\lesssim & O(\frac{1}{N}) .\end{aligned}$$
Now we consider the stopping time $$T=t_0\wedge \inf \left\{t\geq 0:\int _{0}^{t}\alpha(Y^{N}(s))ds>A(N)t_0\right\},$$ and by Proposition \[bound\], we have that $$\mathbb{E}\left(\sup_{t\leq T}|M^{N}(t)|^2\right)\leq 4\mathbb{E}\int_{0}^{T}\alpha(Y^N(t))dt\leq 4A(N)t_0.$$ On $\Omega_2$, we have $T =t_0$, so $\Omega_2 \cap \Omega_3^{c}\subset \{\sup_{t\leq T}|M^{N}_t|>\delta\}$. By Chebyshev’s inequality we have that $$\mathbb{P}(\Omega_2 \cap \Omega_3^{c})\leq \mathbb{P}\left(\sup_{t\leq T}|M^{N}_t|>\delta\right)\leq \frac{\mathbb{E}\left(\sup_{t\leq T}|M^{N}(t)|^2\right)}{\delta^2}\leq 4A(N)t_0/\delta^2.$$ Thus, by Equation (\[e\]), we have the following result, $$\begin{split}
\mathbb{P}\left(\sup_{t\leq t_0}|Y^{N}(t)-y(t)|>\epsilon\right)&\leq \mathbb{P}(\Omega_0^c\cup \Omega_1^c\cup \Omega_3^c)\\
&\leq \mathbb{P}(\Omega_2 \cap \Omega_3^{c})+\mathbb{P}(\Omega_0^{c}\cup \Omega_1^{c}\cup \Omega_2^{c})\\
&\leq 4A(N)t_0/\delta^2+\mathbb{P}(\Omega_0^{c}\cup \Omega_1^{c}\cup \Omega_2^{c})\\
&=36A(N)t_0 e^{2Lt_0}/\epsilon^2+\mathbb{P}(\Omega_0^{c}\cup \Omega_1^{c}\cup \Omega_2^{c}).
\end{split}$$\
Let $A(N)=\frac{4(C+\gamma)}{N}$, then $\Omega_{2}^{c}=\emptyset$. And since $Y_0^N\xrightarrow{p} y(0)$, $\lim_{N\rightarrow \infty}\mathbb{P}(\Omega_{2}^{c})=0$. Therefore we have $$\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{t\leq t_0}|Y^{N}(t)-y(t)|>\epsilon\right)=\lim_{N\rightarrow\infty}\mathbb{P}(\Omega_{1}^{c}) .$$ By Proposition \[drift\], $\lim_{N\rightarrow\infty}\mathbb{P}(\Omega_{1}^{c})=0$. Thus, we proved the final result $$\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{t\leq t_0}|Y^{N}(t)-y(t)|>\epsilon\right)=0.$$
\[bound\] For any stopping time $T$ such that $\mathbb{E}(T)<\infty$, we have $$\mathbb{E}\left(\sup_{t\leq T}|M^{N}(t)|^2\right)\leq 4\mathbb{E}\int_{0}^{T}\alpha(Y^N(t))dt.$$
See proof of Proposition 4.2 (page 46) in @tao2017stochastic.
\[Lipschitz\] The drift function $f(y)$ given in Equation (\[eqn:b\]) is a Lipschitz function with respect to the Euclidean norm in $\mathbb{R}^{K}$.
Denote $\|\cdot\|$ the Euclidean norm in $\mathbb{R}^{K}$. Consider $y,\tilde{y}\in [0,1]^{K}$, $$\begin{aligned}
\|f(y)-f(\tilde{y})\|&\leq& 2\left(\frac{\lambda \gamma \max_i g_i}{\min_i g_i}+\mu\right)\|y-\tilde{y}\|\end{aligned}$$ which proves that $f(y)$ is Lipschitz with respect to Euclidean norm in $\mathbb{R}^{K}$.
\[drift\] Under the assumptions of Theorem \[fluid\_limit\], we have for any $\epsilon>0$ and $s\geq 0$, $$\lim_{N\rightarrow \infty}P(|F^N(Y^{N}(s))-f(Y^{N}(s))|>\epsilon)= 0.$$
$$\begin{aligned}
\left| F^N(Y^{N}(s))-f(Y^{N}(s))\right|&=&\left|\sum_{k=0}^{K-1}\left(\frac{N^*}{N}-\gamma\right)\frac{Y^{N}(s)(k)g_k}{\sum_{i=0}^{K-1}Y^{N}(s)(i)g_i}(\mathbf{1}_{K-1}-\mathbf{1}_{k})\right|\nonumber\\
&\leq & 2\left|\frac{N^*}{N}-\gamma\right|\left|\sum_{k=0}^{K-1}\frac{\max_i g_i}{\min_i g_i} \right|\nonumber\\
&=& 2\frac{K\max_i g_i}{\min_i g_i} \left|\frac{N^*}{N}-\gamma\right| \rightarrow 0\end{aligned}$$
Proof of Central Limit Results {#proof-of-central-limit-results .unnumbered}
------------------------------
\[martingale\_brackets\] $\sqrt{N}M^{N}(t)$ is a family of martingales independent of $D^{N}(0)$ with Doob-Meyer brackets given by $$\begin{aligned}
& &\boldlangle \sqrt{N}M^{N}_k(t),\sqrt{N}M^{N}_j(t)\boldrangle\nonumber\\
&=&\begin{cases}
\int_{0}^{t}\left[\sum_{i=\max(K_U,k+1)}^{K-1}\mu p_{i,k}Y^N_i(s) +\left(\mu(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}\right.\right. & \nonumber\\
+\left.\left.\lambda N^*\frac{ g_k}{N\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_k(s)\right]ds, & k=j<K-1 \nonumber\\
\int_{0}^{t}\left[\mu(1-p_{K-1,K-1})Y_{K-1}^N(s)+\lambda \frac{N^*}{N}\left(1- \frac{g_{K-1}Y_{K-1}^N(s)}{\sum_{i=0}^{K-1}Y_i^N(s)g_i}\right)\right]ds, & k=j=K-1\\
-\int_{0}^{t} \left[\mu p_{k,j}Y^N_k(s)+\lambda N^*\frac{ g_j}{N\sum_{i=0}^{K-1}Y_i^N(s)g_i}Y_j^N(s)\mathbf{1}\{k=K-1\}\right]ds, & j<k, k\geq K_U\nonumber\\
0. & \text{otherwise}
\end{cases}\end{aligned}$$
By Dynkin’s formula, $$\begin{aligned}
\boldlangle \sqrt{N}M_k^{N}(t)\boldrangle&=&\int_{0}^{t}N\sum_{x\neq Y^{N}(s)}|x-Y^{N}(s)|^2 Q(Y^{N}(s),x)(k)ds\nonumber\\
&=&\int_{0}^{t}N\alpha_k(Y^N(s))ds\nonumber\\
&=&\begin{cases}
\int_{0}^{t}\left[\sum_{i=\max(K_U,k+1)}^{K-1}\mu p_{i,k}Y^N_i(s) +\left(\mu(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}\right.\right.\nonumber\\
+\left.\left.\frac{\lambda N^* g_k}{N\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_k(s)\right]ds, & k<K-1\nonumber\\
\int_{0}^{t}\left[\mu(1-p_{K-1,K-1})Y^N_{K-1}(s)+\lambda \frac{N^*}{N}\left(1- \frac{g_{K-1}Y^N_{K-1}(s)}{\sum_{i=0}^{K-1}Y^N_i(s)g_i}\right)\right]ds, & k=K-1
\end{cases}\\
&\triangleq &\int_{0}^{t}(F^N_{+}(Y^{N}(s))(k)+F^N_{-}(Y^{N}(s))(k))ds\end{aligned}$$ where $$F^N_{+}(Y^{N}(s))(k)=\begin{cases}
\sum_{i=\max(K_U,k+1)}^{K-1}\mu p_{i,k}Y^N_i(s), & k<K-1 \\
\lambda \frac{N^*}{N}\left(1- \frac{g_{K-1}Y^N_{K-1}(s)}{\sum_{i=0}^{K-1}Y^N_i(s)g_i}\right), & k=K-1
\end{cases}$$ and $$F^N_{-}(Y^{N}(s))(k)=\begin{cases}
\left(\mu(1-p_{k,k})\mathbf{1}_{\{k\geq K_U\}}+\frac{\lambda N^* g_k}{N\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_k(s), & k<K-1 \\
\mu(1-p_{K-1,K-1})Y^N_{K-1}(s), & k=K-1
\end{cases}$$ To compute $\boldlangle \sqrt{N}M_k^{N}(t),\sqrt{N}M_j^{N}(t)\boldrangle$ for $j<k$ and $k\geq K_U$, since $$\begin{split}
&\boldlangle M_k^{N}(t)+M_j^{N}(t)\boldrangle\\
=&\int_{0}^{t}\sum_{x\neq Y^{N}(s)}\left|x_k+x_j-Y^{N}(s)(k)-Y_{j}^{N}(s)\right|^2 Q(Y^{N}(s),x)ds\\
=&\frac{1}{N}\int_{0}^{t}\left[\mu\left(\sum_{i=\max(j+1,K_U),i\neq k}^{K-1}p_{i,j}+\sum_{i=k+1}^{K-1}p_{i,k}\right)Y^N_i(s) +\left(\mu(1-p_{j,j})+\frac{\lambda \gamma g_j}{\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_j(s) \right.\\
& \left. +\left(\mu(1-p_{k,k}-p_{k,j})+\frac{\lambda \gamma g_k}{\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_k(s)+\lambda \gamma \{k=K-1\}\right]ds
\end{split}$$ We have that $$\begin{split}
&\boldlangle \sqrt{N}M_k^{N}(t),\sqrt{N}M_j^{N}(t)\boldrangle\\
=&\frac{N}{2}\left[\boldlangle M_k^{N}(t)+M_j^{N}(t)\boldrangle-\boldlangle M_k^{N}(t)\boldrangle-\boldlangle M_j^{N}(t)\boldrangle\right]\\
=&\frac{1}{2}\int_{0}^{t}\left[\mu\left(\sum_{i=\max(j+1,K_U),i\neq k}^{K-1}p_{i,j}+\sum_{i=k+1}^{K-1}p_{i,k}\right)Y^N_i(s) +\left(\mu(1-p_{j,j})+\frac{\lambda \gamma g_j}{\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_j(s) \right.\\
& \left. +\left(\mu(1-p_{k,k}-p_{k,j})+\frac{\lambda \gamma g_k}{\sum_{i=0}^{K-1}Y^N_i(s)g_i} \right)Y^N_k(s)+\lambda \gamma \{k=K-1\}\right]ds\\
&-\frac{1}{2}\int_{0}^{t}(F^N_{+}(Y^{N}(s))(k)+F^N_{+}(Y^{N}(s))(j)+F^N_{-}(Y^{N}(s))(k)+F^N_{-}(Y^{N}(s))(j))ds\\
=&-\int_{0}^{t} \left[\mu p_{k,j}Y^N_k(s)+\lambda N^*\frac{ g_j}{N\sum_{i=0}^{K-1}Y_i^N(s)g_i}Y_j^N(s)\mathbf{1}\{k=K-1\}\right] ds.
\end{split}$$ Finally, $M_k^N$ and $M_j^N$ are independent when $j,k\leq K_U$, thus in this case the Doob-Meyer brackets is equal to 0.
\[driftbound\] For any $s\geq 0$, $$\limsup_{N\rightarrow \infty}\sqrt{N}\left|F^N(Y^{N}(s))-f(Y^{N}(s))\right|=0.$$ where $F^N$ is the drift function defined in Equation (\[F\^N\]).
$$\begin{aligned}
& &\limsup_{N\rightarrow \infty}\sqrt{N}\left|F^N(Y^{N}(s))-f(Y^{N}(s))\right|\nonumber\\
&=&\limsup_{N\rightarrow \infty}\sqrt{N}\left|\sum_{k=0}^{K-1}\left(\frac{N^*}{N}-\gamma\right)\frac{Y^{N}_{k}(s)g_k}{\sum_{i=0}^{K-1}Y^{N}_{i}(s)g_i}(\mathbf{1}_{K-1}-\mathbf{1}_{k})\right|\nonumber\\
&\leq & \limsup_{N\rightarrow \infty}2\sqrt{N}2\left|\frac{N^*}{N}-\gamma\right|\left|\sum_{k=0}^{K-1}\frac{\max_i g_i}{\min_i g_i} \right|\nonumber\\
&=& \limsup_{N\rightarrow \infty}2\sqrt{N}\frac{K\max_i g_i}{\min_i g_i} \left|\frac{N^*}{N}-\gamma\right| \nonumber\\
&=&0\end{aligned}$$
\[L2bound\] For any $T\geq 0$, if $$\limsup_{N\rightarrow \infty}\mathbb{E}\left(|D^{N}(0)|^2 \right) < \infty ,$$ then we have $$\limsup_{N\rightarrow \infty}\mathbb{E}\left(\sup_{0\leq t\leq T}|D^{N}(t)|^2 \right) < \infty .$$
By Proposition \[driftbound\], $\sqrt{N}|F^N(Y^{N}(s))-f(Y^{N}(s))|=O(1)$, then $$\begin{split}
|D^{N}(t)|&\leq |D^{N}(0)|+\sqrt{N}|M^{N}(t)|+O(1)t+\int_{0}^{t}\sqrt{N} |f(Y^{N}(s))-f(y(s))|ds\\
&\leq|D^{N}(0)|+\sqrt{N}|M^{N}(t)|+O(1)t+\int_{0}^{t}\sqrt{N}L|Y^{N}(s)-y(s)|ds\\
&=|D^{N}(0)|+\sqrt{N}|M^{N}(t)|+O(1)t+\int_{0}^{t}L|D^{N}(s)|ds.
\end{split}$$ By Gronwall’s Lemma, $$\sup_{0\leq t\leq T}|D^{N}(t)|\leq e^{LT}\left(|D^{N}(0)|+O(1)T+\sup_{0\leq t\leq T}|\sqrt{N}M^{N}(t)|\right),$$ then $$\limsup_{N\rightarrow \infty}\mathbb{E}\left(\sup_{0\leq t \leq T}|D^{N}(t)|^2\right)\leq e^{2LT}\left[\limsup_{N\rightarrow \infty}\mathbb{E}(|D^{N}(0)|)+O(1)T+\limsup_{N\rightarrow \infty}\mathbb{E}\left(\sup_{0\leq t\leq T}\sqrt{N}|M^{N}(t)|\right)\right]^2.$$ We know that $$\left[\mathbb{E}\left(\sup_{0\leq t\leq T}\sqrt{N}|M^{N}(t)|\right)\right]^2\leq N\mathbb{E}\left(\sup_{0\leq t\leq T}|M^{N}(t)|^2\right)\leq 4NA(N)T,$$ and that $A(N)=O(\frac{1}{N})$. Therefore $$\limsup_{N\rightarrow \infty}\mathbb{E}\left(\sup_{0\leq t\leq T}\sqrt{N}|M^{N}(t)| \right)<\infty.$$ Together with our assumption that $\limsup_{N\rightarrow \infty}\mathbb{E}(|D^{N}(0)|^2)<\infty$, we have $$\limsup_{N\rightarrow \infty}\mathbb{E}\left(\sup_{0\leq t \leq T}|D^{N}(t)|^2 \right)<\infty.$$\
\[tightness\] If $(D^{N}(0))_{N=1}^{\infty}$ is tight then $(D^{N})_{N=1}^{\infty}$ is tight and its limit points are continuous.
To prove the tightness of $(D^{N})_{N=1}^{\infty}$ and the continuity of the limit points, we can apply results from @billingsley2013convergence, which implies that we only need to show the following two conditions holds for each $T>0$ and $\epsilon>0$,
- $$\lim_{K\rightarrow \infty}\limsup_{N\rightarrow \infty}\mathbb{P}\left(\sup_{0\leq t\leq T}|D^{N}(t)|>K \right)=0,$$
- $$\lim_{\delta\rightarrow 0}\limsup_{N\rightarrow \infty}\mathbb{P}\left(w(D^{N},\delta,T)\geq \epsilon \right)=0$$
where for $x\in \mathbb{D}^{d}$, $$w(x,\delta,T)=\sup\left\{\sup_{u,v\in[t,t+\delta]}|x(u)-x(v)|:0\leq t\leq t+\delta\leq T\right\}.$$ By Lemma \[L2bound\], there exists $C_{0}>0$ such that $$\begin{aligned}
\lim_{K\rightarrow \infty}\limsup_{N\rightarrow \infty}\mathbb{P} \left(\sup_{0\leq t\leq T}|D^{N}(t)|>K \right) &\leq& \lim_{K\rightarrow \infty}\limsup_{N\rightarrow \infty}\frac{\mathbb{E}\left(\sup_{0\leq t\leq T}|D^{N}(t)|^2 \right)}{K^2} \\
&\leq& \lim_{K\rightarrow \infty}\frac{C_{0}}{K^2} \\
&=&0,\end{aligned}$$ which proves condition (i).
For condition (ii), we have that
$$\begin{aligned}
D^N(u) - D^N(v) &=& \underbrace{\sqrt{N} \cdot ( M^N(u) - M^N(v))}_{\text{first term}} + \underbrace{\int^{u}_{v} \sqrt{N} \left( F^N(Y^N(z)) - f(Y^N(z)) \right) dz}_{\text{second term}} \nonumber \\&+& \underbrace{\int^{u}_{v} \sqrt{N} \left( f(Y^N(z)) - f(y(z)) \right) dz }_{\text{third term}}\end{aligned}$$
for any $0<t\leq u<v\leq t+\delta\leq T$. Now it suffices to show that each of the three terms of $D^N(u) - D^N(v)$ satisfies condition (ii). In what follows, we will show that each of the three terms satisfies condition (ii) to complete the proof of tightness.
For the first term, similar to the proof of Proposition \[drift\], we can show that $$\sup_{t\leq T}\left|F^N_{+}(Y^N(t))-f_{+}(Y^N(t))\right|\xrightarrow{p}0, \quad \sup_{t\leq T}\left|F^N_{-}(Y^N(t))-f_{-}(Y^N(t))\right|\xrightarrow{p}0.$$ And by the proof of Proposition \[Lipschitz\], $f_{+}(y), f_{-}(y)$ are also Lipschitz with constant $L$, then by the fact that the composition of Lipschitz functions are also Lipschitz, $$\begin{aligned}
\max\left\{\sup_{t\leq T}|f_{+}(Y^N(t))-f_{+}(y(t))|,\sup_{t\leq T}|f_{-}(Y^N(t))-f_{-}(y(t))|\right\}\leq L\sup_{t\leq T}|Y^N(t)-y(t)|.\end{aligned}$$ By Theorem \[fluid\_limit\], $$\sup_{t\leq T}|Y^N(t)-y(t)|\xrightarrow{p} 0.$$ Thus for any $\epsilon>0$, $$\begin{aligned}
& &\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{t\leq T}\left|\boldlangle \sqrt{N}M_k^N(t)\boldrangle- \boldlangle M_k(t)\boldrangle\right|>\epsilon\right)\nonumber \\
&=&\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{t\leq T}\left|\int_{0}^{t}\left(F^N_{+}(Y^{N}(s))+F^N_{-}(Y^{N}(s))- f_{+}(y(s))-f_{-}(y(s))\right)ds\right|>\epsilon\right)\nonumber\\
&\leq & \lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{t\leq T}T\left|F^N_{+}(Y^N(t))- f_{+}(Y^{N}_{t})\right|>\epsilon/3\right)+\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{t\leq T}T\left|F^N_{-}(Y^N(t))- f_{-}(Y^{N}_{t})\right|>\epsilon/3\right)\nonumber\\
& &+\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{t\leq T}2LT\left|Y^N(t)- y(t)\right|>\epsilon/3 \right)\nonumber \\
&=& 0,\end{aligned}$$ which implies $$\label{martingale_brackets_convergence}
\sup_{t\leq T}\left|\boldlangle \sqrt{N}M^N_k(t)\boldrangle- \boldlangle M_k(t)\boldrangle\right|\xrightarrow{p} 0.$$ We also know that the jump size of $D^{N}(t)$ is $1/\sqrt{N}$, therefore $$\lim_{N\rightarrow \infty}\mathbb{E}\left[\sup_{0<t\leq T}\left|M^{N}(t)-M^{N}(t-)\right| \right]=0.$$ By Theorem 1.4 in Chapter 7 of @Ethier2009, $\sqrt{N}M^{N}(t)$ converges to the Brownian motion $M(t)$ in distribution in $\mathbb{D}(\mathbb{R}_{+},\mathbb{R}^{K+1})$. By Prohorov’s theorem, $(\sqrt{N}M^{N})_{N=1}^{\infty}$ is tight. And since $M(t)$ is a Brownian motion, its sample path is almost surely continuous.
For the second term, we have by Proposition \[driftbound\] that the quantity $ \sqrt{N} \left( F^N(Y^N(z)) - f(Y^N(z)) \right) $ is bounded for any value of $z\in [0,T]$. Therefore, there exists some constant $C_{1}$ that does not depend on $N$ such that $$\sup_{z\in [0,T]}\sqrt{N} \left| F^N(Y^N(z)) - f(Y^N(z)) \right|\leq C_{1}.$$ Then $$\begin{aligned}
& &\lim_{\delta\rightarrow 0}\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{u,v\in [0,T],|u-v|\leq \delta}\int^{u}_{v} \sqrt{N} \left| F^N(Y^N(z)) - f(Y^N(z)) \right| dz > \epsilon \right)\nonumber \\
&\leq & \lim_{\delta\rightarrow 0}\lim_{N\rightarrow \infty}\mathbb{P}\left(\delta \sup_{z\in [0,T]}\sqrt{N} \left| F^N(Y^N(z)) - f(Y^N(z)) \right| > \epsilon \right)\nonumber \\
&\leq & \lim_{\delta\rightarrow 0}\mathbb{P}\left(\delta C_{1} > \epsilon \right)\nonumber \\
&=& 0.\end{aligned}$$ Thus, we have proved the oscillation bound for the second term. Finally for the third term we have that $$\begin{aligned}
\int^{u}_{v} \sqrt{N} \left| f(Y^N(z)) - f(y(z)) \right| dz & \leq& \int^{u}_{v} \sqrt{N} L\left| Y^N(z) - y(z) \right| dz\nonumber \\
&=& \int^{u}_{v} L \cdot \left|D^N(z)\right| dz \nonumber\\\
&\leq & L\delta \sup_{t\in [0,T]}|D^{N}(t)|.\end{aligned}$$ By Lemma \[L2bound\], $$\begin{aligned}
& &\lim_{\delta\rightarrow 0}\lim_{N\rightarrow \infty}\mathbb{P}\left(\sup_{u,v\in [0,T],|u-v|\leq \delta}\int^{u}_{v} \sqrt{N} \left| f(Y^N(z)) - f(y(z)) \right| dz>\epsilon\right)\nonumber\\
&\leq &\lim_{\delta\rightarrow 0}\lim_{N\rightarrow \infty}\mathbb{P}\left(L\delta\sup_{t\in [0,T]}|D^{N}(t)|>\epsilon\right)\nonumber\\
&\leq & \lim_{\delta\rightarrow 0}\lim_{N\rightarrow \infty}\frac{\mathbb{E}\left(\sup_{t\in [0,T]}|D^{N}(t)|^2\right)}{(\epsilon/L \delta)^2}\nonumber \\
&\leq &\lim_{\delta\rightarrow 0}\frac{C_{0}(L\delta)^2}{\epsilon^2}\nonumber \\
&=& 0,\end{aligned}$$ which implies that the oscillation bound holds for the third term.
|
---
abstract: 'We construct massive open string states around a classical solution in the oscillator formulation of Vacuum String Field Theory. In order for the correct mass spectrum to be reproduced, the projection operators onto the modes of the left- and right-half of the string must have an anomalous eigenvalue $1/2$, and the massive states are constructed using the corresponding eigenvector. We analyze numerically the projection operators by regularizing them to finite size matrices and confirm that they indeed have eigenvalue $1/2$. Beside the desired massive states, we have spurious massive as well as massless states, which are infinitely degenerate. We show that these unwanted states can be gauged away.'
author:
- |
Hiroyuki [Hata]{}[^1], and Hisashi [Kogetsu]{}[^2]\
[*Department of Physics, Kyoto University, Kyoto 606-8502, Japan*]{}
date: 'August, 2002'
title: '\'
---
=17.5pt plus 0.2pt minus 0.1pt
addtoreset[equation]{}[section]{}
Introduction and summary
========================
Vacuum String Field Theory (VSFT) [@RSZ:0012251; @RSZ:0102112; @Rastelli:2001vb; @RSZ:0106010] has been proposed as a string field theory expanded around the tachyon vacuum. In order for VSFT to really be connected to ordinary bosonic string theory on an unstable D25-brane, there must exist a Lorentz and translationally invariant classical solution of VSFT satisfying the following two requirements: the fluctuation modes around the solution reproduce the open string spectrum, and the energy density of the solution is equal to the D25-brane tension. Recently, there has been much progress in understanding the above problem. In particular, a full classical solution of VSFT including the ghost part has been presented in [@HK:0108150] using the oscillator formulation, and the tachyon and the massless vector fluctuation modes have been constructed there. They have shown that the tachyon mass is correctly reproduced. However, the massless vector mode contains an arbitrary vector in the level number space, implying that there are infinite number of massless vector states. This problem was later resolved by Imamura [@Imamura:0204031]: most of the massless vector modes can be gauged away by VSFT gauge transformation leaving only one physical vector mode.
The purpose of this paper is to carry out the oscillator construction of fluctuation modes representing higher level massive modes of open string.[^3] This is in fact a non-trivial and interesting problem. Let us take, as a candidate massive state with mass squared equal to $(k-1)\alpha'$, a state given as $k$ matter creation operators $a_n^{\mu\dagger}$ acting on the tachyon state. It is a natural extension of the tachyon and massless vector states of [@HK:0108150]. However, naive analysis shows that this kind of states are all massless. The wave equation (namely, the linearized equation of motion), ${{\mathcal{Q}}_{\rm B}}\Phi=0$, for the fluctuation $\Phi$ of the above type is reduced to a simple algebraic equation consisting only of the projection operators, $\rho_+$ and $\rho_-$, onto the modes of the left- and right-half of the string [@Rastelli:2001rj]. The masslessness is a consequence of the basic property of projection operators, $\rho_\pm^2 =\rho_\pm$.
We find, however, that the above mode can represent a massive state with the expected mass squared, $(k-1)\alpha'$, if $\rho_\pm$ has an anomalous eigenvalue $1/2$ despite that it is a projection operator. Such an anomalous eigenvalue is of course impossible for projection operators in a finite dimensional space. However, there is a subtle point for $\rho_\pm$ which is an operator in the infinite dimensional space of string level number. In fact, the eigenvector ${\boldsymbol{f}}^{(\kappa)}$ of the matrix representation of the Virasoro algebra $K_1=L_1+ L_{-1}$ (the eigenvalue $\kappa$ is continuous and extending from $-\infty$ to $\infty$) is at the same time the eigenvector of $\rho_\pm$ and the corresponding eigenvalue is the step function $\theta(\pm\kappa)$ [@RSZ:0111281]. Therefore, the vector ${{\boldsymbol{f}}^{(0)}}$ is the eigenvector of $\rho_\pm$ with eigenvalue $\theta(0)$, which is indefinite but could be the desired value $1/2$. To verify whether this expectation is correct, we have to study $\rho_\pm$ with some kind of regularization. In this paper, we analyze numerically the eigenvalue problem of $\rho_\pm$ regularized to finite size matrices to obtain results supporting the above expectation: $\rho_\pm$ has an eigenvalue and an eigenvector which tend to $1/2$ and ${{\boldsymbol{f}}^{(0)}}$, respectively, as the size of the matrices is increased.
Even if $\rho_\pm$ has the expected anomalous eigenvalue $1/2$, there still remains a problem to be solved for the construction of higher level open string modes. Analysis of the wave equation for fluctuations of the above type, $(a^\dagger)^k{| \rm tachyon \rangle}$, shows that there still exists infinite degeneracy of massive states with mass squared equal to $\ell\alpha'$ ($\ell\le k-2$). In addition, we also have spurious massless states mentioned above. We have to show that these unwanted states are not physical ones. This problem is solved in the same manner as in the massless vector case [@Imamura:0204031]: infinite number of spurious states can be gauged away. However, we need gauge transformations of a different kind from that used in [@Imamura:0204031] in order to remove all the unwanted massive states.
Our construction of massive open string states is not complete, and there remain a number of future problems. First, we have to present a rigorous analytic proof of the existence of the anomalous eigenvalue $1/2$ of the projection operators $\rho_\pm$. Second, as we shall see later, we construct only the highest spin states at a given mass level. The construction of lower spin states is our remaining subject. Finally, in our analysis we consider the wave equation ${{\mathcal{Q}}_{\rm B}}\Phi=0$ only in the Fock space of first quantized string states. Namely, our massive modes $\Phi$ satisfy the wave equation in the sense of ${\langle \rm Fock |} {{\mathcal{Q}}_{\rm B}}{| \Phi \rangle} = 0$ for any Fock space element ${\langle \rm Fock |}$. However, analysis of the potential height problem of the D25-brane solution of VSFT shows that we have to consider the equation of motion in a larger space including the states constructed upon the D25-brane solution [@Rastelli:2001wk; @Okawa:0204012; @Reexam]. This is our important future problem.
The organization of the rest of this paper is as follows. In sec. 2, we analyze the equation of motion for our candidate massive modes and argue that $\rho_\pm$ needs an anomalous eigenvalue $1/2$. In sec. 3, we present numerical analysis of the eigenvalue problem of finite size $\rho_\pm$. In sec. 4, we show that the spurious states are unphysical ones which can be removed by gauge transformations. In appendix \[evaluation\], we present technical details used in the text.
Massive modes {#section-massive}
=============
The action of VSFT is given by [@RSZ:0012251; @RSZ:0102112; @RSZ:0106010] $$S=-K{\left}({\frac{1}{2}}\Psi\cdot{\mathcal{Q}}\,\Psi + \frac13\,\Psi\cdot(\Psi*\Psi){\right}),
\label{eq:VSFT-action}$$ where $K$ is a constant and the BRST operator ${\mathcal{Q}}$ of VSFT consists purely of ghost oscillators: $${\mathcal{Q}}=c_0+\sum_{n=1}^\infty f_n\left(c_n +(-1)^n c_n^\dagger\right) .
\label{eq:cQ}$$ The VSFT action (\[eq:VSFT-action\]) is invariant under the gauge transformation, $$\delta_\Lambda\Psi={\mathcal{Q}}\Lambda+\Psi*\Lambda-\Lambda*\Psi.
\label{eq:dLPsi}$$ The D25-brane configuration of VSFT is a translationally and Lorentz invariant solution ${\Psi_\mathrm{c}}$ to the equation of motion: $${\mathcal{Q}}{\Psi_\mathrm{c}}+{\Psi_\mathrm{c}}*{\Psi_\mathrm{c}}=0 .
\label{eq:EOM}$$ Assuming that ${\Psi_\mathrm{c}}$ factorizes into the matter part ${\Psi_\mathrm{c}^\mathrm{m}}$ and the ghost one ${\Psi_\mathrm{c}^\mathrm{g}}$, ${\Psi_\mathrm{c}}={\Psi_\mathrm{c}^\mathrm{m}}\otimes{\Psi_\mathrm{c}^\mathrm{g}}$, (\[eq:EOM\]) is reduced into the following two: $$\begin{aligned}
&{\Psi_\mathrm{c}^\mathrm{m}}={\Psi_\mathrm{c}^\mathrm{m}}*{\Psi_\mathrm{c}^\mathrm{m}},
\label{eq:EOMm}
\\
&{\mathcal{Q}}{\Psi_\mathrm{c}^\mathrm{g}}+{\Psi_\mathrm{c}^\mathrm{g}}*{\Psi_\mathrm{c}^\mathrm{g}}=0 .
\label{eq:EOMg}\end{aligned}$$
Here we shall fix our convention for the Neumann coefficient matrices. The matter part of the three-string vertex defining the $*$-product is given in the oscillator representation by $${| V^{\mathrm{m}} \rangle}_{123}=
\exp\!{\left}(-\sum_{r,s=1}^3\sum_{m,n\geq 0}{\frac{1}{2}}a_m^{(r)\dagger}V_{mn}^{rs}a_n^{(r)\dagger}
{\right}){| p_1 \rangle}{| p_2 \rangle}{| p_3 \rangle} ,
\label{eq:V}$$ with $a_0=\sqrt{2}\,p$ (we are taking the convention of $\alpha'=1$). The Neumann coefficient matrices and the vectors, $M_\alpha$ and ${\boldsymbol{v}}_\alpha$ ($\alpha=0,\pm$), are related to $V^{rs}$ in (\[eq:V\]) as follows: $$\begin{aligned}
&(M_0)_{mn}= (C V^{rr})_{mn},\quad
(M_\pm)_{mn}=(C V^{r,r\pm 1})_{mn} ,
{\nonumber}\\
&({\boldsymbol{v}}_0)_n = V^{rr}_{n0},\quad
({\boldsymbol{v}}_\pm)_n = V^{r,r\pm 1}_{n0} ,
\quad (m,n\ge 1)\end{aligned}$$ where $C$ is the twist matrix, $C_{mn}=(-1)^m\delta_{mn}$.
The matter part solution ${\Psi_\mathrm{c}^\mathrm{m}}$ to (\[eq:EOMm\]) has been obtained as a squeezed state [@KP:0008252; @RSZ:0102112]: $${| {\Psi_\mathrm{c}^\mathrm{m}}\rangle}=\left[\det(1-T{\mathcal{M}})\right]^{13}
\exp\biggl(-{\frac{1}{2}}\sum_{m,n\ge 1}a_m^\dagger (CT)_{mn}a_n^\dagger
\biggr){| 0 \rangle} ,$$ where the matrix $T$ is given in terms of $M_0$ by $$T=\frac{1}{2 M_0}\left(1+M_0-\sqrt{(1-M_0)(1+3M_0)}\right) .
\label{eq:T}$$ The ghost part solution ${\Psi_\mathrm{c}^\mathrm{g}}$ to (\[eq:EOMg\]) has also been obtained by taking the Siegel gauge and assuming the squeezed state form [@HK:0108150]. Beside determining ${\Psi_\mathrm{c}^\mathrm{g}}$, (\[eq:EOMg\]) fixes the coefficients $f_n$ in ${\mathcal{Q}}$ (\[eq:cQ\]) which are arbitrary for the gauge invariance alone.
Let us express the VSFT field $\Psi$ as a sum of ${\Psi_\mathrm{c}}$ and the fluctuation $\Phi$: $$\Psi={\Psi_\mathrm{c}}+ \Phi .
\label{eq:Psi=Psic+Phi}$$ Then the linear part of the equation of motion for $\Phi$ reads $${{\mathcal{Q}}_{\rm B}}\Phi\equiv
{\mathcal{Q}}\Phi+{\Psi_\mathrm{c}}*\Phi+\Phi*{\Psi_\mathrm{c}}=0 ,
\label{eq:wtcQ}$$ where ${{\mathcal{Q}}_{\rm B}}$ is the BRST operator around the classical solution ${\Psi_\mathrm{c}}$. We would like to construct the fluctuation modes $\Phi$ corresponding to higher level open string states and satisfying the wave equation (\[eq:wtcQ\]). We assume the factorization for these modes and that the ghost part is common to that of ${\Psi_\mathrm{c}}$: $$\Phi={\Phi^\mathrm{m}}\otimes{\Psi_\mathrm{c}^\mathrm{g}}.
\label{eq:Phi=PhimPsicg}$$ Then the wave equation for the matter part ${\Phi^\mathrm{m}}$ is given by $${\Phi^\mathrm{m}}={\Psi_\mathrm{c}^\mathrm{m}}*{\Phi^\mathrm{m}}+{\Phi^\mathrm{m}}*{\Psi_\mathrm{c}^\mathrm{m}}.
\label{eq:EOMf}$$ In the following we are interested only in the matter part ${\Phi^\mathrm{m}}$ of the fluctuation modes and omit its superscript $\mathrm{m}$. Eq. (\[eq:EOMf\]) has been solved for the tachyon mode ${\Phi_\mathrm{t}}$ [@HK:0108150]. Explicitly, it is given by $${| {\Phi_\mathrm{t}}\rangle}=\exp\biggl(
-\sum_{n\ge 1}t_n a_n^\dagger a_0 +ip\,\widehat x\biggr)
{| {\Psi_\mathrm{c}^\mathrm{m}}\rangle} ,
\label{eq:tachyon}$$ with $${\boldsymbol{t}}=3(1+T)(1+3M_0)^{-1}{\boldsymbol{v}}_0.$$ It has been shown that ${\Phi_\mathrm{t}}$ (\[eq:tachyon\]) satisfies (\[eq:EOMf\]) when the momentum $p_\mu$ carried by ${\Phi_\mathrm{t}}$ is on the tachyon mass-shell $p_\mu^2=-m_{\rm tachyon}^2=1$.
Now we shall start constructing fluctuation modes at a generic mass level. As a candidate fluctuation mode with mass squared equal to $k-1$, let us take the following one; $k$ creation operators acting on the tachyon mode ${\Phi_\mathrm{t}}$: $${| {\Phi^{(k)}}\rangle}=\sum_{n_1,\cdots,n_k\ge 1}
\beta^{\mu_1\cdots\mu_k}_{{n_1}\cdots{n_k}}
{a_{n_1}^{\mu_1\dagger}}\cdots{a_{n_k}^{\mu_k\dagger}}
{| {\Phi_\mathrm{t}}\rangle} ,
\label{eq:massive-state}$$ where $\beta$ is an unknown coefficient satisfying $$\beta^*_{n_1\cdots n_k}=(-1)^k(-1)^{\sum_{i=1}^k n_i}
\beta_{n_1\cdots n_k} ,
\label{eq:bb*}$$ which is due to the hermiticity constraint of ${\Phi^{(k)}}$ (see appendix \[evaluation\]). Substituting (\[eq:massive-state\]) into (\[eq:EOMf\]), we obtain equations determining $\beta$. The detailed calculation using the oscillator expression of the three-string vertex is presented in appendix \[evaluation\]. Although the assumed state (\[eq:massive-state\]) has fixed number $k$ of creation operators $a^\dagger$ acting on ${\Phi_\mathrm{t}}$, there emerge on the RHS of (\[eq:EOMf\]) states with fewer number of $a^\dagger$ acting on ${\Phi_\mathrm{t}}$ besides those with $k$ $a^\dagger$s. In order to eliminate these unwanted terms, we impose the following transverse and traceless conditions on $\beta$:[^4] $$\begin{aligned}
&p_{\mu_1}\beta^{\mu_1\cdots\mu_k}=0,
\label{eq:transverse-condition}
\\
&\beta_{\mu_1}{}^{\mu_1\cdots\mu_{k-1}}=0 .
\label{eq:traceless-condition}\end{aligned}$$ Then the equation for the coefficient $\beta$ is given by $$\beta_{m_1\cdots m_k}- 2^{-p^2}
\Bigl({(\rho_-)}_{m_1n_1}\cdots {(\rho_-)}_{m_kn_k}
+{(\rho_+)}_{m_1n_1}\cdots {(\rho_+)}_{m_kn_k}
\Bigr)\beta_{n_1\cdots n_k}
=0 ,
\label{eq:mass-eq}$$ with $$\rho_\pm=\frac{TM_\pm+M_\mp}{(1+T)(1-M_0)}.
\label{eq:rhopmcomm}$$ The matrices $\rho_\pm$ are projection operators [@Rastelli:2001rj] satisfying $$\left(\rho_\pm\right)^2=\rho_\pm,\quad
\rho_+\rho_-=\rho_-\rho_+=0,\quad
\rho_+ +\rho_-=1 .
\label{eq:proj}$$
Eqs. (\[eq:mass-eq\]) and (\[eq:proj\]) lead to a disappointing result that our states (\[eq:massive-state\]) can represent only massless states. Namely, multiplying (\[eq:mass-eq\]) by $\rho_{s_1}\!\otimes\!\rho_{s_2}\!\otimes\cdots\otimes\!\rho_{s_k}$ with $s_i=+$ or $-$, we find that the equations for the purely $\rho_+$ component $(\rho_+\otimes\cdots\otimes\rho_+)\beta$ and the purely $\rho_-$ one $(\rho_-\otimes\cdots\otimes\rho_-)\beta$ are reduced to $$\bigl(1- 2^{-p^2}\bigr)(\rho_\pm\otimes\cdots\otimes\rho_\pm)
\beta=0 ,$$ implying that they are massless states. On the other hand, (\[eq:mass-eq\]) tells that the mixed components, for example, $(\rho_+\otimes\rho_-\otimes\cdots\otimes\rho_-)\beta$, are equal to zero. In the case of vector state with $k=1$, eq. (\[eq:mass-eq\]) with $\rho_+ +\rho_-=1$ substituted reproduces the result of [@HK:0108150] that this is a massless state and the coefficient $\beta_n^\mu$ is arbitrary.
The above result is inevitable so long as the basic equations (\[eq:proj\]) are valid. However, there is a subtle point concerning the eigenvalues of $\rho_\pm$. Recall that the eigenvalue problem of the Neumann coefficient matrices $M_0$ and $M_1\equiv M_+ - M_-$ has been solved in [@RSZ:0111281]. They found that these matrices are expressed in terms of a single matrix $K_1$ which is the matrix representation of the Virasoro algebra $L_1+L_{-1}$, and the eigenvalue problem of $M_\alpha$ is reduced to that of $K_1$. Let ${\boldsymbol{f}}^{(\kappa)}$ be the eigenvector of $K_1$ corresponding to the eigenvalue $\kappa$: $$K_1{\boldsymbol{f}}^{(\kappa)}=\kappa{\boldsymbol{f}}^{(\kappa)} .
\label{eq:Kf=kf}$$ The distribution of $\kappa$ is uniform and extending from $-\infty$ to $\infty$. This eigenvector ${\boldsymbol{f}}^{(\kappa)}$ is at the same time that of $M_0$, $M_1$, $T$ and hence $\rho_\pm$. In particular we have $$\rho_\pm{\boldsymbol{f}}^{(\kappa)}=\theta(\pm\kappa){\boldsymbol{f}}^{(\kappa)} ,
\label{eq:rhof=kappaf}$$ where $\theta(\kappa)$ is the step function $$\theta(\kappa)=
\begin{cases}
1 & (\kappa > 0) \\ 0 & (\kappa <0)
\end{cases} .
\label{eq:theta}$$ The subtle point is the eigenvalue of $\rho_\pm$ at $\kappa=0$. In fact, it has been known that there is an eigenvector ${\boldsymbol{f}}^{(\kappa=0)}$, which is twist-odd, $C{\boldsymbol{f}}^{(0)}=-{\boldsymbol{f}}^{(0)}$. However, the eigenvalue $\theta(\kappa=0)$ of $\rho_\pm$ is indefinite.
If we are allowed to set $\theta(0)=1/2$ in (\[eq:rhof=kappaf\]), which would look most plausible, the fluctuation (\[eq:massive-state\]) represents a massive state at the expected mass level, $$p^2=1-k ,$$ by adopting either of the following two choices of $\beta_{n_1\cdots n_k}^{\mu_1\cdots\mu_k}$ concerning its dependence on the level number indices $n_1\cdots n_k$:
- $\beta$ is the tensor product of $k$ ${{\boldsymbol{f}}^{(0)}}$s, $$\beta={{\boldsymbol{f}}^{(0)}}\otimes\cdots\otimes{{\boldsymbol{f}}^{(0)}}.
\label{eq:b=f...f}$$
- $\beta$ is the tensor product of $(k-1)$ ${{\boldsymbol{f}}^{(0)}}$s and one arbitrary vector ${\boldsymbol{w}}$, $$\beta={{\boldsymbol{f}}^{(0)}}\otimes\cdots\otimes{{\boldsymbol{f}}^{(0)}}\otimes{\boldsymbol{w}} .
\label{eq:b=f...fw}$$
In both cases we have to multiply (\[eq:b=f...f\]) and (\[eq:b=f...fw\]) by a transverse and traceless tensor carrying the Lorentz indices, and carry out symmetrization if necessary. Quite similarly, by taking $\beta$ which is a tensor product of $\ell$ ${{\boldsymbol{f}}^{(0)}}$s and $k-\ell$ arbitrary vectors ($\ell \le k-1$), we obtain a state at mass level $p^2=-\ell$.[^5]
Now we have to resolve two problems. One is whether $\rho_\pm$ really has eigenvalue $1/2$. Second, even if this is the case, we have infinite degeneracy of massive as well as massless states which are apparently physical ones. We have to show that these spurious states are gauge artifacts. Analysis of these two questions is the subject of the following two sections.
Numerical analysis of ${\boldsymbol{\rho_\pm}}$
===============================================
As seen in the previous section, the existence of the eigenvalue $1/2$ of the “projection operators” $\rho_\pm$ was essential for the construction of the massive fluctuation modes. The corresponding eigenvector is expected to be ${{\boldsymbol{f}}^{(0)}}$, the zero-mode of $K_1$. It is obvious that we need some regularization for studying this expectation since what we want to know is the value of the step function $\theta(\kappa)$ (\[eq:theta\]) at $\kappa=0$. In the following, we shall show numerically that $\rho_\pm$ has indeed eigenvalue $1/2$ by regularizing them to finite size matrices.
We have solved numerically the eigenvalue problem of the regularized $\rho_+$ obtained by replacing $M_0$ and $M_1$ in it with $L\times L$ ones. Since we have $\rho_-=C\rho_+ C$, we do not need to repeat the analysis for $\rho_-$. The expression (\[eq:rhopmcomm\]) of $\rho_\pm$ was in fact obtained by naively using the non-linear relations among the Neumann coefficient matrices [@GJ1; @GJ2] upon the original expressions, which are given by (see appendix \[evaluation\]) $$\rho_\pm =(M_+,M_-)(1-T{\mathcal{M}})^{-1}
\begin{cases}
{
\left(\begin{smallmatrix} 0\\ 1 \end{smallmatrix}\right)}\\{
\left(\begin{smallmatrix} 1\\ 0 \end{smallmatrix}\right)}
\end{cases} ,
\label{eq:rhopmorig}$$ with $${\mathcal{M}}=\begin{pmatrix}
M_0 & M_+ \\ M_- & M_0
\end{pmatrix} .
\label{eq:cM}$$ Since the non-linear relations no longer hold for regularized $M_\alpha$ and naive use of them may be dangerous near $\kappa=0$ [@Anomaly; @HatMorTer; @Reexam], we have employed the original expression (\[eq:rhopmorig\]) in our numerical analysis.
Tables \[rho+-even-eigenvalue\] – \[rho+-odd-eigenvector\] show the result of our calculations. Since the eigenvalue distributions are qualitatively different between even and odd $L$, we have carried out the analysis for each of these two cases. In the case of even $L$, all the eigenvalues of $\rho_+$ are close to either $0$ or $1$ except two “anomalous” ones, $\lambda^{(1)}$ and $\lambda^{(2)}$ , which are given in table \[rho+-even-eigenvalue\] for various even $L$. Though the raw values of these anomalous eigenvalues are not so close to $1/2$, their values at $L=\infty$ obtained by fitting are surprisingly close to the expected value of $1/2$.
$L$ $\lambda^{(1)}$ $\lambda^{(2)}$
---------- ----------------- -----------------
$50$ $0.771$ $0.259$
$100$ $0.752$ $0.279$
$150$ $0.742$ $0.289$
$200$ $0.735$ $0.295$
$300$ $0.726$ $0.304$
$500$ $0.716$ $0.314$
$\infty$ $0.512$ $0.489$
: Anomalous eigenvalues of $\rho_+$ (\[eq:rhopmorig\]) for various even $L$. The values at $L=\infty$ have been obtained by the fitting function of the form $\sum_{k=0}^5 c_k/\left(\ln L\right)^k$. We use the same fitting function also in other tables \[rho+-even-eigenvector1\] – \[rho+-odd-eigenvector\]. []{data-label="rho+-even-eigenvalue"}
$L$ $a^{(1)}_3$ $a^{(1)}_5$ $a^{(1)}_7$ $a^{(1)}_9$ $a^{(1)}_{11}$ $b^{(1)}_4$ $b^{(1)}_6$ $b^{(1)}_8$ $b^{(1)}_{10}$ $b^{(1)}_{12}$
---------- ------------- ------------- ------------- ------------- ---------------- ------------- ------------- ------------- ---------------- ----------------
$50$ $1.104$ $1.185$ $1.256$ $1.320$ $1.382$ $-1.073$ $1.119$ $-1.030$ $1.208$ $-1.257$
$100$ $1.086$ $1.150$ $1.204$ $1.251$ $1.295$ $-1.040$ $1.055$ $-1.067$ $1.079$ $-1.092$
$150$ $1.077$ $1.134$ $1.183$ $1.223$ $1.260$ $-1.026$ $1.029$ $-1.030$ $1.032$ $-1.035$
$200$ $1.072$ $1.125$ $1.168$ $1.205$ $1.239$ $-1.018$ $1.015$ $-1.010$ $1.007$ $-1.005$
$300$ $1.065$ $1.113$ $1.151$ $1.185$ $1.215$ $-1.009$ $0.998$ $-0.987$ $0.977$ $-0.970$
$500$ $1.058$ $1.100$ $1.134$ $1.163$ $1.189$ $-1.000$ $0.981$ $-0.964$ $0.949$ $-0.937$
$\infty$ $0.999$ $1.003$ $1.008$ $1.012$ $1.008$ $-0.944$ $0.883$ $-0.819$ $0.735$ $-0.606$
: Components of the vectors ${\boldsymbol{a}}^{(1)}$ and ${\boldsymbol{b}}^{(1)}$ for various even $L$. []{data-label="rho+-even-eigenvector1"}
$L$ $a^{(2)}_3$ $a^{(2)}_5$ $a^{(2)}_7$ $a^{(2)}_9$ $a^{(2)}_{11}$ $b^{(2)}_4$ $b^{(2)}_6$ $b^{(2)}_8$ $b^{(2)}_{10}$ $b^{(2)}_{12}$
---------- ------------- ------------- ------------- ------------- ---------------- ------------- ------------- ------------- ---------------- ----------------
$50$ $1.064$ $1.114$ $1.159$ $1.201$ $1.243$ $-1.030$ $1.038$ $-1.047$ $1.058$ $-1.074$
$100$ $1.056$ $1.096$ $1.130$ $1.161$ $1.189$ $-1.014$ $1.006$ $-0.998$ $0.992$ $-0.988$
$150$ $1.052$ $1.089$ $1.112$ $1.146$ $1.170$ $-1.007$ $0.994$ $-0.981$ $0.970$ $-0.961$
$200$ $1.049$ $1.084$ $1.113$ $1.138$ $1.160$ $-1.003$ $0.987$ $-0.972$ $0.958$ $-0.947$
$300$ $1.046$ $1.079$ $1.105$ $1.128$ $1.147$ $-0.999$ $0.979$ $-0.960$ $0.944$ $-0.930$
$500$ $1.043$ $1.072$ $1.096$ $1.117$ $1.134$ $-0.994$ $0.970$ $-0.949$ $0.930$ $-0.913$
$\infty$ $0.991$ $0.988$ $0.987$ $0.983$ $0.970$ $-0.941$ $0.888$ $-0.840$ $0.785$ $-0.704$
: Components of the vectors ${\boldsymbol{a}}^{(2)}$ and ${\boldsymbol{b}}^{(2)}$ for various even $L$. []{data-label="rho+-even-eigenvector2"}
Analysis of the eigenvectors of these anomalous eigenvalues is presented in tables \[rho+-even-eigenvector1\] and \[rho+-even-eigenvector2\]. Let us denote by ${\boldsymbol{u}}^{(1)}$ and ${\boldsymbol{u}}^{(2)}$ the eigenvectors corresponding to the eigenvalues $\lambda^{(1)}$ and $\lambda^{(2)}$, respectively. These eigenvectors do not have a definite twist. We define a twist-odd vector ${\boldsymbol{a}}^{(i)}$ ($i=1,2$) with components $a_{2n+1}^{(i)}=u_{2n+1}^{(i)}/f^{(0)}_{2n+1}$, and a twist-even one ${\boldsymbol{b}}^{(i)}$ with $b_{2n}^{(i)}=u_{2n}^{(i)}/u_{2}^{(i)}$. Here, ${\boldsymbol{u}}^{(i)}$ is normalized so that $u_1^{(i)}=1$, and the components of ${{\boldsymbol{f}}^{(0)}}$ are given by [@RSZ:0111281] $$f^{(0)}_n =
\begin{cases}
\displaystyle
\frac{(-1)^{(n-1)/2}}{\sqrt{n}} & n\mbox{: odd} \\
0 & n\mbox{: even}
\end{cases}
\label{eq:fz_n}$$ Components of ${\boldsymbol{a}}^{(i)}$ and ${\boldsymbol{b}}^{(i)}$ for various even $L$ and their values at $L=\infty$ obtained by fitting are given in tables \[rho+-even-eigenvector1\] and \[rho+-even-eigenvector2\] for $i=1$ and $2$, respectively. They strongly support that all (odd) components of ${\boldsymbol{a}}^{(i)}$ are equal to one and that ${\boldsymbol{b}}^{(1)}$ and ${\boldsymbol{b}}^{(2)}$ are equal to each other.[^6] This implies that ${\boldsymbol{u}}^{(1)}$ and ${\boldsymbol{u}}^{(2)}$ are given as linear combinations of two vectors, ${{\boldsymbol{f}}^{(0)}}$ which is twist-odd and ${\boldsymbol{b}}^{(1)}(={\boldsymbol{b}}^{(2)})$ which is twist-even. Therefore, in the limit $L\to\infty$, the projection operator $\rho_+$ has the eigenvalue $1/2$ which is doubly degenerate, and the corresponding eigenvectors are ${{\boldsymbol{f}}^{(0)}}$ and ${\boldsymbol{b}}^{(1)}$. The norm of ${\boldsymbol{b}}^{(1)}$ seems to have worse divergence than that of ${{\boldsymbol{f}}^{(0)}}$ which is logarithmically divergent. We do not know whether ${\boldsymbol{b}}^{(1)}$, which appears only for even $L$, has any relevance to the construction of fluctuation modes.
$L$ $\lambda$
---------- -----------
$49$ $0.525$
$99$ $0.523$
$149$ $0.522$
$199$ $0.521$
$299$ $0.520$
$499$ $0.519$
$\infty$ $0.501$
: Anomalous eigenvalues of $\rho_+$ (\[eq:rhopmorig\]) for various odd $L$ and their extrapolation to $L=\infty$. []{data-label="rho+-odd-eigenvalue"}
$L$ $a_3$ $a_5$ $a_7$ $a_9$ $a_{11}$
---------- --------- --------- --------- --------- ----------
$49$ $1.104$ $1.187$ $1.262$ $1.332$ $1.402$
$99$ $1.084$ $1.147$ $1.201$ $1.249$ $1.293$
$149$ $1.075$ $1.131$ $1.178$ $1.217$ $1.255$
$199$ $1.069$ $1.121$ $1.163$ $1.200$ $1.232$
$299$ $1.063$ $1.109$ $1.146$ $1.179$ $1.208$
$499$ $1.056$ $1.097$ $1.123$ $1.158$ $1.182$
$\infty$ $0.997$ $1.000$ $1.001$ $0.993$ $0.961$
: Components of the vector ${\boldsymbol{a}}$ for various odd $L$ and their extrapolation to $L=\infty$. []{data-label="rho+-odd-eigenvector"}
In the case of odd $L$, we have only one anomalous eigenvalue $\lambda$ largely deviated from either $0$ or $1$. Table \[rho+-odd-eigenvalue\] shows this eigenvalue $\lambda$ for various odd $L$ and its extrapolation to $L=\infty$. As expected, the table \[rho+-odd-eigenvalue\] supports that $\lambda\to 1/2$ as $L\to\infty$. The eigenvector ${\boldsymbol{u}}$ of this eigenvalue $\lambda$ does not have definite a twist for a finite $L$. However, its twist-even component is negligibly small compared with the twist-odd one. In fact, the norm of the twist-even part, $(1+C){\boldsymbol{u}}/2$, is at most $0.6\%$ of the norm of the whole vector ${\boldsymbol{u}}$.[^7] Therefore, we have studied only the twist-odd vector ${\boldsymbol{a}}$ with components $a_{2n+1}=u_{2n+1}/f^{(0)}_{2n+1}$. The results given in table \[rho+-odd-eigenvector\] again support our expectation that ${\boldsymbol{u}}$ is equal to ${{\boldsymbol{f}}^{(0)}}$.
In summary, in both even and odd $L$ cases, our numerical analysis of the eigenvalue problem of $\rho_\pm$ (\[eq:rhopmorig\]) confirms our expectation that $\rho_\pm$ has an anomalous eigenvalue $1/2$ and the corresponding eigenvector is ${{\boldsymbol{f}}^{(0)}}$.
Gauging away the spurious states
================================
In sec. \[section-massive\] we have shown that we can construct massive fluctuation modes if the projection operator $\rho_\pm$ has eigenvalue $1/2$. This property has been verified numerically in the last section. However, there still remains a problem: we have infinite degeneracy of massive and massless states as we saw in the last part of sec.\[section-massive\]. In this section we shall show that these spurious states can in fact be gauged away. Our argument here is an application of that given in [@Imamura:0204031] for spurious massless vector states.
For simplicity we restrict ourselves to the lowest massive state $\Phi^{(k=2)}$: $${| \Phi^{(2)} \rangle}=\beta_{mn}^{\mu\nu}
a^{\mu\dagger}_m a^{\nu\dagger}_n {| {\Phi_\mathrm{t}}\rangle} ,
\label{eq:Phi2}$$ where $\beta_{mn}^{\mu\nu}$ is traceless and transverse with respect to its Lorentz indices. However, the following argument can straightforwardly be extended to ${\Phi^{(k)}}$ with larger $k$. Fourier-expanding $\beta_{mn}$ in terms of the eigenvector ${\boldsymbol{f}}^{(\kappa)}$ of $K_1$, $$\beta_{mn}^{\mu\nu}=\int\!d\kappa\int\!d\lambda\,
f^{(\kappa)}_m f^{(\lambda)}_n {\beta^{\mu\nu}(\kappa,\lambda)},
\label{eq:intrepbeta}$$ and substituting it into the equation for $\beta$, (\[eq:mass-eq\]) with $k=2$, we obtain the following equation for ${\beta^{\mu\nu}(\kappa,\lambda)}$: $$\left(1 - 2^{-p^2}\left[\theta(-\kappa)\theta(-\lambda)
+\theta(\kappa)\theta(\lambda)\right]\right){\beta^{\mu\nu}(\kappa,\lambda)}=0 .
\label{eq:eombetakl}$$ The construction of massive states given in sec.\[section-massive\] (see the paragraph containing (\[eq:b=f...f\]) and (\[eq:b=f...fw\])) can be restated as follows in terms of the spectral function ${\beta^{\mu\nu}(\kappa,\lambda)}$. First, (\[eq:eombetakl\]) implies that ${\beta^{\mu\nu}(\kappa,\lambda)}$ must have support only in the regions $\{\kappa\ge 0,\lambda\ge 0\}$ and $\{\kappa\le 0,\lambda\le 0\}$. Taking into account that $\theta(0)=1/2$ as we have seen in the previous section, there are three possible types of ${\beta^{\mu\nu}(\kappa,\lambda)}$ giving massive as well as massless states (table \[tab:bkl\]). Since the function $g(\kappa)$ for the type-B state and ${\beta^{\mu\nu}(\kappa,\lambda)}$ for the type-C are arbitrary except that they are smooth (and has support only in the regions stated above for ${\beta^{\mu\nu}(\kappa,\lambda)}$ for type-C), these two types of states are infinitely degenerate. We shall show that the type-B and C states can in fact be removed by gauge transformations of VSFT.
${\beta^{\mu\nu}(\kappa,\lambda)}$ $\mbox{(mass)}^2$
--- ------------------------------------ -------------------
A $\delta(\kappa)\delta(\lambda)$ $1$
B $\delta(\kappa)g(\lambda)$ $1$
C arbitrary smooth function 0
: Three types of $(\kappa,\lambda)$-dependence of ${\beta^{\mu\nu}(\kappa,\lambda)}$ satisfying (\[eq:eombetakl\]) and the corresponding mass squared. The function $g(\kappa)$ is an arbitrary smooth function. The whole ${\beta^{\mu\nu}(\kappa,\lambda)}$ must be symmetric under the exchange of $(\mu,\kappa)$ and $(\nu,\lambda)$. []{data-label="tab:bkl"}
The gauge transformation of VSFT, (\[eq:dLPsi\]), expressed in terms of the fluctuation $\Phi$ (\[eq:Psi=Psic+Phi\]) reads $\delta_\Lambda\Phi={{\mathcal{Q}}_{\rm B}}\Lambda+\Phi *\Lambda-\Lambda *\Phi$ with ${{\mathcal{Q}}_{\rm B}}$ defined by (\[eq:wtcQ\]). We shall consider the inhomogeneous part $$\delta_\Lambda^{\mathrm{I}}\Phi={{\mathcal{Q}}_{\rm B}}\Lambda ,
\label{eq:dLI}$$ of the whole transformation, and in particular take the following type of $\Lambda$: $$\Lambda= \Lambda^{\mathrm{m}}\otimes {\mathcal{I}}^{\mathrm{g}},
\label{eq:trf-form}$$ where ${\mathcal{I}}^{\mathrm{g}}$ is the ghost part of the identity string field ${\mathcal{I}}={\mathcal{I}}^{\mathrm{m}}\otimes{\mathcal{I}}^{\mathrm{g}}$ satisfying ${\mathcal{I}}^{\mathrm{g}}*{\Psi_\mathrm{c}^\mathrm{g}}={\Psi_\mathrm{c}^\mathrm{g}}*{\mathcal{I}}^{\mathrm{g}}={\Psi_\mathrm{c}^\mathrm{g}}$ for ${\Psi_\mathrm{c}^\mathrm{g}}$ of (\[eq:EOMg\]) and ${\mathcal{Q}}\,{\mathcal{I}}^{\mathrm{g}}=0$ [@Kishimoto:0110124; @KO:0112169; @Imamura:0204031]. For this $\Lambda$ and fluctuation $\Phi$ of the factorized form (\[eq:Phi=PhimPsicg\]), we have $$\delta_\Lambda^{\mathrm{I}}{\Phi^\mathrm{m}}= {\Psi_\mathrm{c}^\mathrm{m}}*\Lambda^{\mathrm{m}}-\Lambda^{\mathrm{m}}*{\Psi_\mathrm{c}^\mathrm{m}}.
\label{eq:dLIPhim}$$ As the matter part $\Lambda^{\mathrm{m}}$ of the gauge transformation string field, we take $${| \Lambda^{\mathrm{m}} \rangle}
= \gamma_{mn}^{\mu\nu} a^{\mu\dagger}_m a^{\nu\dagger}_n
{| {\Phi_\mathrm{t}}\rangle} ,
\label{eq:L=gaaPhit}$$ with the coefficient $\gamma_{mn}^{\mu\nu}$ being traceless and transverse with respect to $\mu$ and $\nu$. Then, (\[eq:dLIPhim\]) is given by $$\delta_\Lambda^{\mathrm{I}}{| {\Phi^\mathrm{m}}\rangle} = 2^{-p^2}
\Bigl({(\rho_-)}_{mp}{(\rho_-)}_{nq}-{(\rho_+)}_{mp}{(\rho_+)}_{nq}\Bigr)
\gamma_{pq}^{\mu\nu}a^{\mu\dagger}_m a^{\nu\dagger}_n
{| {\Phi_\mathrm{t}}\rangle} ,
\label{eq:dLIPhimgamma}$$ which is expressed as the following transformation on the spectral function ${\beta^{\mu\nu}(\kappa,\lambda)}$ of (\[eq:intrepbeta\]): $$\delta_\Lambda^{\mathrm{I}}{\beta^{\mu\nu}(\kappa,\lambda)}= 2^{-p^2}
\Bigl(\theta(-\kappa)\theta(-\lambda)-\theta(\kappa)\theta(\lambda)
\Bigr){\gamma^{\mu\nu}(\kappa,\lambda)},
\label{eq:dLbeta}$$ where ${\gamma^{\mu\nu}(\kappa,\lambda)}$ is defined for $\gamma_{mn}^{\mu\nu}$ similarly to (\[eq:intrepbeta\]).
Eq. (\[eq:dLbeta\]) implies that the type-B and C states are unphysical ones which can be eliminated by the present gauge transformation. First, the type-B states are gauged away by taking $\delta(\lambda)g(\kappa)\epsilon(\kappa)$ as the $(\kappa,\lambda)$-dependence of ${\gamma^{\mu\nu}(\kappa,\lambda)}$. Here, $\epsilon(\kappa)$ is the signature function $\epsilon(\kappa)=\theta(\kappa)-\theta(-\kappa)$. Second, ${\gamma^{\mu\nu}(\kappa,\lambda)}$ for gauging away the type-C states is ${\gamma^{\mu\nu}(\kappa,\lambda)}=-2^{p^2}\left[\theta(-\kappa)\theta(-\lambda)-
\theta(\kappa)\theta(\lambda)\right]{\beta^{\mu\nu}(\kappa,\lambda)}$. It is obvious that the type-A states cannot be removed by the present gauge transformation.
Finally in this section we shall comment on the relation between the gauge transformation used in this paper and that in [@Imamura:0204031]. Our gauge transformation string field $\Lambda^{\mathrm{m}}$ (\[eq:L=gaaPhit\]) is of different type from that used in [@Imamura:0204031]; the latter is based on the identity string field instead of the classical solution ${\Psi_\mathrm{c}^\mathrm{m}}$. If we have adopted $\Lambda^{\mathrm{m}}$ of the type of [@Imamura:0204031], we would have obtained (\[eq:dLIPhimgamma\]) with $\rho_\pm$ replaced by $(1+T)\rho_\pm$. This gauge transformation cannot remove the type-B states which are absent for the vector case discussed in [@Imamura:0204031].
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank K. Hashimoto, Y. Imamura, E. Itou, H. Kajiura, I. Kishimoto, Y. Matsuo, S. Moriyama, T. Muramatsu, K. Ohmori, T. Takahashi and S. Teraguchi for valuable discussions and comments. The works of H.H. was supported in part by a Grant-in-Aid for Scientific Research from Ministry of Education, Culture, Sports, Science, and Technology (\#12640264).
Derivation of (\[eq:mass-eq\]) {#evaluation}
==============================
In this appendix we outline a derivation of (\[eq:mass-eq\]) from the equation of motion (\[eq:EOMf\]). First we shall mention the hermiticity constraint (\[eq:bb\*\]). We impose the following hermiticity condition on the matter part ${\Phi^\mathrm{m}}$ of a string field of the type (\[eq:Phi=PhimPsicg\]): $${}_2{\langle {\Phi^\mathrm{m}}|}=\prod_{r=1,2}\int\!\frac{d^{26}p_r}{(2\pi)^{26}}
\,(2\pi)^{26}\delta^{26}(p_1+p_2)
\,{}_{12}{\langle R^{\mathrm{m}}| {\Phi^\mathrm{m}}\rangle}_1 ,
\label{eq:ref-beta-k}$$ where the matter reflector (two-string vertex) ${\langle R^{\mathrm{m}} |}$ is defined by $${}_{12}{\langle R^{\mathrm{m}} |}={\langle p_1 |}{\langle p_2 |}
\exp\biggl(-\sum_{n\geq 1}(-1)^n a^{(1)}_na^{(2)}_n\biggr) .
\label{eq:R}$$ This constraint reduces the number of degrees of freedom in ${\Phi^\mathrm{m}}$ to half and ensures the hermiticity of the action. Eq. (\[eq:bb\*\]) for $\beta$ is immediately obtained by plugging the expression (\[eq:massive-state\]) into (\[eq:ref-beta-k\]) and using that ${\Phi_\mathrm{t}}$ itself satisfies (\[eq:ref-beta-k\]).
The wave equation (\[eq:EOMf\]) for the matter fluctuation ${\Phi^\mathrm{m}}$ is rewritten in the oscillator representation as $${| {\Phi^\mathrm{m}}\rangle}_3-{}_1{\langle {\Psi_\mathrm{c}^\mathrm{m}}|}_2{\langle {\Phi^\mathrm{m}}| {V^{\mathrm{m}}}\rangle}_{123}
-{}_1{\langle {\Phi^\mathrm{m}}|}_2{\langle {\Psi_\mathrm{c}^\mathrm{m}}| {V^{\mathrm{m}}}\rangle}_{123}=0 ,
\label{eq:EOMV}$$ where we have omitted the integrations $\left(\prod_{r=1}^3\int d^{26}p_r/(2\pi)^{26}\right)
(2\pi)^{26}\delta^{26}(p_1+p_2+p_3)$ in the second and third terms. Let us consider the second term of (\[eq:EOMV\]) with ${\Phi^{(k)}}$ (\[eq:massive-state\]) substituted for ${\Phi^\mathrm{m}}$, ${}_1{\langle {\Psi_\mathrm{c}^\mathrm{m}}|}_2{\langle {\Phi^{(k)}}| {V^{\mathrm{m}}}\rangle}_{123}$. The basic formula for this calculation is obtained from the well-known formula valid for any bosonic oscillators satisfying $[a_i,a^\dagger_j]=\delta_{ij}$: $$\begin{aligned}
& {\langle 0 |}\exp {\left}( - {\frac{1}{2}}a_i A_{ij} a_{j}-
K_ia_i {\right})
\exp {\left}( -{\frac{1}{2}}a_i^\dagger B_{ij}
a_{j}^\dagger -J_ia_i^\dagger{\right}){| 0 \rangle}{\nonumber}\\
&\qquad=[\det {\left}(1-AB{\right})]^{-1/2}
\exp {\left}(-{\frac{1}{2}}JPAJ-{\frac{1}{2}}KBPK+JPK {\right}) ,
\label{eq:exp}\end{aligned}$$ with $P=(1-AB)^{-1}$. Letting $(-{\partial}/{\partial}K_{i_1})\cdots(-{\partial}/{\partial}K_{i_k})$ act on (\[eq:exp\]), we get $$\begin{aligned}
&{\langle 0 |}a_{i_1}\cdots a_{i_k}
\exp\!\left(-{\frac{1}{2}}a_i A_{ij} a_{j}-K_ia_i\right)
\exp\!\left(-{\frac{1}{2}}a_i^\dagger B_{ij}
a_{j}^\dagger -J_ia_i^\dagger\right){| 0 \rangle}{\nonumber}\\
&=[\det {\left}(1-AB{\right})]^{-{\frac{1}{2}}}\Biggl\{
\prod_{a=1}^k(KBP-JP)_{i_a}
{\nonumber}\\
&\qquad
+\sum_{a>b}(BP)_{i_a i_b}\prod_{c\ne a,b}(KBP-JP)_{i_c}
+ \ldots \Biggr\}
\exp {\left}(-{\frac{1}{2}}JPAJ-{\frac{1}{2}}KBPK+JPK {\right}) ,
\label{eq:k-differentiation}\end{aligned}$$ where we have omitted terms with more than one $(BP)_{i_a i_b}$ factors. In our applications of (\[eq:k-differentiation\]), the index $i$ represents the level number $n$, the Lorentz index $\mu$ and the string index $r=1,2$. Contracting (\[eq:k-differentiation\]) with $(\beta^{\mu_1\cdots\mu_k}_{n_1\cdots n_k})^*$, the terms containing the $(BP)_{i_a i_b}$ factors drop out due to the traceless condition (\[eq:traceless-condition\]), while due to the transverse condition the $(KBP-JP)_{i}$ factor with $i=(n,\mu,r=2)$ contributes only $-a_m^{\mu (3)\dagger}\!\left(\rho_- C\right)_{mn}$ with $\rho_-$ defined by (\[eq:rhopmorig\]). Therefore, we have $$\begin{aligned}
{}_1{\langle {\Psi_\mathrm{c}^\mathrm{m}}|}_2{\langle {\Phi^{(k)}}| {V^{\mathrm{m}}}\rangle}_{123}
&=2^{-p^2}
(-1)^k a^{\mu_1(3)\dagger}_{m_1}\cdots a^{\mu_k(3)\dagger}_{m_k}
(\rho_-C)_{m_1n_1}\cdots(\rho_-C)_{m_kn_k}
(\beta^{\mu_1\cdots\mu_k}_{n_1\cdots n_k})^*{| {\Phi_\mathrm{t}}\rangle}
{\nonumber}\\
&=2^{-p^2}
a^{\mu_1(3)\dagger}_{m_1}\cdots a^{\mu_k(3)\dagger}_{m_k}
{(\rho_-)}_{m_1n_1}\cdots {(\rho_-)}_{m_kn_k}
\beta^{\mu_1\cdots\mu_k}_{n_1\cdots n_k}{| {\Phi_\mathrm{t}}\rangle} ,\end{aligned}$$ where we have used ${\Psi_\mathrm{c}^\mathrm{m}}*{\Phi_\mathrm{t}}=2^{-p^2}{\Phi_\mathrm{t}}$ and the hermiticity (\[eq:bb\*\]). This gives the $\rho_-$ term of (\[eq:mass-eq\]). Derivation of the $\rho_+$ term of (\[eq:mass-eq\]) from the last term of (\[eq:EOMV\]) is quite similar.
[99]{}
L. Rastelli, A. Sen and B. Zwiebach, “String field theory around the tachyon vacuum,” \[arXiv:hep-th/0012251\]. L. Rastelli, A. Sen and B. Zwiebach, “Classical solutions in string field theory around the tachyon vacuum,” \[arXiv:hep-th/0102112\]. L. Rastelli, A. Sen and B. Zwiebach, “Boundary CFT construction of D-branes in vacuum string field theory,” JHEP [**0111**]{}, 045 (2001) \[arXiv:hep-th/0105168\]. L. Rastelli, A. Sen, and B. Zwiebach, “Vacuum string field theory,” \[arXiv:hep-th/0106010\]. H. Hata and T. Kawano, “Open string states around a classical solution in vacuum string field theory,” JHEP [**0111**]{}, 038 (2001) \[arXiv:hep-th/0108150\]. L. Rastelli, A. Sen and B. Zwiebach, “A note on a proposal for the tachyon state in vacuum string field theory,” JHEP [**0202**]{}, 034 (2002) \[arXiv:hep-th/0111153\]. R. Rashkov and K. S. Viswanathan, “A note on the tachyon state in vacuum string field theory,” arXiv:hep-th/0112202. R. Rashkov and K. S. Viswanathan, “A proposal for the vector state in vacuum string field theory,” arXiv:hep-th/0201229. Y. Okawa, “Open string states and D-brane tension from vacuum string field theory,” JHEP [**0207**]{}, 003 (2002) \[arXiv:hep-th/0204012\]. Y. Imamura, “Gauge transformations on a d-brane in vacuum string field theory,” \[arXiv:hep-th/0204031\]. L. Rastelli, A. Sen and B. Zwiebach, “Half strings, projectors, and multiple D-branes in vacuum string field theory,” JHEP [**0111**]{}, 035 (2001) \[arXiv:hep-th/0105058\]. L. Rastelli, A. Sen and B. Zwiebach, “Star algebra spectroscopy,” JHEP [**0203**]{}, 029 (2002) \[arXiv:hep-th/0111281\]. H. Hata and S. Moriyama, “Reexamining classical solution and tachyon mode in vacuum string field theory,” arXiv:hep-th/0206208. V. A. Kostelecky and R. Potting, “Analytical construction of a nonperturbative vacuum for the open bosonic string,” Phys. Rev. D [**63**]{}, 046007 (2001) \[arXiv:hep-th/0008252\]. D. J. Gross and A. Jevicki, “Operator Formulation Of Interacting String Field Theory,” Nucl. Phys. B [**283**]{}, 1 (1987). D. J. Gross and A. Jevicki, “Operator Formulation Of Interacting String Field Theory. 2,” Nucl. Phys. B [**287**]{}, 225 (1987). H. Hata and S. Moriyama, “Observables as twist anomaly in vacuum string field theory,” JHEP [**0201**]{}, 042 (2002) \[arXiv:hep-th/0111034\]. H. Hata, S. Moriyama and S. Teraguchi, “Exact results on twist anomaly,” JHEP [**0202**]{}, 036 (2002) \[arXiv:hep-th/0201177\]. I. Kishimoto, “Some properties of string field algebra,” JHEP [**0112**]{}, 007 (2001) \[arXiv:hep-th/0110124\]. I. Kishimoto and K. Ohmori, “CFT description of identity string field: Toward derivation of the VSFT action,” JHEP [**0205**]{}, 036 (2002) \[arXiv:hep-th/0112169\].
[^1]: [hata@gauge.scphys.kyoto-u.ac.jp]{}
[^2]: [kogetsu@gauge.scphys.kyoto-u.ac.jp]{}
[^3]: In this paper we shall consider only the oscillator formulation of VSFT. For the construction of the fluctuation modes using boundary conformal field theory, see [@Rastelli:2001wk; @Rashkov:2001js; @Rashkov:2002xz; @Okawa:0204012].
[^4]: Due to these two conditions, (\[eq:transverse-condition\]) and (\[eq:traceless-condition\]), our construction of massive modes is restricted only to the highest spin states at a given mass level.
[^5]: See sec. 4 for precise form of these states in the case $k=2$.
[^6]: If we adopt the fitting by polynomials of $1/L$, we obtain better coincidence between $b_{2n}^{(1)}$ and $b_{2n}^{(2)}$ at $L=\infty$ for larger $n$.
[^7]: In the case of even $L$, the norms of the even and odd parts of ${\boldsymbol{u}}^{(i)}$, $(1\pm C){\boldsymbol{u}}^{(i)}/2$, are of the same order.
|
---
abstract: 'We present a lattice analysis of a confining Yang-Mills theory without Goldstone boson. We have analytically investigated the model by a strong coupling expansion and by an intensive lattice Monte Carlo simulation using standard lattice QCD methods. We show that this theory is an interesting candidate for describing weak bosons as composite particles.'
author:
- |
A. Galli\
[*Paul Scherrer Institute, CH-5231 Villigen PSI, Switzerland*]{}
title: 'Composite Weak Bosons: a Lattice Analysis'
---
The Standard Model [@SM] (SM) describes the strong, weak and electromagnetic interactions by a gauge theory based on the group $G_{SM}=SU(3)\times SU(2)\times U(1)$ which is broken by the Higgs mechanism to $SU(3)\times U(1)$. The theory is essentially determined once the matter fields and their transformation under the local gauge transformations of $G_{SM}$ are specified. The matter fields (leptons and quarks) and the Higgs boson are considered to be elementary. They interact with each other by the exchange of gauge bosons which are also considered to be elementary. The structure of the SM has been phenomenologically confirmed to high accuracy.\
In spite of the beautiful corroboration of the SM by experiments a natural questions arises: How elementary are the leptons, the quarks, the Higgs bosons and the gauge bosons? The idea that the SM itself is an effective theory of another, more fundamental, theory, where quarks, leptons and bosons are composites of more fundamental fields, is almost as old as the SM itself. The idea of quark and lepton compositeness is motivated by the observed connection between quarks and leptons, by the generation puzzle and by the existence of too many parameters in the SM. The Higgs compositeness is motivated by the fine tuning problem. The W and Z compositeness is motivated by their relation to a composite Higgs and by the observation that all short-range interactions are residual interactions of a more fundamental long-range interaction.\
The substructures, the new fundamental fields, are supposed to carry a new internal quantum number (which we refer to [*hypercolor*]{}) and the quarks, leptons and bosons are hypercolorless composite systems of them. The binding of the substructures due to hypercolor is viewed as an analogy with the color confinement mechanism of QCD. However, since the SM spectrum is different from the hadron spectrum, the hypercolor interaction has to be described by a strongly coupled Yang-Mills theory [*different*]{} from QCD.\
Several models treat the quarks, leptons and bosons as composite systems. Today a conspicuous number of theorems exist which have ruled out most of the existing models and radically restricted the possibilities of constructing realistic composite models [@1]. One particular model has survived: The Yang-Mills theory without Goldstone bosons [@11]. This model is a usual confining Yang-Mills theory with $SU(2)$ local hypercolor gauge group, $SU(2)$ global isospin group and generalized Majorana fermions in the fundamental representation of the local and global symmetry groups.\
We consider a gauge theory whose fermion content is represented by a Weyl spinor $F_{\alpha,a}^A(x)$. Here $\alpha$ denotes the (undotted) spinor index ($\alpha=1,2$), $A$ denotes the fundamental representation index of a global SU(2) isospin group ($A=1,2$) and $a$ denotes the fundamental representation index of the local SU(2) hypercolor gauge group ($a=1,2$). We introduce the generalized Majorana spinor $\psi$ starting from the Weyl spinors $F$ and its conjugate $F^\dagger$ $$\psi(x)=\left(\begin{array}{c}F(x)\\QF^\dagger(x)\end{array}\right)=
\left(\begin{array}{cc}1&0\\0&Q\end{array}\right){\varphi}(x)$$ and its adjoint $$\bar\psi(x)=(F^T(x)Q,F^\dagger(x))={\varphi}^T(x)
\left(\begin{array}{cc}Q&0\\0&1\end{array}\right).$$ The matrix $Q$ represents the antisymmetric matrix in spin, hypercolor and isospin space. Of course the fields $\psi$ and $\bar{\psi}$ are not independent fields. The choice of the global isospin group $SU(2)$ and of the local hypercolor group $SU(2)$ allows us to write a gauge invariant mass term for the generalized Majorana fermion fields $$\bar{\psi}\psi=FQF+F^\dagger QF^\dagger.$$ Note that this choice is unique if one deals with Majorana fermions. Because of the existence of the mass term we can define the Yang-Mills action on the lattice in Euclidean space in the form of a Wilson action. For further details we refer to ref. [@13].\
To be precise, this model considers the photon to remain elementary and switched off. The weak gauge bosons $W^{\pm}$ and $Z^0$ then form a mass degenerate triplet. The vector isotriplet bound state of the substructure represents the W-boson triplet. To be viable a composite model of the weak bosons has to reproduce the known weak boson spectrum: the lightest bound states have to be the W-bosons and heavier bound states have to lie in an experimentally unexplored energy range. The only possibility of having a Yang-Mills theory which reproduces the weak boson spectrum is to choose the degrees of freedom in a way that they naturally avoid bound states lighter than the vector isotriplet of the theory which characterizes the W-boson triplet. This is possible if the unwanted light bound states which naturally show up as Goldstone bosons or pseudo Goldstone bosons in many models (like, for example, a pseudoscalar isomultiplet, which would be the pion analogue of QCD) are avoided. The choice of the generalized Majorana fermions in this model avoids the $SU_A(2)$ global chiral symmetry of the Yang-Mills Lagrangian because left- and right-handed degrees of freedom are not independent. The axial current (which would generate the $SU_A(2)$ chiral symmetry) does not exist and it is not possible to have a breaking of $SU_A(2)$ with the related low lying Goldstone bosons. In fact, the pseudoscalar isotriplet vanishes by the Pauli principle (it is a symmetric combination of Grassmann variables).\
Because of the confining character of this theory, we need non-perturbative methods to make predictions. It is important that the fermion theory under discussion can be defined by a gauge invariant lattice regularization. A lattice regularization à la Wilson [@12] is possible because the choice of the isospin group $SU(2)$ allows us to replace the Dirac mass term and the Dirac-type Wilson term by a hypercolor gauge invariant Majorana type expression.\
An extensive strong coupling expansion analysis [@13] of the spectrum of this theory has shown that the spin one isotriplet bound state (the right quantum number to represent the W-boson of the SM) could be the lightest state if the pseudoscalar isosinglet acquires a mass by the chiral anomaly in analogy with the $\eta'$ in QCD.\
We also calculated the spectrum of the lightest bound states by a quenched Monte Carlo simulation [@MC] and we showed that the vector isotriplet bound state of this theory is the lightest one. We have performed our Monte Carlo simulations using the following standard technique of lattice QCD:
1. We used the quenched approximation. As in QCD we assume that the quenched approximation is reasonable also in our model.
2. To generate the quenched gauge configurations we used the heat-bath and over relaxed updating (1 heat-bath sweep for 6 over relaxed sweeps).
3. We used the symmetrized Peskin’s formula [@peskin] and the cooling algorithm [@cool] to measure the topological charge of the gauge configurations. Topological non-trivial configurations are needed to evaluate the chiral anomaly contribution to the mass of the pseudoscalar isosinglet bound state.
4. To invert the fermion matrix we used the minimal residual and the conjugate gradient algorithms.
5. We used the smearing technique [@smearing] (PSI-Wuppertal smearing) to obtain early plateaux in the local masses.
To connect the measured quantities to physics we have to fix the lattice spacing $a$. For setting the lattice spacing $a$ we identify the experimental value of the $W-$boson triplet $M_W=80$ GeV with the value of the mass of the vector isotriplet determined from the simulations and extrapolated to the chiral limit. We expect that the mass of the substructure is much smaller than a typical binding energy, therefore in analogy with QCD we extrapolated the bound state masses to the chiral limit, determined by the critical hopping parameter. In our model there is [*no*]{} Goldstone boson which can identify the chiral limit. However, the critical hopping parameter can be evaluated by assuming that the pseudoscalar isosinglet behaves like a Goldstone boson when the contribution of the chiral anomaly is switched off and therefore it is massless.\
We have performed the simulations on different lattices and with different values of $\beta$ to control finite $a$ and volume effects. In some simulations we have calculated the masses of bound states without the contribution of the chiral anomaly. These simulations confirmed the results of the strong coupling expansion. In other simulations we have used and improved the method of ref. [@ito] to evaluate the chiral anomaly contribution to the mass of the pseudoscalar isosinglet bound state. This computation required the evaluation of disconnected fermion loops and the generation of topological non-trivial gauge configurations. The mass of the pseudoscalar isosinglet turned out to be larger than the vector isotriplet mass.\
As a main result our lattice simulations has shown that the vector isotriplet bound state, which represents the weak boson triplet, is the lightest bound state in our model as it should be for any viable candidate of electroweak composite model. In addition we have also predicted the masses of the first bound states which are heavier than the vector isotriplet representing the weak boson triplet. These bound states are an additional vector isotriplet, a vector isosinglet and a pseudoscalar isosinglet with masses in the range of a few hundred GeV. These predictions open new interesting experimental perspectives at LEPII and LHC.
[99]{} S.L. Glashow, Nucl. Phys. B22 (1961) 579; A. Salam and J.C. Ward, Phys. Lett. 13 (1964) 166; S. Weinberg, Phys. Rev. Lett 19 (1967) 1264 J.M. Gipson, Y. Tosa and R.E. Marshak, Phys. Rev. D32 (1985) 284; D. Weingarten, Phys. Rev. Lett. 51 (1983) 1830; S. Nussinov, Phys. Rev. Lett. 51 (1983) 2081; E. Witten, Phys. Rev. Lett. 51 (1983) 2351; G. ’t Hooft et al., recent Developments in Gauge Theories (Plenum Press, New York 1980); S. Weinberg and E. Witten, Phys. Lett. 96B (1980) 59; C. Vafa and E. Witten, Nucl. Phys. B234 (1984) 173; W. Buchmüller, R.D. Peccei and T. Yanagida, Phys. Lett. 124B (1983) 67; M. Claudson, E. Fahri and R.L. Jaffe, Phys. Rev. D34 (1986) 873, and references therein H. Schlereth, Phys. Lett. B267 (1991) 513 K. Wilson, Phys. Rev. D10 (1974) 2445 A. Galli, PSI preprint PSI-PR-94-08, submitted to Phys. Rev. D A.Galli, PSI preprint PSI-PR-94-20, submitted to Nuc. Phys. B M.Peskin, Cornell University preprint CLNS 395 (1978), Thesis; P. Di Vecchia et al., Nucl. Phys. B192 (1981) 392 B.Berg, Phys. Lett. B104 (1981) 475; M.Teper, Nucl. Phys. B (Proc. Suppl.) 20 (1991) 159 S.Güsken, Nucl. Phys. B (Proc. Suppl.) 20 (1991) 465; C.Alexandrou, et al., Nucl. Phys. B414 (1994) 815 S. Itho, Y. Iwasaki and T. Yoshié, Phys. Lett. 145B (1984) 93;Phys. Lett. 184B (1987) 375;Phys. Rev. D36 (1987) 527
|
---
abstract: 'In turbulent superfluid He II, the quantized vortex lines interact via the classical Biot–Savart law to form a complicated vortex tangle. We show that vortex tangles with the same vortex line density will have different energy spectra, depending on the normal fluid which feeds energy into the superfluid component, and identify the spectral signature of two forms of superfluid turbulence: Kolmogorov tangles and Vinen tangles. By decomposing the superfluid velocity field into local and nonlocal contributions, we find that in Vinen tangles the motion of vortex lines depends mainly on the local curvature, whereas in Kolmogorov tangles the long-range vortex interaction is dominant and leads to the formation of clustering of lines, in analogy to the ’worms‘ of ordinary turbulence.'
author:
- 'L. K. Sherwin-Robson'
- 'C. F. Barenghi'
- 'A. W. Baggaley'
title: Local and nonlocal dynamics in superfluid turbulence
---
Introduction {#section:1}
============
The hydrodynamics of helium II is noteworthy for two reasons: its two-fluid nature (an inviscid superfluid and a viscous normal fluid), and the fact that superfluid vorticity is constrained to thin, discrete vortex lines of fixed (quantized) circulation [@Donnelly]; in ordinary (classical) fluids, by contrast, the vorticity is a continuous field. Turbulence in helium II (called superfluid turbulence, or quantum turbulence) consists of a three-dimensional tangle of interacting vortex lines. The properties of this new form of turbulence and current thinking (in terms of theory and experiments) have been recently reviewed [@Barenghi-Skrbek-Sreeni; @Nemirovskii].
Under certain conditions, it has been argued [@Vinen-Niemela; @Skrbek-Sreeni] that the turbulent tangle is characterized by a single length scale, the average distance $\ell$ between the vortex lines, which is inferred from the experimentally observed vortex line density $L$ (length of vortex line per unit volume) as $\ell \approx L^{-1/2}$. Models based on this property describe fairly well the pioneering experiments of Vinen [@Vinen], in which an applied heat flux drives the superfluid and the normal fluid in opposite directions (thermal counterflow). More recently, such ‘Vinen’ tangles were created at very low temperatures by short injections of ions [@Walmsley-Golov], exhibiting the characteristic decay $L \sim t^{-1/2}$ predicted by Vinen [@Baggaley-ultraquantum].
Under different conditions, however, the experimental evidence is consistent with a more structured vortex tangle [@Vinen-Niemela; @Volovik], where the kinetic energy is distributed over a range of length scales according to the same Kolmogorov law which governs ordinary turbulence. ‘Kolmogorov’ tangles have been generated at high temperatures by stirring liquid helium with grids [@Donnelly-grid] or propellers [@Tabeling; @Salort], and at very low temperatures by an intense injection of ions [@Walmsley-Golov], exhibiting the decay $L \sim t^{-3/2}$ expected from the energy spectrum [@Donnelly-grid; @Baggaley-ultraquantum].
The experimental evidence for these two forms of superfluid turbulence is only indirect and arises from macroscopic observables averaged over the experimental cell, such as pressure [@Tabeling; @Salort] and vortex line density [@Donnelly-grid], not from direct visualization of vortex lines. In a recent paper [@Sherwin] we have characterized the energy spectrum of the two forms of turbulence, and showed that ’Kolmogorov’ turbulence contains metastable, coherent vortex structures [@Baggaley-structures; @Baggaley-Laurie], similar perhaps to the ‘worms’ which are observed in ordinary turbulence [@Frisch]. The aim of this work is to go a step further, and look for the dynamical origin of the reported spectral difference and coherent structures.
It is well known [@Saffman] that in an incompressible fluid the velocity field ${{\bf {v}}}$ is determined by the instantaneous distribution of vorticity ${{\mbox{\boldmath $\omega$}}}$ via the Biot-Savart law:
$${{\bf {v}}}({{\bf {x}}})=\frac{1}{4 \pi} \int \frac{{{\mbox{\boldmath $\omega$}}}({{\bf {x}}}') \times ({{\bf {x}}}-{{\bf {x}}}') }
{\vert {{\bf {x}}}-{{\bf {x}}}' \vert^3 } d^3{{\bf {x}}}'
\label{eq:BS1}$$
where the integral extends over the entire flow. The question which we address is whether the velocity at the point ${{\bf {x}}}$ is mainly determined by the (local) vorticity near ${{\bf {x}}}$ or by (nonlocal) contributions from further away. Since the quantization of the circulation implies that the velocity field around a vortex line is strictly $1/r$ (where $r$ is the radial distance from the line), from the predominance of local effects we would infer that the vorticity is randomly distributed and nonlocal effects cancel each other out; conversely, the predominance of nonlocal effects would suggest the existence of coherence structures.
If the vorticity were a continuous field, the distinction between local and nonlocal would involve an arbitrary distance, however in our problem the concentrated nature of vorticity introduces a natural distinction between local and nonlocal contributions, as we shall see.
Method {#section:2}
======
The two most popular models [@Barenghi-Skrbek-Sreeni] for studying superfluid turbulence are the Gross-Pitaevskii equation (GPE) and the Vortex Filament Model (VFM). Each has advantages and disadvantages. The main advantage of the GPE over the VFM is that vortex reconnections are solutions of the equation of motion and do not require an ad–hoc algorithmic procedure. On the other hand, the experimental context of our interest is liquid helium at intermediate temperatures where the effects of the friction are important (at the chosen value $T=1.9~K$, typical of experiments, the normal fluid fraction is 42 percent). We must keep in mind that, although the GPE is a good quantitative model of weakly interacting atomic gases, it is only an idealized model of liquid helium; moreover, the GPE applies only at very low temperatures, and there is not yet a consensus[@Proukakis2008] on its finite-temperature generalizations. This is why, following the approach of Schwarz[@Schwarz], we choose to use the VFM and model superfluid vortex lines as space curves ${{\bf {s}}}(\xi,t)$ (where $t$ is time and $\xi$ is arc length) of infinitesimal thickness and circulation $\kappa=9.97 \times 10^{-4}~\rm cm^2/s$ which move according to
$$\label{eq:Schwarz}
\frac{ds}{dt} = {{\bf {v}}}_s + \alpha {{\bf {s}}}' \times ({{\bf {v}}}_n^{ext} - {{\bf {v}}}_s)
- \alpha' {{\bf {s}}}' \times [{{\bf {s}}}' \times ({{\bf {v}}}_n^{ext} - {{\bf {v}}}_s)].$$
Here $\alpha$ and $\alpha'$ are temperature-dependent friction coefficients [@BDV; @DB], ${{\bf {v}}}_n$ is the normal fluid velocity, and a prime denotes derivative with respect to arc length (hence ${{\bf {s}}}'=d{{\bf {s}}}/d\xi$ is the local unit tangent vector, and $C=\vert {{\bf {s}}}''\vert$ is the local curvature; the radius of curvature, $R=1/C$, is the radius of the osculating circle at the point ${{\bf {s}}}$). The superfluid velocity consists of two parts: ${{\bf {v}}}_s={{\bf {v}}}_s^{ext}+{{\bf {v}}}_s^{self}$. The former ${{\bf {v}}}_s^{ext}$ represents any externally applied superflow; the latter ${{\bf {v}}}_s^{self}$ the self-induced velocity at the point ${{\bf {s}}}$, results from Equation \[eq:BS1\] in the limit of concentrated vorticity:
$${{\bf {v}}}_s^{self}({{\bf {s}}})=
-\frac{\kappa}{4\pi} \oint_{\mathcal{L}}
\frac{({{\bf {s}}}- {{\bf {r}}})}{|{{\bf {s}}}- {{\bf {r}}}|^3} \times {{\bf {dr}}},
\label{eq:BS2}$$
where the line integral extends over the entire vortex configuration $\mathcal{L}$.
We discretised the vortex lines into a large number of points ${{\bf {s}}}_j$ ($j=1, \cdots N$). The minimal separation $\delta$ between the points is such that the vortex curves are sufficiently smooth (at the temperatures of interest here, the friction, controlled by the parameters $\alpha$ and $\alpha'$, damps out high frequency perturbations called Kelvin waves). The VFM assumes that vortex lines are infinitely thin, thus Eq. (\[eq:BS2\]) is valid only if $\vert {{\bf {s}}}- {{\bf {r}}}\vert >> a_0$ where $a_0$ is the vortex core (the region around the vortex axis where the superfluid density drops from its bulk value to zero). Thus the integral diverges if one attempts to find the velocity at ${{\bf {r}}}={{\bf {s}}}$. As remarked by Schwarz, the occurrence of a similar problem in classical hydrodynamics is not helpful, since the physics of the vortex core is different. The solution to the problem which was proposed by Schwarz [@Schwarz1985] and thereafter adopted in the helium literature is based on Taylor expanding the integrand around the singularity and comparing against the well-know expression for the self-induced velocity of a circular ring. In this way he obtained a decomposition of the self-induced velocity at the point ${{\bf {s}}}_j$ (Eq. \[eq:BS2\]) into the following local and nonlocal contributions:
$${{\bf {v}}}_s^{self}({{\bf {s}}}_j)={{\bf {v}}}_s^{loc}({{\bf {s}}}_j)+{{\bf {v}}}_s^{non}({{\bf {s}}}_j)=
\frac{\kappa}{4\pi} \ln \left( \frac{\sqrt{\Delta \xi_{+}\Delta \xi_{-}}}{a_0} \right)
{{\bf {s}}}'_j \times {{\bf {s}}}''_j
+ \frac{\kappa}{4\pi} \oint_{\mathcal{L}'} \frac{({{\bf {s}}}_j - {{\bf {r}}})}{|{{\bf {s}}}_j - {{\bf {r}}}|^3} \times {{\bf {dr}}}\label{eq:BS3}$$
where $\Delta \xi_{+}$ and $\Delta \xi_{-}$ are the arc lengths of the curves between the point ${{\bf {s}}}_j$ and the adjacent points ${{\bf {s}}}_{j-1}$ and ${{\bf {s}}}_{j+1}$ along the vortex line. $\mathcal{L}'$ is the original vortex configuration $\mathcal{L}$ but now without the section between ${{\bf {s}}}_{j-1}$ and ${{\bf {s}}}_{j+1}$. The superfluid vortex core radius $a_0 \approx 10^{-8}~\rm cm$ acts as cutoff parameter. Details and tests of the numerical techniques against the experimental and the numerical literature are published elsewhere [@Baggaley-reconnections; @Baggaley-stats; @Baggaley-PNAS; @Adachi]. Note that the local contribution is proportional to ${{\bf {s}}}' \times {{\bf {s}}}''$, in the binormal direction.
All calculations are performed in a cubic periodic domain of size $D=0.1$ using an Adams-Bashforth time-stepping method (with typical time step $\Delta t =5 \times 10^{-5}~\rm s$), a tree-method [@Baggaley-tree] with opening angle $\theta=0.4$, and typical minimal resolution $\delta=1.6 \times 10^{-3}~\rm cm$. For example, an increase in the numerical resolution from $\delta =0.0016$ to $\delta =0.0008$ produces a small increase of 2.5% in the importance of the nonlocal contribution in Fig. (\[fig:8\]). Moreover, the tests against experiments mentioned above [@Baggaley-stats; @Baggaley-PNAS; @Adachi], guarantee that the numerical resolution is sufficient, and put the distinction between ${{\bf {v}}}_s^{loc}$ and ${{\bf {v}}}_s^{non}$ on solid ground.
It is known from experiments [@Bewley2008] and from more microscopic models [@Koplik; @Zuccher; @Allen2014] that colliding vortex lines reconnect with each other. An algorithmic procedure is introduced to reconnect two vortex lines if they become sufficiently close to each other. This procedure (although arbitrary, unlike the GPE as mentioned before) has been extensively tested [@Baggaley-reconnections]; various slightly different reconnection algorithms have been proposed and have never been found in disagreement with experimental observations.
We choose a temperature typical of experiments, $T=1.9~\rm K$ (corresponding to $\alpha=0.206$ and $\alpha'=0.00834$). In all three cases, the initial condition consists of a few seeding vortex loops, which interact and reconnect, quickly generating a turbulent vortex tangle which appears independent of the initial condition.
We study the following three different regimes of superfluid turbulence, characterized by the following forms of the normal fluid’s velocity field ${{\bf {v}}}_n^{ext}$:
1. [**Uniform normal flow.**]{} Firstly, to model turbulence generated by a small heat flux at the blocked end of a channel (thermal counterflow), we impose a uniform normal fluid velocity ${{\bf {v}}}_n^{ext}=V_n {\hat {\bf x}}$ in the x-direction (which we interpret as the direction of the channel) which is proportional to the applied heat flux; to conserve mass, we add a uniform superflow ${{\bf {v}}}_s^{ext}= -(\rho_n/\rho_s) V_n {\hat {\bf x}}$ in the opposite direction, where $\rho_n$ and $\rho_s$ are respectively the normal fluid and superfluid densities. Eqs. (\[eq:Schwarz\]) and (\[eq:BS3\]) are solved in the imposed superflow’s reference frame. This model is the most used in the literature, from the pioneering work of Schwarz [@Schwarz] to the recent calculations of Tsubota and collaborators [@Adachi].
2. [**Synthetic turbulence.**]{} To model turbulence generated by pushing helium through pipes or channels [@Salort] using plungers or bellows or by stirring it with grids [@Donnelly-grid] or propellers [@Tabeling], we start from the observation that, due to liquid helium’s small viscosity $\mu$, the normal fluid’s Reynolds number ${\rm Re}=V D/\nu_n$ is usually large (where $V$ is the rms velocity and $\nu_n=\mu/\rho_n$ the kinematic viscosity), hence we expect the normal fluid to be turbulent. We assume ${{\bf {v}}}_s^{ext}={\bf 0}$ and [@Osborne-2006]
$${{\bf {v}}}_n^{ext}({{\bf {s}}},t)=\sum_{m=1}^{M}({\bf A}_m \times {\bf k}_m \cos{\phi_m}
+{\bf B}_m \times {\bf k}_m \sin{\phi_m}),
\label{eq:KS}$$
where $\phi_m={{\bf {k}}}_m \cdot {{\bf {s}}}+ f_m t$, ${{\bf {k}}}_m$ are wavevectors and $f_m=\sqrt{k^3_m E(k_m)}$ are angular frequencies. The random parameters ${\bf A}_m$, ${\bf B}_m$ and ${\bf k}_m$ are chosen so that the normal fluid’s energy spectrum obeys Kolmogorov’s scaling $E(k_m)\propto k_m^{-5/3}$ in the inertial range $k_1<k< k_M$, where $k_1 \approx 2 \pi/D$ and $k_M$ correspond to the outer scale of the turbulence and the dissipation length scale respectively. Then we define the Reynolds number via ${\rm Re}=(k_M/k_1)^{4/3}$. The synthetic turbulent flow defined by Eq. (\[eq:KS\]) is solenoidal, time-dependent, and compares well with Lagrangian statistics obtained in experiments and direct numerical simulations of the Navier-Stokes equation. It is therefore physically realistic and numerically convenient to model current experiments on grid or propeller generated superfluid turbulence.
3. [**Frozen Navier-Stokes turbulence.**]{} Synthetic turbulence, being essentially the superposition of random waves, lacks the intense regions of concentrated vorticity which are typical of classical turbulence [@Frisch]. For this reason we consider a third model: a turbulent flow ${{\bf {v}}}_n^{ext}$ obtained by direct numerical simulation (DNS) of the classical Navier-Stokes equation in a periodic box with no mean flow. Since the simultaneous calculation of superfluid vortices and turbulent normal fluid would be prohibitively expensive, we limit ourselves to a time snapshot of ${{\bf {v}}}_n^{ext}$. In other words, we determine the vortex lines under a ’frozen‘ turbulent normal fluid. Our source for the DNS data is the John Hopkins Turbulence Database [@JHTDB1; @JHTDB2], which consists of a velocity field on a $1024^3$ spatial mesh. The estimated Reynolds number is ${\rm Re} \approx (L_0/\eta_0)^{4/3}=3205$ where $L_0$ is the integral scale and $\eta_0$ the Kolmogorov scale. Although ${\rm Re}=3205$ is not a very large Reynolds number, at $T=1.9~K$ helium’s viscosity is $\mu=1.347 \times 10^{-5}~\rm g/(cm~ s)$, the normal fluid density is $\rho_n=0.0611~\rm g/cm^3$, the kinematic viscosity is $\nu_n=\mu/\rho_n=0.22 \times 10^{-3}~\rm cm/s$ [@DB], and therefore ${\rm Re}=UD/\nu_n=3205$ corresponds to the reasonable speed $U=0.7~\rm cm/s$ in a typical $D=1~\rm cm$ channel. To keep the resulting vortex line density of this model to a computationally practical value, we rescale the velocity components such that they are 60 percent of their original values, thus obtaining the vortex line density $L \approx 20,0000~\rm cm^{-2}$. To obtain Fig. \[fig:11\] the scaling factor is only 45 percent, yielding $L \approx 6,000~\rm cm^{-2}$; because of the nonlinearity of the Navier-Stokes equation, this procedure is clearly an approximation but is sufficient for our aim of driving a less intense of more intense vortex tangle.
Fig. \[fig:1\] shows the magnitude of the normal velocity field plotted (at a fixed time $t$) on the $xy$ plane at $z=0$, corresponding to models 2 and 3 (we do not plot the normal fluid velocity for model 1 because it is uniform). It is apparent that models 1, 2 and 3 represent a progression of increasing complexity of the driving normal flow. In Fig. \[fig:1\], note in particular the localized regions of strong velocity which appear in model 3.
Results
=======
The intensity of the turbulence is measured by the vortex line density $L=\Lambda/V$ (where $\Lambda$ is the superfluid vortex length in the volume $V=D^3$), which we monitor for a sufficiently long time, such that the properties which we report, refer to a statistically steady state of turbulence fluctuating about a certain average $L$ (we choose parameters so that $L$ is typical of experiments).
Fig. \[fig:2\] shows the initial transient of the vortex line density followed by the saturation to statistically steady-states of turbulence corresponding to model 1 (uniform normal flow), model 2 (synthetic normal flow turbulence) and model 3 (frozen Navier-Stokes turbulence). In all cases, the intensity of the drive is chosen to generate approximately the same vortex line density, $L \approx 20,000~\rm cm^{-2}$.
Snapshots of the vortex tangles are shown in Fig. \[fig:3\]. The vortex tangle driven by the uniform normal fluid (model 1, left) appears visually as the most homogeneous; the vortex tangle driven by the frozen Navier-Stokes turbulence (model 3, right) appears as the least homogeneous. What is not apparent in the figure is the mild anisotropy of the tangle driven by the uniform normal fluid. To quantify this anisotropy, we calculate the projected vortex lengths in the three Cartesian directions ($\Lambda_x$, $\Lambda_y$, $\Lambda_z$) and find $\Lambda_x/\Lambda=0.34 < \Lambda_y/\Lambda=\Lambda_z/\Lambda=0.54$ for model 1, confirming a small flattening of the vortices in the yz plane (this effect was discovered by the early investigations of Schwarz [@Schwarz]). In comparison, models 2 and 3 are more isotropic: for model 2 (tangle driven by synthetic turbulence) we find $\Lambda_x/\Lambda=0.50$, $\Lambda_y/\Lambda=0.47$ and $\Lambda_z/\Lambda=0.49$), and for model 3 (tangle driven by frozen Navier-Stokes turbulence) we obtain $\Lambda_x/\Lambda=0.48$, $\Lambda_y/\Lambda=0.50$, $\Lambda_z/\Lambda=0.49$.
Fig. \[fig:4\] and \[fig:5\] shows the average curvature $C=< C_j>$ (sampled over the discretization points $j=1, \cdots N$) and the distributions of local curvatures $C_j=\vert {{\bf {s}}}_j'' \vert$. The tangle generated by model 3 (frozen Navier-Stokes turbulence) has the smallest average curvature: the presence of long lines (large radius of curvature $R=1/\vert {{\bf {s}}}'' \vert$) is indeed visible in Fig. \[fig:2\]. In terms of curvature, the tangles generated by models 1 and 2 are more similar to each other - the average curvature is almost twice as large as for model 3, indicating that the vortex lines are more in the form of small loops.
However, vortex line density (Fig. \[fig:2\]), visual inspection (Fig. \[fig:3\]) and curvature (Fig. \[fig:4\] and \[fig:5\]) do not carry information about the [*orientation*]{} of the vortex lines, a crucial ingredient of the dynamics. Fig. \[fig:6\] shows the energy spectrum $E_s(k)$, defined by $$\frac{1}{V} \int_V \frac{1}{2}{\bf v}_s^2 dV=\int_0^{\infty} E_s(k) dk
\label{eq:spectrum}$$
where $k$ is the magnitude of the three-dimensional wavevector. The energy spectrum describes the distribution of kinetic energy over the length scales. The spectra of the tangles generated by synthetic normal flow turbulence (model 2) and by the frozen Navier-Stokes turbulence (model 3) are consistent with the classical Kolmogorov scaling $E_s(k) \sim k^{-5/3}$ for $k < k_{\ell}=2 \pi/\ell$; the kinetic energy is clearly concentrated at the largest length scales (small $k$). In contrast, the spectrum of the tangle generated by the uniform normal fluid (model 1) peaks at the intermediate length scales, and at large wavenumbers is consistent with the shallower $k^{-1}$ dependence of individual vortex lines.
A natural question to ask is whether our results are affected by the particular vortex reconnection algorithm used. In principle, both the large $k$ region and the small $k$ region of the spectrum could be affected: the former, because vortex reconnections involve changes of the geometry of the vortices at small length scales, the latter because the energy flux may be affected. To rule out this possibility we have performed simulations using the reconnection algorithm of Kondaurova et al. [@Kondaurova], which tests whether vortex filaments would cross each others path during the next time step (for details, see also ref. [@Baggaley-reconnections]). Fig. (\[fig:7\]) is very similar to Fig. (\[fig:6\]), confirming that the shape of the energy spectra does not depend on the reconnection algorithm.
It is also instructive to examine the spatial distribution of the superfluid energy densities arising from the three normal fluid models: Fig. \[fig:8\] displays the superfluid energy density $\epsilon_s=\vert {{\bf {v}}}_s \vert^2/2$ on the $xy$ plane averaged over $z$. The left panel (model 1, uniform normal flow) shows that $\epsilon_s$ is approximately constant, that is to say the vortex tangle is homogeneous; the middle and right panels (model 2 and 3 for synthetic normal flow turbulence and frozen Navier-Stokes turbulence) show that the energy density is increasingly nonhomogeneous, particularly model 3. The localized regions of large energy density correspond to vortex lines which are locally parallel to each other, reinforcing each other’s velocity field rather than cancelling it out.
The natural question which we ask is what is the cause of the spectral difference shown in Fig. \[fig:6\]. To answer the question we examine the local and nonlocal contributions to the superfluid velocity, defined according to Equation \[eq:BS3\]. Fig. \[fig:9\] shows the fraction $v^{non}/v^{self}$ of the superfluid velocity which arises from nonlocal contributions, where $v^{non}=<\vert {{\bf {v}}}_{non}({{\bf {s}}}_j)\vert>$ and $v^{self}=<\vert {{\bf {v}}}_{self}({{\bf {s}}}_j)\vert>$ are sampled over the discretization points $j=1, \cdots N$ at a given time $t$. The difference is striking. Nonlocal effects are responsible for only 25 percent of the total superfluid velocity field in the tangle generated by the uniform normal fluid (model 1), for 45 percent in the tangle generated by synthetic normal flow turbulence (model 2), and for more than 60 percent in the tangle generated by frozen Navier-Stokes turbulence (model 3).
Finally, we explore the dependence of the result on the vortex line density $L$ by generating statistically steady states of turbulence driven by uniform normal fluid (model 1) and synthetic turbulence (model 2) with different values of $L$, see Fig. \[fig:10\]. Fig. \[fig:11\] shows that for model 1 (uniform normal flow) the relative importance of nonlocal contributions remains constant at about 25 percent over a wide range of vortex line density, from $L \approx 6000$ to $L \approx 20,000~ {\rm cm^{-2}}$, whereas for model 2 (synthetic normal flow turbulence) and 3 (frozen Navier–Stokes turbulence), it increases with $L$.
Conclusion
==========
Turbulent vortex tangles can be produced in the laboratory using various means: by imposing a flux of heat, by pushing liquid helium II through pipes, or by stirring it with moving objects. The numerical experiments presented here show that reporting the vortex line density $L$ is not enough to characterize the nature of the superfluid turbulence which can be generated in helium II. Vortex tangles with the same value of $L$ may have very different energy spectra, depending on the normal fluid flow which feeds energy into the vortex lines, as shown in Fig. \[fig:6\]. If the normal fluid is turbulent, energy is contained in the large eddies and is distributed over the length scales consistent with the classical $k^{-5/3}$ Kolmogorov law at large $k$, suggesting the presence of a Richardson cascade. If the normal fluid is uniform, most of the energy is contained at the intermediate length scales, and the energy spectrum scales consistently with $k^{-1}$ at large $k$. Using a terminology already in the literature, we identify these two forms of superfluid turbulence as ‘Kolmogorov tangles’ and ‘Vinen tangles’ respectively.
The superfluid velocity field is determined by the instantaneous configuration of vortex lines. Since the superfluid velocity field decays only as $1/r$ away from the axis of a quantum vortex line, the interaction between vortex lines is long-ranged, at least in principle. By examining the ratio of local and nonlocal contributions to the total velocity field, we have determined that in Vinen tangles far-field effects tend to cancel out ($v_s^{non}/v_s^{self} \approx 25 \%$ independently of $L$), the motion of a vortex line is mainly determined by its local curvature, and the vortex tangle is homogeneous. In Kolmogorov tangles, on the contrary, nonlocal effects are dominant and increase with the vortex line density; this stronger vortex-vortex interaction leads to the clustering of vortex lines, for which the vortex tangle is much less homogeneous and contains coherent vorticity regions, in analogy to what happens in ordinary turbulence. The presence of intermittency effects such as coherent structures in the driving normal fluid, which we have explored with model 3, enhances the formation of superfluid vortex bundles, resulting in a more inhomogeneous superfluid energy and in larger nonlocal contributions to the vortex lines’ dynamics.
Future work will explore the problem for turbulence with nonzero mean flow.
We thank Kalin Kanov for help with DNS data.
[10]{}
R.J. Donnelly (1991) [*Quantized Vortices In Helium II*]{} (Cambridge University Press, Cambridge, UK).
C.F. Barenghi, L. Skrbek and K.R. Sreenivasan, Proc. Nat. Acad. Sci. USA [**111**]{}, 4647 (2014).
S. Nemirovskii, Phys. Rep. [**524**]{} 85 (2013)
W. F. Vinen and J. J. Niemela, J. Low Temp. Phys. [**128**]{}, 167 (2002).
L. Skrbek and K.R. Sreenivasan, Phys. Fluids [**24**]{}, 0113 (2012).
W.F. Vinen, Proc. R. Soc. A [**240**]{} 114 (1957).
P.M. Walmsley and A.I. Golov, Phys. Rev. Lett. [**100**]{}, 245301 (2008).
A.W. Baggaley, C.F. Barenghi, and Y.A. Sergeev, Phys. Rev. B [**85**]{}, 060501 (2012).
G.E. Volovik J. Low Temp. Phys. [**136**]{}, 309 (2004).
M.R. Smith, R.J. Donnelly, N. Goldenfeld, and W.F. Vinen, Phys. Rev. Lett. [**71**]{} 2583 (1993).
J. Maurer, and P. Tabeling, Europhys. Lett. [**43**]{}, 29 (1998).
J. Salort J, et al., Phys. Fluids [**22**]{}, 125102 (2010).
A.W. Baggaley, L.K. Sherwin, C.F. Barenghi, and Y.A. Sergeev, Phys. Rev. B [**86**]{}, 104501 (2012).
A.W. Baggaley, C.F. Barenghi, A. Shukurov, and Y.A. Sergeev, Europhys. Lett. [**98**]{}, 26002 (2012).
A.W. Baggaley, J. Laurie, and C.F. Barenghi, Phys. Rev. Lett. [**109**]{}, 205304 (2012).
U. Frisch, [*Turbulence. The legacy of A.N. Kolmogorov*]{}, Cambridge University Press, Cambridge (1995).
P.G. Saffman, [*Vortex Dynamics*]{} (Cambridge University Press, Cambridge (1992).
N.P. Proukakis and B. Jackson, J. Phys. B At. Mol. Opt. [**40**]{}, 203002 (2008)
K.W. Schwarz, Phys. Rev. B [**38**]{}, 2398 (1988).
K.W. Schwarz, Phys. Rev. B [**31**]{}, 5782 (1985).
C. F. Barenghi, R. J. Donnelly, W. F. Vinen, J. Low Temp. Physics [**52**]{}, 189 (1983).
R. J. Donnelly, C. F. Barenghi, J. Phys. Chem. Reference Data [**27**]{}, 1217 (1998).
J. Koplik and H. Levine, Phys. Rev. Lett. [**71**]{}, 1375 (1993).
G.P. Bewley, M.S. Paoletti, K.R. Sreenivasan, and D.P. Lathrop, Proc. Nat. Acad. Sci. USA [**105**]{}, 13707 (2008).
S. Zuccher, M. Caliari, and C.F. Barenghi (2012), Phys. Fluids (24): 125108.
A.J. Allen, S. Zuccher, M. Caliari, N.P. Proukakis, N.G. Parker, and C.F. Barenghi, Phys. Rev. A [**90**]{}, 013601 (2014).
A.W. Baggaley, J. Low Temp. Phys. [**168**]{}, 18 (2012).
A.W. Baggaley, and C.F. Barenghi, Phys. Rev. E [**84**]{}, 067301 (2011).
A.W. Baggaley and R. Hänninen, Proc. Nat. Acad. Sci. USA [**111**]{}, 4667 (2014)
H. Adachi, S. Fujiyama, and M. Tsubota, Phys. Rev. B [**81**]{}, 104511 (2010).
A.W. Baggaley and C.F. Barenghi, J. Low Temp. Phys. [**66**]{}, 3 (2012).
D.R. Osborne, J.C. Vassilicos, K. Sung, and J.D. Haigh (2006), Phys. Rev. E [**74**]{}, 036309 (2006).
The data set is available at the website [*turbulence.pha.jhu.edu*]{}
Y. Li, E. Perlman, M. Wan, Y. Yang, R. Burns, C. Meneveau, R. Burns, S. Chen, A. Szalay & G. Eyink. J. Turbulence [**9**]{}, No. 31 (2008).
L.P. Kondaurova, V.A. Andryuschenko, and S.K. Nemirovskii, J. Low Temp. Phys. [**150**]{}, 415 (2008).
![(color online). Magnitude of the driving normal fluid velocity field, $\vert {{\bf {v}}}_n^{ext} \vert$, plotted on the $xy$-plane at $z=0$ corresponding to model 2 (synthetic normal flow turbulence, left) and model 3 (frozen Navier-Stokes turbulence, right). The velocity scales (cm/s) are shown at right of each panel. Note the more localized, more intense regions of velocity which are present in model 3. []{data-label="fig:1"}](figure1a.eps "fig:"){width="45.00000%"} ![(color online). Magnitude of the driving normal fluid velocity field, $\vert {{\bf {v}}}_n^{ext} \vert$, plotted on the $xy$-plane at $z=0$ corresponding to model 2 (synthetic normal flow turbulence, left) and model 3 (frozen Navier-Stokes turbulence, right). The velocity scales (cm/s) are shown at right of each panel. Note the more localized, more intense regions of velocity which are present in model 3. []{data-label="fig:1"}](figure1b.eps "fig:"){width="45.00000%"}
![ (color online). Evolution of the vortex line density $L$ (cm$^{-2}$) vs time $t$ (s) for model 1 (red line, uniform normal flow), model 2 (black line, synthetic normal fluid turbulence) and model 3 (dashed blue line, frozen Navier-Stokes turbulence). The inset displays the oscillations of $L$ vs $t$ in more detail. Parameters: temperature $T=1.9~\rm K$, $V_n=1~\rm cm/s$ (for model 1), ${\rm Re}= 79.44$ (for model 2), and ${\rm Re}= 3025$ (for model 3). []{data-label="fig:2"}](figure2.eps "fig:"){width="55.00000%"}\
![ Snapshot of the vortex tangle for model 1 (uniform normal fluid, left), model 2 (synthetic turbulence, middle) and model 3 (frozen Navier-Stokes turbulence, right) at time $t=20~\rm s$ (parameters as in Fig. (\[fig:2\]). []{data-label="fig:3"}](figure3a.eps "fig:"){width="30.00000%"} ![ Snapshot of the vortex tangle for model 1 (uniform normal fluid, left), model 2 (synthetic turbulence, middle) and model 3 (frozen Navier-Stokes turbulence, right) at time $t=20~\rm s$ (parameters as in Fig. (\[fig:2\]). []{data-label="fig:3"}](figure3b.eps "fig:"){width="30.00000%"} ![ Snapshot of the vortex tangle for model 1 (uniform normal fluid, left), model 2 (synthetic turbulence, middle) and model 3 (frozen Navier-Stokes turbulence, right) at time $t=20~\rm s$ (parameters as in Fig. (\[fig:2\]). []{data-label="fig:3"}](figure3c.eps "fig:"){width="30.00000%"}
![ (color online). Average curvature $C$ ($\rm cm^{-1}$) vs time $t$ ($\rm s$) for model 1 (uniform normal flow, red line), model 2 (synthetic turbulence, black line) and model 3 (frozen Navier-Stokes turbulence, dashed blue line). Parameters as in Fig. (\[fig:2\]). []{data-label="fig:4"}](figure4.eps "fig:"){width="55.00000%"}\
![ (color online). Left: Probability density function of the curvature, ${\rm PDF}(C)$, vs curvature, $C$ ($\rm cm^{-1}$), corresponding to model 1 (uniform normal flow, red line), model 2 (synthetic turbulence, black line) and model 3 (frozen Navier-Stokes turbulence, dashed blue line). Right: the same data plotted on a log log scale, where the matching slopes on the plot illustrate that we have the same Kelvin waves in all three models (because they are all at the same temperature). Parameters as in Fig. \[fig:2\].[]{data-label="fig:5"}](figure5a.eps "fig:"){width="45.00000%"} ![ (color online). Left: Probability density function of the curvature, ${\rm PDF}(C)$, vs curvature, $C$ ($\rm cm^{-1}$), corresponding to model 1 (uniform normal flow, red line), model 2 (synthetic turbulence, black line) and model 3 (frozen Navier-Stokes turbulence, dashed blue line). Right: the same data plotted on a log log scale, where the matching slopes on the plot illustrate that we have the same Kelvin waves in all three models (because they are all at the same temperature). Parameters as in Fig. \[fig:2\].[]{data-label="fig:5"}](figure5b.eps "fig:"){width="43.00000%"}
![ (color online). Energy spectrum $E(k)$ (arbitrary units) vs wavenumber $k$ ($\rm cm^{-1}$) (time averaged over the saturated regime) corresponding to vortex tangles generated by uniform normal fluid (model 1, top), synthetic normal fluid turbulence (model 2, middle) and frozen Navier-Stokes turbulence (model 3, bottom). The dashed lines indicate the $k^{-1}$ (top) and the $k^{-5/3}$ dependence (middle and bottom), respectively. Parameters as in Fig. (\[fig:2\]). The compensated spectra $k E_s(k)$ and $k^{5/3} E_s(k)$ in the insets show the regions of $k$–space where the approximate scalings $k^{-1}$ and $k^{-5/3}$ apply. []{data-label="fig:6"}](figure6a.eps "fig:"){width="42.00000%"}\
![ (color online). Energy spectrum $E(k)$ (arbitrary units) vs wavenumber $k$ ($\rm cm^{-1}$) (time averaged over the saturated regime) corresponding to vortex tangles generated by uniform normal fluid (model 1, top), synthetic normal fluid turbulence (model 2, middle) and frozen Navier-Stokes turbulence (model 3, bottom). The dashed lines indicate the $k^{-1}$ (top) and the $k^{-5/3}$ dependence (middle and bottom), respectively. Parameters as in Fig. (\[fig:2\]). The compensated spectra $k E_s(k)$ and $k^{5/3} E_s(k)$ in the insets show the regions of $k$–space where the approximate scalings $k^{-1}$ and $k^{-5/3}$ apply. []{data-label="fig:6"}](figure6b.eps "fig:"){width="42.00000%"}\
![ (color online). Energy spectrum $E(k)$ (arbitrary units) vs wavenumber $k$ ($\rm cm^{-1}$) (time averaged over the saturated regime) corresponding to vortex tangles generated by uniform normal fluid (model 1, top), synthetic normal fluid turbulence (model 2, middle) and frozen Navier-Stokes turbulence (model 3, bottom). The dashed lines indicate the $k^{-1}$ (top) and the $k^{-5/3}$ dependence (middle and bottom), respectively. Parameters as in Fig. (\[fig:2\]). The compensated spectra $k E_s(k)$ and $k^{5/3} E_s(k)$ in the insets show the regions of $k$–space where the approximate scalings $k^{-1}$ and $k^{-5/3}$ apply. []{data-label="fig:6"}](figure6c.eps "fig:"){width="42.00000%"}
![ (color online). Energy spectrum $E(k)$ (arbitrary units) vs wavenumber $k$ ($\rm cm^{-1}$) (time averaged over the saturated regime) as in Fig. (\[fig:6\]), but the simulations are performed using the reconnection algorithm of Kondaurova et al. [@Kondaurova]. Note that there is no significant difference from spectra obtained using our standard algorithm, see Fig. (\[fig:6\]). Vortex tangles generated by uniform normal fluid (model 1, top), synthetic normal fluid turbulence (model 2, middle) and frozen Navier-Stokes turbulence (model 3, bottom). The dashed lines indicate the $k^{-1}$ (top) and the $k^{-5/3}$ dependence (middle and bottom), respectively. Parameters: temperature $T=1.9~\rm K$, $V_n=0.75~\rm cm/s$ (for model 1), ${\rm Re}= 81.59$ (for model 2), and ${\rm Re}= 3025$ (for model 3).[]{data-label="fig:7"}](figure7a.eps "fig:"){width="42.00000%"}\
![ (color online). Energy spectrum $E(k)$ (arbitrary units) vs wavenumber $k$ ($\rm cm^{-1}$) (time averaged over the saturated regime) as in Fig. (\[fig:6\]), but the simulations are performed using the reconnection algorithm of Kondaurova et al. [@Kondaurova]. Note that there is no significant difference from spectra obtained using our standard algorithm, see Fig. (\[fig:6\]). Vortex tangles generated by uniform normal fluid (model 1, top), synthetic normal fluid turbulence (model 2, middle) and frozen Navier-Stokes turbulence (model 3, bottom). The dashed lines indicate the $k^{-1}$ (top) and the $k^{-5/3}$ dependence (middle and bottom), respectively. Parameters: temperature $T=1.9~\rm K$, $V_n=0.75~\rm cm/s$ (for model 1), ${\rm Re}= 81.59$ (for model 2), and ${\rm Re}= 3025$ (for model 3).[]{data-label="fig:7"}](figure7b.eps "fig:"){width="42.00000%"}\
![ (color online). Energy spectrum $E(k)$ (arbitrary units) vs wavenumber $k$ ($\rm cm^{-1}$) (time averaged over the saturated regime) as in Fig. (\[fig:6\]), but the simulations are performed using the reconnection algorithm of Kondaurova et al. [@Kondaurova]. Note that there is no significant difference from spectra obtained using our standard algorithm, see Fig. (\[fig:6\]). Vortex tangles generated by uniform normal fluid (model 1, top), synthetic normal fluid turbulence (model 2, middle) and frozen Navier-Stokes turbulence (model 3, bottom). The dashed lines indicate the $k^{-1}$ (top) and the $k^{-5/3}$ dependence (middle and bottom), respectively. Parameters: temperature $T=1.9~\rm K$, $V_n=0.75~\rm cm/s$ (for model 1), ${\rm Re}= 81.59$ (for model 2), and ${\rm Re}= 3025$ (for model 3).[]{data-label="fig:7"}](figure7c.eps "fig:"){width="42.00000%"}
![ Superfluid energy density $\epsilon_s=\vert {{\bf {v}}}_s \vert^2/2$ smoothed over the average inter-vorex spacing $\ell$, plotted on the $xy$ plane and averaged over $z$. Left: model 1 (uniform normal fluid); middle: model 2 (synthetic turbulence); right: model 3 (frozen Navier-Stokes turbulence). Parameters as in Fig. (\[fig:2\]). []{data-label="fig:8"}](figure8.eps){width="100.00000%"}
![(color online). Ratio of nonlocal to total self-induced velocity as a function of time $t$ $\rm (s) $ for tangles generated by uniform normal fluid (model 1, red line, bottom), synthetic normal fluid turbulence (model 2, black line, middle) and frozen Navier-Stokes turbulence (model 3, blue line, top). Parameters as in Fig. (\[fig:2\]). []{data-label="fig:9"}](figure9.eps "fig:"){width="65.00000%"}\
![ (color online). Evolution of the vortex line density $L$ (cm$^{-2}$) vs time $t$ (s) for model 1 (red line, uniform normal flow, respectively at $V_n=1~\rm (cm/s)$ (top), $V_n=0.75~\rm (cm/s)$ (middle) and $V_n=0.55~\rm (cm/s)$ (bottom)), model 2 (black line, synthetic normal fluid turbulence, respectively at $Re=79.44$ (top), $Re=81.59$ (middle) and $Re=83.86$ (bottom)), and model 3 (dashed blue line, frozen Navier-Stokes turbulence, at $Re=3025$). []{data-label="fig:10"}](figure10.eps){width="60.00000%"}
![(color online). Ratio $v^{non}/v^{self}$ as a function of vortex line density $L$ $(\rm cm^{-2})$ corresponding to model 1 (uniform normal flow, red circles), model 2 (synthetic normal flow turbulence, black crosses) and model 3 (frozen Navier-Stokes equation, blue stars). []{data-label="fig:11"}](figure11.eps){width="65.00000%"}
|
---
abstract: 'In this paper we study the distribution of the non-trivial zeros of the zeta-function $\zeta(s)$ (and other L-functions) under Montgomery’s pair correlation approach. We use semidefinite programming to improve the asymptotic bounds for $N^*(T)$, $N_d(T)$ and $N({\lambda},T)$.'
address:
- 'IMPA - Instituto Nacional de Matemática Pura e Aplicada - Estrada Dona Castorina, 110, Rio de Janeiro, RJ, Brazil 22460-320'
- 'Hausdorff Center for Mathematics, Universität Bonn, Endenicher Allee 60, 53115 Bonn, Germany'
- 'Massachusetts Institute of Technology, Department of Mathematics, 77 Massachusetts Avenue, Simons building (Building 2), Room 2-241'
author:
- 'Andrés Chirre, Felipe Gonçalves and David de Laat'
title: 'Pair Correlation Estimates for the Zeros of the Zeta-Function via Semidefinite Programming'
---
Introduction
============
In this paper we give improved asymptotic bounds for several quantities related to the zeros of the zeta-function (and other functions) under Montgomery’s pair correlation approach [@M]. The key idea is to replace the usual bandlimited auxiliary functions by the class of functions used in the linear programming bounds developed by Cohn and Elkies [@CE] for the sphere packing problem. The advantage of this framework is that it reduces the problems to convex optimization problems that can be solved numerically via semidefinite programming. For all problems we considered this produces better bounds than any bandlimited construction.
Background
----------
Let $\zeta(s)$ be Riemann’s zeta-function. It is well-known that all non-trivial zeros of $\zeta(s)$ are located in the critical strip $0<{{\rm Re}\,}s<1$, and the Riemann Hypothesis (RH) is the statement that all these zeros are aligned in the line ${{\rm Re}\,}s=1/2$. Let $N(T)$ count the number of zeros $\rho = \beta + i \gamma$ of $\zeta(s)$, repeated according the multiplicity, such that $0<\beta<1$ and $0 <\gamma \leq T$. The Riemann-von Mangoldt formula (in its weaker form) states that $$\begin{aligned}
\label{RvMformula}
N(T) = (1+o(1)) \frac{T}{2\pi} \log T.\end{aligned}$$ Let $$N^{*}(T):=\displaystyle\sum_{0<\gamma\leq T}m_{\rho},$$ where the sum is over the non-trivial zeros of $\zeta(s)$ counting multiplicities[^1] and $m_\rho$ is the multiplicity of $\rho$. In addition to RH, it is also conjectured that all zeros of $\zeta(s)$ are simple, and therefore it is [*conjectured*]{} that $$\begin{aligned}
\label{NstarNconjecure}
N^{*}(T) \sim N(T).\end{aligned}$$ To study the distribution of the zeros of zeta, Montgomery defined the pair correlation function $$\label{paircorrelfunction}
N(x,T):=\sum_{\substack{0< {\gamma},{\gamma}'\leq T \\ 0<{\gamma}'-{\gamma}\leq \frac{2\pi x}{\log T}}} 1$$ and [*conjectured*]{} that $$N(x,T) \sim N(T)\int_0^x \bigg(1-\frac{\sin^2(\pi y)}{(\pi y)^2}\bigg){\,\text{\rm d}}y.$$ Note that by the average gap between zeros is $\frac{2\pi}{\log T}$, hence $N(x,T)$ is counting zeros not greater than $x$ times the average gap.
One line of research to understand and give evidence for the conjectures above is to produce bounds of the form $$\label{prob1}
N^*(T) \leq (1+c) N(T),$$ and $$\label{prob2}
N(x,T) \gg N(T),$$ with $c,x>0$ as small as possible, as $T\to\infty$. These two problems have been widely studied with several improvements being made over the years. One of the approaches is to use some suitable explicit formula (relating sums with integrals) with an auxiliary function $f$ in some class ${\mathcal{A}}$ and produce an inequality relating the quantity we are interested to bound with some functional ${\mathcal{Q}}(f)$ over ${\mathcal{A}}$. Minimizing (or maximizing) the functional over the class ${\mathcal{A}}$ would then produce the best bound one can possibly get with that specific approach. Nowadays, this idea is a standard technique in analytic number theory (introduced first by Beurling and Selberg) and the following are some references (clearly not a complete list) where the main approach is exactly that: Large sieve inequalities [@GV; @HV]; Erdös-Turán inequalities [@CV2; @Va]; Hilbert-type inequalities [@CL3; @CLV; @CV2; @GMK; @GV; @Va]; Tauberian theorems [@GV]; Bounds in the theory of the Riemann zeta-function and $L$-functions [@CC; @CChi; @CChiM; @CCLM; @CCM; @CCM2; @CF; @CS; @Chi; @Ga; @GG]; Prime gaps [@CMS].
From our point of view, our main contribution connects here. So far the only class ${\mathcal{A}}$ used for problems and was some Paley-Wiener space of bandlimited functions. We relax the bandlimited condition by requiring only certain sign conditions on the auxiliary function that match exactly with the very same conditions required by the linear programming bounds for the packing problem (see Section \[prelims\] for a detailed explanation). This relation is what ultimately inspired and allowed us to perform numerical computations to find good test functions for the functionals we derive in Section \[prelims\]. Furthermore, as far as we know, it is the first time this method is used in the zeta-function theory.
Main Results
============
\[thmNstar\] Assuming RH we have $$\begin{aligned}
N^*(T)\leq ({1.3208}+o(1))N(T).
\end{aligned}$$ Assuming GRH we have $$\begin{aligned}
N^*(T)\leq ({1.3155}+o(1))N(T).
\end{aligned}$$
Montgomery [@M] was the first to show the constant $1.3333...$. This result was later improved to $1.3275$ by Cheer and Goldston [@CG]. Assuming the generalized Riemann Hypothesis (GRH), Goldston, Gonek, Özlük and Snyder [@GGOS] improved it to $1.3262$. To the best of our knowledge, our bounds are the current best.
Theorem \[thmNstar\] has an important application to estimating the quantity of simple zeros of $\zeta(s)$. Let $$N_s(T):=\displaystyle\sum_{\substack{0<\gamma\leq T\\ m_\rho=1}}1.$$ Using the fact that $${N_s(T)\geq \displaystyle\sum_{0<{\gamma}\leq T}(2-m_\rho)}=2N(T)-N^{*}(T).$$ we obtain the following corollary.
Assuming RH we have $${N_s(T)}\geq ({0.6792} + o(1)){N(T)}.$$ Assuming GRH we have $${N_s(T)}\geq ({0.6845} + o(1)){N(T)}. $$
Under the pair correlation approach the previous best result known is due by Cheer and Goldston [@CG] showing that $67.27\%$ of the zeros are simple. Assuming GRH, Goldston, Gonek, Özlük and Snyder [@GGOS] showed that $67.38\%$ are simple. In this way, we improved all these bounds. However, by a different technique, still assuming RH, Bui and Heath-Brown [@BHB] improved the result to $70.37\%$, which currently is the best.
Combining the above result of Bui and Heath-Brown with Theorem \[thmNstar\] and an argument of Ghosh, we can bound the proportion of distinct zeros of zeta. Let $$N_d(T) := \sum_{0<\gamma\leq T} \frac{1}{m_\rho},$$ be the number of distinct zeros of $\zeta(s)$ with $0<{\gamma}\leq T$. Using the inequality $$2N_s(T)\leq \displaystyle\sum_{0<\gamma\leq T}\dfrac{(m_\rho-2)(m_\rho-3)}{m_\rho}=N^{*}(T)-5N(T)+6N_d(T).$$ in conjunction with the estimate $$N_s(T) \geq (0.7037+o(1))N(T)$$ and Theorem \[thmNstar\], we deduce the following corollary.
Assuming RH we have $${N_d(T)}\geq ({0.8477} + o(1)){N(T)}.$$ Assuming GRH we have $${N_d(T)}\geq ({0.8486} + o(1)){N(T)}. $$
Using the pair correlation approach, the best previous result known is due to Farmer, Gonek and Lee [@FGL] with constant $0.8051$. By a different technique, assuming RH, Bui and Heath-Brown [@BHB] improved the constant to $0.8466$. To the best of our knowledge, our new bounds are the current best.
We also obtain improved results for Montgomery’s pair correlation function.
\[thmpaircorreltation\] Assuming RH and we have $$N({0.6039},T)\gg N(T).$$ Assuming GRH and we have $$N(0.5769,T)\gg N(T).$$
Montgomery [@M] showed that $N(0.68..., T)\gg N(T)$, and in [@GGOS] it is pointed out that it is not difficult to modify Montgomery’s argument to derive the sharper constant $0.6695$. This result was improved by Goldston, Gonek, Özlük and Snyder [@GGOS] with constant $0.6072$. Later, Carneiro, Chandee, Littmann and Milinovich [@CCLM] improved the constant to $0.6068...$. Assuming GRH and , Goldston, Gonek, Özlük and Snyder showed the constant $0.5781...$. To the best of our knowledge, our new bounds are the current best.
Results for zeros of Dirichlet L-functions
------------------------------------------
To obtain averaged bounds for the percentage of simple zeros of primitive Dirichlet L-functions we use the framework established by Chandee, Lee, Liu and Radziwiłł [@CLLR]. Let $\Phi$ be a real-valued smooth function supported in the interval $[a,b]$ with $0<a<b<\infty$. Define its Mellin transform by $$\mathcal{M}\Phi(s)=\int_{0}^{\infty}\Phi(x)x^{s-1}{\,\text{\rm d}}x.$$ For a character $\chi$ mod $q$, let $L(s,\chi)$ be its associated Dirichlet L-function. Under GRH, all non-trivial zeros of $L(s,\chi)$ lie on the critical line ${{\rm Re}\,}s = 1/2$. Let $$N_{\Phi}(Q):=\displaystyle\sum_{Q\leq q\leq 2Q}\dfrac{W(q/Q)}{\varphi(q)}\displaystyle\sum_{\substack{\chi \ (\mathrm{mod} \ q) \\ \text{primitive}}} \displaystyle\;\sum_{\gamma_\chi}\big|\mathcal{M}\Phi(i\gamma_\chi)\big|^2,$$ where $W$ is a non-negative smooth function supported in $(1,2)$, and where the last sum is over all non-trivial zeros ${\tfrac12}+i\gamma_\chi$ of the Dirichlet $L$-function $L(s,\chi)$. In [@CLLR Lemma 2.1] it is shown that $$N_{\Phi}(Q)\sim\dfrac{A}{2\pi}Q\log Q \int_{-\infty}^{\infty}\big|\mathcal{M}\Phi(ix)\big|^2{\,\text{\rm d}}x,$$ where $$A={\mathcal{M}}W(1)\displaystyle\prod_{p\ \text{prime}}\bigg(1-\dfrac{1}{p^2}-\dfrac{1}{p^3}\bigg).$$ Let $$N_{\Phi,s}(Q)=\displaystyle\sum_{Q\leq q\leq 2Q}\dfrac{W(q/Q)}{\varphi(q)}\displaystyle\sum_{\substack{\chi \ (\mathrm{mod} \ q) \\ \text{primitive}}} \displaystyle\;\sum_{\substack{\gamma_\chi \\ \text{simple}}}\big|\mathcal{M}\Phi(i\gamma_\chi)\big|^2.$$ The quantity $$\liminf_{Q\to\infty}\frac{N_{\Phi,s}(Q)}{N_{\Phi}(Q)}$$ then measures (in average) the proportion of simple zeros among all primitive Dirichlet L-functions.
In addition, for the following theorem, we require that $\Phi(x)$ and ${\mathcal{M}}\Phi(ix)$ are non-negative functions. We note that we can also further relax the conditions on $\Phi$ so to include the function given by ${\mathcal{M}}\Phi(ix) = (\sin x/x)^2$, as was established in [@CLLR] and [@S].
\[thmsimplezerosofdir\] Assuming GRH we have $$N_{\Phi,s}(Q)\geq (0.9350+o(1))N_{\Phi}(Q).$$
Using the pair correlation approach, the best previous result known is due to Sono [@S], showing that $93.22\%$ of the zeros are simple. To the best of our knowledge, our bound is the current best.
Results for zeros of $\xi'(s)$
------------------------------
We can extend our analysis to the zeros of $\xi'(s)$, where $$\begin{aligned}
\xi(s)=\frac{1}{2}s(s-1)\pi^{-\frac{s}{2}}\Gamma\bigg(\frac{s}{2}\bigg)\zeta(s).\end{aligned}$$ It is known that $\xi'(s)$ has only zeros in the critical strip $0<{{\rm Re}\,}s<1$ and that RH implies that all its zeros satisfy ${{\rm Re}\,}s=1/2$. Let $N_1(T)$ count the number of zeros $\rho_1=\beta_1+i\gamma_1$ of $\xi'(s)$ (with multiplicity) such that $0< \gamma_1 \leq T$. It is also known that $$\begin{aligned}
N_1(T) = (1+o(1)) \frac{T}{2\pi} \log T.\end{aligned}$$ We can then similarly define the function $$N_{1}^{*}(T):=\displaystyle\sum_{0<\gamma_1\leq T}m_{\rho_1},$$ where $m_{\rho_1}$ is the multiplicity of the zero $\rho_1$.
\[thmn1star\] Assuming RH we have $$\begin{aligned}
N_1^*(T)\leq (1.1175+o(1))N_1(T).
\end{aligned}$$
To the best of our knowledge, this bound is the current best.
Defining the functions $N_{1,s}(T)$ and $N_{1,d}(T)$ (quantity of simple and distinct zeros respectively) for $\xi'(s)$ and using the inequalities $${N_{1,s}(T) \geq 2 N_1(T) - N_1^*(T)}$$ and $${N_{1,d}(T) \geq \frac{3}{2}N_1(T)-\frac{1}{2}N_1^*(T)},$$ that can be derived the same way as for $\zeta(s)$, we obtain the following corollary.
Assuming RH we have $$N_{1,s}(T)\geq (0.8825+o(1))N_1(T). $$ and $$N_{1,d}(T)\geq (0.9412+o(1))N_1(T). $$
The best previous result is due to Farmer, Gonek and Lee [@FGL], showing that more than $85.83\% $ of the zeros of $\xi'(s)$ are simple. To the best of our knowledge, our new bounds are the current best.
Derivation of the optimization problems {#prelims}
=======================================
Let ${\mathcal{A}}_{LP}$ be the class of even continuous functions $f\in L^1({\mathbb{R}})$ satisfying the following conditions:
1. ${\widehat}f(0)=f(0)=1$;
2. ${\widehat}f\geq 0$;
3. $f$ is eventually non-positive.
By eventually non-positive we mean that $f(x)\leq 0$ for all sufficiently large $|x|$. We then define the last sign change of $f$ by $$r(f)=\inf \big\{r>0: f(x)\leq 0 \text{ for } |x|\geq r\big\}.$$ It is easy to show that if $f\in A_{LP}$, then ${\widehat}f \in L^1({\mathbb{R}})$.
A remarkable breakthrough in the sphere problem was achieved by Cohn and Elkies in [@CE], where they showed that if $\Delta({\mathbb{R}}^d)$ is the highest sphere packing density in ${\mathbb{R}}^d$ then $$\Delta({\mathbb{R}}^d) \leq {\mathcal{Q}}(f)$$ for any $f\in {\mathcal{A}}_{LP}({\mathbb{R}}^d)$ (this is the analogous class in higher dimensions defined for radial functions $f$), where $${\mathcal{Q}}(f) = \frac{\pi^{d/2}}{(d/2)!2^{d}}r(f)^d.$$ With this approach they generated numerical upper bounds, called linear programming bounds, for the packing density for dimensions up to $36$ (nowadays it goes much higher) that improved every single upper bound known at the time and still are the current best. These upper bounds in dimensions $8$ and $24$ revealed to be extremely close to the lower bounds given by the $E_8$ root lattice and the $\Lambda_{24}$ Leech lattice, revealing that in these special dimensions the linear programming approach could exactly act as the dual problem. This is what inspired Viazovska [@V; @CKMRV] to follow their program and solve the sphere packing problem in dimensions $8$ and $24$. What is interesting and surprising to us is that the same space ${\mathcal{A}}_{LP}$ can be used (but with a functional different than ${\mathcal{Q}}(f)$) to produce numerical bounds in analytic number theory.
The general strategy to study problems and is based on Montgomery’s function $$\begin{aligned}
F(x,T)=\frac{1}{N(T)}\sum_{0<\gamma, \gamma'\leq T}T^{ix(\gamma-\gamma')}w(\gamma-\gamma'),\end{aligned}$$ where the sum is over pairs of ordinates of zeros (with multiplicity) of $\zeta(s)$ and $w(u)=\frac{4}{4+u^2}$. The first step is to use Fourier inversion to obtain $$\label{explicit_formula}
\displaystyle\sum_{0<\gamma, \gamma'\leq T}g\bigg((\gamma-\gamma')\dfrac{\log T}{2\pi}\bigg)w(\gamma-\gamma')=N(T)\int_{-\infty}^{\infty}\widehat{g}(x)F(x,T){\,\text{\rm d}}x,$$ for suitable functions $g$, and use some known asymptotic estimate for $F(x,T)$ as $T\to\infty$ (which is proven only under RH or GRH). Secondly, after a series of inequalities, we produce a minimization problem over ${\mathcal{A}}_{LP}$ for some functional ${\mathcal{Z}}$. We then approach the problem numerically, using the class of functions used for the sphere packing problem in [@CE] and sum-of-squares/semidefinite programming techniques to optimize over these functions. The same basic strategy can be, in principle, carried out for other functions where we have a pair correlation approach. Indeed, we will also derive functionals related to the zeros of $\xi'(s)$ and a certain average of primitive Dirichlet L-functions.
Bounding $N^*(T)$ and $N(x,T)$
------------------------------
Ultimately, the functionals we need to define depend on the asymptotic behavior of $F(x,T)$. To analyze the function $N^{*}(T)$ we define the functionals $${\mathcal{Z}}(f) =r(f) + \frac{2}{r(f)}\int_0^{r(f)}f(x)x{\,\text{\rm d}}x$$ and $${\widetilde}{\mathcal{Z}}(f)=r(f) + \frac{2}{r(f)}\int_0^{r(f)}f(x)x{\,\text{\rm d}}x + 3\int_{r(f)}^{\frac{3}{2}r(f)}f(x){\,\text{\rm d}}x - \frac{2}{r(f)}\int_{r(f)}^{\frac{3}{2}r(f)}f(x)x{\,\text{\rm d}}x.$$
\[thmsimplezerosofxi\] Let $f\in {\mathcal{A}}_{LP}$. Assuming RH we have $$N^*(T)\leq ({\mathcal{Z}}(f)+o(1))N(T).$$ Assuming GRH, for every fixed small $\delta>0$ we have $$N^*(T)\leq ({\widetilde}{\mathcal{Z}}(f) + O(\delta) + o(1))N(T).$$
We start assuming only RH. Refining the original work of Montgomery [@M], Goldston and Montgomery [@GM Lemma 8] stated that $$\label{Fasymptotics}
F(x,T) = \big(T^{-2|x|}\log T+|x|\big)(1+o(1)),$$ uniformly for $|x|\leq 1$. Let $f\in {\mathcal{A}}_{LP}$ and let $g(x)={\widehat}f(x/r(f))/r(f)$. We can then use the explicit formula in conjunction with the asymptotic formula above to obtain $$\begin{aligned}
\sum_{0<\gamma, \gamma'\leq T} g\bigg((\gamma-\gamma')\dfrac{\log T}{2\pi}\bigg)w(\gamma-\gamma') = N(T)\left[\widehat{g}(0) +\int_{-1}^{1}{{\widehat}g}(x)|x|{\,\text{\rm d}}x + \int_{|x|>1}{{\widehat}g}(x)F(x,T){\,\text{\rm d}}x + o(1)
\right],\end{aligned}$$ where the $o(1)$ above is justified since ${\widehat}g$ is continuous and $T^{-2|x|}\log T\to{{\boldsymbol}\delta}_0(x)$ as $T\to\infty$ (in the distributional sense). Moreover, since $F(x,T)$ is non-negative and ${\widehat}g(x)\leq 0$ for $|x|\geq 1$ we deduce that $$\begin{aligned}
\sum_{0<\gamma, \gamma'\leq T}g\bigg((\gamma-\gamma')\dfrac{\log T}{2\pi}\bigg)w(\gamma-\gamma') & \leq N(T)\left[ {\widehat}g(0)+2\int_{0}^{1}{{\widehat}g}(x)x{\,\text{\rm d}}x + o(1)\right] = N(T)\left[\frac{{\mathcal{Z}}(f)}{r(f)} + o(1)\right].\end{aligned}$$ On the other hand, clearly we have $$\begin{aligned}
\sum_{0<\gamma, \gamma'\leq T} g\bigg((\gamma-\gamma')\dfrac{\log T}{2\pi}\bigg)w(\gamma-\gamma') \geq g(0)\sum_{0<\gamma\leq T} m_\rho = \frac{N^*(T)}{r(f)}.\end{aligned}$$ Combining these results we show the first inequality in the theorem.
Assume now GRH. It is then shown in [@GGOS] that for any fixed and sufficiently small $\delta >0$ we have $$\label{Fasymptoticmore}
F(x,T) \geq \frac{3}{2}-|x| - o(1),$$ uniformly for $1\leq |x|\leq \frac{3}{2}-\delta$ as $T\to \infty$. Using this estimate and the fact that ${\widehat}g(x)\leq 0$ for $|x|\geq 1$ we obtain $$\begin{aligned}
\sum_{ 0<\gamma, \gamma'\leq T}{\widehat}g\bigg((\gamma-\gamma')\dfrac{\log T}{2\pi}\bigg)w(\gamma-\gamma')& \leq N(T)\left[{{\widehat}g}(0) + 2\int_{1}^{\frac{3}{2}-\delta}{{\widehat}g}(x)\left(\frac{3}{2}-x\right){\,\text{\rm d}}x + o(1)
\right] \\
& = N(T)\left[\frac{{\widetilde}{\mathcal{Z}}(f)}{r(f)}+o(1) +O(\delta)\right].\end{aligned}$$ Arguing as before we finish the proof.
To analyze $N(x,T)$ we define the [function]{} $${\mathcal{P}}(f) = \inf \big\{{\lambda}>0: p_f({\lambda}) > 0 \big\},$$ where $$p_f({\lambda})=-1+\frac{{\lambda}}{r(f)}+ \frac{2r(f)}{{\lambda}}\int_0^{\frac{{\lambda}}{r(f)}}{\widehat}f(x)x{\,\text{\rm d}}x,$$ and the [function]{} $${\widetilde}{\mathcal{P}}(f) = \inf\big\{{\lambda}>0: {\widetilde}p_f({\lambda}) > 0 \big\},$$ where $${\widetilde}p_f({\lambda})=-1+\frac{{\lambda}}{r(f)}+ \frac{2r(f)}{{\lambda}}\int_0^{\frac{{\lambda}}{r(f)}}{\widehat}f(x)x{\,\text{\rm d}}x +3\int_{\frac{{\lambda}}{r(f)}}^{\frac{3{\lambda}}{2r(f)}}{\widehat}f(x){\,\text{\rm d}}x - \frac{2r(f)}{{\lambda}}\int_{\frac{{\lambda}}{r(f)}}^{\frac{3{\lambda}}{2r(f)}}{\widehat}f(x)x{\,\text{\rm d}}x.$$ Note that these functions are well defined since $p_f$ and ${\widetilde}p_f$ are $C^1$ functions that assume $-1$ at ${\lambda}=0$, and using the fact that $\widehat{f}\in L^1(\mathbb{R})$ one can show $$\lim_{{\lambda}\to\infty} \frac{p_f({\lambda})}{{\lambda}} = \lim_{{\lambda}\to\infty} \frac{{\widetilde}p_f({\lambda})}{{\lambda}} =\frac{1}{r(f)} > 0.$$
\[thmsmallgapszerosofxi\] Let $f\in {\mathcal{A}}_{LP}$ and ${\varepsilon}>0$. Assuming RH and we have $$N({\mathcal{P}}(f)+\varepsilon,T)\gg N(T).$$ Assuming GRH we have $$N({\widetilde}{\mathcal{P}}(f)+\varepsilon,T)\gg N(T).$$
In the following we only exhibit the proof assuming RH since under GRH the proof is very similar, and the only extra information needed is in . Let $f\in{\mathcal{A}}_{LP}$ and $\lambda>0$. Applying the explicit formula for $g(x)= f(r(f)x/{\lambda})$ in conjunction with we obtain $$\begin{aligned}
\sum_{0<\gamma, \gamma'\leq T}g\bigg((\gamma-\gamma')\dfrac{\log T}{2\pi}\bigg)w(\gamma-\gamma') = N(T)\int_{-\infty}^{\infty}\widehat{g}(x)F(x,T){\,\text{\rm d}}x
& \geq N(T)\left[{\widehat}g(0)+2\int_{0}^{1}{\widehat}g(x)x{\,\text{\rm d}}x+o(1)\right]
\\ & = N(T)\left[1+p_f({\lambda})+o(1)\right].\end{aligned}$$ Since ${\widehat}f\geq 0$, we have $\|f\|_{\infty}=f(0)=1$. Recall now the pair correlation function $N(x,T)$ defined in . We have $$\begin{aligned}
\sum_{0<\gamma, \gamma'\leq T}g\bigg((\gamma-\gamma')\dfrac{\log T}{2\pi}\bigg)w(\gamma-\gamma') & = N^*(T) +2\sum_{\substack{0<\gamma, \gamma'\leq T\\ 0<\gamma-\gamma'}}f\bigg((\gamma-\gamma')\dfrac{r(f)\log T}{2\pi{\lambda}}\bigg)w(\gamma-\gamma') \\
& \leq N^*(T) +2\sum_{\substack{0<\gamma, \gamma'\leq T\\ 0<\gamma-\gamma' \leq \frac{2\pi {\lambda}}{\log T}}}f\bigg((\gamma-\gamma')\dfrac{r(f)\log T}{2\pi{\lambda}}\bigg)w(\gamma-\gamma') \\
& \leq N^*(T)+2N({\lambda},T) \\ & = (1+o(1))N(T) + 2N({\lambda},T),\end{aligned}$$ where in the last step we have used . Then, we obtain $$\begin{aligned}
\frac{N({\lambda},T)}{N(T)}\geq \frac{p_f({\lambda})}{2}+o(1).\end{aligned}$$ Noting that $N({\lambda},T)$ increases with ${\lambda}$, we can then choose ${\lambda}$ arbitrarily close to ${\mathcal{P}}(f)$ and obtain the desired result.
Bounding $N_{\Phi,s}(Q)$
------------------------
Define the following functional over ${\mathcal{A}}_{LP}$: $${\mathcal{L}}(f) = \frac{r(f)}{2} + \dfrac{4}{r(f)}\int_{0}^{\frac{r(f)}{2}}x f(x){\,\text{\rm d}}x+2\int_{\frac{r(f)}{2}}^{r(f)}f( x){\,\text{\rm d}}x.$$ We have the following theorem.
\[thmnphi\] Let $f\in{\mathcal{A}}_{LP}$. Assuming GRH, for every fixed small $\delta>0$ we have $$N_{\Phi,s}(Q) \geq (2- {\mathcal{L}}(f) + O(\delta) + o(1))N_{\Phi}(Q).$$
For $Q>1$ and $x\in{\mathbb{R}}$, we define the pair correlation function $F_\Phi$ by $$\begin{aligned}
F_\Phi(Q^{x},W)=\dfrac{1}{N_{\Phi}(Q)}\displaystyle\sum_{Q\leq q\leq 2Q}\dfrac{W(q/Q)}{\varphi(q)}\displaystyle\sum_{\substack{\chi \ (\mathrm{mod} \ q)\\ \mathrm{primitive}}}\bigg|\displaystyle\sum_{\gamma_\chi}\mathcal{M}\Phi(i\gamma_\chi)Q^{i\gamma_\chi{x}}\bigg|^2.\end{aligned}$$ Using the asymptotic large sieve, Chandee, Lee, Liu and Radziwiłł[@CLLR] showed the following asymptotic formula under GRH $$\begin{aligned}
\label{FPhiasymp}
& F_{\Phi}(Q^{{x}},W) \\ & = (1+o(1))\bigg[1-(1-|x|)_++\Phi\big(Q^{-|{x}|}\big)^2\log Q \bigg(\dfrac{1}{2\pi}\int_{-\infty}^{\infty}\big|\mathcal{M}\Phi(it)\big|^2{\,\text{\rm d}}t\bigg)^{-1}\bigg] + O\Big(\Phi(Q^{-|{x}|})\log^{1/2} Q\Big)\nonumber,\end{aligned}$$ which holds uniformly for $|{x}|\leq 2-\delta$ as $Q\to\infty$, for any fixed and sufficiently small $\delta>0$. Let $$N^*_{\Phi}(Q):=\displaystyle\sum_{Q\leq q\leq 2Q}\dfrac{W(q/Q)}{\varphi(q)}\displaystyle\sum_{\substack{\chi \ (\mathrm{mod} \ q) \\ \text{primitive}}}\displaystyle\sum_{\gamma_\chi}m_{\rho_\chi}\big|\mathcal{M}\Phi(i\gamma_\chi)\big|^2,$$ where $m_{\rho_\chi}$ denote the multiplicity of the nontrivial zero $\rho_\chi={\tfrac12}+i\gamma_\chi$ of $L(s,\chi)$. Since $$\displaystyle\sum_{\substack{\gamma_\chi\\\text{simple}}}\big|\mathcal{M}\Phi(i\gamma_\chi)\big|^2\geq \displaystyle\sum_{\gamma_\chi}(2-m_{\rho_\chi})\big|\mathcal{M}\Phi(i\gamma_\chi)\big|^2$$ we obtain $$\begin{aligned}
\label{ineqNstarwithN}
N_{\Phi,s}(Q)\geq 2N_{\Phi}(Q)-N^{*}_{\Phi}(Q).\end{aligned}$$ For any $g\in L^1({\mathbb{R}})$ with ${\widehat}g\in L^1({\mathbb{R}})$ we have the following explicit formula (Fourier inversion) $$\begin{aligned}
\sum_{Q\leq q\leq 2Q} \dfrac{W(q/Q)}{\varphi(q)}\displaystyle\sum_{\substack{\chi \ (\mathrm{mod} \ q) \\ \text{primitive}}}\displaystyle\;\sum_{\gamma_\chi, \gamma'_\chi}\mathcal{M}\Phi(i\gamma_\chi)\mathcal{M}\Phi(i\gamma'_\chi)\,\widehat{g}\bigg(\dfrac{(\gamma_\chi-\gamma'_\chi)\log Q}{2\pi}\bigg)
=N_{\Phi}(Q)\int_{-\infty}^{\infty}g(x)F_{\Phi}(Q^x,W){\,\text{\rm d}}x .\end{aligned}$$ Letting $f\in\mathcal{A}_{LP}$ and $g(x)=f(r(f)x/(2-\delta))$, for any primitive character $\chi \ (\mathrm{mod} \ q)$ we obtain $$\begin{aligned}
&\sum_{\gamma_\chi, \gamma'_\chi}\mathcal{M}\Phi(i\gamma_\chi)\mathcal{M}\Phi(i\gamma'_\chi)\,\widehat{g}\bigg(\dfrac{(\gamma_\chi-\gamma'_\chi)\log Q}{2\pi}\bigg) \\ & = \displaystyle\sum_{\gamma_\chi}m_{\rho_\chi}\big|\mathcal{M}\Phi(i\gamma_\chi)\big|^2\,\widehat{g}(0) \nonumber + \displaystyle\sum_{\gamma_\chi\neq \gamma'_\chi}\mathcal{M}\Phi(i\gamma_\chi)\mathcal{M}\Phi(i\gamma'_\chi)\,\widehat{g}\bigg(\dfrac{(\gamma_\chi-\gamma'_\chi)\log Q}{2\pi}\bigg) \nonumber \\
& \geq \dfrac{2-\delta}{r(f)}\,\displaystyle\sum_{\gamma_\chi}m_{\rho_\chi}\big|\mathcal{M}\Phi(i\gamma_\chi)\big|^2.
\end{aligned}$$ This implies that $$\begin{aligned}
& \sum_{Q\leq q\leq 2Q} \dfrac{W(q/Q)}{\varphi(q)}\displaystyle\sum_{\substack{\chi \ (\mathrm{mod} \ q) \\ \text{primitive}}}\displaystyle\;\sum_{\gamma_\chi, \gamma'_\chi}\mathcal{M}\Phi(i\gamma_\chi)\mathcal{M}\Phi(i\gamma'_\chi)g\bigg(\dfrac{(\gamma_\chi-\gamma'_\chi)\log Q}{2\pi}\bigg) \geq \dfrac{2-\delta}{r(f)}N^*_{\Phi}(Q).\end{aligned}$$ On the other hand, observing that $$\Phi\big(Q^{-|{x}|}\big)^2\log Q \bigg(\dfrac{1}{2\pi}\int_{-\infty}^{\infty}\big|\mathcal{M}\Phi(it)\big|^2{\,\text{\rm d}}t\Big)\to {{\boldsymbol}\delta}(x),$$ as $Q\to \infty$ (in the distributional sense) and that $$\log^{1/2} Q \int_{-(2-\delta)}^{2-\delta} g(x) \Phi(Q^{-|x|}){\,\text{\rm d}}x \leq 2\log^{-1/2} Q \int_{Q^{-({2-\delta})}}^1 \Phi(t)\frac{{\,\text{\rm d}}t}{t} = O(\log^{-1/2} Q),$$ we can use the asymptotic estimate to obtain $$\begin{aligned}
\int_{-\infty}^{\infty}g({x})F_{\Phi}(Q^{x},W){\,\text{\rm d}}{x}
& \leq \int_{-(2-\delta)}^{2-\delta}g({x})F_{\Phi}(Q^{x},W){\,\text{\rm d}}{x} \\ & = g(0)+ \int_{-(2-\delta)}^{2-\delta}g({x})(1-(1-|x|)_+){\,\text{\rm d}}x + O(\log^{-1/2} Q) + o(1)
\\ & =\frac{2 {\mathcal{L}}(f)}{r(f)} + O(\delta) + o(1).\end{aligned}$$ We then conclude that $$N^*_{\Phi}(Q) \leq N_{\Phi}(Q) \left({\mathcal{L}}(f) + O(\delta) + o(1) \right).$$ Using we finish the proof.
Bounding $N_1^*(T)$
-------------------
Similarly to the case of the Riemann zeta-function, the functionals that we need to define depend on the asymptotic behavior of the function $F_1(x,T)$ defined by $$\begin{aligned}
\label{5_9_18:57pm}
F_1(x,T)=N_1(T)^{-1}\displaystyle\sum_{0<\gamma_1, \gamma'_1 \leq T}T^{i\alpha(\gamma_1-\gamma'_1)}w(\gamma_1-\gamma'_1),\end{aligned}$$ where $x\in\mathbb{R}$, $T>0$ and the sum is over pairs of ordinates of zeros (with multiplicity) of $\xi'(s)$. To analyze $N_1^{*}(T)$ we define the following functional $$\begin{aligned}
{\mathcal{Z}}_{1}(f) =r(f)+\dfrac{2}{r(f)}\int_{0}^{r(f)}x\, f(x){\,\text{\rm d}}x-\dfrac{8}{r(f)^2}\int_{0}^{r(f)}x^2\, f(x){\,\text{\rm d}}x +\displaystyle\sum_{k=1}^{\infty}\dfrac{2c_k}{r(f)^{2k+1}}\int_{0}^{r(f)}x^{2k+1}\, f(x){\,\text{\rm d}}x,\end{aligned}$$ where $c_k=2^{2k+1}\frac{(k-1)!}{(2k)!}$.
\[thmsimplezerosofxiprime\] Let $f\in {\mathcal{A}}_{LP}$. Assuming RH, for every fixed small $\delta >0$ we have $$N_1^*(T)\leq ({\mathcal{Z}}_1(f) + O(\delta) + o(1))N_1(T).$$
A result similar to for the function $F_1(x,T)$ defined in is also known (see [@FGL Theorem 1.1]), which is the following: for any fixed small $\delta>0$ we have $$F_1(x,T)=T^{-2|x|}\log T+|x| - 4|x|^2+\displaystyle\sum_{k=1}^{\infty}c_k |x|^{2k+1}+ o(1)(1+T^{-2|x|}\log T),$$ uniformly for $|x|\leq 1-\delta$ as $T\to\infty$, where $c_k=2^{2k+1}\frac{(k-1)!}{(2k)!}$. The proof then follows the same strategy as the proof for $\zeta(s)$ and we leave the details to the reader.
Numerically optimizing the bounds
=================================
Going back to the sphere packing problem, since we obviously have $\Delta({\mathbb{R}}^1)=1$, this shows $r(f)\geq 1$ for all $f\in {\mathcal{A}}_{LP}$. The last sign change equals $1$ for two (suspiciously) well-known functions: the hat function $$H(x)=(1-|x|)_+,$$ whose Fourier transform is ${\widehat}H(x)=\frac{\sin^2(\pi x)}{(\pi x)^2}$, and Selberg’s function $$S(x)=\frac{\sin^2(\pi x)}{(\pi x)^2(1-x^2)},$$ whose Fourier transform is supported in $[-1,1]$ and given by ${\widehat}S(x)=1-|x|+\frac{\sin(2\pi x)}{2\pi}$ for $|x|<1$. In particular, we can use these two functions to evaluate the functionals derived in Section \[prelims\] to obtain bounds, but this does not result in the best possible bounds. To obtain better bounds we use the class of functions used in the linear programming bounds by Cohn and Elkies [@CE] for sphere packing. That is, we consider the subspace $\mathcal A_{LP}(d)$ consisting of the functions $f \in \mathcal A_{LP}$ of the form $$\label{eq:subspace}
f(x) = p(x) e^{-\pi x^2},$$ where $p$ is an even polynomial of degree $2d$.
In [@CE], optimization over a closely related class of functions is done by specifying the functions by their real roots and optimizing the root locations. For the sphere packing problem this works very well, where in ${\mathbb{R}}^{24}$ it leads to a density upper bound that is sharp to within a factor $1 + 10^{-51}$ of the optimal configuration [@CM]. We have also tried this approach for the optimization problems in this paper, but this did not work very well because the optimal functions seem to have very few real roots, which produces a strange effect in the numerical computations, where the last forced root tends to diverge when you increase the degree of the polynomial[^2]. Instead we use sum-of-squares characterizations and semidefinite programming, as was done in [@dLOV] for the binary sphere packing problem.
*Semidefinite programming* is the optimization of a linear functional over the intersection of a cone of positive semidefinite matrices (real symmetric matrices with nonnegative eigenvalues) and an affine space. A semidefinite program is often given in block form, which can be written as $$\begin{aligned}
\text{minimize} \; \sum_{i=1}^I \mathrm{tr}(X_iC_i) : \; & \sum_{i=1}^I \mathrm{tr}(X_iA_{i,j}) = b_j \text{ for } j \in [m],\\
& \,X_1,\ldots,X_I \in \mathbb R^{n \times n} \text{ positive semidefinite},\end{aligned}$$ where $I\in {\mathbb{N}}$ gives the number of blocks, $\{C_i\} \subseteq {\mathbb{R}}^{n \times n}$ is the objective, and $\{A_{i,j}\} \subseteq {\mathbb{R}}^{n \times n}$, $b \in {\mathbb{R}}^m$ give the linear constraints (for notational simplicity we take all blocks to have the same size). Semidefinite programming is a broad generalization of linear programming (which we recover by setting $n=1$ in the above formulation), and, as for linear programming, there exist efficient algorithms for solving them. The reason semidefinite programming comes into play here, is that we can model polynomial inequality constraints as sum-of-squares constraints, which in turn can be written as semidefinite constraints; see, e.g., [@Bl].
Proof of Theorems \[thmNstar\], \[thmsimplezerosofdir\], and \[thmn1star\]
--------------------------------------------------------------------------
To obtain the first part of Theorem \[thmNstar\] from Theorem \[thmsimplezerosofxi\] we need to minimize the functional ${\mathcal{Z}}$ over the space $\mathcal A_{LP}(d)$. We can see this as a bilevel optimization problem, where we optimize over scalars $R \geq 1$ in the outer problem, and over functions $f \in \mathcal A_{LP}(d)$ satisfying $r(f) = R$ in the inner problem. The outer problem is a simple one dimensional optimization problem for which we use Brent’s method [@Br]. The inner problem can be written as a semidefinite program as we discuss below. The numerical results suggest that the optimal $R$ goes to $1$ as $d \to \infty$ (which is itself intriguing and so far we have no explanation), but for fixed $d$ we need to optimize $R$ to obtain a good bound.
A polynomial $p$ that is nonnegative on $[R,\infty)$ can be written as $s_1(x) + (x-R) s_2(x)$, where $s_1$ and $s_2$ are sum-of-squares polynomials with $\mathrm{deg}(s_1),\mathrm{deg}(s_2(x))+1 \leq \deg(p)$; see, e.g., [@PS]. This shows that functions of the form that are non-positive on $[R,\infty)$ can be written as $$f(x) = -\big(s_1(x^2) + (x^2-R^2) s_2(x^2)\big) e^{-\pi x^2}.$$
Let $v(x)$ be a vector whose entries form a basis of the univariate polynomials of degree at most $d$. The polynomials $s_1$ and $s_2$ are sum-of-squares if and only if they can be written as $s_i(x) = v(x)^{\sf T} X_i v(x)$ for some positive semidefinite matrices $X_i$ of size $d+1$. That is, we can parameterize functions of the form that are non-positive on $[R,\infty)$ by two positive semidefinite matrices $X_1$ and $X_2$ of size $d+1$.
The space of functions of the form is invariant under the Fourier transform. Since a polynomial of degree $2d$ that is nonnegative on $[0,\infty)$ can be written as $s_3(x) + x s_4(x)$, where $s_i(x) = v(x)^{\sf T} X_i v(x)$ for $i=3,4$ are sum-of-squares polynomials of degree $2d$, we have that $\hat f$ is of the form $${\widehat}{f}(x) = \big(s_3(x^2) + x^2 s_4(x^2)\big) e^{-\pi x^2}.$$
Let $\mathcal T$ be the operator that maps $x^{2k}$ to the function $\frac{k!}{\pi^k} L_k^{-1/2}(\pi x^2)$, where $L_k$ is the Laguerre polynomial of degree $k$ with parameter $-1/2$. Then, for $p$ an even polynomial, we have that $(\mathcal Tp)(x) e^{-\pi x^2}$ is the Fourier transform of $p(x) e^{-\pi x^2}$. We can now describe the functions of the form that are non-positive on $[R,\infty)$ and have nonnegative Fourier transform by positive semidefinite matrices $X_1,\ldots,X_4$ of size $d+1$ whose entries satisfy the linear relations coming from the identity $I(X_1,\ldots,X_4) = 0$, where $$I(X_1,\ldots,X_4) = \mathcal T\big(-s_1(x^2) - (x^2-R^2) s_2(x^2)\big) - \big(s_3(x^2) + x^2 s_4(x^2)\big).$$ Here $\mathcal T(-s_1(x^2) - (x^2-R^2) s_2(x^2))$ is a polynomial whose coefficients are linear combinations in the entries of $X_1$ and $X_2$, and the same for $s_3(x^2) + x^2 s_4(x^2)$ with $X_3$ and $X_4$. The linear constraints on the entries of $X_1,\ldots,X_4$ are then obtained by expressing $I(X_1,\ldots,X_4)$ in some polynomial basis and setting the coefficients to zero.
The conditions $f(0) = 1$ and $f(R) = 0$ are linear in the entries of $X_1$ and $X_2$, and the condition ${\widehat}f(0)=1$ is a linear condition on the entries of $X_3$ and $X_4$. Finally, the objective ${\mathcal{Z}}(f)$ is a linear combination in the entries of $X_1$ and $X_2$, which can be implemented by using the identity $$\int x^m e^{-\pi x^2} {\,\text{\rm d}}x = -\frac{1}{2\pi^{m/2+1/2}} \Gamma\Big(\frac{m+1}{2}, \pi x^2\Big),$$ where $\Gamma$ is the upper incomplete gamma function. Hence, the problem of minimizing ${\mathcal{Z}}(f)$ over functions $f \in \mathcal A_{LP}(d)$ that satisfy $r(f) = R$ is a semidefinite program.
To obtain the second part of Theorem \[thmNstar\] from Theorem \[thmsimplezerosofxi\] and to obtain Theorem \[thmsimplezerosofdir\] from \[thmnphi\] we use the same approach with a different functional. To obtain Theorem \[thmn1star\] from Theorem \[thmsimplezerosofxiprime\] we also do the same as above, but now truncate the series in the functional ${\mathcal{Z}}_{1}$ at $k=15$ and add the easy to compute upper bound $10^{-10}$ on the remainder of the terms.
### Implementation and numerical issues {#sec:rig}
In implementing the above as a semidefinite program we have to make two choices for the polynomial basis that we use: the basis defining the vector $v(x)$, and the basis to enforce the identity $I(X_1,\ldots,X_4) = 0$. This choice of bases is important for the numerical conditioning of the resulting semidefinite program. Following [@dLOV] we choose the Laguerre basis $\{L_n^{-1/2}(2\pi x^2)\}$, as this seems natural and performs well in practice (it multiplied by $e^{-\pi x^2}$ is the complete set of even eigenfunctions of the Fourier transform). We solve the semidefinite programs using sdpa-gmp [@Nakata], which is a primal-dual interior point solver using high precision floating point arithmetic. For the code to generate the semidefinite programs and to perform the post processing we use Julia [@BEKV], Nemo [@FHHJ], and Arb [@J] (where we use Arb for the ball arithmetic used in the verification procedure). For all computations we use $d=40$. In solving the systems we observe that $X_1$ can be set to zero everywhere without affecting the bounds, so that $r(f) = R$ holds exactly for the function $f(x) = (R^2 - x^2) v(x^2)^{\sf T} X_2 v(x^2) e^{-\pi x^2}$ defined by $X_2$.
The above optimization approach uses floating point arithmetic and a numerical interior point solver. This means the identity $I(0, X_2, X_3, X_4) = 0$ will not be satisfied exactly, and, moreover, because the solver can take infeasible steps the matrices $X_2$, $X_3$, and $X_4$ typically have some eigenvalues that are slightly negative. In practice this leads to incorrect upper bounds if the floating point precision is not high enough in relation to the degree $d$. Here we explain the procedure we use to obtain bounds that are guaranteed to be correct. This is an adaptation of the method from [@Lo] and [@dLOV].
We first solve the above optimization problem numerically to find $R$ and $f$ for which we have a good objective value $v = {\mathcal{L}}(f)$. Then we solve the semidefinite program again for the same value of $R$, but now we solve it as a feasibility problem with the additional constraint ${\mathcal{L}}(f) \leq v + 10^{-6}$. The interior point solver will try to give the analytical center of the semidefinite program, so that typically the matrices are all positive definite; that is, the eigenvalues are all strictly positive. Then we use interval arithmetic to check rigorously that $X_2$, $X_3$, and $X_4$ are positive definite, and we compute a rigorous lower bound $b$ on the smallest eigenvalues of $X_3$ and $X_4$.
Using interval arithmetic we compute an upper bound $B$ on the largest coefficient of $I(0, X_2, X_3,X_4)$ in the basis given by the $2d+1$ entries on the diagonal and upper diagonal of the matrix $(R^2 - x^2) v(x^2) v(x^2)^{\sf T}$. If $b \geq (1+2d)B$, then it follows that it is possible to modify the corresponding entries in $X_3$ and $X_4$ such that these matrices stay positive definite and such that $I(0,X_2,X_3,X_4) = 0$ holds exactly [@Lo]. This shows that the Fourier transform of the function $f(x) = (R^2 - x^2) v(x^2)^{\sf T} X_2 v(x^2)e^{-\pi x^2}$ is nonnegative.
We use interval arithmetic to compute $f(0) = R^2 s_2(0)$, $\mathcal T((R^2-x^2) s_2(x^2))(0)$, and ${\mathcal{Z}}(f)$, $\tilde {\mathcal{Z}}(f)$, ${\mathcal{Z}}_1(f)$, or ${\mathcal{L}}(f)$. We can then compute rigorous bounds by observing that, for example, the first part of Theorem \[thmNstar\] can be written as follows: Suppose $f$ is a continuous $L^1({\mathbb{R}})$ function with $f(x) \leq 0$ for $|x| \geq R$ and with nonnegative Fourier transform, then $$N^*(T) \leq \left(\frac{f(0)}{\hat f(0)} {\mathcal{Z}}(f) + o(1)\right) N(T).$$
In the arXiv version of this paper we attach the files ‘Z-$40$.txt’, ‘tildeZ-$40$.txt’, ‘L-$40$.txt’, and ‘Z1-$40$.txt’ that contain the value of $R$ on the first line and the matrices $X_2, X_3$ and $X_4$ on the next $3$ lines (all in $100$ decimal floating point values). For convenience it also contains the coefficients of $f$ in the monomial basis on the last line (but these are not used in the verification procedure). We include a script to perform the above verification and compute the bounds rigorously, as well as the code for setting up the semidefinite programs, using a custom semidefinite programming specification library.
Proof of Theorem \[thmpaircorreltation\]
----------------------------------------
To obtain the first part of Theorem \[thmpaircorreltation\] from Theorem \[thmsmallgapszerosofxi\] we need to minimize the function ${\mathcal{P}}$ over the space $\mathcal A_{LP}$. We can formulate this as a bilevel optimization problem in which we optimize over $R \geq 1$ in the outer problem. In the inner problem we perform a binary search over $\Lambda$ to find the smallest $\Lambda$ for which there exists a function $f \in \mathcal A_{LP}(d)$ that satisfies $f(R) = 0$, $f(x) \leq 0$ for $|x| \geq R$, and $p_f(\Lambda) \geq 0$.
To get a bound whose correctness we can verify rigorously we replace the constraints $f(0)=1$, ${\widehat}f(0) = 1$, and $p_f(\Lambda) \geq 0$ by $f(0) = 1-10^{-10}$, ${\widehat}f(0) = 1+10^{-10}$, and $p_f(\Lambda) \geq 10^{-10}$. We then use the above optimization approach to find good values for $R$ and $\Lambda$. We then add $10^{-6}$ to $\Lambda$ and solve the feasibility problem again to get the strictly feasible matrices $X_2,X_3$, and $X_4$. By performing the same procedure as in \[sec:rig\] we can verify that the Fourier transform of the function $f$ defined by $X_2$ is nonnegative everywhere, and using interval arithmetic we can check that the inequalities $f(0) \leq 1$, ${\widehat}f(0) \geq 1$, and $p_f(\Lambda) > 0$ all hold. Note that this verification procedure does not actually check that $\Lambda$ is equal to or even close to ${\mathcal{P}}(f)$, but the proof of Theorem \[thmsmallgapszerosofxi\] also works if we replace ${\mathcal{P}}(f)$ by any $\Lambda$ for which $p_f(\Lambda)$ is strictly positive. To obtain the second part of the theorem, we do the same except that we replace $p_f$ by ${\widetilde}p_f$.
In the arXiv version of this paper we attach the files ‘P-$40$.txt’, ‘tildeP-$40$.txt’, that have the same layout as the files mentioned in \[sec:rig\], with an additional line containing the value of $\Lambda$. We again include the code to perform the verification and to produce the files.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are very thankful to Emanuel Carneiro and Micah Milinovich for the helpful discussions, suggestions and references.
[100]{}
J. Bezanson, A. Edelman, S. Karpinski and V.B. Shah, , SIAM Rev. 59 (2017), 65–98.
G. Blekherman, P. Parrilo, R. Thomas, ,
R. P. Brent, Algorithms for Minimization Without Derivatives, Englewood Cliffs, NJ: Prentice-Hall, 1973. Ch. 3-4.
H. M. Bui and D. R. Heath-Brown, , Bull. Lond. Math. Soc. 45 (2013), no. 5, 953–961.
E. Carneiro and V. Chandee, Bounding $\zeta(s)$ in the critical strip, J. Number Theory 131 (2011), no. 3, 363–384.
E. Carneiro and A. Chirre, Bounding $S_n(t)$ on the Riemann hypothesis, Math. Proc. Cambridge Philos. Soc. 164 (2018), no. 2, 259–283.
E. Carneiro, A. Chirre and M. B. Milinovich, Bandlimited approximations and estimates for the Riemann zeta-function, Publ. Mat., to appear.
E. Carneiro, V. Chandee, F. Littmann and M. B. Milinovich, Hilbert spaces and the pair correlation of zeros of the Riemann zeta-function, J. Reine Angew. Math. 725 (2017), 143–182.
E. Carneiro, V. Chandee and M. B. Milinovich, Bounding $S(t)$ and $S_1(t)$ on the Riemann Hypothesis, Math. Ann. 356 (2013), no. 3, 939–968.
E. Carneiro, V. Chandee and M. Milinovich, A note on the zeros of zeta and $L$-functions, Math. Z. 281 (2015), no. 1-2, 315–332.
E. Carneiro and R. Finder, On the argument of $L$-functions, Bull. Braz. Math. Soc. (N.S.) 46 (2015), no. 4, 601–-620.
E. Carneiro and F. Littmann, Extremal functions in de Branges and Euclidean spaces, Adv. Math. 260 (2014), 281–349.
E. Carneiro, F. Littmann, and J. D. Vaaler, Gaussian subordination for the Beurling-Selberg extremal problem, Trans. Amer. Math. Soc. 365 (2013), no. 7, 3493–3534.
E. Carneiro, M.B. Milinovich, K. Soundararajan, [ Fourier optimization and prime gaps]{}, Comment. Math. Helv., to appear.
E. Carneiro and J. D. Vaaler, Some extremal functions in Fourier analysis II, Trans. Amer. Math. Soc. 362 (2010), no. 11, 5803–5843.
V. Chandee, Y. Lee, S.-C. Liu and M. Radziwiłł, Simple zeros of primitive Dirichlet $L$-functions and the asymptotic large sieve, Q. J. Math 65 (2014), no. 1, 63–87.
V. Chandee and K. Soundararajan, Bounding $|\zeta(1/2+it)|$ on the Riemann hypothesis, Bull. London Math. Soc. 43 (2011), no. 2, 243–250.
A. Y. Cheer and D. A. Goldston, Simple zeros of the Riemann zeta-function, Proc. Amer. Math. Soc. 118 (1993), no. 2, 365–372.
A. Chirre, A note on Entire $L$-functions, Bull. Braz. Math. Soc., to appear.
H. Cohn and N. Elkies, , Ann. of Math. (2) 157 (2003), no. 2, 689–714.
H. Cohn and F. Gonçalves, An optimal uncertainty principle in twelve dimensions via modular forms, arXiv:1712.04438 \[math.CA\].
H. Cohn and S. D. Miller, , .
H. Cohn, A. Kumar, S. D. Miller, D. Radchenko, and M. Viazovska, The sphere packing problem in dimension $24$, Ann. of Math. (2) [185]{} (2017), no. 3, 1017–1033.
D. Farmer, S. Gonek and Y. Lee, Pair correlation of the zeros of the derivative of the Riemann $\xi$-function, J. Lond. Math. Soc.(2) 90 (2014), no. 1, 241–269.
C. Fieker, W. Hart, T. Hofmann and F. Johansson, ,
P. X. Gallagher, Pair correlation of zeros of the zeta function, J. Reine Angew. Math. 362 (1985), 72–86.
D. A. Goldston and S. M. Gonek, A note on S(t) and the zeros of the Riemann zeta-function, Bull. Lond. Math. Soc. 39 (2007), no. 3, 482–486.
D. A. Goldston, S. M. Gonek, A. E. Özlük and C. Snyder, On the pair correlation of zeros of the Riemann zeta-function, Proc. London Math. Soc. (3) 80 (2000), no. 1, 31–49.
D. A. Goldston and H. L. Montgomery, Pair correlation of zeros and primes in short intervals, in [*Analytic number theory and Diophantine problems*]{} (ed. A. C. Adolphson, J. B. Conrey, A. Ghosh and R. I. Yager), Birkhaüser, Boston (1987), 183–203.
F. Gonçalves, M. Kelly and J. Madrid, One-sided band-limited approximations of some radial functions, Bull. Braz. Math. Soc. (N.S.) 46 (2015), no. 4, 563–599.
S. W. Graham and J. D. Vaaler, A class of extremal functions for the Fourier transform, Trans. Amer. Math. Soc. 265 (1981), no. 1, 283–302.
J. Holt and J. D. Vaaler, The Beurling-Selberg extremal functions for a ball in the Euclidean space, Duke Math. J. 83 (1996), no. 1, 202–248.
F. Johansson, Arb: efficient arbitrary-precision midpoint-radius interval arithmetic, IEEE Transactions on Computers, 66 (2017), no. 8, 1281–1292.
D. de Laat, F. Oliveira, F. Vallentin, Upper bounds for packings of spheres of several radii, Forum Math. Sigma 2 (2014), e23, 42 pp.
J. Löfberg, Pre- and post-processing sums-of-squares programs in practice IEEE Trans. Automat. Control 54 (2009), 1007–1011.
H. L. Montgomery, The pair correlation of zeros of the zeta function, Proc. Symp. Pure Math. 24, Providence (1973), 181–193.
M. Nakata, A numerical evaluation of highly accurate multiple-precision arithmetic version of semidefinite programming solver: SDPA-GMP,-QD and-DD, In Computer-Aided Control System Design (CACSD), 2010 IEEE International Symposium on, pages 29–34. IEEE, 2010.
G. Pólya, G. Szegö, Problems and theorems in analysis. II, Classics in Mathematics, Springer-Verlag, 1998; Translated from German by C. E. Billigheimer; Reprint of the 1976 English translation.
K. Sono, A note on simple zeros of primitive Dirichlet L-functions, Bull. Aust. Math. Soc. 93 (2016), no. 1, 19–30.
J. D. Vaaler, Some extremal functions in Fourier analysis, Bull. Amer. Math. Soc. (N.S.) 12 (1985), no. 2, 183–216.
M. S. Viazovska, The sphere packing problem in dimension $8$, Ann. of Math. (2) [185]{} (2017), no. 3, 991–1015.
[^1]: For every sum over zeros in this article the involved quantities should be repeated according to the multiplicity of the zero.
[^2]: It is worth mentioning that, in a related uncertainty problem, Cohn and Gonçalves [@CGo] discovered the same kind of instability in low dimensions.
|
---
abstract: |
A function $F:\mathbb{F}_{p^n}\rightarrow \mathbb{F}_{p^m},$ is a vectorial $s$-plateaued function if for each component function $F_{b}(\mu)=Tr_n(\alpha F(x)), b\in \mathbb{F}_{p^m}^*$ and $\mu \in \mathbb{F}_{p^n}$, the Walsh transform value $|\widehat{F_{b}}(\mu)|$ is either $0$ or $ p^{\frac{n+s}{2}}$. In this paper, we explore the relation between (vectorial) $s$-plateaued functions and partial geometric difference sets. Moreover, we establish the link between three-valued cross-correlation of $p$-ary sequences and vectorial $s$-plateaued functions. Using this link, we provide a partition of $\mathbb{F}_{3^n}$ into partial geometric difference sets. Conversely, using a partition of $\mathbb{F}_{3^n}$ into partial geometric difference sets, we constructed ternary plateaued functions $f:\mathbb{F}_{3^n}\rightarrow \mathbb{F}_3$. We also give a characterization of $p$-ary plateaued functions in terms of special matrices which enables us to give the link between such functions and second-order derivatives using a different approach.
partial geometric designs, partial geometric difference sets, plateaued functions, three-valued cross-corelation function.\
address: 'Department of Mathematics, Faculty of Science, Ankara University, Tandogan, Ankara, 06100, Turkey.'
author:
- Ayça Çeşmelioğlu
- Oktay Olmez
title: Graphs of Vectorial Plateaued Functions as Difference Sets
---
[^1]
Introduction
============
A (block) design is a pair $(\mathcal{P}, \mathcal{B})$ consisting of a finite set $\mathcal{P}$ of points and a finite collection $\mathcal{B}$ of nonempty subsets of $\mathcal{P}$ called blocks. Designs serve as a fundamental tool to investigate combinatorial objects. Also designs have attracted many researchers from different fields for solutions of applications problems including binary sequences with $2$-level autocorrelation, optical orthogonal codes, low density parity check codes, synchronization, radar, coded aperture imaging, and optical image alignment, distributed storage systems and cryptographic functions with high nonlinearity[@chung1989; @delsarte; @dillon1974; @ding2015; @olmez2].
One of the main construction method of designs is called difference set method. This method served as a powerful tool to construct symmetric designs, error correcting codes, graphs and cryptographic functions [@assmus1992; @bernasconi2001; @bjl; @brouwer2012; @buratti2011; @dingbook; @dingbook2; @pott; @pott2011]. This paper will focus on the links between designs and a family of a function known as plateaued functions from cryptography. Especially we will investigate the connections between partial geometric difference sets and graph of plateaued functions.
A function from the field $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_p$ is called a $p$-ary function. If $p=2$ then the function is simply called as Boolean. $p$-ary functions with various characteristics have been an active research subject in cryptography. Bent functions and plateaued functions are two well-known families which has prominent properties in this field [@carlet2003; @CCC; @Carlet; @Carlet1; @Carlet2; @CarM; @Mes1; @Mes2]. These two families of functions can be characterized by their Walsh spectrum. A function $f$ from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_p$ is called an $s$-plateaued function if the Walsh transform $|\widehat{f}(\mu)|\in \{0,p^{\frac{n+s}{2}}\}$ for each $\mu \in\mathbb{F}_{p^n}$. A $0$-plateaued function $f$ is called as bent and its Walsh transform satisfies $|\widehat{f}(\mu)|= p^{\frac{n}{2}}$ for each $\mu \in\mathbb{F}_{p^n}$. Plateaued functions and bent functions play a significant role in cryptography, coding theory and sequences for communications [@Carlet1; @Carlet2; @CarM].
Boolean bent functions were introduced by Rothaus in [@rot1976]. These functions have optimal nonlinearity and can only exist when $n$ is even. In [@dillon1974], it is shown that the existence of Boolean bent functions is equivalent to the existence of a family of difference sets known as Hadamard difference sets. Boolean plateaued functions are introduced by Zheng and Zhang as a generalization of bent functions in [@ZZ]. Boolean plateaued functions have attracted the attention of researchers since these functions provide some suitable candidates that can be used in cryptosystems. A difference set characterization of these functions was recently provided by the second author. In [@olmez2], it is shown that the existence of Boolean plateaued functions is equivalent to the existence of partial geometric difference sets.
In arbitrary characteristic, the graph of $f:\mathbb{F}_{p^n}\rightarrow \mathbb{F}_{p^m}$, $G_{f}=\{(x,f(x)):x\in \mathbb{F}_{p^{n}}\}$, plays an important role for the relation to difference sets, [@pott1; @tan2010]. For instance, the graph of a $p$-ary bent function can be recognized as a relative difference set. In general, a characterization of plateaued functions in terms of difference sets is not known. A partial result in this direction is provided in [@cmt] for partially bent functions which is a subfamily of plateaued functions.
There are recent result concerning explicit characterization of plateaued functions in odd characteristics through their second order derivatives in [@CMOS; @mos1; @mos2].
In this paper, we first investigate the link between the graph of a plateaued function and partial geometric difference set. We also provide several characterization of plateaued functions in terms of associated difference set properties. By using these characterization we provide a family of vectorial plateaued functions which has an interesting connection to three-valued cross correlation functions.
The organization of the paper is as follows. In Section 2, we provide preliminary results concerning partial geometric difference sets. In Section 3, we mainly provide the links between vectorial plateaued functions and partial geometric difference sets. We also provide a construction as a result of our characterizations. In Section 4, we focus on $p-$ary plateaued function. We provide several characteristics which are obtained from Butson-Hadamard-like matrices. This section also provides results concerning partially bent functions and partial geometric designs.
Preliminaries
=============
Let $G$ be a group of order $v$ and let $S \subset G$ be a $k$-subset. For each $g \in G$, we define $$\delta(g):=|\{(s,t)\in S \times S \colon g=st^{-1}\}|.$$ Next we define the difference sets of our interest.
Let $v, k$ be positive integers with $v>k>2$. Let $G$ be a group of order $v$. A $k$-subset $S$ of $G$ is called a partial geometric difference set (PGDS) in $G$ with parameters $(v, k; \alpha,\beta)$ if there exist constants $\alpha$ and $\beta$ such that, for each $x\in G$, $$\sum\limits_{y\in S}\delta(xy^{-1})=\left \{\begin{array}{ll} \alpha & \mbox{if } x\notin S,\\
\beta & \mbox{if } x\in S.\\ \end{array} \right .$$
There are two subclasses of PGDS namely difference sets and semiregular relative difference sets which have deep connections with coding theory, and cryptography [@assmus1992; @dingbook; @dingbook2]:
- A $(v,k,\lambda)$-difference set (DS) in a finite group $G$ of order $v$ is a $k$-subset $D$ with the property that $\delta(g)=\lambda$ for all nonzero elements of $G$.
- A $(m,u,k,\lambda)$-relative difference set (RDS) in a finite group $G$ of order $m$ relative to a (forbidden) subgroup $U$ is a $k$-subset $R$ with the property that $$\delta(g)=\begin{cases}
k & g=1_G \\
\lambda & g\in G \backslash U\\
0 & otherwise \\
\end{cases}$$ The RDS is called [*semiregular*]{} if $m = k = u \lambda$.
Clearly a $(v,k,\lambda)$-DS is a $(v,k; k\lambda, n+k\lambda)$-PGDS and an $(m,u,k,\lambda)$ semiregular RDS is a $(mu,k; \lambda(k-1), k(\lambda+1)-\lambda)$-PGDS [@olmez1].
Group characters are powerful objects to investigate various types of difference sets. A [*character*]{} $\chi$ of an abelian group $G$ is a homomorphism from $G$ to the multiplicative group of the complex numbers. The character $\chi_0$ defined by $\chi_0(g) = 1$ for all $g \in G$ is called the [*principal character*]{}; all other characters are called [*nonprincipal*]{}. We define the character sum of a subset $S$ of an abelian group $G$ as $\chi(S):= \sum_{s \in S} \chi(s)$.
\[chartheoryPGDS\] A $k$-subset $S$ of an abelian group $G$ is a partial geometric difference set in $G$ with parameters $(v,k;\alpha,\beta)$ if and only if $|\chi (S)|=\sqrt{\beta-\alpha}$ or $\chi (S)=0$ for every non-principal character $\chi$ of $G$.
For instance, let $f$ be a $p$-ary bent function from the field $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p}$. The set $G_f=\{(x,f(x)): x\in \mathbb{F}_{p^{n}} \}$ is called graph of $f$. Any non-principal character $\chi$ of the additive group of $\mathbb{F}_{p^{n}}\times \mathbb{F}_p$ satisfies $|\chi(G_f)|^2=p^n$ or $0$. This observation yields that $G_f$ is a $(p^{n+1},p^n, p^{2n-1}-p^{n-1},p^{2n-1}-p^{n-1}+p^n)$-PGDS in $H=\mathbb{F}_{p^{n}}\times\mathbb{F}_{p}$.
Walsh transform provides interesting connections between $p$-ary functions and difference sets. For a prime $p$, we define a primitive complex $p$-th root of unity $\zeta_p=e^{\frac{2\pi i}{p}}$. Let $f$ be a function from the field $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_p$ and let $F(x)=\zeta_p^{f(x)}$. The Walsh transform of $f$ is defined as follows: $$\displaystyle \widehat{f}(\mu)=\sum_{x\in \mathbb{F}_{p^{n}}} \zeta_p^{f(x)-Tr_n(\mu x)},~~~\mu \in \mathbb{F}_{p^{n}}$$ where $$Tr_n(z)=\sum_{i=0}^{n-1}z^{p^i}.$$
The convolution of $F$ and $G$ is defined by $$(F * G)(a)=\sum_{x\in \mathbb{F}_{p^n} } F(x)G(x-a).$$ We will also take advantage of the convolution theorem of Fourier analysis. The Fourier transform of a convolution of two functions is $$\widehat{F * G}=\widehat{F}\cdot \widehat{G}.$$
Results on Vectorial Functions
==============================
Let $F$ be a vectorial function from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p^{m}}$. For every $b\in \mathbb{F}^*_{p^{m}}$, the component function $F_b$ of $F$ from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_p$ is defined as $F_{b}(x)=Tr_m(bF(x))$. A vectorial function is called *vectorial plateaued* if all its nonzero component functions are plateaued. If the nonzero component functions of a vectorial plateaued function are s-plateaued for the same $0 \leq s \leq n$ then $F$ is called as *s-plateaued* following the terminology in [@mos1].
The set $G_{F}=\{(x,F(x)):x\in \mathbb{F}_{p^{n}}\}$ is called the graph of $F$. Next we will characterize vectorial functions by their graphs.
\[plateaued-PGDS\] Let $F:\mathbb{F}_{p^{n}} \rightarrow \mathbb{F}_{p^{m}}$ be a vectorial function. Then the graph of $F$ is a $(p^{n+m},p^n; \alpha,\beta)$ partial geometric difference set in $H=\mathbb{F}_{p^{n}}\times\mathbb{F}_{p^m}$ satisfying $\beta-\alpha=\theta$ if and only if $|\widehat{F_b}(a)| \in \{0, \sqrt{\theta}\}$ for all non zero $b \in \mathbb{F}_{p^{m}}$ and $a \in \mathbb{F}_{p^{n}}$. In particular, $\alpha=p^{2n-m}-p^{n+s-m}$ and $\beta=p^{n+s}+p^{2n-m}-p^{n+s-m}$.
A non-principal character of $\mathbb{F}_{p^{n}}\times \mathbb{F}_{p^{m}}$ can be written as $\chi_{(a,b)}(x,y)=\zeta_p^{Tr_{n}(ax)+Tr_m(by)}$ for a nonzero $(a,b) \in \mathbb{F}_{p^{n}}\times \mathbb{F}_{p^{m}}$. For any nonzero $b \in {{\mathbb F}}_{p^n}$, the Walsh transform of $F_b$ is $$\begin{aligned}
\widehat{F_b}(a)=&\sum_{x\in \mathbb{F}_{p^{n}}}\zeta_p^{-Tr_{n}(ax)+Tr_m(bF(x))}=\sum_{x\in \mathbb{F}_{p^{n}}}\chi_{(-a,b)}(x,F_b(x)) \\
=&\chi_{(-a,b)}(G_F)
\end{aligned}$$ for any $a \in F_{p^m}$. Therefore $|\widehat{F_b}(a)| \in \{0, \sqrt{\theta}\} |$ for all non zero $b \in \mathbb{F}_{p^{m}}$ if and only if $G_{F}=\{(x,F(x)):x\in \mathbb{F}_{p^{n}}\}$ is a $(p^{n+m},p^n;\alpha,\beta)$ partial geometric difference set satisfying $\beta-\alpha=\theta$. Using the well-known *Parseval identiy*, one immediately sees that $\theta=p^{n+s}$ and $|\{a\in \mathbb{F}_{p^{n}}:\widehat{F_b}(a)\neq 0\}|=p^{n-s}$ for some $0 \leq s \leq n$. The parameters of a partial geometric difference set satisfies the relation in [@olmez1] and hence we have $$p^{3n}=(\beta-\alpha)p^n+\alpha\nu=p^{n+s}p^n+\alpha\nu.$$ Then we see that $\alpha=p^{2n-m}-p^{n+s-m}$ and $\beta=p^{n+s}+p^{2n-m}-p^{n+s-m}$.
Theorem \[plateaued-PGDS\] implies that a vectorial function $F:\mathbb{F}_{p^{n}} \rightarrow \mathbb{F}_{p^{m}}$ is s-plateaued if and only if its graph is a partial geometric difference set with the parameters $(p^{n+m},p^n;p^{2n-m}-p^{n+s-m},p^{n+s}+p^{2n-m}-p^{n+s-m})$. Note that Theorem \[plateaued-PGDS\] is also valid for $m=1$, i.e. the case of p-ary functions. Since for an $s$-plateaued $p$-ary function $f:\mathbb{F}_{p^{n}}\rightarrow \mathbb{F}_{p}$, the function $bf(x)$ is $s$-plateaued for each $b\in \mathbb{F}_{p}^*$, we can consider $f$ as a vectorial $s$-plateaued function. The case $s=0$ is the case of vectorial bent functions and if we additionally have $m=n$, these vectorial functions are known as *planar* functions [@ColM].
Next we will investigate links between vectorial $s-$plateaued functions and partial geometric difference sets.
\[diff-plateauedprop\] Let $F:\mathbb{F}_{p^{n}}\rightarrow \mathbb{F}_{p^{m}}$ be a vectorial function. Then $F$ is $s$-plateaued if and only if $$\sum\limits_{a \in \mathbb{F}_{p^{n}}}|\{s \in \mathbb{F}_{p^{n}}\colon y=F(s+x-a)-F(s)+F(a)\}|=\left \{\begin{array}{ll} \alpha & \mbox{if } y\neq F(x),\\
\beta & \mbox{if } y=F(x)\\ \end{array} \right .$$
$$\begin{split}
\delta((x,y))&=|\{((s_1,t_1),(s_2,t_2))\in G_F \times G_F \colon x=s_1-s_2, y=t_1-t_2=F(s_1)-F(s_2)\}|\\
&=|\{s_2 \in \mathbb{F}_{p^{n}}\colon y=F(s_2+x)-F(s_2)\}|\\
\end{split}$$ So the criteria for PGDS is given by $$\sum\limits_{a \in \mathbb{F}_{p^{n}}}\delta((x-a,y-F(a)))=\left \{\begin{array}{ll} \alpha & \mbox{if } y\neq F(x),\\
\beta & \mbox{if } y=F(x)\\ \end{array} \right .$$ and hence $$\sum\limits_{a \in \mathbb{F}_{p^{n}}}|\{s \in \mathbb{F}_{p^{n}}\colon y=F(s+x-a)-F(s)+F(a)\}|=\left \{\begin{array}{ll} \alpha & \mbox{if } y\neq F(x),\\
\beta & \mbox{if } y=F(x)\\ \end{array} \right .$$
The above result can be associated with the derivative of an $s-$plateaued function. The derivative of a vectorial function is defined by $$D_aF(x)=F(x+a)-F(x).$$
To see the connection let us first replace $y$ in the expression $$y=F(s+x-a)-F(s)+F(a)$$ by $F(x)-c$ for $c \in \mathbb{F}_{p^n}$. Hence we have $$c=F(s)-F(a)-F(s+x-a)+F(x)=D_{s-a}F(a)-D_{s-a}F(x).$$ This observation yields $$\begin{aligned}
&\sum\limits_{a \in \mathbb{F}_{p^{n}}}|\{s \in \mathbb{F}_{p^{n}}\colon y=F(s+x-a)-F(s)+F(a)\}|\\
=&\sum\limits_{a \in \mathbb{F}_{p^{n}}}|\{s \in \mathbb{F}_{p^{n}}\colon D_{s-a}F(a)-D_{s-a}F(x)=c\}|\\
=&\sum\limits_{a \in \mathbb{F}_{p^{n}}}|\{t \in \mathbb{F}_{p^{n}}\colon D_tF(a)-D_tF(x)=c\}|\\
=&|\{(t,a) \in \mathbb{F}_{p^{n}}\times \mathbb{F}_{p^{n}}\colon D_tF(a)-D_tF(x)=c\}|\\
=&N_F(c, x)
\end{aligned}$$ where $N_F(c, x)$ represents the number of pairs $(t,a) \in \mathbb{F}_{p^{n}}\times \mathbb{F}_{p^{n}}$ such that $$D_tF(a)-D_tF(x)=c$$ as in Section 2 of [@mos1]. Thus we will have the following result concerning the derivative and PGDS parameters.
\[diff2-plateaued\] Let $F$ be a function from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p^m}$. Then the set $G_F$ is a PGDS with parameters $(p^{n+m},p^n ;\alpha, \beta)$ if and only if $$N_F(c, x)= \begin{cases}
\alpha, & c\ne 0 \\
\beta, & c = 0 \\
\end{cases}$$ for all $x \in \mathbb{F}_{p^n}$ and some constants $\alpha$ and $\beta$.
Using Theorem \[diff2-plateaued\], Proposition \[diff-plateauedprop\] and Theorem \[plateaued-PGDS\], we are able to prove Theorem 8i. in [@mos1] with a different approach using the properties of partial geometric difference sets. This gives an interesting relation between the parameters of a PGDS and the second order derivatives of (vectorial) plateaued functions.
A family of vectorial $s-$plateaued functions
---------------------------------------------
In this section, we will discuss the link between vectorial $s$-plateaued functions $F(x)=x^d$ from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p^{n}}$ and the cross-correlation function between two $p$-ary $m$-sequences that differ by a decimation $d$. An $m$-sequence and its decimation is defined by $u(t)=\sigma^t$ and $v(t)=u(dt)$ where $\sigma$ is a primitive element of the finite field. The cross-correlation between the sequences $u$ and $v$ is defined by $$\theta(\tau)=\sum_{t=0}^{p^n-2}\zeta_p^{u(t+\tau)-v(t)}.$$ It can be shown that $$\theta(\tau)=-1+\sum_{x \in \mathbb{F}_{p^{n}} }\zeta_p^{Tr_n(ax+x^d)}$$ where $a=-\sigma^\tau$. Therefore for $F_1(x)=Tr(x^d)$, we have $$\label{CC-Walsh}
\theta(\tau)=-1+\widehat{F_1}(-a).$$
\[cross-plateaued\] Let $F(x)=x^{d}$ be a vectorial function from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p^{n}}$ with $gcd(d,p^n-1)=1$. If the cross-correlation of the $p$-ary $m$-sequences that differ by decimation $d$ takes three values, namely $-1, -1+p^{\frac{n+s}{2}}$ and $-1-p^{\frac{n+s}{2}}$, then $F$ is a vectorial $s$-plateaued function.
For each $b\in \mathbb{F}_{p^{n}}^*$, we denote by $F_b(x)$ the function $F_b(x)=Tr(bF(x))=Tr(bx^d)$. The walsh transform $$\widehat{F_b}(0)=\sum_{x \in \mathbb{F}_{p^{n}}} \zeta_p^{Tr_n(bx^d)}=0$$ since $x^d$ is a permutation. For each $a\in \mathbb{F}_{p^{n}}^*$, the Walsh transform of $F_b(x)$ $$\begin{aligned}
\widehat{F_b}(a)&=\sum_{x \in \mathbb{F}_{p^{n}}} \zeta_p^{Tr_n(bx^d-ax)}\\
&=\sum_{x \in\mathbb{F}_{p^{n}} }\zeta_p^{Tr_n(c^dx^d-\frac{a}{c}cx)}\\
&=\sum_{y \in\mathbb{F}_{p^{n}} }\zeta_p^{Tr_n(y^d-\mu y)}\\
&=\widehat{F_1}(\mu)
\end{aligned}$$ where $b=c^d$ and $\mu=a/c$. Note that any $b\in \mathbb{F}_{p^{n}}^*$ can be written as $b=c^d$ for some $c \in \mathbb{F}_{p^{n}}^*$. Then using Equation (\[CC-Walsh\]), we immediately obtain the result that $F(x)$ is vectorial s-plateaued.
Let $p$ be an odd prime and $n,k$ be positive integers with $gcd(n,k)=s$. If $ n/s$ is odd then $\gcd{(p^n-1,d)}=1$ for $d=(p^{2k}+1)/2$ and $d=p^{2k}-p^k+1$.
There are only finitely many known functions with a three-valued cross-correlation. Trachtenberg proved the following in his thesis [@trachtenberg]. Let $n$ be an odd integer and $k$ be an integer such that $gcd(n,k)=s$. Then for each of the decimations $d=\frac{p^{2k}+1}{2}$ and $d=p^{2k}-p^k+1$ the cross-correlation function $\theta_d(\tau)$ takes the values $-1,-1\pm p^{\frac{n+s}{2}}$.
This result is generalized later by Helleseth in Theorem 4.9 in [@helleseth]. He showed that if $gcd(n,k)=s$ and $ n/s$ is odd then for the same decimations, $\theta_d(\tau)$ has the values $-1,-1\pm p^{\frac{n+s}{2}}$. Our result implies that the corresponding vectorial functions are $s-$plateaued.
Pott et. al. provided a classification of weakly regular bent functions via partial difference sets, [@tan2010]. In this classification authors showed that a function from $\mathbb{F}_3^n$ to $\mathbb{F}_3$ when $n$ is even satisfying $f(-x) = f(x)$ and $f(0)=0$ is weakly regular if and only if a $D_1=\{x: f(x)=1\}$ and $D_2=\{x: f(x)=2\}$ are partial difference sets. Later in [@olmez3], the author provided a similar classification for weakly regular bent functions from $\mathbb{F}_3^n$ to $\mathbb{F}_3$ when $n$ is odd via partial geometric difference sets. Next we will also show that our vectorial functions have a similar classification in the case of $p=3$ and $d=\frac{3^{2k}+1}{2}$.
\[partitioned\]. Let $n \ge 3$ be an integer and $d=\frac{3^{2k}+1}{2}$ with $\gcd{(n,k)}=s$ and $n/s$ is odd. For $i=0,1,2$ the sets $D_i=\{x: F(x)=Tr_n(x^d)=i\}$ are $(3^n, 3^{n-1}, 3^{2n-3}-3^{n-2}, 3^{n-1}+3^{2n-3}-3^{n-2})$ partial geometric difference sets in the additive group of $\mathbb{F}_{3^{n}}$.
For each $a\in {{\mathbb F}}_{3^n}^*$, Suppose that $\chi_a(D_1)=x_a+y_a \zeta_3$ with $x_a,y_a\in \mathbb{R}$. Actually, when we are calculating $\chi_a(D_1)$, we are summing powers of $\zeta_3$ as many times as the number of elements in $D_1$ for which $Tr_{n}(ax)=0,1$ or $2$. In other words, if we set $$D_{1,i}=\{x\in D_1:Tr(x)=i\}, i=0,1,2$$ we have $$\chi_a(D_1)=|D_{1,0}|+|D_{1,1}|\zeta_3+|D_{1,2}|\zeta_3^2=(|D_{1,0}|-|D_{1,2}|)+(|D_{1,1}|-|D_{1,2}|)\zeta_3$$ and hence $$x_a=|D_{1,0}|-|D_{1,2}|,y_a=|D_{1,1}|-|D_{1,2}|$$ are both in $\mathbb{Z}$. Also note that for any $x \in \mathbb{F}_{3^n}$, $$F_1(-x)=Tr((-x)^d)=Tr(-x^d)=-F_1(x)$$ and that gives us $$x\in D_1 \Longleftrightarrow 2x \in D_2.$$ As a consequence, $\chi_a(D_2)=\overline{\chi_a(D_1)}=x_a+y_a \zeta_3^2$. Since $D_i$’s form a partition of the additive group of $\mathbb{F}_{3^{n}}$ $$\chi_a(D_0)+ \chi_a(D_1)+\chi_a(D_2)=0,$$ and we obtain $$\chi_a(D_0)=y_a-2x_a$$ Consider the Walsh transform values $\widehat{F_1}(a), \widehat{F_1}(-a)$ of $F_1(x)=Tr_n(x^d)$. $$\begin{aligned}
\widehat{F_1}(-a)&=\sum_{x\in \mathbb{F}_{3^{n}}}\zeta_3^{Tr_{n}(ax+x^d)}\\
&=\chi_a(D_0)+\zeta_3\chi_a(D_1)+\zeta_3^2\chi_a(D_2)\\
&=y_a-2x_a+\zeta_3(x_a+y_a \zeta_3)+\zeta_3^2(x_a+y_a \zeta_3^2)\\
&=y_a-2x_a+\zeta_3(x_a+y_a)+\zeta_3^2(x_a+y_a)=-3x_a\\
\end{aligned}$$ and $$\begin{aligned}
\widehat{F_1}(a)&=\sum_{x\in \mathbb{F}_{3^{n}}}\zeta_3^{Tr_{n}(-ax+x^d)}\\
&=\chi_{-a}(D_0)+\zeta_3\chi_{-a}(D_1)+\zeta_3^2\chi_{-a}(D_2)\\
&=\overline{\chi_a(D_0)}+\zeta_3\overline{\chi_a(D_1)}+\zeta_3^2\overline{\chi_a(D_2)}\\
&=\overline{\chi_a(D_0)}+\zeta_3\chi_a(D_2)+\zeta_3^2\chi_a(D_1)\\
&=y_a-2x_a+\zeta_3(x_a+y_a \zeta_3^2)+\zeta_3^2(x_a+y_a \zeta_3)\\
&=3(y_a-x_a).
\end{aligned}$$ Since $\widehat{F_1}(a),\widehat{F_1}(-a)\in \{0,\pm3^{(n+s)/2}\}$, $$(x_a,y_a) \in \{(0,0),(0,C),(0,-C),(C,C),(C,0),(C,2C),(-C,-C),(-C,-2C),(-C,0)\}$$ where $C=3^{(n+s-2)/2}$. In the following table values of $\chi_a(D_0),\chi_a(D_1),\chi_a(D_2), W_{F}(a),W_{F}(-a)$ corresponding to each possible $(x_a,y_a)$ tuple is given.
[ | c | c | c | c | c | c | c | c | c | c |]{} & $(0,0)$ & $(0,C)$ & $(0,-C)$ & $(C,C)$ & $(C,0)$ & $(C,2C)$ & $(-C,-C)$ & $(-C,-2C)$ & $(-C,0)$\
$\chi_a(D_0)$ & $0$ & $C$ & $-C$ & $-C$ & $-2C$ & $0$ & $C$ & $0$ & $2C$\
$\chi_a(D_1)$ & $0$ & $C\zeta_3$ & $-C\zeta_3$ & $-C\zeta_3^2$ & $C$ & $C\zeta_3-C\zeta_3^2$ & $C\zeta_3^2$ & $C\zeta_3^2-C\zeta_3$ & $-C$\
$\chi_a(D_2)$ & $0$ & $C\zeta_3^2$ & $-C\zeta_3^2$ & $-C\zeta_3$ & $C$ & $C\zeta_3^2-C\zeta_3$ & $C\zeta_3$ & $C\zeta_3-C\zeta_3^2$ & $-C$\
$\widehat{F_1}(a)$ & $0$ & $0$ & $0$ & $-3C$ & $-3C$ & $-3C$ & $3C$ & $3C$ & $3C$\
$\widehat{F_1}(-a)$ & $0$ & $-3C$ & $3C$ & $-0$ & $-3C$ & $3C$ & $0$ & $-3C$ & $3C$\
The tuples $(C,2C)$ and $(-C,-2C)$ are impossible since $$|\chi_a(D_1)|=\sqrt{x_a^2-x_ay_a+y_a^2}=\sqrt{3C^2}=\sqrt{3^{n+s-1}}$$ which contradicts the fact that $|\chi_a(D_1)|\in \mathbb{Z}$ since $n+s-1$ is odd. Therefore the sets $D_1,D_2$ are PGDS.
Next we will show that the tuples $(C,0)$ and $(-C,0)$ are also impossible.
First note that $\widehat{F_1}(0)=0$ since the function $Tr(x^d)$ is balanced. For a non-zero element $a$ the following holds $$\begin{split}
\widehat{F_1}(a)\widehat{F_1}(-a) =& \chi_a(D_0)\chi_a(D_0)+\zeta_p\chi_a(D_0)\chi_a(D_2)+\zeta_p^2(\chi_a(D_0)\chi_a(D_1))\\
&+\zeta_p\chi_a(D_0)\chi_a(D_1)+\zeta_p^2\chi_a(D_1)\chi_a(D_2)+(\chi_a(D_1)\chi_a(D_1))\\
&+\zeta_p^2(\chi_a(D_0)\chi_a(D_2))+(\chi_a(D_2)\chi_a(D_2)+\zeta_p\chi_a(D_1)\chi_a(D_2)\\
&=\left( \chi_a(D_0)\chi_a(D_0)+ \chi_a(D_0)\chi_a(D_1)+ \chi_a(D_1)\chi_a(D_1)\right) (2-\zeta_p-\zeta_p^2)\\
&=3\left( \chi_a(D_0)\chi_a(D_0)+ \chi_a(D_0)\chi_a(D_1)+ \chi_a(D_1)\chi_a(D_1)\right)\\
&=3\left( \chi_a(D_0)\chi_a(D_0)- \chi_a(D_1)\chi_a(D_2)\right)\\
\end{split}$$ We need the following auxiliary lemma.
Let $S$ be a $k-$subset of an abelian group $G$ of order $v$. Then $$\displaystyle \sum_{i=0}^{v-1}\chi_i(SS^{-1})= \sum_{i=0}^{v-1}\chi_i(ke+\sum_{g\in G-{e}}a_gg)=vk.$$
In the group ring $\mathbb{Z}G$ the product $SS^{-1}=k\cdot e+\sum_{g\in G-{e}}a_g\cdot g$ where $a_g \in \mathbb{Z}$. Then $$\begin{split}
\displaystyle \sum_{i=0}^{v-1}\chi_i(SS^{-1}) &= \sum_{i=0}^{v-1}\chi_i(ke+\sum_{g\in G-{e}}a_gg)\\
&= \sum_{i=0}^{v-1}k\cdot \chi_i(e)+\sum_{g\in G-{e}}a_g\cdot \sum_{i=0}^{v-1}\chi_i(g)\\
&=vk
\end{split}$$ This holds since for any $g\in G-{e}$ we have $\sum_{i=0}^{v-1}\chi_i(g)=0$ and $\sum_{i=0}^{v-1}\chi_i(e)=v.$
Using the previous lemma for $S=D_0,D_1$ separately, we obtain $$\begin{split}
\sum_{a\in \mathbb{F}_{3^n}}\widehat{F_1}(a)\widehat{F_1}(-a)&=3\sum_{a\in \mathbb{F}_{3^n}}\left( \chi_a(D_0)\chi_a(D_0)- \chi_a(D_1)\chi_a(D_2)\right)\\
&=3\sum_{a\in \mathbb{F}_{3^n}} \chi_a(D_0D_0^{-1})- 3\sum_{a\in \mathbb{F}_{3^n}}\chi_a(D_1D_1^{-1})\\
&=3\cdot 3^{n}\cdot 3^{n-1}-3\cdot 3^{n}\cdot 3^{n-1}\\
&=0
\end{split}$$ On the other hand, using the values from the table, we can also write $$\sum_{a\in \mathbb{F}_{3^n}}\widehat{F_1}(a)\widehat{F_1}(-a)= 9(\Lambda_1 +\Lambda_2)C^2$$ where $\Lambda_1=|\{a \in {{\mathbb F}}_{3^n}^*:(x_a,y_a)=(C,0)\}|, \Lambda_2=|\{a \in {{\mathbb F}}_{3^n}^*:(x_a,y_a)=(-C,0)\}|$. This implies that for any $a\in {{\mathbb F}}_{3^n}^*$, $(x_a,y_a) \ne (C,0), (-C,0)$ and hence $D_0$ is also a PGDS.
By Theorem \[cross-plateaued\] we have PGDS with parameters $(v=p^{2n}, k=p^n, \alpha=p^{n}-p^s, \beta=p^{s+n}+p^{n}-p^s)$. If $p=3$ then by Theorem \[partitioned\] we also have PGDS with parameters $(v=3^{n}, k=3^{n-1}, \alpha=3^{2n-3}-3^{n-2}, \beta=3^{n-1}+3^{2n-3}-3^{n-2})$. Here we also note that not all decimation will lead such a partition of the finite field $ {{\mathbb F}}_{3^n}$. For instance, let $D_i=\{x: F(x)=Tr_n(x^d)=i\}$ for $d=3^{2}-3+1$. Then computational results imply that none of the $D_i$’s is a partial geometric difference set in ${{\mathbb F}}_{3^5}$. In general, it is a challenging task to characterize all functions which can be used to obtain a partition of a group into partial geometric difference sets.
If there is such a partition we can define a set of complex vectors $$z_{a}=(\chi_a(D_0),\chi_a(D_1),\chi_a(D_2))$$ for any $a\in \mathbb{F}_{3^n}$. Let $$e=(1,\zeta_3, \zeta_3^2).$$ Then norm of the complex inner product of any vector $z_{a}$ and $e$ is either $0$ or $C$ for some integer $C$.
Let $D_0$, $D_1$ and $D_2$ be a partition of $\mathbb{F}_{3^n}$ and $\lambda=3^{\frac{n+s-1}{2}}$ be an integer. Suppose for $i=0,1,2$ each $D_i$ is a partial geometric difference set such that $\chi_a(D_i) \in \{0, \pm \lambda,\pm \lambda \zeta_3,\pm \lambda \zeta_3^2\}$ for each nonprincipal character $\chi_a$. If one of the cases holds
- $|D_0|=|D_1|=|D_2|$,
- $|D_i|=|D_j|=3^{n-1}-3^{\frac{n+s-2}{2}}$ and $|D_k|=3^{n-1}+2.3^{\frac{n+s-2}{2}}$,
- $|D_i|=|D_j|=3^{n-1}+3^{\frac{n+s-2}{2}}$ and $|D_k|=3^{n-1}-2.3^{\frac{n+s-2}{2}}$.
and $|<z_a,e>|^2$ is either $0$ or $3\lambda^2$ then $$f(x)=\begin{cases}
0, & x \in D_0\\
1, & x\in D_1\\
2, & x\in D_2\\
\end{cases}$$ is an $s$-plateaued function.
The Walsh transform $$\widehat{f}(0)=\sum_{x \in \mathbb{F}_{p^{n}}} \zeta_3^{f(x)}=|D_0|+\zeta_3|D_1|+\zeta_3^2|D_2|.$$ With the conditions given in the theorem, we obtain $$\widehat{f}(0)=\begin{cases}
0 & \text{if}\; |D_0|=|D_1|=|D_2|,\\
3^{\frac{n+s}{2}} & \text{if}\;|D_i|=|D_j|=3^{n-1}-3^{\frac{n+s-2}{2}}, |D_k|=3^{n-1}+2.3^{\frac{n+s-2}{2}},\\
-3^{\frac{n+s}{2}} & \text{if}\; |D_i|=|D_j|=3^{n-1}+3^{\frac{n+s-2}{2}}, |D_k|=3^{n-1}-2.3^{\frac{n+s-2}{2}}.
\end{cases}$$ For each $a\in \mathbb{F}_{3^n}^*$, the Walsh transform of $f(x)$ is $$\begin{aligned}
\widehat{f}(a)&=\sum_{x \in \mathbb{F}_{p^{n}}} \zeta_3^{f(x)-Tr_n(ax)}\\
&=\sum_{x \in D_0} \zeta_3^{Tr_n(-ax)} + \zeta_3\sum_{x \in D_1} \zeta_3^{Tr_n(-ax)} + \zeta_3^2\sum_{x \in D_2} \zeta_3^{Tr_n(-ax)}\\
&=\chi_{-a}(D_0)+\zeta_3 \chi_{-a}(D_1) + \zeta_3^2\chi_{-a}(D_2).
\end{aligned}$$ Then with the assumptions of the theorem and after some easy calculations one can show that $|\widehat{f}(a)|^2 \in \{0,3\lambda^2\}$.
Results on p-ary Functions
==========================
In this section we will develop some tools to characterize $p-$ary $s-$plateaued functions. Our results are mimicking the results of [@olmez2].
\[Matrix-Eq\] Let $f$ be a function from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p}$ and $M=(m_{x,y})$ be a $p^n \times p^n$ matrix where $m_{x,y}=\zeta_p^{f(x+y)}$. Then, $f$ is an $s$-plateaued function if and only if $$MM^*M=p^{n+s}M \label{Mfequation}$$ where $M^*$ is the adjoint of the matrix $M$.
Suppose $f$ is an plateaued function with $|\widehat{f}(x)| \in \{0, p^{(n+s)/2}\}$ for all $x \in \mathbb{F}_{p^{n}}$. Then, $$\begin{split}
(MM^*M)_{x,y}&=\sum_{z\in\mathbb{F}_{p^{n}}} \left(\sum_{c\in \mathbb{F}_{p^{n}}} m_{x,c}\overline{m_{z,c}}\right)m_{z,y} \\
&=\sum_{z\in \mathbb{F}_{p^{n}}} \left(\sum_{c\in \mathbb{F}_{p^{n}}} F(x+c)\overline{F(z+c)}\right)F(z+y) \\
&=\sum_{c\in \mathbb{F}_{p^{n}}} F(x+c)\left(\sum_{w\in \mathbb{F}_{p^{n}} } \overline{F(w)}F(w+y-c)\right)\\
&=\sum_{c\in \mathbb{F}_{p^{n}}} (\overline{F}*F)(c-y)F(x+c) \\
&=\sum_{u\in \mathbb{F}_{p^{n}}} (F*\overline{F})(u-x-y)F(u) \\
&=((F*\overline{F})*F)(x+y).
\end{split}$$ Let $A=(F*\overline{F})*F$. Then, the Fourier transform of $A$ is $\widehat{A}=\widehat{F}\cdot\widehat{\overline{F}}\cdot\widehat{F}$. Now by Fourier inversion $$\begin{split}
A(x+y)&=\frac{1}{p^{n}}\sum_{\beta \in \mathbb{F}_{p^{n}}} \widehat{F}(\beta)\widehat{\overline{F}}(\beta)\widehat{F}(\beta)\zeta_p^{Tr((x+y)\beta)}\\
&=p^{n+s}\frac{1}{p^{n}}\sum_{\beta \in \mathbb{F}_{p^{n}}} \widehat{F}(\beta)\zeta_p^{Tr((x+y) \beta)}\\
&=p^{n+s}F(x+y).\\
\end{split}$$ Hence the equation holds.
Suppose $MM^*M=p^{n+s}M$. This implies $((F*\overline{F})*F)(x)=p^{n+s}F(x)$ for all $x \in \mathbb{F}_{p^n}$. Apply the Fourier transform on both of the sides. Then, $$\widehat{F}(x)(\widehat{F}(x)\cdot \overline{\widehat{F}(x)}-p^{n+s})=0$$ for all $x \in \mathbb{F}_{p^n}$. Hence, $|\widehat{F}(x)| \in \{0, p^{(n+s)/2}\}$ for all $x \in \mathbb{F}_{p^{n}}.$
An $n \times n$ complex matrix $M$ is called a Butson-Hadamard matrix if $$MM^* =nI_n.$$ It is easy to see that a $q^n \times q^n $ Butson-Hadamard matrix $M$ also satisfies $$MM^*M=q^nM.$$ Our result implies that $M$ can be associated with $0-$plateaued function. This indicates the well-known connection between Butson-Hadamard matrices and bent functions.
As a corollary of Lemma \[Matrix-Eq\], we can characterize $s$-plateaued functions with their first and second derivatives. First and second derivative of a $p-$ary function is defined by $$D_af(x)=f(x+a)-f(x)$$ and $$D_aD_bf(x)=f(x+a+b)+f(x)-f(x+a)-f(x+b)$$ respectively.
$f$ is an $s$-plateaued function from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p}$ if and only if the expression $\sum_{a,b\in \mathbb{F}_{p^{n}}} \zeta_p ^{D_aD_b f(u)}$ does not depend on $u \in \mathbb{F}_{p^{n}} $. This constant expression equals to $p^{n+s}$.
Since the equation $$MM^*M=p^{n+s}M$$ holds, $$M^*MM^*=p^{n+s}M^*$$ holds too. Fix two non-zero elements $x$ and $y$ of $\mathbb{F}_{p^{n}}$ and let $u=x+y$. $$\begin{split}
\sum_{z\in\mathbb{F}_{p^{n}}} \left(\sum_{c\in \mathbb{F}_{p^{n}}} \overline{m_{x,c}}m_{z,c}\right)\overline{m_{z,y}} &=\sum_{z\in \mathbb{F}_{p^{n}}} \left(\sum_{c\in \mathbb{F}_{p^{n}}} \zeta_p^{-f(x+c)}\zeta_p^{f(z+c)}\right)\zeta_p^{-f(z+y)} \\
&=\sum_{c,z\in \mathbb{F}_{p^{n}}} \zeta_p ^{-f(x+c)+f(z+c)-f(z+y)}\\
&=p^{n+s}\zeta_p ^{-f(x+y)}
\end{split}$$ Now let $z=a+x$ and $c=b+y$. Then $$p^{n+s}=\sum_{a,b\in \mathbb{F}_{p^{n}}} \zeta_p ^{-f(x+y+b)+f(a+b+x+y)-f(a+x+y)+f(x+y)}$$ holds. Thus, $$p^{n+s}=\sum_{a,b\in \mathbb{F}_{p^{n}}} \zeta_p ^{D_aD_b f(u)}.$$
Let $\Delta_f(a)=\sum_{x \in \mathbb{F}_{p^n}} \zeta_p^{D_af(x)}$. $f$ is an $s$-plateaued function from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p}$ if and only if $\sum_{a\in \mathbb{F}_{p^{n}}} \overline{\Delta_f(a)}\Delta_f(a)=p^{2n+s}$.
We have $$MM^*MM^*=p^{n+s}MM^*.$$Let $$N=MM^*.$$ Then $$\begin{split}
(N)_{0,a}&=\sum_{x\in \mathbb{F}_{p^{n}}} \zeta_p^{f(x)-f(a+x)} \\
&=\overline{\Delta_f(a)}\\
\end{split}$$ and $$\begin{split}
(N)_{a,0}&=\sum_{x\in \mathbb{F}_{p^{n}}} \zeta_p^{f(a+x)-f(x)} \\
&=\Delta_f(a)\\
\end{split}$$ Therefore $$\begin{split}
(N^2)_{0,0}&=\sum_{a\in \mathbb{F}_{p^{n}}} \overline{\Delta_f(a)}\Delta_f(a)\\
&=p^{n+s}(N)_{0,0}\\
&=p^{2n+s}\\
\end{split}$$
Now we will use our characterization to provide a simple construction of $s$-plateaued functions. If $A$ is an $m \times n$ matrix and $B$ is an $s \times t$ matrix, then the Kronecker product $A \otimes B$ is the $ms \times nt$ block matrix:
$${\displaystyle \mathbf {A} \otimes \mathbf {B} ={\begin{bmatrix}a_{11}\mathbf {B} &\cdots &a_{1n}\mathbf {B} \\\vdots &\ddots &\vdots \\a_{m1}\mathbf {B} &\cdots &a_{mn}\mathbf {B} \end{bmatrix}},}$$
Let $f: \mathbb{F}_{p^{n}} \to \mathbb{F}_{p} $ and $g : \mathbb{F}_{p^{m}} \to \mathbb{F}_{p}$ be $s_1$-plateaued and $s_2$-plateaued functions respectively. Let $M$ and $N$ be the matrices whose entries determined by $m_{x,y}=\zeta_p^{f(x+y)}$ and $n_{a,b}=\zeta_p^{g(a+b)}$. Let $P=M \otimes N$ be the Kronecker product of $M$ and $N$. Then $$PP^*P=p^{n+m+s_1+s_2}P$$ holds.
Let $P=M \otimes N$. Then $$\begin{split}
PP^*P&=\left( M \otimes N\right) \left( M^* \otimes N^*\right)\left( M \otimes N\right) \\
&=\left( MM^*M\right) \otimes \left( NN^*N \right)\\
&=\left( p^{n+s_1}M\right) \otimes \left( p^{m+s_2}N \right)\\
&=p^{n+m+s_1+s_2} \left( M\otimes N\right) \\
&=p^{n+m+s_1+s_2}P\\
\end{split}$$
Let $f: \mathbb{F}_{p^{n}} \to \mathbb{F}_{p} $ and $g : \mathbb{F}_{p^{m}} \to \mathbb{F}_{p}$ be two $s_1$-plateaued and $s_2$-plateaued functions respectively. Then there exists an $(s_1+s_2)$-plateaued function from $\mathbb{F}_{p^{n+m}}$ to $\mathbb{F}_{p}$.
Let $M$ and $N$ be matrices associated with $f$ and $g$ such that $m_{x,y}=\zeta_p^{f(x+y)}$ and $n_{a,b}=\zeta_p^{g(a+b)}$. Let $P=M \otimes N$. We need to show that there exist a function $h$ such that the entries of $P$ can be associated with $h$.
We will index the rows and columns of $P$ by the elements of $\mathbb{F}_{p^{n+m}}$. First note that there is a subgroup $H$ of the additive group of $\mathbb{F}_{p^{n+m}}$ which is isomorphic to the additive group of $\mathbb{F}_{p^{m}}$. Let us fix a transversal $T=\{\gamma_1,\gamma_2,\dots, \gamma_{p^n} \}$ of this subgroup in $\mathbb{F}_{p^{n+m}}$. Now order the rows and columns of $P$ by $\gamma_1+H$, $\gamma_2+H$, ..., $\gamma_{p^n}+H$ in the block form. Here we have isomorphisms namely $$\phi_1: H \to \mathbb{F}_{p^{m}}$$ and $$\phi_2: \{\gamma_1+H,\gamma_2+H,\dots, \gamma_{p^n}+H \} \to \mathbb{F}_{p^{n}}$$
If $x,y \in \mathbb{F}_{p^{n+m}}$ then $x=\gamma_i+u$ and $y=\gamma_j+v$ for unique elements $u,v \in H$. Now let us examine the $xy$-th entry of $P$. $$P_{x,y}=\zeta_p^{f(\phi_2(\gamma_i+H)+\phi_2(\gamma_j+H))} \cdot \zeta_p^{g(\phi_1(u)+\phi_1(v))}= \zeta_p^{h(x+y)}$$ where $h$ is the desired $(s_1+s_2)$-plateaued function from $\mathbb{F}_{p^{n+m}}$ to $\mathbb{F}_{p}.$ Moreover if $y=0$ then $\gamma_j=v=0$. Thus $$h(x)= f(\phi_2(\gamma_i+H))+g(\phi_1(u)).$$
Partially bent functions
-------------------------
This section is devoted to investigate a family of $s-$plateaued functions known as partially bent functions. A $p-$ary function is called partially bent if the derivative $D_af$ is either balanced or constant for any $a \in \mathbb{F}_{p^n}$. Here we will provide some characterization via their associated designs.
Let $f$ be an $s$-plateaued function and for $a \in \mathbb{F}_{p^{n}} $ define the set $$T_a=\{x+a,f(x)+f(a): x \in \mathbb{F}_{p^{n}} \}.$$ Fix an element $a$ of $\mathbb{F}_{p^{n}} $, then the graph of $f$ can be also written as $$G_f=\{(x,f(x)): x \in \mathbb{F}_{p^{n}}\}=\{(x+a,f(x+a)): x \in \mathbb{F}_{p^{n}}\}.$$
Let $f$ be an $s$-plateaued function with $f(0)=0$. If $f$ has a linear structure $\Lambda$ then $T_a=G_f$ for all $a \in \Lambda$.
Let $f$ be an $s$-plateaued function with $f(0)=0$ and linear structure $\Lambda$ of dimension $m$. Then the incidence matrix $A$ of the design associated with the partial geometric difference set $G_f$ can be written as a Kronecker product of $1\times p^m$ all-ones matrix $j$ and an incidence matrix $N$ of a partial geometric design.
Let $j$ be the $1\times p^m$ all-ones matrix. Let $D$ be the block design associated with $G_f$ where the point set is $\mathbb{F}_{p^{n}} \times \mathbb{F}_{p} $ and the blocks are the translates of the graph of $f$. Suppose $\textbf{B}$ is a block in $D$. Note that $\textbf{B}=(u,v)+G_f$ for some $(u,v) \in \mathbb{F}_{p^{n}} \times \mathbb{F}_{p}.$ Then for each $a\in \Lambda$ we have $$\textbf{B}+(a,f(a))=\textbf{B}$$ since $$(u,v)+G_f=\{(x+u+a,f(x+a)+v): x \in \mathbb{F}_{p^{n}}\}.$$ Thus each block is repeated $p^m$ many times. Therefore, $A=j\otimes N$ for some incidence matrix $N$. Now we are going to show that $N$ is an incidence matrix of a partial geometric design. Since $A$ is an incidence matrix of a partial geometric design, $$\begin{split}
AA^tA&=(\beta-\alpha) j \otimes N+\alpha J_{p^n \times p^n}\\
&=j \otimes (\beta-\alpha) N+\alpha j \otimes J_{p^n \times p^{n-m}}\\
&= j\otimes (\beta-\alpha) N+\alpha J_{p^n \times p^{n-m}}.\\
\end{split}$$ We also have
$$\begin{split}
AA^tA&=jj^tj\otimes NN^tN\\
&=p^m j\otimes NN^tN\\
&=j\otimes p^m NN^tN.\\
\end{split}$$ By comparing the left hand sides we can conclude that the equation $$NN^tN=\frac{(\beta-\alpha)}{p^m} N+\frac{\alpha}{p^m} J_{p^n \times p^{n-m}}$$ holds.
Let $f$ be an $s$-plateaued function from $\mathbb{F}_{p^{n}}$ to $\mathbb{F}_{p}$. Then the set $D_f$ is a PGDS with parameters $(p^{n+1},p^n ; p^{2n-1}-p^{n+s-1}, p^{2n-1}-p^{n+s-1}+p^{n+s})$. If $f$ is a partially bent function then there is an integer $s\geq 0$ such that $f$ is $s$-plateaued and the linear space of $f$ has dimension $s$. Our observation yields the following result.
If $f$ is partially bent function then $f$ can be associated with a partial geometric design with parameters $$v=p^{n+1}, b=p^{n+1-s}, k=p^n, r=p^{n-s}, \alpha= p^{2n-1-s}-p^{n-1}, \beta=p^n+p^{2n-1-s}-p^{n-1}.$$
[1]{}
E. F. Assmus and J. D. Key. Designs and their Codes. Cambridge University Press, No. 103. (1992).
A. Bernasconi, B. Codenotti, and J. M. Vanderkam. A characterization of bent functions in terms of strongly regular graphs. IEEE Transactions on Computers, Sep 1(9): pp. 984-5 (2001).
T. Beth, D. Jungnickel, and H. Lenz. Design Theory, Cambridge University Press, second edition, (1999).
A. E. Brouwer, O. Olmez and S. Y. Song. Directed strongly regular graphs from $1\frac{1}{2}$-designs. European Journal of Combinatorics, 33(6): pp. 1174 -1177 (2012).
M. Buratti, Y. Wei, D. Wu, P. Fan, and M. Cheng: Relative difference families with variable block sizes and their related OOCs. IEEE Transactions on Information Theory, 57: pp. 7489-7497 (2011).
C. Carlet and E. Prouff. On plateaued functions and their constructions. In Fast Software Encryption, Springer: pp. 54-73 (2003).
A. Canteaut, C. Carlet, P. Charpin, C. Fontaine. On cryptographic properties of the cosets of $r (1, m)$. IEEE Transactions on Information Theory 47(4), pp. 1494-1513 (2001).
C. Carlet. Partially-bent functions. Designs, Codes and Cryptography 3(2), pp. 135-145 (1993).
C. Carlet. Boolean functions for cryptography and error correcting codes.Boolean models and methods in mathematics, computer science, and engineering 2, pp. 257-397 (2010).
C. Carlet. Boolean and vectorial plateaued functions and APN functions. IEEE Transactions on Information Theory 61(11), pp. 6272-6289 (2015).
C. Carlet and S. Mesnager. Four decades of research on bent functions. Designs, Codes and Cryptography 78(1), 5-50 (2016).
C. Carlet, S. Mesnager, F. Özbudak, A. S[i]{}nak. Explicit characterizations for plateauedness of $p-$ary (vectorial) functions. In: Second International Conference on Codes, Cryptology and Information Security (C2SI-2017), In Honor of Claude Carlet. pp. 328-345 (2017).
A. Çeşmelioğlu, W. Meidl, A. Topuzoğlu. Partially bent functions and their properties. In: Larcher, G., Pillichshammer, F., Winterhof, A., Xing, C. (eds.) Applications of Algebra and Number Theory, pp. 22-40. Cambridge University Press, Cambridge (2014).
R.S. Coulter, R.W. Matthews. Bent polynomials over finite fields. Bulletin of the Australian Mathematical Society 56(3), pp. 429-437 (1997).
F. R. K. Chung, J. A. Salehi, and V. K. Wei. Optical orthogonal codes: design, analysis, and applications. IEEE Transactions on Information Theory, 35: pp. 595-604 (1989).
Ph. Delsarte. Weights of linear codes and strongly regular normed spaces. Discrete Mathematics, 3(1): pp. 47-64 (1972).
J. F. Dillon. Elementary Hadamard difference sets. PhD thesis, University of Maryland, (1974).
T. W. Cusick, C. Ding and A. Renvall. Stream Ciphers and Number Theory. North-Holland Mathematical Library, The Netherlands: North-Holland/Elsevier, Amsterdam, vol. 55. (1998).
C. Ding. Linear codes from some 2-designs. IEEE Transactions on Information Theory, 61(6): pp. 3265-3275 (2015).
C. Ding. Codes from difference sets. World Scientific Publishing Company, (2015).
T. Helleseth. Some results about the cross-correlation function between two maximal linear sequences, Discrete Mathematics 16, pp. 209-232 (1976).
S. Mesnager. Characterizations of plateaued and bent functions in characteristic p. In: International Conference on Sequences and Their Applications. pp. 72-82 (2014).
S. Mesnager. Bent functions: Fundamentals and Results. Switzerland, Springer, pp. 1-544 (2016).
S. Mesnager, F. Özbudak, A. S[i]{}nak. Results on characterizations of plateaued functions in arbitrary characteristic. Cryptography and information security in the Balkans, BalkanCryptSec 2015, Koper, Slovenia, Revised Selected Papers. In: Pasalic E., Knudsen L.R. (eds.) LNCS 9540, pp. 17-30. Springer, Berlin (2016)
S. Mesnager, F. Özbudak, A. S[i]{}nak. On the $p$-ary (cubic) bent and plateaued (vectorial) functions. Des. Codes Cryptogr., vol 86 (8), pp 1865-1892 (2018).
O. Olmez. Symmetric $1\frac12$-designs and $1\frac12$-difference sets. J. Combin. Designs. vol. 22, No. 6, pp. 252-269, (2014).
O. Olmez. Plateaued functions and one-and-half difference sets. Designs, Codes and Cryptography, Sep 1 vol. 76, No. 3, pp. 537-49, (2015).
O. Olmez. A link between combinatorial designs and three-weight linear codes. Designs, Codes and Cryptography, 86(4), pp. 817-833 (2018).
A. Pott. Finite Geometry and Character Theory, Springer, 1995.
A. Pott. Nonlinear functions in abelian groups and realtive difference sets. Discrete Applied Mathematics, 138 (1-2), pp.177-193 (2004).
A. Pott, Y. Tan, T. Feng, and S. Ling. Association schemes arising from bent functions. Designs, Codes and Cryptography, 59(1), pp. 319-331 (2011).
O. S. Rothaus. On bent functions. Journal of Combinatorial Theory, Series A, 20(3), pp. 300-305 (1976).
H. M. Trachtenberg. On the cross-correlation functions of maximal linear sequences. Ph. D Dissertation, University of Southern California, Los Angeles, CA (1970).
Y. Tan, A. Pott, and T. Feng. Strongly regular graphs associated with ternary bent functions. Journal of Combinatorial Theory, Series A, 117(6), pp. 668-682 (2010).
Y. Zheng, X. M. Zhang. Plateaued functions. In: ICICS. vol. 99, pp. 284-300 (1999).
[^1]: Oktay Olmez’s research was supported by TUBITAK Research Grant Proj. No. 115F064.
|
---
abstract: 'We discuss that observational constraints on neutrino cooling processes may restrict the spectrum of quark matter phases admissible for compact star interiors.'
author:
- |
D. Blaschke$^{1,2}$ and J. Berdermann$^3$\
\
$^1$Institute for Theoretical Physics, University of Wroclaw, Poland\
$^2$Bogoliubov Laboratory for Theoretical Physics, JINR Dubna, Russia\
$^3$DESY, Platanenallee 6, D-15738 Zeuthen, Germany
title: Neutrinos in dense quark matter and cooling of compact stars
---
Introduction
============
The existence of quark matter in compact star (CS) interiors is debated controversially since observations of the thermal emission from objects like RXJ 1856.5-3754 [@Trumper:2003we] indicate large masses and/or luminosity radii requiring a stiff equation of state (EoS) at high densities. Although this seemingly excludes a phase transition accompanied with a softening of the EoS, it has been demonstrated [@Alford:2006vz] that modern QCD motivated theories could fulfill such observational constraints, e.g., due to the presence of vector mean fields which stiffen the quark matter EoS [@Klahn:2006iw]. Investigations of color superconducting quark matter phases (for a recent review, see [@Alford:2007xm]) show that large diquark pairing gaps of $\sim 10 - 100$ MeV are possible, which lower the critical density for the chiral symmetry restoration due to their competition with the chiral condensate. This entails early quark deconfinement unless color forces remain strong enough to confine even (almost) massless quarks in a new phase of chirally symmetric hadronic matter, the hypothetic quarkyonic phase [@McLerran:2007qj; @Andronic:2009gj]. Here, we will not yet consider quarkyonic matter, but discuss the viability of three-flavor (CFL), two-flavor (2SC) and one-flavor (d-CSL) color superconducting phases under constraints for mass and radius as well as CS cooling.
Spin-0 phases which pair quarks of different flavor and opposite spins with large gaps (2SC and CFL) might be too fragile to withstand the stress of flavor asymmetry and strong magnetic fields in neutron stars. It is not clarified yet whether the occurrence of the celebrated CFL phase in the core of a hybrid star leads necessarily to a gravitational instability [@Klahn:2006iw] as a result of the strong softening of the EoS (for an exception, see [@Pagliara:2007ph]). Disregarding the hadronic shell even interesting mass twin sequences are possible [@Sandin:2007zr]. If hybrid stars with CFL quark matter cores may form a third CS family, this might allow for a new paradigm to explain the Janus-faced CS phenomenology: large luminosity radii vs. millisecond rotation periods, fast cooling vs. slow cooling etc.
After the new color superconductivity phases with large Spin-0 pairing gaps (2SC and CFL) were suggested in 1997, a discussion of late time cooling scenarios has been performed in [@Blaschke:1999qx; @Page:2000wt], followed by a full transport calculation with nontrivial temperature profile evolution in hybrid stars with CSC quark matter cores [@Blaschke:2000dy]. These calculations showed that the pairing pattern where all quarks are paired with large gaps delay the cooling and may be excluded by the observation of fast coolers like the Vela pulsar, unless they form a third family.
In these works as well as in the detailed study in Ref. [@Jaikumar:2005hy] the fact was ignored that in the 2SC phase the quarks with one of the colors (e.g., the blue ones) remain unpaired so that the fast direct Urca (DU) process is operative, entailing too fast cooling. For heuristic purposes, a residual single-color pairing of the blue quarks (X-gap) has been introduced [@Grigorian:2004jq] and it has been found that the best cooling phenomenology would be obtained for hybrid stars with superconducting quark matter cores in the 2SC+X phase when this smallest gap is in the range between 100 keV and 1 MeV, with a decreasing density dependence [@Blaschke:2005dc; @Popov:2005xa]. An alternative to the postulated 2SC+X phase which provides a microscopic approach to the pairing pattern in accordance with cooling phenomenology is the isotropic color-spin-locking (CSL) phase [@Aguilera:2005tg] for which quarks of the same flavor are paired in a spin-1 state. This phase is thus rather inert against the neutron star stress.
Recently, it has been argued that the different quark flavors could occur sequentially, i.e. at different threshold densities in neutron star matter [@Blaschke:2008br; @Blaschke:2008vh]. In analogy to the neutron drip in the crust, a down-quark drip density can occur in the core, when d-quarks undergo a chiral restoration and partial nucleon dissociation leads to the formation of a superconducting single-flavor subphase (d-CSL) immersed in nuclear matter. This mixture of d-CSL and nuclear matter can extend from the interior to the crust core boundary in stable hybrid star configurations which fulfill the stringent mass-radius constraints mentioned above. It also bears interesting perspectives for the neutrino transport and cooling properties [@Blaschke:2008br; @Blaschke:2007bv] which we discuss now a bit more in detail.
Neutrino emissivity and deep crustal heating
============================================
According to [@Stejner:2006tj], two interesting phenomena, the superbursters and the soft X-ray transients may be probe the existence of CS deep crustal heating processes subject to the following constraints\
1. A thin baryonic crust of the star with a width between 100 to 400 m.\
2. An energy release of 1 to 100 MeV per accreted nucleon by conversion at the crust-core boundary.\
3. The thermal conductivity has to be in the range of $10^{19}-10^{22}~{\rm erg~cm^{-1}~s^{-1}~K^{-1}}$.\
4. The fast DU neutrino emissivity process should be strongly suppressed or not operational.\
Following the suggestion of Refs. [@Page:2005ky; @Cumming:2005kk], the conversion of nuclear matter to CFL strange quark matter, powered by continuous accretion in a LMXB may provide a mechanism for the deep crustal heating. As the nuclei penetrate into the quark matter core a conversion energy between 1 to 100 MeV is released per accreted nucleon. This heats the core until an equilibrium temperature is reached and the heating is balanced by neutrino emission. As the DU neutrino emission process is strongly suppressed by large CFL gaps, the heat produced at the crust-core boundary is not radiated away but conducted so that the fusion ignition condition gets fulfilled.\
In [@Blaschke:2008br] we have shown that a d-quark CSL/nuclear phase mixture can also fulfill the above constraints on the deep crustal heating mechanism. In particular, we estimated the heat per accreted nucleon from the d-quark drip due to partial chiral restoration to the order of $10$ MeV and the DU process which we exclude in the nuclear subphase does also not occur in d-CSL subphase since the Fermi sea of up quarks is closed. An interesting question for the d-CSL/nuclear matter phase to be clarified is the possible role of confining interactions between the colored d-quarks which would energetically forbid a rather dilute d-quark admixture and sets a threshold for the d-quark drip. It may well be that the proper account for confining interactions within a chiral model theory would lead to a modification of the nucleon properties rather than to their partial dissociation at the chiral restoration for the down quarks. A modified quarkyonic phase for isospin asymmetric matter could be suggested.
Perspectives
============
Hybrid Stars with a mixture of nuclear matter and d-CSL quark matter phases fulfill not only stringent constraints for large radii and masses but provide also a viable deep crustal heating process. This mixed phase should probably better described as a kind of quarkyonic matter, once this gets accessible to a theoretical description. Further investigation needs the idea that CFL quark core stars form a third CS family providing a possible explanation to a different class of CS phenomena as, e.g., GRBs [@Ouyed:2005tm; @Berdermann:2006rk], SGRs, AXPs and XDINs [@Niebergal:2009yb].
Acknowledgements {#acknowledgements .unnumbered}
----------------
We are grateful to our colleagues for discussions and collaboration. D.B. acknowledges an EPS Fellowship, and partial support by MNiSW grant No. N N 202 231837, by RFBR grant No. 08-02-01003-a and by CompStar, a Research Networking Programme of the European Science Foundation.
[99]{}
-2pt
J. E. Trümper, V. Burwitz, F. Haberl and V. E. Zavlin, Nucl. Phys. Proc. Suppl. [**132**]{}, 560 (2004). M. Alford, D. Blaschke, A. Drago, T. Klähn, G. Pagliara and J. Schaffner-Bielich, Nature [**445**]{}, E7 (2007) \[arXiv:astro-ph/0606524\]. T. Klähn, D. Blaschke, F. Sandin, Ch. Fuchs, A. Faessler, H. Grigorian, G. Röpke and J. Trümper, [Phys. Lett. B]{} [**654**]{}, 170 (2006). M. G. Alford, A. Schmitt, K. Rajagopal and T. Schäfer, Rev. Mod. Phys. [**80**]{}, 1455 (2008). L. McLerran and R. D. Pisarski, Nucl. Phys. A [**796**]{}, 83 (2007). A. Andronic [*et al.*]{}, arXiv:0911.4806 \[hep-ph\]. G. Pagliara and J. Schaffner-Bielich, Phys. Rev. D [**77**]{}, 063004 (2008). F. Sandin and D. Blaschke, Phys. Rev. D [**75**]{}, 125013 (2007). D. Blaschke, T. Klähn and D. N. Voskresensky, Astrophys. J. [**533**]{}, 406 (2000). D. Page, M. Prakash, J. M. Lattimer and A. Steiner, Phys. Rev. Lett. [**85**]{}, 2048 (2000). D. Blaschke, H. Grigorian and D. N. Voskresensky, Astron. Astrophys. [**368**]{}, 561 (2001). P. Jaikumar, C. D. Roberts and A. Sedrakian, Phys. Rev. C [**73**]{}, 042801 (2006). H. Grigorian, D. Blaschke and D. Voskresensky, Phys. Rev. C [**71**]{}, 045801 (2005). D. Blaschke, D. N. Voskresensky and H. Grigorian, Nucl. Phys. A [**774**]{}, 815 (2006). S. Popov, H. Grigorian and D. Blaschke, Phys. Rev. C [**74**]{}, 025803 (2006). D. N. Aguilera, D. Blaschke, M. Buballa and V. L. Yudichev, Phys. Rev. D [**72**]{}, 034008 (2005). D. Blaschke, F. Sandin, T. Klähn and J. Berdermann, Phys. Rev. [**C**]{} (in press), arXiv:0807.0414 \[nucl-th\]; AIP Conf. Proc. [**1038**]{}, 183 (2008). D. Blaschke, F. Sandin and T. Klähn, J. Phys. G [**35**]{}, 104077 (2008). D. B. Blaschke and J. Berdermann, AIP Conf. Proc. [**964**]{}, 290 (2007). M. Stejner and J. Madsen, [Astron. Astrophys.]{} [**458**]{}, 523 (2006). D. Page and A. Cumming, [Astrophys. J.]{} [**635**]{}, L157 (2005). A. Cumming, J. Macbeth, J. J. M. in ’t Zand and D. Page, Astrophys. J. [**646**]{}, 429 (2006). R. Ouyed, R. Rapp and C. Vogt, Astrophys. J. [**632**]{}, 1001 (2005) \[arXiv:astro-ph/0503357\]. J. Berdermann et al., Prog. Part. Nucl. Phys. [**57**]{} (2006) 334. B. Niebergal, R. Ouyed, R. Negreiros and F. Weber, arXiv:0906.3043 \[astro-ph.HE\].
|
---
author:
- Alain Dresse
- |
Marc Henneaux[Also at Centro de Estudios Científicos de Santiago, Casilla 16443, Santiago 9, Chile]{}\
Faculté des Sciences, Université Libre de Bruxelles,\
Campus Plaine C.P. 231, B-1050 Bruxelles (Belgium)
title: BRST Structure of Polynomial Poisson Algebras
---
\#1\#2[[\#1\#2]{}]{} =0 ${\global\advance\parenthesis by1\left(}
\def$[by-1)]{} $${\relax} \def$$ \#1 \#1[0=\#1sp0 by 30]{}
Introduction
============
Polynomial algebras with a Lie bracket fulfilling the derivation property $$[f g, h] = f [g,h] + [f,h]g$$ are called polynomial Poisson algebras and play an increasingly important role in various areas of theoretical physics [@Nak:; @Pri:; @Skl:; @Zam:; @FatZam:; @Oh:; @TarTakFad:; @BakMat:; @BhaRam:; @GraZhe:]. In terms of a set of independent generators $G_a$, $a = 1, \ldots, n$, the brackets are given by $$\label{basic_bracket}
[G_a, G_b] = C_{ab}(G)$$ where $C_{ab} = - C_{ba}$ are polynomials in the $G$’s[^1]. If the polynomials $C_{ab}(G)$ vanish when the $G$’s are set equal to zero, i.e. if they have no constant part, the polynomial algebra is said to be first class, in analogy with the terminology for constrained Hamiltonian systems (see e.g. [@HenTei:QuaGauSys]). An important class of first class Poisson algebras are symmetric algebras over a finite dimensional Lie algebra. In that case, the bracket (\[basic\_bracket\]) belongs to the linear span of the $G_a$’s, i.e. the $C_{ab}(G)$ are homogeneous of degree one in the $G$’s, $[G_a, G_b] = C_{ab}{}^c G_c$. We shall call this situation the “Lie algebra case”, and refer to the non Lie algebra case as the “open algebra case” using again terminology from the theory of first class constrained systems [@HenTei:QuaGauSys][^2].
The purpose of this paper is to investigate the BRST structure of first class Poisson algebras. The BRST formalism has turned out recently to be the arena of a fruitful interplay between physics and mathematics (see e.g. [@HenTei:QuaGauSys] and references therein). A crucial ingredient of BRST theory is the recursive pattern of homological perturbation theory [@Sta:] which allows one to construct the BRST generator step by step. In most applications, however, this recursive construction collapses almost immediately, and, to our knowledge, no example has been given so far for which the full BRST machinery is required (apart from the field-theoretical membrane models [@Hen:PhyLet; @FujKub:]). We show in this paper that Poisson algebras—actually, already quadratic Poisson algebras—offer splendid examples illustrating the complexity of the BRST construction. While Lie algebras yield a BRST generator of rank 1 (see e.g. [@HenTei:QuaGauSys]), the BRST charge for quadratic Poisson algebras can be of arbitrarily high rank. We also point out that BRST concepts provide intrinsic characterizations of Poisson algebras.
In the next section, we briefly review the BRST construction. We then discuss how it applies to Poisson algebras, even when the generators $G_a$ are not realized as phase space functions of some dynamical system. We analyze the BRST cohomology and introduce the concepts of covariant and minimal ranks, for which an elementary theorem is proven. Quadratic algebras are then shown to provide models with arbitrarily high rank. These contain “self-reproducing” algebras for which the bracket of $G_a$ with $G_b$ is proportional to the product $G_a G_b$. The first few terms in the BRST generator are also computed for more general algebras by means of a program written in REDUCE. The paper ends with some concluding remarks on the quantum case.
A Brief Survey of the BRST Formalism
====================================
We follow the presentation of [@HenTei:QuaGauSys], to which we refer for details and proofs. Given a set of independent functions $G_a(q,p)$ defined in some phase space $P$ with local coordinates $(q^i, p_i)$ and fulfilling the first class property $[G_a, G_b] \approx 0$, where $\approx$ denotes equality on the surface $G_a(q,p) = 0$, one can introduce an odd generator $\Omega$ (“the BRST generator”) in an extended phase space containing further fermionic conjugate pairs $(\eta^a, {{\cal P}}_a)$ (the “ghost pairs”) which has the following properties : $$\begin{aligned}
[\Omega, \Omega] &=& 0 \label{nilpotency} \\
\Omega &=& G_a \eta^a + \mbox{``more''}.\end{aligned}$$ Here, “more” stands for terms containing at least one ghost momentum ${{\cal P}}_a$. We take the ghosts $\eta^a$ to be real and their momenta imaginary, with graded Poisson bracket $$[{{\cal P}}_a, \eta^b] = - \delta_a{}^b$$
The BRST derivation $s$ in the extended phase space is generated by $\Omega$, $$s \bullet = [ \bullet, \Omega]$$ and is a differential ($s^2 = 0$) because of (\[nilpotency\]). One also introduces a grading, the “ghost number” by setting $${\mbox{gh}}\eta^a = - {\mbox{gh}}{{\cal P}}_a = 1, \quad {\mbox{gh}}q^i = {\mbox{gh}}p_i = 0.$$ The ghost number of the BRST generator is equal to 1.
The BRST generator $\Omega$ is constructed recursively as follows. One sets $$\Omega = {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}} + {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ \Omega \end{array}}} + \cdots$$ where ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}}$ contains $k$ ghost momenta. One has ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}} = G_a \eta^a$. The nilpotency condition becomes, in terms of ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}}$, $$\label{delta-om=d}
\delta {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+1)} \\ \Omega \end{array}}} + {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}} = 0$$ where ${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}}$ involves only the lower order ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (s)} \\ \Omega \end{array}}}$ with $s \leq p$ and is defined by $$\label{d-p}
{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}} = 1 / 2 \left[
\sum^p_{k=0} [{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} , {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p-k)} \\ \Omega \end{array}}}]_{\mbox{orig}} +
\sum^{p-1}_{k=0} [ {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k+1)} \\ \Omega \end{array}}} ,
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p-k)} \\ \Omega \end{array}}}]_{{{\cal P}}, \eta}
\right].$$
Here, the bracket $[\; , \;]_{\mbox{orig}}$ refers to the Poisson bracket in the original phase space, which only acts on the $q^i$ and $p_i$, and not on the ghosts, whereas $[\; , \;]_{{{\cal P}}, \eta}$ refers to the Poisson bracket acting only on the ghost and ghost momenta arguments and not on the original phase space variables. The “Koszul” differential $\delta$ in (\[delta-om=d\]) is defined by $$\label{koszul}
\delta q^i = \delta p_i = 0, \quad \delta \eta^a = 0, \quad \delta
{{\cal P}}_a = - G_a$$ and is extended to arbitrary functions on the extended phase space as a derivation. One easily verifies that $\delta^2 = 0$.
Given ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (s)} \\ \Omega \end{array}}}$ with $s \leq p$, one solves (\[delta-om=d\]) for ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+1)} \\ \Omega \end{array}}}$. This can always be done because $\delta
{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}} = 0$, and because $\delta$ is acyclic in positive degree. One then goes on to ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+2)} \\ \Omega \end{array}}}$ etc... until one reaches the complete expression for $\Omega$. The last function ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}}$ that can be non zero is ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (n-1)} \\ \Omega \end{array}}}$ where $n$ is the number of constraints. Indeed, the product $\eta^{a_1} \cdots \eta^{a_n} \eta^{a_{n+1}}$ of $n+1$ anticommuting ghost variables in ${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (n)} \\ \Omega \end{array}}$ is zero. The function ${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+1)} \\ \Omega \end{array}}$ is determined by (\[delta-om=d\]) up to a $\delta$-exact term. This amounts to making a canonical transformation in the extended phase space.
First Class Polynomial Poisson algebras
=======================================
The standard BRST construction recalled in the previous section assumes that the $G_a$’s are realized as functions on some phase space, and allows the $C^c{}_{ab}$ in $$[G_a, G_b] = C^c{}_{ab} G_c$$ to be functions of $q^i$ and $p_i$. However, when the $C^c{}_{ab}$’s depend on the $q$’s and $p$’s only through the $G_a$’s themselves, as is the case when the $G_a$’s form a first class polynomial Poisson algebra, one can define the BRST generator directly in the algebra $\Bbb{C}\,({{\cal P}}_a) \otimes \Bbb{C}\,(G_a) \otimes \Bbb{C}\,(\eta^a)$ of polynomials in the $G$’s, the $\eta$’s and the ${{\cal P}}$’s without any reference to the explicit realization of the $G$’s as phase space functions[^3]. That is, the BRST generator can be associated with the Poisson algebra itself.
The reason for which this can be done is that both the Koszul differential $\delta$ defined by (\[koszul\]) [*and*]{} the ${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}}$ in (\[d-p\]) involve only $G_a$ and not $q^i$ or $p_i$ individually. Thus, ${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+1)} \\ \Omega \end{array}}$ can be taken to depend only on $G_a$. The BRST generator is defined accordingly in the algebra $\Bbb{C}\,({{\cal P}}_a)
\otimes \Bbb{C}\,(G_a) \otimes \Bbb{C}\,(\eta_a)$.
One can give an explicit solution of (\[delta-om=d\]) in terms of the homotopy $\sigma$ defined on the generators by $$\sigma G_a = - {{\cal P}}_a, \quad \sigma {{\cal P}}_a = \sigma G_a = 0$$ and extended to the algebra $\Bbb{C}\,({{\cal P}}_a) \otimes \Bbb{C}\,(G_a) \otimes
\Bbb{C}\,(\eta_a)$ as a derivation, $$\label{sigma}
\sigma = - {{\cal P}}_a \frac{\partial}{\partial G_a}.$$ One has $$\sigma \delta + \delta \sigma = N$$ where $N$ counts the degree in the $G$’s and the ${{\cal P}}$’s. Hence, if ${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D_m \end{array}}$ is the term of degree $m$ in $(G, {{\cal P}})$ of ${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}}$, a solution of (\[delta-om=d\]) is given by $$\label{om-p}
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+1)} \\ \Omega \end{array}}} = - \sum_m 1/m \left( \sigma {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}}_m
\right)$$ since $\delta {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}}=0$ [@HenTei:QuaGauSys] and $m >
0$ (one has $m \geq p$ and for $p=0$, $m \geq 1$ because $[{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}},
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}}]$ contains $G_a$ by the first class property).
It should be stressed that the partial derivations $\partial/\partial
G_a$ in(\[sigma\]) are well defined because the functions on which they act depend only on on $G_a$. For an arbitrary function of $q^i,
p_i$, $\partial F / \partial G_a$ would not be well defined even if the constraints $G_a$ are independent (i.e. irreducible) as here. One must specify what is kept fixed. For example, if there is one constraint $p_1 = 0$ on the four-dimensional phase space $(q^1, p_1),
(q^2, p_2)$, then $\partial p_2 / \partial p_1 = 0$ if one keeps $q^1,
q^2$ and $p_2$ fixed, but $\partial p_2 / \partial p_1 = 1$ if one keeps $q^1, q^2$ and $p_2 - p_1$ fixed. Note that the subsequent developments require only that the $C^a{}_{b c}$ be functions of the $G_a$, but not that these functions be polynomials. We consider here the polynomial case for the sole sake of simplicity.
As mentioned earlier, the solution (\[om-p\]) of the equation (\[delta-om=d\]) is not unique. We call it the “covariant solution” because the homotopy $\sigma$ defined by (\[sigma\]) is invariant under linear redefinitions of the generators.
[**Example:**]{} for a Lie algebra $$[G_a, G_b] = C^c{}_{ab} G_c$$ the covariant BRST generator is given by $$\label{L-A-omega}
\Omega = G_a \eta^a - 1/2 {{\cal P}}_a C^a{}_{bc} \eta^c \eta^b.$$ Its nilpotency expresses the Jacobi identity for the structure constants $C^a{}_{bc}$. One has ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ \Omega \end{array}}} = 0$ for $p \geq
2$.
In general, the BRST generator $\Omega$ for a generic Poisson algebra contains higher order terms whose calculation may be quite cumbersome. However, because the procedure is purely algorithmic, it can be performed by means of an algebraic program like REDUCE.
The cohomology of the Poisson algebra may be defined to be the cohomology of the BRST differential $s$ in the algebra $\Bbb{C}\,({{\cal P}}_a)
\otimes \Bbb{C}\,(G_a) \otimes \Bbb{C}\,(\eta_a)$. Because $s$ contains $\delta$ as its piece of lowest antighost number (with ${\mbox{antigh}}({{\cal P}}_a) = 1, {\mbox{antigh}}(\mbox{anything else}) = 0$), and because $\delta$ provides a resolution of the zero-dimensional point $G_a =
0$, standard arguments show that the cohomology of $s$ is isomorphic to the cohomology of the differential $s'$ in $\Bbb{C}\,(\eta^a)$, $$\label{eq:sPrime}
s' \eta^a = 1/2 C^a{}_{bc} \eta^b \eta^c$$ where $C^a{}_{bc}$ is defined by $$C^a{}_{bc} = \left.\frac{\partial C_{bc}}{\partial G_a}\right|_{G = 0}$$
The $C^a{}_{bc}$ fulfill the Jacobi identity so that $s'^2 = 0$. Hence, they are the structure constants of a Lie algebra, which is called the Lie algebra underlying the given Poisson algebra.
Because of (\[eq:sPrime\]), the BRST cohomology of a Poisson algebra is isomorphic to the cohomology of the underlying Lie algebra. For a different and more thorough treatment of Poisson cohomology, see [@Hue:].
Rank
====
Again in analogy with the terminology used in the theory of constrained systems, we shall call [*“covariant rank”*]{} of a first class polynomial Poisson algebra the degree in ${{\cal P}}_a$ of the covariant BRST generator. This concept is invariant under linear redefinitions of the generators because the covariant BRST generator is itself invariant if one transforms the ghosts and their momenta as $$\begin{aligned}
G_a &\rightarrow& \bar{G}_a = A_a{}^b G_b \\
{{\cal P}}_a &\rightarrow& \bar{{{\cal P}}}_a = A_a{}^b {{\cal P}}_b \\
\eta^a &\rightarrow& \bar{\eta}^a = (A^{-1})_b{}^a \eta^b\end{aligned}$$
We shall call [*“minimal rank”*]{} the degree in ${{\cal P}}_a$ of the solution of $[ \Omega, \Omega] = 0$ of lowest degree in ${{\cal P}}$ (i.e., one chooses at each stage ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+1)} \\ \Omega \end{array}}}$ in such a way that $\Omega$ has lowest possible degree in ${{\cal P}}$). It is easy to see that for a Lie algebra, the concepts of covariant and minimal ranks coincide. As we shall see on an explicit example below, they do not in the general case.
Now, for a Lie algebra, the rank is not particularly interesting in the sense that it does not tell much about the structure of the algebra : the rank of a Lie algebra is equal to zero if and only if the algebra is abelian. It is equal to one otherwise. For non linear Poisson algebras, the rank is more useful. All values of the rank compatible with the trivial inequality $$rank \leq n-1$$ may occur. Thus, the rank of the BRST generator provides a non trivial characterization of Poisson algebras. Conversely, non linear Poisson algebras yield an interesting illustration of the full BRST machinery where higher order terms besides ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ \Omega \end{array}}}$ are required in $\Omega$ to achieve nilpotency.
Upper bound on the rank
=======================
One can understand the fact that the rank of a Lie algebra is at most equal to one by introducing a degree in $\Bbb{C}\,({{\cal P}}_a) \otimes
\Bbb{C}\,(G_a) \otimes \Bbb{C}\,(\eta_a)$ different from the ghost degree as follows.
Assume that one can assign a “degree” $n_a \geq 1$ to the generators $G_a$ in such a way that the bracket decreases the degree by at least one, $$\label{deg-g=n}
\deg G_a = n_a, \; \deg([G_a, G_b]) \leq n_a + n_b - 1.$$ Then, one can bound the covariant and minimal ranks of the algebra by $\sum_a (n_a - 1) + 1$, $$r \leq \sum_a (n_a - 1) + 1$$
In the case of a Lie algebra, one can take $n_a = 1$ for all the generators since $\deg([G_a, G_b])$ is then equal to one and fulfills (\[deg-g=n\]). The theorem then states that the rank is bounded by one, in agreement with (\[L-A-omega\]).
Assign the following degrees to $\eta^a$ and ${{\cal P}}_a$, $$\deg \eta^a = - n_a + 1, \; \deg {{\cal P}}_a = n_a - 1$$ If $\delta A = B$ and $\deg B = b$, then $\deg A = b - 1$ since $\delta$ increases the degree by one. Now ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}} = G_a
\eta^a$ is of degree one. It follows that $[{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}},
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}}] = [{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}},{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}}]_{\mbox{orig}}$ is of degree $\leq 1$ and hence, by (\[delta-om=d\]) and (\[d-p\]), $\deg {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ \Omega \end{array}}
\leq 0$. More generally, one has $\deg {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} \leq -k + 1$. Indeed, if this relation is true up to order $k-1$, then it is also true at order $k$ because in $$\delta {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} \sim [{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (r)} \\ \Omega \end{array}}},{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (s)} \\ \Omega \end{array}}}]_{\mbox{orig}} + [{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (r')} \\ \Omega \end{array}}},
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (s')} \\ \Omega \end{array}}}]_{{{\cal P}}, \eta}$$ ($r+s = k-1, \; r' + s' = k$), the right hand side is of degree $\leq
-k+2$. Thus $\deg {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} \leq -k + 2 - 1 = -k + 1$.
But the element with most negative degree in the algebra is given by the product of all the $\eta$’s, which has degree $-\sum_a(n_a - 1)$. Accordingly, ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}}$ is zero whenever $-k+1 >= - \sum_a(n_a - 1)$, which implies $r \leq \sum_a(n_a-1)+1$ as stated in the theorem.
[**Remarks:**]{}
1. One can improve greatly the bound by observing that the $\eta$’s do not come alone in ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}}$. There are also $k$ momenta ${{\cal P}}_a$ which carry positive degree. This remark will, however, not be pursued further here.
2. One can actually assign degrees smaller than one to the generators $G_a$. For instance, in the case of an Abelian Lie algebra, one may take $deg G_a = 1/2, \; \deg \eta^a = 1/2, \deg {{\cal P}}_a = -
1/2$. Because the degree of a ghost number one object is necessarily greater than or equal to $1/2$, the condition $\deg {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} \leq -k+1$ (if ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} \neq 0$) implies ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} = 0$ for $k > 0$.
Self-reproducing algebras
=========================
While Lie algebras are characterized by the existence of a degreee that is decreased by the bracket, one may easily construct examples of Poisson algebras for which such a degree does not exist. The simplest ones are quadratic algebras for which $[G_a, G_b]$ is proportional to $G_a, G_b$ $$[G_a, G_b] = M_{ab} G_a G_b \quad\quad\mbox{no summation on $a,b$}$$ with $M_{ab} = -M_{ba}$. The Jacobi identity is fulfilled for arbitrary $M$’s. Since $\deg(G_a G_b) = n_a + n_b$, the inequality (\[deg-g=n\]) is violated for any choice of $n_a$. Because $[G_a, G_b]$ is proportional to $G_a G_b$, we shall call these algebras “self-reproducing algebras”.
The most general self-reproducing algebra with three generators is given by $$\begin{aligned}
[G_1, G_2] &=& \alpha\, G_1 G_2 \\{}
[G_2, G_3] &=& \beta \, G_2 G_3 \\{}
[G_3, G_1] &=& \gamma \, G_1 G_3.\end{aligned}$$ This Poisson algebra can be realized on a six-dimensional phase space by setting $$G_1 = \exp(p_2 + \alpha q_3), G_2 = \exp(p_3 + \beta q_1), G_3 =
\exp(p_1 + \gamma q_2).$$ The covariant BRST charge for this model is equal to $$\begin{aligned}
\Omega &=&\eta^1 \, G_1 + \eta^2 \, G_2 + \eta^3 \, G_3 + \\
\nonumber
& & 1/2 \,
(\alpha \,\eta^{2}\,\eta^{1}\,{{\cal P}}_{2}\,G_{1}
-\alpha \,\eta^{2}\,\eta^{1}\,{{\cal P}}_{1}\,G_{2}
-\beta \,\eta^{3}\,\eta^{2}\,{{\cal P}}_{3}\,G_{2} \\ \nonumber
& &\mbox{~~~~~~}
-\beta \,\eta^{3}\,\eta^{2}\,{{\cal P}}_{2}\,G_{3}
+\gamma \,\eta^{3}\,\eta^{1}\,{{\cal P}}_{3}\,G_{1}
+\gamma \,\eta^{3}\,\eta^{1}\,{{\cal P}}_{1}\,G_{3}) + \\ \nonumber
& & 1/12 \, (
( - \alpha \,\beta +2\,\alpha \,\gamma -\beta \,\gamma )
\,\eta^{3}\,\eta^{2}\,\eta^{1}\,{{\cal P}}_{3}\,{{\cal P}}_{2}\,G_{1} + \\
\nonumber & &\mbox{~~~~~~~~~}
( -2\,\alpha \,\beta +\alpha \,\gamma +\beta \,\gamma )
\,\eta^{3}\,\eta^{2}\,\eta^{1}\,{{\cal P}}_{3}\,{{\cal P}}_{1}\,G_{2} +\\
\nonumber & &\mbox{~~~~~~~~~}
( -\alpha \,\beta -\alpha \,\gamma +2\,\beta \,\gamma )
\,\eta^{3}\,\eta^{2}\,\eta^{1}\,{{\cal P}}_{2}\,{{\cal P}}_{1}\,G_{3}
)\end{aligned}$$ and is of rank 2 (the maximum possible rank) unless $\alpha = \beta =
\gamma$, or $\alpha = \beta = 0$, $\gamma \neq 0$, in which case it is of rank 1.
Examples
========
We now give the BRST charge (or the first terms of the BRST charge) for some particular Poisson algebras. The examples have been treated using REDUCE, using the treatment of summation over dummy indices developed in [@Dre:CanExp; @Dre:Imacs]. Details of the implementation of the BRST algorithm can be found in [@BurCapDre:]. All dummy variables are noted as $d_i$ where $i$ is an integer. Unless stated otherwise, there is an implicit summation on all dummy variables. For the examples in which the Jacobi identity is not trivially satisfied, the expressions have been normalized so that no combinations of terms in a polynomial belongs to the polynomial ideal generated by the left hand side of the Jacobi identity. In particular, polynomials in this ideal are represented by identically null expressions.
Self-Reproducing Algebras
-------------------------
As we have just defined, the basic Poisson brackets for the generators $G_d$ of the [*self-reproducing algebra*]{} are given by $$[G_{d_1}, G_{d_2}] = M_{d_1 d_2} G_{d_1} G_{d_2}$$ without summation over the dummy variables $d_1$ and $d_2$. The matrix $M$ is antisymmetric, but otherwise arbitrary.
The seven first orders of the covariant BRST charge are given by $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}}=
\[G_{d_{1}}\,\eta^{d_{1}}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ \Omega \end{array}}}=
\[\frac{G_{d_{1}}\,M_{d_{1}d_{2}}\,\eta^{d_{1}}\,
\eta^{d_{2}}\,{{\cal P}}_{d_{2}}}{
2}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (2)} \\ \Omega \end{array}}}=
\[\frac{-
\(G_{d_{1}}\,M_{d_{1}d_{2}}\,\eta^{d_{1}}\,
\eta^{d_{2}}\,\eta^{d_{3}}\,{{\cal P}}_{d_{2}}\,
{{\cal P}}_{d_{3}}\,
\(M_{d_{1}d_{3}}
+M_{d_{2}d_{3}}
\)
\)
}{
12}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (3)} \\ \Omega \end{array}}}=
\[\frac{-
\(G_{d_{1}}\,M_{d_{1}d_{2}}\,M_{d_{1}d_{4}}\,
M_{d_{2}d_{3}}\,\eta^{d_{1}}\,\eta^{d_{2}}\,
\eta^{d_{3}}\,\eta^{d_{4}}\,{{\cal P}}_{d_{2}}\,
{{\cal P}}_{d_{3}}\,{{\cal P}}_{d_{4}}
\)
}{
24}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (4)} \\ \Omega \end{array}}}=
\[\(G_{d_{1}}\,\eta^{d_{1}}\,\eta^{d_{2}}\,
\eta^{d_{3}}\,\eta^{d_{4}}\,\eta^{d_{5}}\,
{{\cal P}}_{d_{2}}\,{{\cal P}}_{d_{3}}\,{{\cal P}}_{d_{4}}\,
{{\cal P}}_{d_{5}}\,\nl
\off{3499956}
\(-
\(M_{d_{1}d_{2}}\,M_{d_{1}d_{3}}\,M_{d_{1}d_{4}}\,
M_{d_{1}d_{5}}
\)
+4\,M_{d_{1}d_{2}}\,M_{d_{1}d_{4}}\,M_{d_{1}d_{5}}\,
M_{d_{2}d_{3}}\nl
\off{3827636}
+2\,M_{d_{1}d_{2}}\,M_{d_{1}d_{4}}\,M_{d_{2}d_{3}}\,
M_{d_{4}d_{5}}
+M_{d_{1}d_{2}}\,M_{d_{1}d_{5}}\,M_{d_{2}d_{3}}\nl
\off{3827636}
\,M_{d_{2}d_{4}}
-M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{2}d_{5}}
-M_{d_{1}d_{4}}\,M_{d_{1}d_{5}}\,\nl
\off{3827636}
M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}
-2\,M_{d_{1}d_{4}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{4}d_{5}}
+M_{d_{1}d_{5}}\,\nl
\off{3827636}
M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,M_{d_{2}d_{5}}
-M_{d_{1}d_{5}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{4}d_{5}}
\)
\)
/720
\]
\Nl}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (5)} \\ \Omega \end{array}}}=
\[\(G_{d_{2}}\,M_{d_{1}d_{2}}\,M_{d_{3}d_{4}}\,
\eta^{d_{1}}\,\eta^{d_{2}}\,\eta^{d_{3}}\,
\eta^{d_{4}}\,\eta^{d_{5}}\,\eta^{d_{6}}\,
{{\cal P}}_{d_{1}}\,{{\cal P}}_{d_{3}}\,{{\cal P}}_{d_{4}}\,
{{\cal P}}_{d_{5}}\,{{\cal P}}_{d_{6}}\nl
\off{3499956}
\,
\(-
\(M_{d_{2}d_{3}}\,M_{d_{2}d_{5}}\,M_{d_{2}d_{6}}
\)
+2\,M_{d_{2}d_{3}}\,M_{d_{2}d_{5}}\,M_{d_{5}d_{6}}
+M_{d_{2}d_{3}}\nl
\off{4100703}
\,M_{d_{2}d_{6}}\,M_{d_{3}d_{5}}
-M_{d_{2}d_{3}}\,M_{d_{3}d_{5}}\,M_{d_{3}d_{6}}
-M_{d_{2}d_{5}}\nl
\off{4100703}
\,M_{d_{2}d_{6}}\,M_{d_{3}d_{5}}
-2\,M_{d_{2}d_{5}}\,M_{d_{3}d_{5}}\,M_{d_{5}d_{6}}\nl
\off{4100703}
+M_{d_{2}d_{6}}\,M_{d_{3}d_{5}}\,M_{d_{3}d_{6}}
-M_{d_{2}d_{6}}\,M_{d_{3}d_{5}}\,M_{d_{5}d_{6}}
\)
\)
/1440
\]
\Nl}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (6)} \\ \Omega \end{array}}}=
\[\(G_{d_{3}}\,\eta^{d_{1}}\,\eta^{d_{2}}\,
\eta^{d_{3}}\,\eta^{d_{4}}\,\eta^{d_{5}}\,
\eta^{d_{6}}\,\eta^{d_{7}}\,{{\cal P}}_{d_{1}}\,
{{\cal P}}_{d_{2}}\,{{\cal P}}_{d_{4}}\,{{\cal P}}_{d_{5}}\,
{{\cal P}}_{d_{6}}\,{{\cal P}}_{d_{7}}\,\nl
\off{3499956}
\(-
\(M_{d_{1}d_{2}}\,M_{d_{1}d_{3}}\,M_{d_{1}d_{6}}\,
M_{d_{2}d_{4}}\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}
\)
-M_{d_{1}d_{2}}\,M_{d_{1}d_{3}}\nl
\off{3827636}
\,M_{d_{2}d_{4}}\,M_{d_{2}d_{6}}\,M_{d_{4}d_{5}}\,
M_{d_{6}d_{7}}
+2\,M_{d_{1}d_{2}}\,M_{d_{1}d_{3}}\,M_{d_{2}d_{4}}\,
\nl
\off{3827636}
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}
+M_{d_{1}d_{2}}\,M_{d_{1}d_{3}}\,M_{d_{2}d_{6}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\nl
\off{3827636}
\,M_{d_{6}d_{7}}
-2\,M_{d_{1}d_{2}}\,M_{d_{1}d_{3}}\,M_{d_{2}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
+2\,M_{d_{1}d_{2}}\,M_{d_{1}d_{3}}\,M_{d_{2}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
-2\,M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{2}d_{6}}\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
-13\,M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,
M_{d_{2}d_{4}}\,M_{d_{3}d_{6}}\,M_{d_{4}d_{5}}\,
M_{d_{6}d_{7}}\nl
\off{3827636}
+4\,M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
+2\,M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{6}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
-4\,M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
+4\,M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
-2\,M_{d_{1}d_{2}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{4}}\,
M_{d_{3}d_{6}}\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
+M_{d_{1}d_{3}}\,M_{d_{1}d_{4}}\,M_{d_{1}d_{6}}\,
M_{d_{2}d_{6}}\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}
-2\,M_{d_{1}d_{3}}\nl
\off{3827636}
\,M_{d_{1}d_{4}}\,M_{d_{2}d_{4}}\,M_{d_{4}d_{5}}\,
M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}
+5\,M_{d_{1}d_{3}}\,M_{d_{1}d_{4}}\,\nl
\off{3827636}
M_{d_{2}d_{7}}\,M_{d_{3}d_{7}}\,M_{d_{4}d_{5}}\,
M_{d_{4}d_{6}}
+M_{d_{1}d_{3}}\,M_{d_{1}d_{6}}\,M_{d_{2}d_{3}}\,
M_{d_{2}d_{4}}\nl
\off{3827636}
\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}
-2\,M_{d_{1}d_{3}}\,M_{d_{1}d_{6}}\,M_{d_{2}d_{6}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\nl
\off{3827636}
\,M_{d_{6}d_{7}}
-5\,M_{d_{1}d_{3}}\,M_{d_{1}d_{6}}\,M_{d_{2}d_{7}}\,
M_{d_{3}d_{7}}\,M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\nl
\off{3827636}
+M_{d_{1}d_{3}}\,M_{d_{1}d_{7}}\,M_{d_{2}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}
-M_{d_{1}d_{3}}\,M_{d_{1}d_{7}}\nl
\off{3827636}
\,M_{d_{2}d_{7}}\,M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,
M_{d_{6}d_{7}}
-M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{2}d_{6}}\,\nl
\off{3827636}
M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}
-8\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{3}d_{6}}\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
+2\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{4}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
+M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{2}d_{6}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}
-2\,M_{d_{1}d_{3}}\nl
\off{3827636}
\,M_{d_{2}d_{3}}\,M_{d_{2}d_{7}}\,M_{d_{4}d_{5}}\,
M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}
+2\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\nl
\off{3827636}
\,M_{d_{2}d_{7}}\,M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,
M_{d_{6}d_{7}}
-2\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{4}}\,
\nl
\off{3827636}
M_{d_{3}d_{5}}\,M_{d_{3}d_{6}}\,M_{d_{3}d_{7}}
+12\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,
M_{d_{3}d_{4}}\,M_{d_{3}d_{6}}\nl
\off{3827636}
\,M_{d_{3}d_{7}}\,M_{d_{4}d_{5}}
-6\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{4}}\,
M_{d_{3}d_{6}}\,M_{d_{4}d_{5}}\nl
\off{3827636}
\,M_{d_{6}d_{7}}
-5\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{4}}\,
M_{d_{3}d_{7}}\,M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\nl
\off{3827636}
+5\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{4}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
+5\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{6}}\,
M_{d_{3}d_{7}}\,M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\nl
\off{3827636}
+13\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,
M_{d_{3}d_{6}}\,M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,
M_{d_{6}d_{7}}\nl
\off{3827636}
-5\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
+5\,M_{d_{1}d_{3}}\,M_{d_{2}d_{3}}\,M_{d_{3}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
+2\,M_{d_{1}d_{3}}\,M_{d_{2}d_{4}}\,M_{d_{3}d_{4}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
-8\,M_{d_{1}d_{3}}\,M_{d_{2}d_{6}}\,M_{d_{3}d_{4}}\,
M_{d_{3}d_{6}}\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
+2\,M_{d_{1}d_{3}}\,M_{d_{2}d_{6}}\,M_{d_{3}d_{6}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}\nl
\off{3827636}
-M_{d_{1}d_{3}}\,M_{d_{2}d_{7}}\,M_{d_{3}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}\nl
\off{3827636}
+M_{d_{1}d_{3}}\,M_{d_{2}d_{7}}\,M_{d_{3}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}
-2\,M_{d_{1}d_{4}}\,\nl
\off{3827636}
M_{d_{2}d_{4}}\,M_{d_{3}d_{4}}\,M_{d_{4}d_{5}}\,
M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}
+M_{d_{1}d_{4}}\,M_{d_{2}d_{6}}\,M_{d_{3}d_{4}}\nl
\off{3827636}
\,M_{d_{3}d_{6}}\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}
+5\,M_{d_{1}d_{6}}\,M_{d_{2}d_{6}}\,M_{d_{3}d_{4}}\,
M_{d_{3}d_{6}}\nl
\off{3827636}
\,M_{d_{4}d_{5}}\,M_{d_{6}d_{7}}
-8\,M_{d_{1}d_{6}}\,M_{d_{2}d_{6}}\,M_{d_{3}d_{6}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\nl
\off{3827636}
\,M_{d_{6}d_{7}}
-2\,M_{d_{1}d_{7}}\,M_{d_{2}d_{7}}\,M_{d_{3}d_{6}}\,
M_{d_{3}d_{7}}\,M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\nl
\off{3827636}
+6\,M_{d_{1}d_{7}}\,M_{d_{2}d_{7}}\,M_{d_{3}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{4}d_{7}}
-6\,\nl
\off{3827636}
M_{d_{1}d_{7}}\,M_{d_{2}d_{7}}\,M_{d_{3}d_{7}}\,
M_{d_{4}d_{5}}\,M_{d_{4}d_{6}}\,M_{d_{6}d_{7}}
\)
\)
/60480
\]
\Nl}$$
These expressions are not particularly illuminating but are of interest because they generically do not vanish and hence, define higher order BRST charges. This can be seen by means of the following example, in which only the brackets of the first generator with the other ones are non vanishing, and taken equal to $$\begin{aligned}
\label{eq:maxquad}
[G_1, G_{\alpha}] &=& G_1 G_{\alpha} = - [ G_{\alpha}, G_1]
\quad (\alpha = 2,3,\ldots,n),\\{}
[G_{\alpha}, G_{\beta}] &=& 0.\end{aligned}$$ For this particular self-reproducing algebra, all orders of the covariant BRST charge can be explicitly computed. One finds $$\begin{aligned}
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}} &=& \eta^a G_a \\
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}} &=& \alpha_k ({\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ T_1 \end{array}} + (-)^{k+1} {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ T_2 \end{array}})\end{aligned}$$ where $$\begin{aligned}
{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ T_1 \end{array}} &=& G_1 \eta^{\alpha_1} \cdots \eta^{\alpha_k} \eta^1
{{\cal P}}_{\alpha_1} \cdots {{\cal P}}_{\alpha_k} \\
{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ T_2 \end{array}} &=& G_{\alpha_k} \eta^{\alpha_1} \cdots \eta^{\alpha_k} \eta^1
{{\cal P}}_1 {{\cal P}}_{\alpha_1} \cdots {{\cal P}}_{\alpha_{k-1}}\end{aligned}$$ and $$\begin{aligned}
\alpha_1 &=& -1/2, \; \alpha_2 = -1/12, \; \alpha_3 = 0 \\
\alpha_k &=& -\frac{1}{k+1} \sum_{l=1}^{k-3} \alpha_{l+1} \alpha_{k-l-1}
\quad \mbox{for $k > 3$}\label{eq:AlpRec}.\end{aligned}$$
This can be seen as the only non zero brackets involved in the construction of the BRST charge are $$\begin{aligned}
[{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ T_1 \end{array}}, {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}}]_{\mbox{orig}} &=& (-)^{k+1}{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ S \end{array}} \\{}
[{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ T_1 \end{array}}, {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (l)} \\ T_2 \end{array}}]_{\eta{{\cal P}}} &=&
(-)^{l(k+1)}{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k+l-1)} \\ S \end{array}},\end{aligned}$$ where $${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ S \end{array}} = G_1 G_\alpha \eta^{\alpha_1} \cdots \eta^{\alpha_k} \eta^\alpha
\eta^1
{{\cal P}}_{\alpha_1} \cdots {{\cal P}}{\alpha_k}.$$ We further have $$\label{eq:sigmaS}
\sigma {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ S \end{array}} = (-)^{k+1} {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k+1)} \\ T_2 \end{array}} - {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k+1)} \\ T_1 \end{array}}$$ Given these relations, it is straightforward to verify (39). First, one easily checks that (39) is correct for $k=1$ with $\alpha_1$ equal to 1/2. Let us then assume that (39) is true for $k = 0, 1 ...$ up to $p$. One then obtains $${\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ D \end{array}} = \beta_p {\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p)} \\ S \end{array}}$$ with $\beta_p$ given by $$\beta_p = (-)^{p+1} \alpha_p - \sum_{k=0}^{p-1} (-)^{p(k+1)} \alpha_{k+1}
\alpha_{p-k} = - \sum_{k=1}^{p-2} (-)^{p(k+1)} \alpha_{k+1} \alpha_{p-k}$$ from which one gets, using (\[eq:sigmaS\]), that ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (p+1)} \\ \Omega \end{array}}}$ is indeed given by (39) with $\alpha_{p+1}$ equal to $$\alpha_{p+1} = \frac{\beta_p}{p+2}$$ Observe now that $\alpha_k = 0$ for $k$ odd, $k \neq 1$. This can again be shown by recurrence. First note that $\alpha_3$ = 0. Now let $p$ be even, $p > 3$. Suppose $\alpha_k = 0$ for $k$ odd, $1 < k < p$. All terms in the relation defining $\beta_{p}$ are proportional to an $\alpha_m$ with $m$ odd, $1 < m < p$, since $k+1$ and $p-k$ have opposite parities. Therefore, $\beta_p = 0 = \alpha_{p+1}$ and thus $\alpha_k
= 0$ for $k$ odd, $k > 1$. Accordingly only $\alpha_k $ with $k$ even can be different from zero. The expression for $\alpha_k$ reduces then to (43) since $k+1$ must be even in (49).
Although $\alpha_k = 0$ for $k$ odd, $k > 1$, one easily sees that $\alpha_k < 0$ for $k$ even. This is true for $k =
2$ as $\alpha_2 = -1/12$. Let $p$ be even, and suppose $\alpha_m < 0$ for $1 < m < p$, $m$ even. Then, all terms in the sum in the recurrence relation (\[eq:AlpRec\]) are strictly positive, so that $\alpha_p < 0$. Since $\alpha_p \neq 0$, the quadratic algebra (\[eq:maxquad\]) provides examples of systems with arbitrarily high covariant rank.
Note also that the minimal rank is equal to one: indeed, the non covariant BRST charge given by $$\tilde{\Omega}={{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}} + {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ \Omega \end{array}}} + ({{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ T \end{array}}}_1 - {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ T \end{array}}}_2)/2 =
\eta^a G_a - G_a \eta^a \eta^1 {{\cal P}}_1$$ is nilpotent and $$\delta({{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ T \end{array}}}_1 - {{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ T \end{array}}}_2) = 0$$ so $\tilde{\Omega}$ is indeed a valid BRST charge. This shows that the minimal and covariant ranks are in general different.
Finally, it is easy to modify slightly the basic brackets so as to induce non zero covariant ${{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (k)} \\ \Omega \end{array}}}$ with $k$ odd. One simply replaces (\[eq:maxquad\]) by $$\begin{aligned}
[G_{n-1}, G_n] &=& G_{n-1} G_n = - [G_n, G_{n-1}] \\{}
[G_1, G_n] &=& - G_1 G_n, \\{}
[G_1, G_{\alpha}] &=& G_1 G_{\alpha} \quad (\alpha \neq n).\end{aligned}$$
Purely Quadratic Algebras
-------------------------
A generalization of the above is the pure quadratic algebra. The basic Poisson brackets are then given by $$[G_{d_1}, G_{d_2}] = D_{d_1 d_2}^{d_3 d_4} G_{d_3} G_{d_4}$$ where $D_{d_1 d_2}^{d_3 d_4}$ is antisymmetric in $d_1, d_2$ and symmetric in $d_3, d_4$. The Jacobi identity implies that $$D_{d_4 d_1}^{d_5 d_6} D_{d_2 d_3}^{d_4 d_7} + \mbox{symm}(d_5, d_6,
d_7) + \mbox{cyclic}(d_1,d_2,d_3) = 0$$
The first orders of the covariant BRST charge are given by $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (0)} \\ \Omega \end{array}}}=
\[G_{d_{1}}\,\eta^{d_{1}}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (1)} \\ \Omega \end{array}}}=
\[\frac{D_{d_{1}d_{2}}^{
d_{4}d_{3}}\,G_{d_{4}}\,\eta^{d_{1}}\,
\eta^{d_{2}}\,{{\cal P}}_{d_{3}}}{
2}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (2)} \\ \Omega \end{array}}}=
\[\frac{D_{d_{2}d_{3}}^{
d_{6}d_{5}}\,D_{d_{6}d_{1}}^{
d_{7}d_{4}}\,G_{d_{7}}\,
\eta^{d_{1}}\,\eta^{d_{2}}\,\eta^{d_{3}}\,
{{\cal P}}_{d_{4}}\,{{\cal P}}_{d_{5}}}{
6}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (3)} \\ \Omega \end{array}}}=
\[\frac{-
\(D_{d_{2}d_{3}}^{
d_{8}d_{7}}\,D_{d_{4}d_{1}}^{
d_{9}d_{5}}\,D_{d_{8}d_{9}}^{
d_{10}d_{6}}\,G_{d_{10}}\,\eta^{d_{1}}\,
\eta^{d_{2}}\,\eta^{d_{3}}\,\eta^{d_{4}}\,
{{\cal P}}_{d_{5}}\,{{\cal P}}_{d_{6}}\,{{\cal P}}_{d_{7}}
\)
}{
24}
\]
\cr}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (4)} \\ \Omega \end{array}}}=
\[\(\eta^{d_{1}}\,\eta^{d_{2}}\,\eta^{d_{3}}\,
\eta^{d_{4}}\,\eta^{d_{5}}\,{{\cal P}}_{d_{6}}\,
{{\cal P}}_{d_{7}}\,{{\cal P}}_{d_{8}}\,{{\cal P}}_{d_{9}}\,
\nl
\off{3499956}
\(3\,D_{d_{1}d_{2}}^{
d_{13}d_{6}}\,D_{d_{4}d_{5}}^{
d_{12}d_{9}}\,
D_{d_{10}d_{3}}^{
d_{11}d_{8}}\,D_{d_{12}d_{13}}^{
d_{10}d_{7}}\,G_{d_{11}}
+4\,D_{d_{3}d_{4}}^{
d_{10}d_{9}}\,D_{d_{5}d_{2}}^{
d_{13}d_{6}}\,
D_{d_{10}d_{11}}^{
d_{12}d_{8}}\nl
\off{3827636}
\,D_{d_{13}d_{1}}^{
d_{11}d_{7}}\,G_{d_{12}}
-4\,D_{d_{4}d_{5}}^{
d_{12}d_{9}}\,D_{d_{10}d_{3}}^{
d_{11}d_{8}}\,
D_{d_{12}d_{1}}^{
d_{13}d_{7}}\,D_{d_{13}d_{2}}^{
d_{10}d_{6}}\,G_{d_{11}}
\)
\)
/360
\]
\Nl}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (5)} \\ \Omega \end{array}}}=
\[\(D_{d_{12}d_{13}}^{
d_{14}d_{10}}\,G_{d_{14}}\,\eta^{d_{1}}\,
\eta^{d_{2}}\,\eta^{d_{3}}\,\eta^{d_{4}}\,
\eta^{d_{5}}\,\eta^{d_{6}}\,{{\cal P}}_{d_{7}}\,
{{\cal P}}_{d_{8}}\,{{\cal P}}_{d_{9}}\,{{\cal P}}_{d_{10}}\,
{{\cal P}}_{d_{11}}\nl
\off{3499956}
\,
\(-
\(2\,D_{d_{1}d_{2}}^{
d_{16}d_{7}}\,D_{d_{5}d_{6}}^{
d_{15}d_{11}}\,
D_{d_{15}d_{3}}^{
d_{12}d_{8}}\,D_{d_{16}d_{4}}^{
d_{13}d_{9}}
\)
+3\,D_{d_{1}d_{3}}^{
d_{16}d_{8}}\,D_{d_{4}d_{5}}^{
d_{12}d_{11}}\,
\nl
\off{4100703}
D_{d_{6}d_{2}}^{
d_{15}d_{7}}\,D_{d_{15}d_{16}}^{
d_{13}d_{9}}
-4\,D_{d_{4}d_{5}}^{
d_{12}d_{11}}\,D_{d_{6}d_{1}}^{
d_{15}d_{7}}\,
D_{d_{15}d_{3}}^{
d_{16}d_{9}}\,D_{d_{16}d_{2}}^{
d_{13}d_{8}}
\)
\)
/720
\]
\Nl}$$ $$\displaylines{\qdd
{{\renewcommand{\arraystretch}{.7} \begin{array}[b]{@{}c@{}} {\scriptscriptstyle (6)} \\ \Omega \end{array}}}=
\[\(\eta^{d_{1}}\,\eta^{d_{2}}\,\eta^{d_{3}}\,
\eta^{d_{4}}\,\eta^{d_{5}}\,\eta^{d_{6}}\,
\eta^{d_{7}}\,{{\cal P}}_{d_{8}}\,{{\cal P}}_{d_{9}}\,
{{\cal P}}_{d_{10}}\,{{\cal P}}_{d_{11}}\,{{\cal P}}_{d_{12}}\,
{{\cal P}}_{d_{13}}\,\nl
\off{3499956}
\(6\,D_{d_{1}d_{2}}^{
d_{17}d_{8}}\,D_{d_{6}d_{7}}^{
d_{16}d_{13}}\,
D_{d_{14}d_{5}}^{
d_{15}d_{12}}\,D_{d_{16}d_{17}}^{
d_{18}d_{11}}\,D_{d_{18}d_{4}}^{
d_{19}d_{10}}\,D_{d_{19}d_{3}}^{
d_{14}d_{9}}\,G_{d_{15}}\nl
\off{3827636}
-9\,D_{d_{1}d_{2}}^{
d_{18}d_{8}}\,D_{d_{3}d_{4}}^{
d_{19}d_{9}}\,
D_{d_{6}d_{7}}^{
d_{16}d_{13}}\,D_{d_{14}d_{5}}^{
d_{15}d_{12}}\,D_{d_{16}d_{17}}^{
d_{14}d_{10}}\,D_{d_{18}d_{19}}^{
d_{17}d_{11}}\,G_{d_{15}}\nl
\off{3827636}
+12\,D_{d_{1}d_{2}}^{
d_{18}d_{8}}\,D_{d_{3}d_{4}}^{
d_{19}d_{9}}\,
D_{d_{6}d_{7}}^{
d_{17}d_{13}}\,D_{d_{14}d_{15}}^{
d_{16}d_{12}}\,D_{d_{17}d_{18}}^{
d_{15}d_{11}}\,D_{d_{19}d_{5}}^{
d_{14}d_{10}}\,G_{d_{16}}\nl
\off{3827636}
-10\,D_{d_{1}d_{2}}^{
d_{18}d_{8}}\,D_{d_{6}d_{7}}^{
d_{16}d_{13}}\,
D_{d_{14}d_{5}}^{
d_{15}d_{12}}\,D_{d_{16}d_{3}}^{
d_{17}d_{9}}\,D_{d_{17}d_{19}}^{
d_{14}d_{11}}\,D_{d_{18}d_{4}}^{
d_{19}d_{10}}\,G_{d_{15}}\nl
\off{3827636}
+24\,D_{d_{1}d_{2}}^{
d_{18}d_{8}}\,D_{d_{6}d_{7}}^{
d_{16}d_{13}}\,
D_{d_{14}d_{5}}^{
d_{15}d_{12}}\,D_{d_{16}d_{17}}^{
d_{14}d_{11}}\,D_{d_{18}d_{4}}^{
d_{19}d_{10}}\,D_{d_{19}d_{3}}^{
d_{17}d_{9}}\,G_{d_{15}}\nl
\off{3827636}
-4\,D_{d_{1}d_{2}}^{
d_{18}d_{8}}\,D_{d_{6}d_{7}}^{
d_{17}d_{13}}\,
D_{d_{14}d_{15}}^{
d_{16}d_{12}}\,D_{d_{17}d_{4}}^{
d_{15}d_{9}}\,D_{d_{18}d_{5}}^{
d_{19}d_{11}}\,D_{d_{19}d_{3}}^{
d_{14}d_{10}}\,G_{d_{16}}\nl
\off{3827636}
+4\,D_{d_{1}d_{2}}^{
d_{19}d_{8}}\,D_{d_{6}d_{7}}^{
d_{16}d_{13}}\,
D_{d_{14}d_{5}}^{
d_{15}d_{12}}\,D_{d_{16}d_{17}}^{
d_{18}d_{11}}\,D_{d_{18}d_{3}}^{
d_{14}d_{9}}\,D_{d_{19}d_{4}}^{
d_{17}d_{10}}\nl
\off{3827636}
\,G_{d_{15}}
-6\,D_{d_{1}d_{3}}^{
d_{18}d_{9}}\,D_{d_{5}d_{6}}^{
d_{14}d_{13}}\,
D_{d_{7}d_{2}}^{
d_{17}d_{8}}\,D_{d_{14}d_{15}}^{
d_{16}d_{12}}\,D_{d_{17}d_{18}}^{
d_{19}d_{11}}\nl
\off{3827636}
\,D_{d_{19}d_{4}}^{
d_{15}d_{10}}\,G_{d_{16}}
-4\,D_{d_{1}d_{3}}^{
d_{19}d_{9}}\,D_{d_{5}d_{6}}^{
d_{14}d_{13}}\,
D_{d_{7}d_{2}}^{
d_{17}d_{8}}\,D_{d_{14}d_{15}}^{
d_{16}d_{12}}\,\nl
\off{3827636}
D_{d_{17}d_{18}}^{
d_{15}d_{11}}\,D_{d_{19}d_{4}}^{
d_{18}d_{10}}\,G_{d_{16}}
-16\,D_{d_{5}d_{6}}^{
d_{14}d_{13}}\,D_{d_{7}d_{1}}^{
d_{17}d_{8}}\,
D_{d_{14}d_{15}}^{
d_{16}d_{12}}\nl
\off{3827636}
\,D_{d_{17}d_{4}}^{
d_{18}d_{11}}\,D_{d_{18}d_{3}}^{
d_{19}d_{10}}\,
D_{d_{19}d_{2}}^{
d_{15}d_{9}}\,G_{d_{16}}
+16\,D_{d_{6}d_{7}}^{
d_{16}d_{13}}\,D_{d_{14}d_{5}}^{
d_{15}d_{12}}\nl
\off{3827636}
\,D_{d_{16}d_{4}}^{
d_{17}d_{11}}\,D_{d_{17}d_{3}}^{
d_{19}d_{10}}\,
D_{d_{18}d_{1}}^{
d_{14}d_{8}}\,D_{d_{19}d_{2}}^{
d_{18}d_{9}}\,G_{d_{15}}
\)
\)
/15120
\]
\Nl}$$
Again, these expressions are not particularly illuminating. The point emphasized here is that the calculation of the BRST charge is purely algorithmic and follows a general, well-established pattern.
Since homogeneous quadratic algebras contain the self-reproducing algebras as special case, they are generically of maximal covariant rank. More on this in [@ADresse3].
L-T algebras
------------
We now consider adding a linear term to the quadratic algebra above. The basic Poisson brackets for the generators $G_d$ are given by $$[G_{d_1}, G_{d_2}] = C_{d_1 d_2}^{d_3} \, G_{d_3} +
D_{d_1 d_2}^{d_3 d_4} G_{d_3} G_{d_4}$$ where $ C_{d_1 d_2}^{d_3} $ and $D_{d_1 d_2}^{d_3 d_4}$ are antisymmetric in $d_1, d_2$, and $D_{d_1 d_2}^{d_3 d_4}$ is symmetric in $d_3, d_4$. A particular instance of such an algebra is given by Zamolodchikov algebras [@Zam:]. We will start with a specific example, and consider general quadratically nonlinear Poisson algebras next.
The generators in the example are assumed to split into $L_{a}$ and $T_b$, $a = 1, \ldots n_1$, $b = n_1 + 1, \ldots, n$, with the brackets $$\begin{aligned}
[L_{a_1}, L_{a_2}] &=& \tilde{C}_{a_1 a_2}^{a_3} L_{a_3} \nonumber \\{}
[L_{a_1}, T_{b_1}] &=& \tilde{C}_{a_1 b_1}^{a_2} L_{a_2} +
\tilde{C}_{a_1 b_1}^{b_2} T_{b_2}
\label{eq:zam} \\{}
[T_{b_1}, T_{b_2}] &=& \tilde{C}_{b_1 b_2}^{a_1} L_{a_1} +
\tilde{C}_{b_1 b_2}^{b_3} T_{b_3} +
\tilde{D}_{b_1 b_2}^{a_1 a_2} L_{a_1} L_{a_2}
\nonumber\end{aligned}$$ so that contractions of $\tilde{D}$ are impossible.
Going back to the notations $G_{d_i} = \{L_{a}, T_{b}\}$, $d = 1,
\ldots, n$ the Jacobi identity imply $$C_{d_1 d_2}^{d_4} C_{d_3 d_4}^{d_5} + \mbox{cyclic}(d_1,d_2,d_3) = 0$$ $$\{D_{d_1 d_2}^{d_4 d_5} C_{d_3 d_4}^{d_6} + \mbox{symm}(d_5, d_6) \} + \\
C_{d_1 d_2}^{d_4} D_{d_3 d_4}^{d_5 d_6} + \mbox{cyclic}(d_1,d_2,d_3) = 0$$ and contractions of $D$ vanish.
For instance, the conditions (\[eq:zam\]) are fulfilled if one takes for the $L$’s the generators of a semi-simple Lie algebra and take the $T$’s to commute with the $L$’s and to close on the Casimir element: $$\begin{aligned}
[L_a, T_b] &=& 0 \\{}
[T_{b_1}, T_{b_2}] &=& \delta_{b_1 b_2} k^{a_1 a_2} L_{a_1} L_{a_2}\end{aligned}$$ where $k^{a_1 a_2}$ is the Killing bilinear form. The Jacobi identity is verified because the Casimir element commutes with the $L$’s.
The previous theorem on the rank yields, by taking $n(l) = 1$ and $n(T) = 3/2$, that the rank is bounded by $1/2 m + 1$, where $m$ is the number of T-generators. Actually, the rank is much lower, since the covariant BRST charge is computed to be $$\displaylines{\qdd
\Omega =
\[\frac{1}{2}
\,C_{d_{1}d_{2}}^{
d_{3}}\,\eta^{d_{1}}\,\eta^{d_{2}}\,
{{\cal P}}_{d_{3}}
+
\frac{1}{24}
\,C_{d_{8}d_{9}}^{
d_{6}}\,D_{d_{1}d_{2}}^{
d_{8}d_{7}}\,D_{d_{3}d_{4}}^{
d_{9}d_{5}}\,\eta^{d_{1}}\,\eta^{d_{2}}\nl
\off{2695321}
\,\eta^{d_{3}}\,\eta^{d_{4}}\,{{\cal P}}_{d_{5}}\,
{{\cal P}}_{d_{7}}\,{{\cal P}}_{d_{6}}
+
\frac{1}{2}
\,D_{d_{1}d_{2}}^{
d_{4}d_{3}}\,G_{d_{4}}\,\eta^{d_{1}}\,
\eta^{d_{2}}\,{{\cal P}}_{d_{3}}
+G_{d_{1}}\,\eta^{d_{1}}
\]
\Nl}$$ which is identical to the result in [@SchSevNie:QuaBRSChaQua].
Generalizations
---------------
The previous L-T algebras can be generalized in various directions. One may consider the general quadratic non homogeneous Poisson structure $$\begin{aligned}
& [G_{d_1}, G_{d_2}] = C_{d_1 d_2}^{d_3} \, G_{d_3} +
D_{d_1 d_2}^{d_3 d_4} G_{d_3} G_{d_4} & \\
& C_{d_1 d_2}^{d_4} C_{d_3 d_4}^{d_5} + \mbox{cyclic}(d_1,d_2,d_3)
= 0 & \\
& \{D_{d_1 d_2}^{d_4 d_5} C_{d_3 d_4}^{d_6} + \mbox{symm}(d_5, d_6) \} +
C_{d_1 d_2}^{d_4} D_{d_3 d_4}^{d_5 d_6} + \mbox{} \quad\quad &
\nonumber \\
& \quad \quad \mbox{cyclic}(d_1,d_2,d_3) = 0 &\\
& D_{d_4 d_1}^{d_5 d_6} D_{d_2 d_3}^{d_4 d_7} + \mbox{symm}(d_5, d_6,
d_7) + \mbox{cyclic}(d_1,d_2,d_3) = 0&\end{aligned}$$ with $ C_{d_1 d_2}^{d_3} $ and $D_{d_1 d_2}^{d_3 d_4}$ antisymmetric in $d_1, d_2$, and $D_{d_1 d_2}^{d_3 d_4}$ symmetric in $d_3, d_4$. One may also include higher order terms in the bracket while preserving the existence of a degree decreased by the bracket, as in the so called spin 4 algebra : $$\begin{aligned}
[L_{a_1}, L_{a_2}] &=& C_{a_1 a_2}^{a_3} L_{a_3} \\{}
[L_{a_1}, T_{b_1}] &=& C_{a_1 b_1}^{b_2} T_{b_2} \\{}
[L_{a_1}, W_{c_1}] &=& C_{a_1 c_1}^{c_2} W_{c_2} \\{}
[T_{b_1}, T_{b_2}] &=& C_{b_1 b_2}^{a_1} L_{a_1} + C_{b_1 b_2}^{b_3} T_{b_3} +
D_{b_1 b_2}^{a_1 a_2} L_{a_1} L_{a_2} \\{}
[T_{b_1}, W_{c_1}] &=& C_{b_1 c_1}^{b_2} T_{b_2} +
D_{b_1 c_1}^{a_1 b_2} L_{a_1} T_{b_2} \\{}
[W_{c_1}, W_{c_2}] &=& C_{c_1 c_2}^{a_1} L_{a_1} +
D_{c_1 c_2}^{a_1 a_2} L_{a_1} L_{a_2} +
E_{c_1 c_2}^{a_1 a_2 a_3} L_{a_1} L_{a_2} L_{a_3}.\end{aligned}$$ If one sets $n(L) = 1, n(T) = 3/2, L(W) = 2$, one gets $n([A,B]) \leq
n(A) + n(B)$.
We have checked, using REDUCE, that in both cases the first seven terms in $\Omega$ are generically non zero.
Conclusion
==========
We have shown in this paper that polynomial Poisson algebras provide a rich arena in which the perturbative features of the BRST construction are perfectly illustrated. We believe this to be of interest because models of higher rank are rather rare and are usually thought not to arise in practice. Non polynomial Poisson algebras(e.g. of the type arising in the study of quantum groups) can also be analyzed along the same BRST lines and should provide further models of higher rank.
We have not discussed the quantum realization of Poisson algebras, and whether the nilpotency condition for the BRST generator is maintained quantum-mechanically. This is a difficult question, which is model-dependent. Indeed, while the ghost contribution to $\Omega^2$ can be evaluated independently of the specific form of the $G_a$’s in terms of the canonical variables $(q^i, p_i)$ (once a representation of the ghost anticommutation relations is chosen), the “matter” contribution to $\Omega^2$ depends on the “anomaly” terms in $[G_a, G_b]$, which, in turn, depend on the specific form of the $G_a$’s. It would be interesting to pursue this question further.
Acknowledgments
===============
We are grateful to Jim Stasheff and Claudio Teitelboim for fruitful discussions at the early stages of this research. This work has been supported in part by research funds from FNRS (Belgium) and by a research contract with the Commission of the European Communities.
[10]{}
M. Nakamura. , 37:195, 1967.
S.B. Priddy. , 152:39, 1970.
E.K. Sklyanin. , 16:263, 1982.
A. B. Zamolodchikov. , 65:1205, 1985.
V.A. Fateev and A. B. Zamolodchikov. , B280:644, 1987.
Y.G. Oh. , 12:87, 1986.
V.O. Tarasov, L.A. Takhatadzlupan, and L.D. Faddeev. , 57:1059, 1983.
I Bakas and P. Mathieu. , 208B:101, 1988.
K. H. Bhaskara and K. Rama. , 32:2319, 1991.
Ya. I. Granovskii, A. S. Zhedanov, and I. M. Lutsenko. , 91:474, 1992.
M. Henneaux and C. Teitelboim. . Princeton University Press, 1992.
V.K.A.M. Gugenheim and J.D. Stasheff. , 38:237, 1986; J.D. Stasheff. ,19:287, 1988 and references therein.
M. Henneaux. , 120B:75, 1983.
K. Fujikawa and J. Kubo. , 199B:75, 1987.
J. Huebschmann. Extensions of Lie algebras. Heidelberg preprint, 1989.
A. Dresse. Canonical form of expressions involving dummy variables. Submitted to J. Symb. Comp.
A. Dresse. Treatment of dummy variables and BRST theory in computer algebra. In [*Proceedings of the IMACS Symposium SC–1993*]{}, 1993.
A. Burnel, H. Caprasse, and A. Dresse. in preparation.
A. Dresse. Ph.D. Thesis, U.L.B. (in preparation).
K. Schoutens, A. Sevrin, and P. van Nieuwenhuizen. , 124:87, 1989.
[^1]: We shall restrict here the analysis to ordinary polynomial algebras with commuting generators, but one can easily extend the study to the graded case with both commuting and anticommuting generators.
[^2]: It should be stressed that the polynomial algebra generated by the $G$’s, equipped with the bracket (\[basic\_bracket\]) is [*always*]{} an infinite-dimensional Lie algebra, even in the “open algebra” case.
[^3]: In agreement with the notations of [@HenTei:QuaGauSys], we denote the algebra of polynomials in the anticommuting variables ${{\cal P}}_a$ with complex coefficients by $\Bbb{C}\,({{\cal P}}_a)$, and not by the more familiar notation $\Lambda({{\cal P}}_a)$. A typical element of $\Bbb{C}\,({{\cal P}}_1)$ is $a + b {{\cal P}}_1$ with $a, b \in \Bbb{C}\,$ since $({{\cal P}}_1)^2 = 0$.
|
---
abstract: 'We obtain a simple presentation of the hyperelliptic mapping class group ${{\mathcal{M}}^h}(N)$ of a nonorientable surface $N$. As an application we compute the first homology group of ${{\mathcal{M}}^h}(N)$ with coefficients in $H_1(N;{{\mathbb{Z}}})$.'
address: ' Institute of Mathematics, University of Gdańsk, Wita Stwosza 57, 80-952 Gdańsk, Poland '
author:
- Michał Stukow
title: A finite presentation for the hyperelliptic mapping class group of a nonorientable surface
---
[^1]
Introduction
============
Let $N_{g,s}^n$ be a smooth, nonorientable, compact surface of genus $g$ with $s$ boundary components and $n$ punctures. If $s$ and/or $n$ is zero, then we omit it from the notation. If we do not want to emphasise the numbers $g,s,n$, we simply write $N$ for a surface $N_{g,s}^n$. Recall that $N_{g}$ is a connected sum of $g$ projective planes, and $N_{g,s}^n$ is obtained from $N_g$ by removing $s$ open disks and specifying the set $\Sigma=\{{{z}_1,\dotsc,{z}_{n}}\}$ of $n$ distinguished points in the interior of $N_g$.
Let ${\textrm{Diff}}(N)$ be the group of all diffeomorphisms ${h\colon N\to N}$ such that $h$ is the identity on each boundary component and $h(\Sigma)=\Sigma$. By ${{\mathcal{M}}}(N)$ we denote the quotient group of ${\textrm{Diff}}(N)$ by the subgroup consisting of maps isotopic to the identity, where we assume that isotopies fix $\Sigma$ and are the identity on each boundary component. ${{\mathcal{M}}}(N)$ is called the *mapping class group* of $N$.
The mapping class group ${{\mathcal{M}}}(S_{g,s}^n)$ of an orientable surface is defined analogously, but we consider only orientation preserving maps. If we include orientation reversing maps, we obtain the so-called *extended mapping class group* ${{\mathcal{M}}}^{\pm}(S_{g,s}^n)$.
Suppose that the closed orientable surface $S_g$ is embedded in ${{\mathbb{R}}}^3$ as shown in Figure \[r01\], in such a way that it is invariant under reflections across $xy,yz,xz$ planes. Let ${{\varrho}\colon S_g\to S_g}$ be the *hyperelliptic involution*, i.e. the half turn about the $y$-axis.
![Surface $S_g$ embedded in ${{\mathbb{R}}}^3$.[]{data-label="r01"}](R01){width="90.00000%"}
The *hyperelliptic mapping class group* ${{{\mathcal{M}}}}^h(S_g)$ is defined to be the centraliser of ${\varrho}$ in ${{{\mathcal{M}}}}(S_g)$. In a similar way we define the *extended hyperelliptic mapping class group* ${{{\mathcal{M}}}}^{h\pm}(S_g)$ to be the centraliser of ${\varrho}$ in ${{{\mathcal{M}}}}^{\pm}(S_g)$.
Background
----------
The hyperelliptic mapping class group turns out to be a very interesting and important subgroup of the mapping class group. Its algebraic properties have been studied extensively – see [@Bir-Hil; @Kawazumi1] and references there. Although ${{{\mathcal{M}}}}^h(S_g)$ is an infinite index subgroup of ${{{\mathcal{M}}}}(S_g)$ for $g\geqslant 3$, it plays surprisingly important role in studying its algebraic properties. For example Wajnryb’s simple presentation [@Wajn_pre] of the mapping class group ${{{\mathcal{M}}}}(S_g)$ differs from the presentation of the group ${{{\mathcal{M}}}}^h(S_g)$ by adding one generator and a few relations. Another important phenomenon is the fact, that every finite cyclic subgroup of maximal order in ${{{\mathcal{M}}}}(S_g)$ is conjugate to a subgroup of ${{{\mathcal{M}}}}^h(S_g)$ [@MaxHyp].
Homological computations play a prominent role in the theory of mapping class groups. Let us mention that in the case of the hyperelliptic mapping class group, Bödigheimer, Cohen and Peim [@BoedCoh] computed $H^*({{{\mathcal{M}}}}^h(S_g);{{\mathbb{K}}})$ with coefficients in any field ${{\mathbb{K}}}$. Kawazumi showed in [@Kawazumi1] that if $\textrm{ch}({{\mathbb{K}}})\neq 2$ then $H^*({{{\mathcal{M}}}}^h(S_g);H^1(S_g;{{\mathbb{K}}}))=0$. For the integral coefficients, Tanaka [@Tanaka] showed that $H_1({{{\mathcal{M}}}}^h(S_g);H_1(S_g;{{\mathbb{Z}}}))\cong{{\mathbb{Z}}}_2$. Let us also mention that Morita [@MoritaJacFou] showed that in the case of the full mapping class group, $H_1({{{\mathcal{M}}}}(S_g);H_1(S_g,{{\mathbb{Z}}}))\cong{{\mathbb{Z}}}_{2g-2}$.
Main results
------------
The purpose of this paper is to extend the notion of the hyperelliptic mapping class group to the nonorientable case. We define this group ${{{\mathcal{M}}}}^h(N)$ in Section \[sec:def:hiper\] and observe that it contains a natural subgroup ${{{\mathcal{M}}}}^{h+}(N)$ of index 2 (Remark \[rem:positiv:subg\]).
Then we obtain simple presentations of these groups (Theorems \[th:004\] and \[tw:pres:pos\]). By analogy with the orientable case, these presentations may be thought of as the first approximation of a presentation of the full mapping class group ${{{\mathcal{M}}}}(N)$. In fact, for $g=3$ the hyperelliptic mapping class group ${{{\mathcal{M}}}}^h(N)$ coincide with the full mapping class group ${{{\mathcal{M}}}}(N)$ (see Corollary \[Presen:g3:Bir\]). If $g\geq 4$, then Paris and Szepietowski [@SzepParis] obtained a simple presentation of ${{{\mathcal{M}}}}(N)$, which can be rewritten (Proposition 3.3 and Theorem 3.5 of [@StukowSimpSzepPar]) so that it has the hyperelliptic involution ${\varrho}$ as one of the generators, and the hyperelliptic relations (Theorem \[th:004\]) appear among defining relations.
As an application of obtained presentations we compute the first homology groups of ${{{\mathcal{M}}}}^h(N)$ and ${{{\mathcal{M}}}}^{h+}(N)$ with coefficients in $H_1(N;{{\mathbb{Z}}})$ (Theorems \[tw:hom:positive\] and \[tw:hom:full\]).
Definitions of ${{{\mathcal{M}}}}^h(N_g)$ and ${{{\mathcal{M}}}}^{h+}(N_g)$ {#sec:def:hiper}
===========================================================================
Let $S_{g-1}$ be a closed oriented surface of genus $g-1\geqslant 2$ embedded in ${{\mathbb{R}}}^3$ as shown in Figure \[r01\], in such a way that it is invariant under reflections across $xy,yz,xz$ planes, and let ${j\colon S_{g-1}\to S_{g-1}}$ be the symmetry defined by $j(x,y,z)=(-x,-y,-z)$. Denote by $C_{{{{\mathcal{M}}}}^{\pm}(S_{g-1})}(j)$ the centraliser of $j$ in ${{{\mathcal{M}}}}^{\pm}(S_{g-1})$. The orbit space $S_{g-1}/{\langle j \rangle}$ is a nonorientable surface $N_{g}$ of genus $g$ and it is known (Theorem 1 of [@BirChil1]) that there is an epimorphism $${\pi_j\colon C_{{{{\mathcal{M}}}}^{\pm}(S_{g-1})}(j)\to {{{\mathcal{M}}}}(N_{g})}$$ with kernel $\ker \pi_j={\langle j \rangle}$. In particular $${{{\mathcal{M}}}}(N_{g})\cong C_{{{{\mathcal{M}}}}^{\pm}(S_{g-1})}(j)/{\langle j \rangle}.$$ Observe that the hyperelliptic involution ${\varrho}$ is an element of $C_{{{{\mathcal{M}}}}^{\pm}(S_{g-1})}(j)$. Hence the following definition makes sense.
Define the *hyperelliptic mapping class group* ${{{\mathcal{M}}}}^h(N)$ of a closed nonorientable surface $N$ to be the centraliser of $\pi_j({\varrho})$ in the mapping class group ${{{\mathcal{M}}}}(N)$. We say that $\pi_j({\varrho})$ is the *hyperelliptic involution* of $N$ and by abuse of notation we write ${\varrho}$ for $\pi_j({\varrho})$.
In order to have a little more straightforward description of ${\varrho}$ observe, that the orbit space $S_{g-1}/{\langle j \rangle}$ gives the model of $N_g$, where $N_g$ is a connected sum of an orientable surface $S_r$ and a projective plane (for $g$ odd) or a Klein bottle (for $g$ even) – see Figure \[r02\].
![Nonorientable surface $N_g$.[]{data-label="r02"}](R02){width="96.00000%"}
To be more precise, $N_g$ is the left half of $S_{g-1}$ embedded in ${{\mathbb{R}}}^3$ as in Figure \[r01\] with boundary points identified by the map $(x,y,z)\mapsto (-x,-y,-z)$. Note that $g=2r+1$ for $g$ odd and $g=2r+2$ for $g$ even. In such a model, ${{\varrho}\colon N_g\to N_g}$ is the map induced by the half turn about the $y$-axis.
Observe that the set of fixed points of ${{\varrho}\colon N_g\to N_g}$ consists of $g$ points $\{p_1,p_2,\ldots,p_g\}$ and the circle $p$. Therefore ${{{\mathcal{M}}}}^h(N)$ consists of isotopy classes of maps which must fix the set $\{p_1,p_2,\ldots,p_g\}$ and map the circle $p$ to itself. Moreover, the orbit space $N_g/{\langle {\varrho}\rangle}$ is the sphere $S_{0,1}^g$ with one boundary component corresponding to $p$ and $g$ distinguished points corresponding to $\{p_1,p_2,\ldots,p_g\}$. Since elements of ${{{\mathcal{M}}}}^h(N_g)$ may not fix $p$ point–wise, it is more convenient to treat $p$ as the distinguished puncture $p_{g+1}$, hence we will identify $N_g/{\langle {\varrho}\rangle}$ with the sphere $S_0^{g,1}$ with $g+1$ punctures. The notation $S_0^{g,1}$ is meant to indicate that maps of $S_0^{g,1}$ (and their isotopies) could permute the punctures $p_1,\ldots,p_g$, but must fix $p_{g+1}$.
The main goal of this section is to prove the following theorem.
\[tw:002\] If $g\geqslant 3$ then the projection $N_g\to N_g/{\langle {\varrho}\rangle}$ induces an epimorphism $${\pi_{\varrho}\colon {{{\mathcal{M}}}}^h(N_g)\to {{{\mathcal{M}}}}^{\pm}(S_0^{g,1})}$$ with $\ker \pi_{\varrho}={\langle {\varrho}\rangle}$.
Consider the following diagram $$\xymatrix@C=3pc@R=3pc{
C_{{{{\mathcal{M}}}}^{\pm}(S_{g-1})}({\langle j,{\varrho}\rangle})\ar^{\pi_{\varrho}}[r]\ar^{\pi_j}[d]&C_{{{{\mathcal{M}}}}^{\pm}(S_0^{2g})}(j)\ar^{\pi_j}[d]\\
{{{\mathcal{M}}}}^h(N_g)\ar@{-->}^{\pi_{\varrho}}[r] \ar@/^1pc/@{->}[u]^{i_j} &{{{\mathcal{M}}}}^{\pm}(S_0^{g,1})
}$$ The left vertical map is the restriction of the projection $${\pi_j\colon C_{{{{\mathcal{M}}}}^{\pm}(S_{g-1})}(j)\to {{{\mathcal{M}}}}(N_g)}$$ to the subgroup consisting of elements which centralise ${\varrho}$. The nice thing about $\pi_j$ is that it has a section $${i_j\colon {{{\mathcal{M}}}}(N_g)\to C_{{{{\mathcal{M}}}}^{\pm}(S_{g-1})}(j)}.$$ In fact, for any $h\in{{{\mathcal{M}}}}(N_g)$ we can define $i_j(h)$ to be an orientation preserving lift of $h$.
The upper horizontal map is the restriction of the homomorphism $${\pi_{\varrho}\colon {{{\mathcal{M}}}}^{h\pm}(S_{g-1})\to {{{\mathcal{M}}}}^{\pm}(S_0^{2g})}$$ induced by the orbit projection $S_{g-1}\to S_{g-1}/{\langle {\varrho}\rangle}$. The fact that this map is a homomorphism was first observed by Birman and Hilden [@Bir-Hil]. The kernel of this map is equal to ${\langle {\varrho}\rangle}$.
The right vertical map is again the homomorphism induced by the orbit projection $S^{2g}_0\to
S^{2g}_0/{\langle j \rangle}$. However now ${j\colon S^{2g}_0\to S^{2g}_0}$ is a reflection with a circle of fixed points. The existence of $\pi_j$ in such a case follows from the work of Zieschang (Proposition 10.3 of [@Zies1]).
Hence there is the homomorphism $${\pi_{\varrho}\colon {{{\mathcal{M}}}}^h(N_g)\to {{{\mathcal{M}}}}^{\pm}(S_0^{g,1})}$$ defined as the composition $$\pi_{\varrho}=\pi_j\circ \pi_{\varrho}\circ i_j.$$ Moreover, $$\begin{aligned}
\ker \pi_{\varrho}&=\ker (\pi_j\circ \pi_{\varrho}\circ i_j)=(\pi_j\circ \pi_{\varrho}\circ i_j)^{-1}(id)\\
&=i_j^{-1}(\pi_{\varrho}^{-1}(\pi_j^{-1}(id)))=i_j^{-1}(\pi_{\varrho}^{-1}({\langle j \rangle}))=i_j^{-1}({\langle j,{\varrho}\rangle})={\langle {\varrho}\rangle}.
\end{aligned}$$
Theorem \[tw:002\] is not true if $N=N_2$. This corresponds to the fact that the Birman-Hilden theorem does not hold for the closed torus $S=S_1$.
\[rem:positiv:subg\] Theorem \[tw:002\] shows that the group ${{{\mathcal{M}}}}^h(N_g)$ contains a very natural subgroup of index 2, namely $${{{\mathcal{M}}}}^{h+}(N_g)=\pi_{\varrho}^{-1}({{{\mathcal{M}}}}(S_0^{g,1})).$$ Geometrically, the subgroup ${{{\mathcal{M}}}}^{h+}(N_g)$ consists of these elements, which preserve the orientation of the circle $p$ (the circle fixed by ${\varrho}$). As we will see later (see Remark \[uw:perspective\]), it seems that the group ${{{\mathcal{M}}}}^{h+}(N)$ corresponds to ${{{\mathcal{M}}}}^h(S)$, whereas ${{{\mathcal{M}}}}^{h}(N)$ corresponds to ${{{\mathcal{M}}}}^{h\pm}(S)$.
Presentations for groups ${{{\mathcal{M}}}}(S_0^{g,1})$ and ${{{\mathcal{M}}}}^{\pm}(S_0^{g,1})$
================================================================================================
Let $w_1,w_2,\ldots,w_g$ be simple arcs connecting punctures $p_1,\ldots,p_{g+1}$ on a sphere $S_0^{g+1}$ as shown in Figure \[r03\].
![Sphere $S_0^{p+1}$ and elementary braid ${\sigma}_i$.[]{data-label="r03"}](R03){width="80.00000%"}
Recall that to each such arc $w_i$ we can associate the elementary braid ${\sigma}_i$ which interchanges punctures $p_i$ and $p_{i+1}$ – see Figure \[r03\]. The following theorem is due to Magnus [@Mag_sfer]. It is also proved in Chapter 4 of [@Bir-BLMCG].
\[th:00\] If $g\geqslant 1$, then ${{{\mathcal{M}}}}(S_0^{g+1})$ has the presentation with generators ${\sigma}_1,\ldots,{\sigma}_g$ and defining relations: $$\begin{aligned}
&{\sigma}_k{\sigma}_j={\sigma}_j{\sigma}_k\quad\text{for $|k-j|>1$},\\
&{\sigma}_j{\sigma}_{j+1}{\sigma}_j={\sigma}_{j+1}{\sigma}_j{\sigma}_{j+1}\quad\text{for $j=1,\ldots,g-1$},\\
&{\sigma}_1\cdots {\sigma}_{g-1}{\sigma}_g^2{\sigma}_{g-1}\cdots {\sigma}_1=1,\\
&({\sigma}_1{\sigma}_2\cdots {\sigma}_{g})^{g+1}=1.
\end{aligned}$$
In order to avoid unnecessary complications, from now on assume that $g\geqslant 3$. Recall that we denote by ${{{\mathcal{M}}}}(S_0^{g,1})$ the subgroup of ${{{\mathcal{M}}}}(S_0^{g+1})$ consisting of maps which fix $p_{g+1}$.
\[th:001\] If $g\geqslant 3$, then ${{{\mathcal{M}}}}(S_0^{g,1})$ has the presentation with generators ${\sigma}_1,\ldots,{\sigma}_{g-1}$ and defining relations:
1. ${\sigma}_k{\sigma}_j={\sigma}_j{\sigma}_k$ for $|k-j|>1$ and $k,j<g$,
2. ${\sigma}_j{\sigma}_{j+1}{\sigma}_j={\sigma}_{j+1}{\sigma}_j{\sigma}_{j+1}$ for $j=1,2,\ldots,g-2$,
3. $({\sigma}_1\cdots {\sigma}_{g-1})^g=1$.
By Lemma 2.2 of [@Big-lin], $${{{\mathcal{M}}}}(S_0^{g,1})\cong B_{g}/{\langle \Delta^2 \rangle},$$ where $B_{g}={{{\mathcal{M}}}}(S_{0,1}^{g})$ is the braid group on $g$ strands, and $$\Delta^2=({\sigma}_1\cdots {\sigma}_{g-1})^g$$ is the generator of the center of $B_g$. Since $B_g$ has the presentation with generators ${\sigma}_1,\ldots,{\sigma}_{g-1}$ and defining relations (A1), (A2), this completes the proof.
Theorem \[th:001\] can be also algebraically deduced from Theorem \[th:00\]. Since ${{{\mathcal{M}}}}(S_0^{g,1})$ is a subgroup of index $g+1$ in ${{{\mathcal{M}}}}(S_0^{g+1})$, for the Schreier transversal we can take $$(1,{\sigma}_{g},{\sigma}_{g}{\sigma}_{g-1},\ldots,{\sigma}_{g}{\sigma}_{g-1}\cdots {\sigma}_1).$$ If we now apply Reidemeister-Schreier process, as generators for ${{{\mathcal{M}}}}(S_0^{g,1})$ we get ${\sigma}_1,\ldots,{\sigma}_{g-1}$ and additionally $\tau_1,\ldots,\tau_g$ where $$\tau_k=\begin{cases}
{\sigma}_g\cdots{\sigma}_{k+1}{\sigma}_k^2{\sigma}_{k+1}^{-1}\cdots{\sigma}_g^{-1}&\text{for $k=1,\ldots,g-1$}\\
{\sigma}_g^2&\text{for $k=g$}.
\end{cases}$$ As defining relations we get
1. ${\sigma}_k{\sigma}_j={\sigma}_j{\sigma}_k$ for $|k-j|>1$ and $k,j<g$,
2. ${\sigma}_k\tau_j=\tau_j{\sigma}_k$ for $j\neq k,k+1$,
3. ${\sigma}_j{\sigma}_{j+1}{\sigma}_j={\sigma}_{j+1}{\sigma}_j{\sigma}_{j+1}$ for $j=1,2,\ldots,g-2$,
4. ${\sigma}_k\tau_{k+1}{\sigma}_{k}^{-1}=\tau_{k+1}^{-1}\tau_k\tau_{k+1}$ for $k=1,2,\ldots,g-1$,
5. ${\sigma}_k\tau_k{\sigma}_{k}^{-1}=\tau_{k+1}$ for $k=1,2,\ldots,g-1$,
6. $\tau_1\tau_2\cdots \tau_g=1$,
7. ${\sigma}_{g-1}\cdots{\sigma}_2{\sigma}_1\tau_1{\sigma}_1{\sigma}_2\cdots {\sigma}_{g-1}=1$,
8. $\left({\sigma}_{g-1}{\sigma}_{g-2}\cdots{\sigma}_{1}\tau_1\right)^g=1$.
If we now remove generators $\tau_1,\ldots,\tau_g$ from the above presentation, we obtain the presentation given by Theorem \[th:001\]. The computations are lengthy, but completely straightforward.
Recall that by ${{{\mathcal{M}}}}^{\pm}(S_0^{g,1})$ we denote the extended mapping class group of the sphere $S_0^{g,1}$, that is the extension of degree 2 of ${{{\mathcal{M}}}}(S_0^{g,1})$. Suppose that the sphere $S_0^{g,1}$ is the metric sphere in ${{\mathbb{R}}}^3$ with origin $(0,0,0)$ and that punctures $p_1,\ldots,p_g$ are contained in the $xy$-plane. Let ${{\sigma}\colon {S_0^{g,1}}\to {S_0^{g,1}}}$ be the map induced by the reflection across the $xy$-plane. We have the short exact sequence. $$\xymatrix{
1\ar[r]&{{{\mathcal{M}}}}(S_0^{g,1})\ar[r]&{{{\mathcal{M}}}}^{\pm}(S_0^{g,1})\ar[r]&{\langle {\sigma}\rangle}\ar[r]&1.
}$$ Moreover, ${\sigma}{\sigma}_i{\sigma}^{-1}={\sigma}_i^{-1}$ for $i=1,\ldots,g-1$. Therefore Theorem \[th:001\] implies the following.
\[th:003\] If $g\geqslant 3$, then ${{{\mathcal{M}}}}^{\pm}(S_0^{g,1})$ has the presentation with generators ${\sigma}_1,\ldots,{\sigma}_{g-1},{\sigma}$ and defining relations:
1. ${\sigma}_k{\sigma}_j={\sigma}_j{\sigma}_k$ for $|k-j|>1$ and $k,j<g$,
2. ${\sigma}_j{\sigma}_{j+1}{\sigma}_j={\sigma}_{j+1}{\sigma}_j{\sigma}_{j+1}$ for $j=1,2,\ldots,g-2$,
3. $({\sigma}_1\cdots {\sigma}_{g-1})^g=1$,
4. ${\sigma}^2=1$,
5. ${\sigma}{\sigma}_i{\sigma}={\sigma}_i^{-1}$ for $i=1,2,\ldots,g-1$.
Presentations for groups ${{{\mathcal{M}}}}^h(N_g)$ and ${{{\mathcal{M}}}}^{h+}(N_g)$
=====================================================================================
By Theorem \[tw:002\] there is a short exact sequence. $$\xymatrix{
1\ar[r]&{\langle {\varrho}\rangle}\ar[r]&{{{\mathcal{M}}}}^h(N_g)\ar[r]^{\pi_{\varrho}}&{{{\mathcal{M}}}}^{\pm}(S_0^{g,1})\ar[r]&1.
}$$ Moreover, it is known that as lifts of braids ${\sigma}_1,\ldots,{\sigma}_{g-1}\in{{{\mathcal{M}}}}^{\pm}(S_0^{g,1})$ we can take Dehn twists $t_{a_1},\ldots,t_{a_{g-1}}\in {{{\mathcal{M}}}}^h(N_g)$ about circles $a_1,\ldots,a_{g-1}$ – cf Figure \[r02\] (small arrows in this picture indicate directions of twists). As a lift of ${\sigma}$ we take the symmetry $s$ across the $xy$ plane (the second lift of ${\sigma}$ is the symmetry ${\varrho}s$, that is the symmetry across the $yz$ plane).
To obtain a presentation for the group ${{{\mathcal{M}}}}^h(N_g)$ we need to lift relations (B1)–(B5) of Theorem \[th:003\]. Each relation of the form $$w({\sigma}_1,\ldots,{\sigma}_{g-1},{\sigma})=1$$ lifts either to $w(t_{a_1},\ldots,t_{a_{g-1}},s)=1$ or to $w(t_{a_1},\ldots,t_{a_{g-1}},s)={\varrho}$. In order to determine which of these two cases does occur it is enough to check whether the homeomorphism $w(t_{a_1},\ldots,t_{a_{g-1}},s)$ changes the orientation of the circle $a_1$ or not. This can be easily done and as a result we obtain the following theorem.
\[th:004\] If $g\geqslant 3$, then ${{{\mathcal{M}}}}^h(N_g)$ has the presentation with generators $t_{a_1},\ldots,t_{a_{g-1}},s,{\varrho}$ and defining relations:
1. $t_{a_k}t_{a_j}=t_{a_j}t_{a_k}$ for $|k-j|>1$ and $k,j<g$,
2. $t_{a_j}t_{a_{j+1}}t_{a_j}=t_{a_{j+1}}t_{a_j}t_{a_{j+1}}$ for $j=1,2,\ldots,g-2$,
3. $(t_{a_{1}}\cdots t_{a_{g-1}})^g=\begin{cases}
1&\text{for $g$ even}\\
{\varrho}&\text{for $g$ odd,}\\\end{cases}$
4. $s^2=1$,
5. $st_{a_j}s=t_{a_j}^{-1}$ for $j=1,2,\ldots,g-1$,
6. ${\varrho}^2=1$,
7. ${\varrho}t_{a_j}{\varrho}=t_{a_j}$ for $j=1,2,\ldots,g-1$,
8. ${\varrho}s{\varrho}=s$.
\[cr:001\] If $g\geqslant 3$, then $$H_1({{{\mathcal{M}}}}^h(N_g))=\begin{cases}
{{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2&\text{for $g$ odd}\\
{{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2&\text{for $g$ even.}
\end{cases}$$
Relation (C2) implies that the abelianization of the group ${{{\mathcal{M}}}}^h(N_g)$ is an abelian group generated by $t_{a_1},s,{\varrho}$. Defining relations take form $$\begin{aligned}
&t_{a_1}^{(g-1)g}=\begin{cases}
1&\text{for $g$ even}\\
{\varrho}&\text{for $g$ odd}\\\end{cases}\\
&s^2=1,\quad t_{a_1}^2=1,\quad {\varrho}^2=1.
\end{aligned}$$ Hence $H_1({{{\mathcal{M}}}}^h(N_g))={\langle t_{a_1},s \rangle}\simeq {{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2$ for $g$ odd and $H_1({{{\mathcal{M}}}}^h(N_g))={\langle t_{a_1},s,{\varrho}\rangle}\simeq
{{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2$ for $g$ even.
The main theorem of [@Chil] implies that the group ${{{\mathcal{M}}}}(N_3)$ is generated by $a_1,a_2$ and a crosscap slide which commutes with ${\varrho}$. Hence ${{{\mathcal{M}}}}^h(N_3)={{{\mathcal{M}}}}(N_3)$ and Theorem \[th:004\] implies the following.
\[Presen:g3:Bir\] The group ${{{\mathcal{M}}}}(N_3)$ has the presentation with generators $t_{a_1},t_{a_2},s$ and defining relations:
1. $t_{a_1}t_{a_2}t_{a_1}=t_{a_2}t_{a_1}t_{a_2}$,
2. $(t_{a_1}t_{a_2}t_{a_1})^4=1$,
3. $s^2=1$,
4. $st_{a_j}s=t_{a_j}^{-1}$ for $j=1,2$.
By Theorem \[th:004\], the group ${{{\mathcal{M}}}}(N_3)$ is generated by $t_{a_1},t_{a_2},{\varrho},s$ with defining relations:
1. $t_{a_1}t_{a_{2}}t_{a_1}=t_{a_{2}}t_{a_1}t_{a_{2}}$,
2. $(t_{a_{1}}t_{a_2})^3={\varrho}$,
3. $s^2=1$,
4. $st_{a_j}s=t_{a_j}^{-1}$ for $j=1,2$,
5. ${\varrho}^2=1$,
6. ${\varrho}t_{a_j}{\varrho}=t_{a_j}$ for $j=1,2$,
7. ${\varrho}s{\varrho}=s$.
Using (C2), we can rewrite (C3) in the form $$\begin{aligned}
{\varrho}=t_{a_{1}}t_{a_2}t_{a_{1}}(t_{a_2}t_{a_{1}}t_{a_2})=t_{a_{1}}t_{a_2}t_{a_{1}}(t_{a_1}t_{a_{2}}t_{a_1
})=(t_{a_1}t_{a_{2}}t_{a_1})^2.
\end{aligned}$$ Hence we can remove ${\varrho}$ from the generating set and then (C6) will transform into (D2). It remains to check that relations (C7) and (C8) are superfluous. Let start with (C7). $$\begin{aligned}
t_{a_1}{\varrho}t_{a_1}^{-1}&=t_{a_1}(t_{a_1}t_{a_{2}}t_{a_1})(t_{a_1}t_{a_{2}}t_{a_1})t_{a_1}^{-1}\\
&=t_{a_1}(t_{a_2}t_{a_{1}}t_{a_2})(t_{a_1}t_{a_{2}}t_{a_1})t_{a_1}^{-1}=
(t_{a_1}t_{a_2}t_{a_{1}})(t_{a_1}t_{a_2}t_{a_{1}})={\varrho}\\
t_{a_2}{\varrho}t_{a_2}^{-1}&=t_{a_2}(t_{a_1}t_{a_{2}}t_{a_1})(t_{a_1}t_{a_{2}}t_{a_1})t_{a_2}^{-1}\\
&=t_{a_2}(t_{a_1}t_{a_{2}}t_{a_1})(t_{a_2}t_{a_{1}}t_{a_2})t_{a_2}^{-1}=
(t_{a_1}t_{a_2}t_{a_{1}})(t_{a_1}t_{a_2}t_{a_{1}})={\varrho}\end{aligned}$$ Now we check (C8). $$\begin{aligned}
s{\varrho}s&=
s(t_{a_1}t_{a_{2}}t_{a_1})^2s=(t_{a_1}^{-1}t_{a_{2}}^{-1}t_{a_1}^{-1})^2=(t_{a_1}t_{a_2}t_{a_1})^{-2
} = (t_{a_1}t_{a_2}t_{a_1})^{2}={\varrho}.
\end{aligned}$$
By restricting homomorphism ${\pi_{\varrho}\colon {{{\mathcal{M}}}}^{h}(N_g)\to {{{\mathcal{M}}}}^{\pm}(S_0^{g,1})}$ to the subgroup ${{{\mathcal{M}}}}^{h+}(N_g)$ we obtain the exact sequence $$\xymatrix{
1\ar[r]&{\langle {\varrho}\rangle}\ar[r]&{{{\mathcal{M}}}}^{h+}(N_g)\ar[r]^{\pi_{\varrho}}&{{{\mathcal{M}}}}(S_0^{g,1})\ar[r]&1.
}$$ Now if we lift the presentation from Theorem \[th:001\], we get the following.
\[tw:pres:pos\] If $g\geqslant 3$, then ${{{\mathcal{M}}}}^{h+}(N_g)$ has the presentation with generators $t_{a_1},\ldots,t_{a_{g-1}},{\varrho}$ and defining relations:
1. $t_{a_k}t_{a_j}=t_{a_j}t_{a_k}$ for $|k-j|>1$ and $k,j<g$,
2. $t_{a_j}t_{a_{j+1}}t_{a_j}=t_{a_{j+1}}t_{a_j}t_{a_{j+1}}$ for $j=1,2,\ldots,g-2$,
3. $(t_{a_{1}}\cdots t_{a_{g-1}})^g=\begin{cases}
1&\text{for $g$ even}\\
{\varrho}&\text{for $g$ odd,}\\\end{cases}$
4. ${\varrho}^2=1$,
5. ${\varrho}t_{a_j}{\varrho}=t_{a_j}$ for $j=1,2,\ldots,g-1$.
\[cr:002\] If $g\geqslant 3$, then $$H_1({{{\mathcal{M}}}}^{h+}(N_g))=\begin{cases}
{{\mathbb{Z}}}_{2(g-1)g}&\text{for $g$ odd}\\
{{\mathbb{Z}}}_{(g-1)g}\oplus{{\mathbb{Z}}}_2&\text{for $g$ even.}
\end{cases}$$
Relation (E2) implies that the abelianization of the group ${{{\mathcal{M}}}}^{h+}(N_g)$ is an abelian group generated by $t_{a_1},{\varrho}$. Defining relations take form: $$\begin{aligned}
&t_{a_1}^{(g-1)g}=\begin{cases}
1&\text{for $g$ even}\\
{\varrho}&\text{for $g$ odd,}\\\end{cases}\\
&{\varrho}^2=1.
\end{aligned}$$ Hence $H_1({{{\mathcal{M}}}}^{h+}(N_g))={\langle t_{a_1} \rangle}\simeq {{\mathbb{Z}}}_{2(g-1)g}$ for $g$ odd and $H_1({{{\mathcal{M}}}}^{h+}(N_g))={\langle t_{a_1},{\varrho}\rangle}\simeq
{{\mathbb{Z}}}_{(g-1)g}\oplus{{\mathbb{Z}}}_2$ for $g$ even.
\[uw:perspective\] To put Corollaries \[cr:001\] and \[cr:002\] into perspective, recall that in the oriented case (Theorem 8 of [@Bir-Hil]), $$\begin{aligned}
{{{\mathcal{M}}}}^{h}(S_g)=\langle& t_{a_1},\ldots,t_{a_{2g+1}},{\varrho}{\;|\;}t_{a_k}t_{a_j}=t_{a_j}t_{a_k},\ t_{a_j}t_{a_{j+1}}t_{a_j}=t_{a_{j+1}}t_{a_j}t_{a_{j+1}},\\
&(t_{a_1}t_{a_2}\cdots t_{a_{2g+1}})^{2g+2}=1,\ {\varrho}=t_{a_1}t_{a_2}\cdots t_{a_{2g+1}}t_{a_{2g+1}}\cdots t_{a_2}t_{a_1},\\ &{\varrho}^2=1, {\varrho}t_{a_1}{\varrho}=t_{a_1}\rangle,\quad
\text{where $j=1,2,\ldots,2g$, $|k-j|>1$}.
\end{aligned}$$ The presentation for the group ${{{\mathcal{M}}}}^{h\pm}(S_g)$ is obtained from the above presentation by adding one generator $s$ and three relations: $$s^2=1,\ st_{a_1}s=t_{a_1}^{-1},\ {\varrho}s{\varrho}=s.$$ Consequently, $H_1({{{\mathcal{M}}}}^{h\pm}(S_g))={{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2$ and $$H_1({{{\mathcal{M}}}}^{h}(S_g))=\begin{cases}
{{\mathbb{Z}}}_{4g+2}&\text{for $g$ even}\\
{{\mathbb{Z}}}_{8g+4}&\text{for $g$ odd.}\\\end{cases}$$ This suggests that algebraically the group ${{{\mathcal{M}}}}^{h+}(N)$ corresponds to ${{{\mathcal{M}}}}^h(S)$, whereas ${{{\mathcal{M}}}}^{h}(N)$ corresponds to ${{{\mathcal{M}}}}^{h\pm}(S)$.
Computing $H_1({{{\mathcal{M}}}}^{h+}(N_g); H_1(N_g;{{\mathbb{Z}}}))$ and $H_1({{{\mathcal{M}}}}^{h}(N_g); H_1(N_g;{{\mathbb{Z}}}))$
====================================================================================================================================
Homology of groups
------------------
Let us briefly review how to compute the first homology of a group with twisted coefficients. Our exposition follows [@Brown; @Tanaka].
For a given group $G$ and $G$-module $M$ (that is ${{\mathbb{Z}}}G$-module) we define the *bar resolution* which is a chain complex $(C_n(G))$ of $G$-modules, where $C_n(G)$ is the free $G$-module generated by symbols $[h_1|\cdots|h_n]$, $h_i\in G$. For $n=0$, $C_0(G)$ is the free module generated by the empty bracket $[\cdot]$. Our interest will restrict to groups $C_2(G),C_1(G),C_0(G)$ for which the boundary operator ${\partial_n\colon C_n(G)\to C_{n-1}(G)}$ is defined by formulas: $$\begin{aligned}
\partial_2([h_1|h_2])&=h_1[h_2]-[h_1h_2]+[h_1],\\
\partial_1([h])&=h[\cdot]-[\cdot].
\end{aligned}$$
The homology of $G$ with coefficients in $M$ is defined as the homology groups of the chain complex $(C_n(G)\otimes M)$, where the chain complexes are tensored over ${{\mathbb{Z}}}G$. In particular, $H_1(G;M)$ is the first homology group of the complex $$\xymatrix@C=3pc@R=3pc{C_2(G)\otimes M\ar[r]^{\ \partial_2\otimes {\rm id}}&C_1(G)\otimes M
\ar[r]^{\ \partial_1\otimes {\rm id}}&C_0(G)\otimes M.}$$ For simplicity, we denote $\partial\otimes {\rm id}={\overline{\partial}}$ henceforth.
If the group $G$ has a presentation $G=\langle X\,|\,R\rangle$, denote by $${\langle {\overline{X}} \rangle}={\langle [x]\otimes m{\;|\;}x\in X, m\in M \rangle}{\subseteq}C_1(G)\otimes M.$$ Then, using the formula for $\partial_2$, one can show that $H_1(G;M)$ is a quotient of ${\langle {\overline{X}} \rangle}\cap \ker{\overline{\partial}}_1$.
The kernel of this quotient corresponds to relations in $G$ (that is elements of $R$). To be more precise, if $r\in R$ has the form $x_1\cdots x_k=y_1\cdots y_n$ and $m\in M$, then $r$ gives the relation (in $H_1(G;M)$) $${\overline{r}}\otimes m\!:\ \sum_{i=1}^{k}x_1\cdots x_{i-1}[x_i]\otimes m=\sum_{i=1}^{n}y_1\cdots y_{i-1}[y_i]\otimes m.\label{eq_rew_rel}$$ Then $$H_1(G;M)={\langle {\overline{X}} \rangle}\cap \ker{\overline{\partial}}_1/{\langle {\overline{R}} \rangle},$$ where $${\overline{R}}=\{{\overline{r}}\otimes m{\;|\;}r\in R,m\in M\}.$$
Action of ${{{\mathcal{M}}}}^{h}(N_g)$ on $H_1(N_g;{{\mathbb{Z}}})$
-------------------------------------------------------------------
Let $c_1,\ldots,c_g$ be one-sided circles indicated in Figure \[r04\]. In this figure surface $N_g$ is represented as the sphere with $g$ crosscaps (the shaded disks represent crosscaps, hence their interiors are to be removed and then the antipodal points on each boundary component are to be identified).
![Surface $N_g$ as a sphere with crosscaps.[]{data-label="r04"}](R04){width="80.00000%"}
The same set of circles is also indicated in Figure \[r02\] – for a method of transferring circles between two models of $N_g$ see Section 3 of [@Stukow_HomTw].
Recall that $H_1(N_g;{{\mathbb{Z}}})$ as a ${{\mathbb{Z}}}$-module is generated by $\gamma_1=[c_1],\ldots,\gamma_g=[c_g]$ with respect to the single relation $$2(\gamma_1+\gamma_2+\cdots+\gamma_g)=0.$$ There is a ${{\mathbb{Z}}}_2$-valued intersection paring $\langle\,,\rangle$ on $H_1(N_g;{{\mathbb{Z}}})$ defined as the symmetric bilinear form (with values in ${{\mathbb{Z}}}_2$) satisfying $\langle\gamma_i,\gamma_j\rangle=\delta_{ij}$ for $1\leqslant i,j\leqslant g$. The mapping class group ${{{\mathcal{M}}}}(N_g)$ acts on $H_1(N_g;{{\mathbb{Z}}})$ via automorphisms which preserve $\langle\,,\rangle$, hence there is a representation $${\psi\colon {{{\mathcal{M}}}}(N_g)\to {\rm Iso}(H_1(N_g;{{\mathbb{Z}}}))}.$$ In fact it is known that this representation is surjective – see [@Pinkal; @Gadgil].
Since we have very simple geometric definitions of $t_{a_i},s,{\varrho}\in {{{\mathcal{M}}}}^h(N_g)$ it is straightforward to check that $$\begin{aligned}
\psi(t_{a_i})&=I_{i-1}\oplus \begin{bmatrix}
0&1\\
-1&2\end{bmatrix}\oplus I_{g-i-1}\\
\psi(t_{a_i}^{-1})&=I_{i-1}\oplus \begin{bmatrix}
2&-1\\
1&0
\end{bmatrix}\oplus I_{g-i-1}\\
\psi(s)&=\begin{bmatrix}
-1&2&-2&2&\ldots&(-1)^{g}\cdot 2\\
0&1&-2&2&\ldots&(-1)^{g}\cdot 2\\
0&0&-1&2&\ldots&(-1)^{g}\cdot 2\\
0&0&0&1&\ldots&(-1)^{g}\cdot 2\\
\vdots&\vdots&\vdots&\vdots&\ddots&\vdots\\
0&0&0&0&\ldots& (-1)^{g}\cdot 1
\end{bmatrix}
\\
\psi({\varrho})&=-I_g
\end{aligned}$$ where $I_k$ is the identity matrix of rank $k$.
The above matrices are written with respect to the generating set $(\gamma_1,\gamma_2,\ldots,\gamma_g)$. Note that $H_1(N_g;{{\mathbb{Z}}})$ is not free, hence one has to be careful with matrices – two different matrices may represent the same element.
Computing ${\langle {\overline{X}} \rangle}\cap \ker {\overline{\partial}}_1$
-----------------------------------------------------------------------------
Observe that if $G={{{\mathcal{M}}}}^h(N_g)$, $M=H_1(N_g;{{\mathbb{Z}}})$ and $h\in G$ then $${\overline{\partial}}_1([h]\otimes \gamma_j)=(h-1)[\cdot]\otimes \gamma_j=[\cdot]\otimes(\psi(h)^{-1}-I_g)\gamma_j.$$ If we identify $C_0(G)\otimes M$ with $M$ by the map $[\cdot]\otimes m\mapsto m$, this formula takes form $${\overline{\partial}}_1([h]\otimes \gamma_j)=(\psi(h)^{-1}-I_g)\gamma_j.$$ Let us denote $[{\varrho}]\otimes \gamma_j, [s]\otimes \gamma_j,
[t_{a_i}]\otimes \gamma_j$ respectively by ${\varrho}_j,s_j$ and $t_{i,j}$. Using the above formula, we obtain $$\begin{aligned}
{\overline{\partial}}_1({\varrho}_j)&=-2\gamma_j\\
{\overline{\partial}}_1(s_j)&=\begin{cases}
-2\sum_{k=1}^{j}\gamma_k &\text{for $j$ odd}\\
-{\overline{\partial}}_1(s_{j-1}) &\text{for $j$ even}
\end{cases}
\\
{\overline{\partial}}_1(t_{i,j})&=\begin{cases}
\gamma_i+\gamma_{i+1}&\text{for $j=i$}\\
-\gamma_i-\gamma_{i+1}&\text{for $j=i+1$}\\
0&\text{otherwise.}
\end{cases}
\end{aligned}$$
Let $g\geqslant 3$ and $G={{{\mathcal{M}}}}^{h+}(N_g)$ then ${\langle {\overline{X}} \rangle}\cap\ker{\overline{\partial}}_1$ is the abelian group which admits the presentation with generators:
1. $t_{i,j}$, where $i=1,\ldots,g-1$ and $j=1,\ldots,i-1,i+2,\ldots,g$
2. $t_{j,j}+t_{j,j+1}$, where $j=1,\ldots,g-1$
3. $2t_{j,j}+{\varrho}_j+{\varrho}_{j+1}$ , where $j=1,\ldots,g-1$
4. $\begin{cases}
2t_{1,1}+2t_{3,3}+\cdots+2t_{g-2,g-2}-{\varrho}_g&\text{for $g$ odd}\\
2t_{1,1}+2t_{3,3}+\cdots+2t_{g-1,g-1}&\text{for $g$ even}
\end{cases}$
and relations $$\begin{aligned}
r_{t_j}\!&:0=2t_{j,1}+\cdots +2(t_{j,j}+t_{j,j+1})+\cdots+2t_{j,g}\quad\text{for $j=1,\ldots,g-1$}\\
r_{{\varrho}}\!&:\begin{cases}
2(2t_{1,1}+{\varrho}_1+{\varrho}_2)+\cdots+2(2t_{g-2,g-2}+{\varrho}_{g-2}+{\varrho}_{g-1})\\
\qquad\qquad =2(2t_{1,1}+2t_{3,3}+\cdots+2t_{g-2,g-2}-{\varrho}_g)\quad \text{for $g$ odd}\\
2(2t_{1,1}+{\varrho}_1+{\varrho}_2)+\cdots+2(2t_{g-1,g-1}+{\varrho}_{g-1}+{\varrho}_{g})\\
\qquad\qquad\qquad =2(2t_{1,1}+2t_{3,3}+\cdots+2t_{g-1,g-1})\quad \text{for $g$ even.}
\end{cases}
\end{aligned}$$
By Theorem \[tw:pres:pos\], ${\langle {\overline{X}} \rangle}$ is generated by $t_{i,j}$ and ${\varrho}_j$. Using formulas for ${\overline{\partial}}_1(t_{i,j})$ and ${\overline{\partial}}_1({\varrho}_j)$ it is straightforward to check that elements (F1)–(F4) are elements of $\ker{\overline{\partial}}_1$. Moreover, $$2t_{j,1}+2t_{j,2}+\cdots+2t_{j,g}=[t_{a_j}]\otimes 2(\gamma_1+\cdots +\gamma_g)=0,$$ hence $r_{t_j}$ is indeed a relation. Similarly we check that $r_{{\varrho}}$ is a relation.
Observe that using relations $r_{t_j}$ and $r_{{\varrho}}$ we can substitute for $2t_{j,g}$ and $2{\varrho}_1$ respectively, hence each element in ${\langle {\overline{X}} \rangle}$ can be written as a linear combination of $t_{i,j},{\varrho}_j$, where each of $t_{1,g},t_{2,g},\ldots,t_{g-1,g},{\varrho}_1$ has the coefficient 0 or 1. Moreover, for a given $x\in {\langle {\overline{X}} \rangle}\subset C_1(G)\otimes H_1(N_g;{{\mathbb{Z}}})$ such a combination is unique. Hence for the rest of the proof we assume that linear combinations of $t_{i,j},{\varrho}_j$ satisfy this condition.
Suppose that $h\in{\langle {\overline{X}} \rangle}\cap \ker{\overline{\partial}}_1$. We will show that $h$ can be uniquely expressed as a linear combination of generators (F1)–(F4).
First observe that $h=h_1+h_2$, where $h_1$ is a combination of generators (F1)–(F2), and $h_2$ does not contain generators of type (F1) nor elements $t_{j,j+1}$. Moreover, $h_1$ and $h_2$ are uniquely determined by $h$.
Next we decompose $h_2=h_3+h_4$, where $h_3$ is a combination of generators (F3) and $h_4$ does not contain ${\varrho}_j$ for $j<g$. As before, $h_3$ and $h_4$ are uniquely determined by $h_2$.
Element $h_4$ has the form $$h_4=\sum_{j=1}^{g-1} \alpha_jt_{j,j}+\alpha {\varrho}_g,$$ for some integers $\alpha,\alpha_1,\ldots,\alpha_{g-1}$. Hence $$0={\overline{\partial}}_1(h_4)=\alpha_1\gamma_1+(\alpha_1+\alpha_2)\gamma_2+\cdots+(\alpha_{g-2}+
\alpha_{g-1})\gamma_{g-1}+(\alpha_{g-1}-2\alpha)\gamma_g.$$ If $g$ is odd this implies that $$\alpha_1=\alpha_3=\cdots=\alpha_{g-2}=2k,\ \alpha_2=\alpha_4=\cdots=\alpha_{g-1}=0,\ \alpha=-k,$$ for some $k\in{{\mathbb{Z}}}$. For $g$ even we get $$\alpha_1=\alpha_3=\cdots=\alpha_{g-1}=2k,\ \alpha=\alpha_2=\alpha_4=\cdots=\alpha_{g-2}=0.$$ In each of these cases $h_4$ is a multiple of the generator (F4).
By an analogous argument we get
\[gen\_ker\_roz\] Let $g\geqslant 3$ and $G={{{\mathcal{M}}}}^h(N_g)$ then ${\langle {\overline{X}} \rangle}\cap\ker{\overline{\partial}}_1$ is the abelian group which admits the presentation with generators: (F1)–(F4),
1. $s_j+s_{j-1}$, where $j$ is even,
2. $s_j-{\varrho}_1-{\varrho}_2-\cdots-{\varrho}_j$, where $j$ is odd.
The defining relations are $r_{t_j}$, $r_{{\varrho}}$ and $$r_{s}\!:\begin{cases}
0=2(s_2+s_1)+2(s_4+s_3)+\cdots+2(s_{g-1}+s_{g-2})\\
\qquad\qquad\qquad\qquad\ +2(s_{g}-{\varrho}_1-{\varrho}_2-\cdots-{\varrho}_g)\quad\text{for $g$ odd}\\
0=2(s_2+s_1)+2(s_4+s_3)+\cdots+2(s_g+s_{g-1})\quad\text{for $g$ even.}
\end{cases}$$
Rewriting relations
-------------------
Using formula we rewrite relations (E1)–(E5) as relations in $H_1({{{\mathcal{M}}}}^{h+}(N_g); H_1(N_g;{{\mathbb{Z}}}))$.
Relation (E1) is symmetric with respect to $k$ and $j$, hence we can assume that $j+1<k$. This relation gives $$\begin{aligned}r^{(E1)}_{k,j:i}\!:0&=([t_{a_k}]+t_{a_k}[t_{a_j}]-[t_{a_j}]-t_{a_j}[t_{a_k}])\otimes \gamma_i\\
&=t_{k,i}+[t_{a_j}]\otimes \psi(t_{a_k}^{-1})\gamma_i-t_{j,i}-[t_{a_k}]\otimes \psi(t_{a_j}^{-1})\gamma_i\\
&=\pm\begin{cases}
0&\text{if $i\neq k,k+1,j,j+1$}\\
t_{j,k}+t_{j,k+1}&\text{if $i=k$ or $i=k+1$}\\
t_{k,j}+t_{k,j+1}&\text{if $i=j$ or $i=j+1$.}
\end{cases}
\end{aligned}$$ Relation (E2) gives $$\begin{aligned}r^{(E2)}_{j:i}\!:0=&
([t_{a_j}]+t_{a_{j}}[t_{a_{j+1}}]+t_{a_{j}}t_{a_{j+1}}[t_{a_{j}}]\\
&-[t_{a_{j+1}}]-t_{a_{j+1}}[t_{a_{j}}]-t_{a_{j+1}}t_{a_{j}}[t_{a_{j+1}}])\otimes \gamma_i\\
=&\begin{cases}
t_{j,i}-t_{j+1,i}&\text{\hspace{-1.5cm}if $i\neq j,j+1,j+2$}\\
t_{j,j+2}-t_{j+1,j}&\text{\hspace{-1.5cm}if $i=j+2$}\\
(*)+2(t_{j,j}+t_{j,j+1})&\text{\hspace{-1.5cm}if $i=j$}\\
(*)-(t_{j,j}+t_{j,j+1})-(t_{j+1,j+1}+t_{j+1,j+2})&\text{if $i=j+1$.}
\end{cases}
\end{aligned}$$ In the above formula $(*)$ denotes some expression homologous to 0 by previously obtained relations. Carefully checking relations $r^{(E1)}_{k,j:i}$ and $r^{(E2)}_{j:i}$ we conclude that generators (F1) generate a cyclic group, and generators (F2) generate a cyclic group of order at most 2.
We next turn to the relation (E5). It gives $$\begin{aligned}r^{(E5)}_{j:i}\!:0=&([{\varrho}]+{\varrho}[t_{a_j}]-[t_{a_j}]-t_{a_j}[{\varrho}])\otimes \gamma_i\\
&=\begin{cases}
-2t_{j,i}&\text{if $i\neq j,j+1$}\\
-2t_{j,j}-{\varrho}_j-{\varrho}_{j+1}&\text{if $i=j$}\\
({\varrho}_j+{\varrho}_{j+1}+2t_{j,j})-2(t_{j,j}+t_{j,j+1})&\text{if $i=j+1$.}
\end{cases}
\end{aligned}$$ These relations imply that generators (F3) are homologically trivial, and generators (F1) generate at most ${{\mathbb{Z}}}_2$.
We now turn to the most difficult relation, namely (E3). This relation gives $$\begin{aligned}
r^{(E3)}_{i}\!:0&=\sum_{k=0}^{g-1}\sum_{n=1}^{g-1}(t_{a_1}\cdots t_{a_{g-1}})^kt_{a_1}\cdots t_{a_{n-1}}[t_{a_n}]\otimes \gamma_i-{\varepsilon}{\varrho}_i\\
&=\sum_{n=1}^{g-1}[t_{a_n}]\otimes \psi(t_{a_1}\cdots t_{a_{n-1}})^{-1} \sum_{k=0}^{g-1}\psi(t_{a_1}\cdots t_{a_{g-1}})^{-k}\gamma_i-{\varepsilon}{\varrho}_i\\
&=\sum_{n=1}^{g-1}[t_{a_n}]\otimes Y_{n} \sum_{k=0}^{g-1}Y_{g}^k\gamma_i-{\varepsilon}{\varrho}_i.
\end{aligned}$$ Where ${\varepsilon}=0$ for $g$ even, ${\varepsilon}=1$ for $g$ odd, and $Y_n=\psi(t_{a_1}\cdots t_{a_{n-1}})^{-1}$. Using the matrix formula for $\psi(t_{a_i}^{-1})$, we obtain $$Y_n\gamma_i=\begin{cases}
-\gamma_{i-1}&\text{if $2\leq i\leq n$}\\
\gamma_{i}&\text{if $i> n$}\\
2\gamma_1+\cdots+2\gamma_{n-1}+\gamma_n&\text{if $i=1$.}
\end{cases}$$ In particular $$Y_g^k\gamma_i=(-1)^k\gamma_{i-k},$$ where we subtract indexes modulo $g$. Therefore we have $$r^{(E3)}_{i}\!:0=\sum_{n=1}^{g-1}[t_{a_n}]\otimes Y_{n}\sum_{k=0}^{g-1}(-1)^k\gamma_{i-k}-{\varepsilon}{\varrho}_i.$$ In order to simplify computations we replace relations: $$r_1^{(E3)},\ r_2^{(E3)},\ldots,\ r_g^{(E3)}$$ with relations: $$r_1^{(E3)}+r_2^{(E3)},\ r_2^{(E3)}+r_3^{(E3)},\ldots,r_{g-1}^{(E3)}+r_g^{(E3)},\ r_g^{(E3)}.$$ Let us begin with $r_g^{(E3)}$. $$\begin{aligned}
r_g^{(E3)}\!:0=&\sum_{n=1}^{g-1}[t_{a_n}]\otimes Y_{n}\sum_{k=0}^{g-1}(-1)^k\gamma_{g-k}-{\varepsilon}{\varrho}_{g}\\
=&\sum_{n=1}^{g-1}[t_{a_n}]\otimes \left(\sum_{k=0}^{g-n-1}(-1)^k\gamma_{g-k}+\sum_{k=g-n}^{g-2}(-1)^{k+1}\gamma_{g-k-1}\right.\\
&\qquad + (-1)^{g-1}(2\gamma_1+\cdots+2\gamma_{n-1}+\gamma_n)\bigg)-{\varepsilon}{\varrho}_{g}.
\end{aligned}$$ Since all generators of type (F1) are homologous to a single generator, say $t$, and $2t=0$, the above relation can be rewritten as $$r_g^{(E3)}\!:0=
(g-1)(g-2)t+\sum_{n=1}^{g-1}[t_{a_n}]\otimes \left((-1)^{g-n-1}\gamma_{n+1}+(-1)^{g-1}\gamma_n\right)-{\varepsilon}{\varrho}_{g}.$$ If $g$ is even, this gives the relation $$\begin{aligned}
r_g^{(E3)}\!:0=&(-t_{1,1}+ t_{1,2})+(-t_{2,2}-t_{2,3})+\cdots+(-t_{g-1,g-1}+t_{g-1,g})\\
=&(t_{1,1}+ t_{1,2})-(t_{2,2}+t_{2,3})+\cdots+(t_{g-1,g-1}+t_{g-1,g})\\
&-2(t_{1,1}+t_{3,3}+\cdots+t_{g-1,g-1}).
\end{aligned}$$ If $g$ is odd, we have $$\begin{aligned}
r_g^{(E3)}\!:0=&(t_{1,1}- t_{1,2})+(t_{2,2}+t_{2,3})+\cdots+(t_{g-1,g-1}+t_{g-1,g})-{\varrho}_g\\
=&-(t_{1,1}+ t_{1,2})+(t_{2,2}+t_{2,3})-\cdots+(t_{g-1,g-1}+t_{g-1,g})\\
&+2(t_{1,1}+t_{3,3}+\cdots+t_{g-2,g-2})-{\varrho}_g.
\end{aligned}$$ In both cases relation $r_g^{(E3)}$ implies that generator (F4) is superfluous.
Now we concentrate on the relation $r_i^{(E3)}+r_{i+1}^{(E3)}$. $$\begin{aligned}
r_i^{(E3)}+r_{i+1}^{(E3)}&\!:0=\sum_{n=1}^{g-1}[t_{a_n}]\otimes Y_{n}
\sum_{k=0}^{g-1}(-1)^k\left(\gamma_{i-k}+\gamma_{i+1-k}\right)-{\varepsilon}({\varrho}_i+{\varrho}_{i+1})\\
&=\sum_{n=1}^{g-1}[t_{a_n}]\otimes Y_{n}
\left(\gamma_{i+1}+(-1)^{g-1}\gamma_{i+1}\right)-{\varepsilon}({\varrho}_i+{\varrho}_{i+1}).
\end{aligned}$$ If $g$ is even, this relation is trivial, and if $g$ is odd it gives $$\begin{aligned}
r_i^{(E3)}&+r_{i+1}^{(E3)}\!:0=2\sum_{n=1}^{g-1}[t_{a_n}]\otimes Y_{n}
\left(\gamma_{i+1}\right)-({\varrho}_i+{\varrho}_{i+1})\\
&=2(t_{1,i+1}+\cdots+t_{i,i+1}-t_{i+1,i}-\cdots-t_{g-1,i})-({\varrho}_i+{\varrho}_{i+1})\\
&=(*)+2(t_{i,i}+t_{i,i+1})-(2t_{i,i}+{\varrho}_i+{\varrho}_{i+1}).
\end{aligned}$$ Hence this relation gives no new information.
Relation $(E4)$ gives no new information, hence we proved the following theorem.
\[tw:hom:positive\] If $g\geqslant 3$, then $$H_1({{{\mathcal{M}}}}^{h+}(N_g); H_1(N_g;{{\mathbb{Z}}}))={{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2.$$
Computing $H_1({{{\mathcal{M}}}}^{h}(N_g); H_1(N_g;{{\mathbb{Z}}}))$
--------------------------------------------------------------------
If $G={{{\mathcal{M}}}}^{h}(N_g)$, then by Proposition \[gen\_ker\_roz\] the kernel ${\langle {\overline{X}} \rangle}\cap\ker{\overline{\partial}}_1$ has two more types of generators: (F5), (F6), and by Theorem \[th:004\] there are three additional relations: (C4),(C5),(C8). $$\begin{aligned}
r_i^{(C4)}\!:0&=[s]\otimes \gamma_i+s[s]\otimes\gamma_i=s_i+[s]\otimes\psi(s)\gamma_i\\
&=2(-1)^i\left(s_1+s_2+\cdots+s_{i-1}+\frac{1+(-1)^i}{2}s_i\right).
\end{aligned}$$ This (inductively) implies that each generator of type (F5) has order at most 2. $$\begin{aligned}
r_i^{(C8)}\!:0&=([{\varrho}]+{\varrho}[s]-[s]-s[{\varrho}])\otimes\gamma_i={\varrho}_i-2s_i-[{\varrho}]\otimes \psi(s)\gamma_i\\
&={\varrho}_i-2s_i-(-1)^i(2{\varrho}_1+2{\varrho}_2+\cdots +2{\varrho}_{i-1}+{\varrho}_i)\\
&=\begin{cases}
-2(s_i-{\varrho}_1-\cdots-{\varrho}_i)&\text{ for $i$ odd}\\
-2(s_{i-1}+s_i)+2(s_{i-1}-{\varrho}_1-\cdots -{\varrho}_{i-1})&\text{ for $i$ even.}
\end{cases}
\end{aligned}$$ This implies that generator (F6) has also order at most 2. $$\begin{aligned}
r_i^{(C5)}\!:0&=([t_{a_j}]+t_{a_j}[s]+t_{a_j}s[t_{a_j}]-[s])\otimes \gamma_i\\
&=t_{j,i}+[s]\otimes\psi(t_{a_j}^{-1})\gamma_i+[t_{a_j}]\otimes \psi(s)\psi(t_{a_j}^{-1})\gamma_i-s_i.
\end{aligned}$$ If $i\neq j$ and $i\neq j+1$, then $$r_i^{(C5)}\!:0=(-1)^i\left(2t_{j,1}+\cdots+2t_{j,i-1}+(1+(-1)^i)t_{j,i}\right),$$ which gives no new information. If $i=j$ or $i=j+1$ and $j$ is odd, then $$r_i^{(C5)}\!:0=(*)\pm [(s_j+s_{j+1})+(t_{j,j}+t_{j,j+1})],$$ where as usual $(*)$ denotes homologically trivial element. This relation implies that generators (F5) are superfluous.
Finally, if $i=j$ or $i=j+1$ and $j$ is even then $$r_i^{(C5)}\!:0=(*)\pm[(s_{j+1}-{\varrho}_1-\cdots-{\varrho}_{j+1})-(s_{j-1}-{\varrho}_1-\cdots-{\varrho}_{j-1})].$$ This implies that all generators of type (F6) are homologous, hence we proved the following.
\[tw:hom:full\] If $g\geqslant 3$, then $$H_1({{{\mathcal{M}}}}^{h}(N_g); H_1(N_g;{{\mathbb{Z}}}))={{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2\oplus{{\mathbb{Z}}}_2.$$
Acknowledgements {#acknowledgements .unnumbered}
================
The author wishes to thank the referee for his/her helpful suggestions.
[10]{}
S. J. Bigelow and R. D. Budney. The mapping class group of a genus two surface is linear. , 1:699–708, 2001.
J. S. Birman. . Number 82 in Ann. of Math. Studies. Princeton Univ. Press, 1974.
J. S. Birman. and D. R. J. Chillingworth. On the homeotopy group of a non–orientable surface. , 71:437–448, 1972.
J. S. Birman and H. M. Hilden. On mapping class groups of closed surfaces as covering spaces. In [*Advances in the theory of [R]{}iemann surfaces*]{}, number 66 in Ann. of Math. Studies, pages 81–115. Princeton Univ. Press, 1971. Proc. Conf., Stony Brook, N.Y., 1969.
C. F. B[ö]{}digheimer, F. R. Cohen, and M. D. Peim. Mapping class groups and function spaces. In [*Homotopy Methods in Algebraic Topology*]{}, volume 271 of [ *Contemp. Math.*]{}, pages 17–39, 2001. Proc. Conf., Boulder, 1999.
K. S. Brown. . Number 87 in Grad. Texts in [M]{}ath. Springer–[V]{}erlag, 1982.
D. R. J. Chillingworth. A finite set of generators for the homeotopy group of a non–orientable surface. , 65:409–430, 1969.
S. Gadgil and D. Pancholi. Homeomorphisms and the homology of non–orientable surfaces. , 115(3):251–257, 2005.
N. Kawazumi. Homology of hyperelliptic mapping class groups for surfaces. , 76:203–216, 1997.
W. Magnus. ber automorphismen von [F]{}undamentalgruppen beraneter [F]{}lachen. , 109:617–646, 1934.
S. Morita. Families of jacobian manifolds and characteristic classes of surface boundles [I]{}. , 39:777–810, 1989.
L. Paris and B. Szepietowski. A presentation for the mapping class group of a nonorientable surface. arXiv:1308.5856v1 \[math.GT\], 2013.
U. [P]{}inkal and J. D. Mc[C]{}arthy. Representing homology automorphisms of nonorientable surfaces. Max Planc Inst. preprint MPI/SFB 85-11. Available at <http://www.math.msu.edu/~mccarthy>, 2004.
M. [S]{}tukow. Conjugacy classes of finite subgroups of certain mapping class groups. , 28(2):101–110, 2004.
M. Stukow. The twist subgroup of the mapping class group of a nonorientable surface. , 46(3):717–738, 2009.
M. [S]{}tukow. A finite presentation for the mapping class group of a nonorientable surface with [D]{}ehn twists and one crosscap slide as generators. arXiv:1310.2722v2 \[math.GT\], 2013.
A. Tanaka. The first homology group of the hyperelliptic mapping class group with twisted coefficients. , 115:19–42, 2001.
B. Wajnryb. A simple presentation for the mapping class group of an orientable surface. , 45(2–3):157–174, 1983.
H. Zieschang. On the [H]{}omeotopy [G]{}roups of [S]{}urfaces. , 206:1–21, 1973.
[^1]: Supported by MNiSW grant N201 366436.
|
---
abstract: 'We propose the eigenvalue problem of an anisotropic diffusion operator for image segmentation. The diffusion matrix is defined based on the input image. The eigenfunctions and the projection of the input image in some eigenspace capture key features of the input image. An important property of the model is that for many input images, the first few eigenfunctions are close to being piecewise constant, which makes them useful as the basis for a variety of applications such as image segmentation and edge detection. The eigenvalue problem is shown to be related to the algebraic eigenvalue problems resulting from several commonly used discrete spectral clustering models. The relation provides a better understanding and helps developing more efficient numerical implementation and rigorous numerical analysis for discrete spectral segmentation methods. The new continuous model is also different from energy-minimization methods such as geodesic active contour in that no initial guess is required for in the current model. The multi-scale feature is a natural consequence of the anisotropic diffusion operator so there is no need to solve the eigenvalue problem at multiple levels. A numerical implementation based on a finite element method with an anisotropic mesh adaptation strategy is presented. It is shown that the numerical scheme gives much more accurate results on eigenfunctions than uniform meshes. Several interesting features of the model are examined in numerical examples and possible applications are discussed.'
author:
- 'Jingyue Wang[^1]'
- 'Weizhang Huang[^2]'
title: 'Image Segmentation with Eigenfunctions of an Anisotropic Diffusion Operator [^3] '
---
[**Key Words.**]{} eigenvalue problem, image segmentation, finite-element schemes, mesh adaptation, anisotropic diffusion.
[**AMS 2010 Mathematics Subject Classification.**]{} 65N25, 68U10, 94A08
Introduction
============
We are concerned with image segmentation using the eigenvalue problem of an anisotropic linear diffusion operator, $$-\nabla \cdot (\mathbb{D}\nabla u)=\lambda u, \quad \text{ in }\Omega
\label{eq:HW-eigen}$$ subject to a homogeneous Dirichlet or Neumann boundary condition, where the diffusion matrix $\mathbb{D}$ is symmetric and uniformly positive definite on $\Omega$. In this study, we consider an anisotropic diffusion situation where $\mathbb{D}$ has different eigenvalues and is defined based on the gray level of the input image.
A method employing an eigenvalue problem to study image segmentation is referred to as a spectral clustering method in the literature. This type of methods have extracted great interest from researchers in the past decade; e.g., see [@Grady-01; @shi_normalized_2000; @Wu-Leahy-1993]. They are typically derived from a minimum-cut criterion on a graph. One of the most noticeable spectral clustering methods is the normalized cut method proposed by Shi and Malik [@shi_normalized_2000] (also see Section \[SEC:relation-1\] below) which is based on the eigenvalue problem $$(D-W) {\bf u} = \lambda D {\bf u} ,
\label{eq:Malik-Shi-eigen}$$ where ${\bf u}$ is a vector representing the gray level value on the pixels, $W$ is a matrix defining pairwise similarity between pixels, and $D$ is a diagonal matrix formed with the degree of pixels (cf. Section 2.2 below). The operator $L=D-W$ corresponds to the graph Laplacian in graph spectral theory. An eigenvector associated with the second eigenvalue is used as a continuous approximation to a binary or $k$-way vector that indicates the partitions of the input image. Shi and Malik suggested that image segmentation be done on a hierarchical basis where low level coherence of brightness, texture, and etc. guides a binary (or $k$-way) segmentation that provides a big picture while high level knowledge is used to further partition the low-level segments.
While discrete spectral clustering methods give impressive partitioning results in general, they have several drawbacks. Those methods are typically defined and operated on a graph or a data set. Their implementation cost depends on the size of the graph or data set. For a large data set, they can be very expensive to implement. Moreover, since they are discrete, sometimes their physical and/or geometrical meanings are not so clear. As we shall see in Section \[SEC:relation-1\], the normalized cut method of Shi and Malik [@shi_normalized_2000] is linked to an anisotropic diffusion differential operator which from time to time can lead to isotropic diffusion.
The objective of this paper is to investigate the use of the eigenvalue problem (\[eq:HW-eigen\]) of an anisotropic diffusion operator for image segmentation. This anisotropic model can be viewed as a continuous, improved anisotropic generalization of discrete spectral clustering models such as (\[eq:Malik-Shi-eigen\]). The model is also closely related to the Perona-Malik anisotropic filter. The advantages of using a continuous model for image segmentation include (i) It has a clear physical interpretation (heat diffusion or Fick’s laws of diffusion in our case); (ii) Many well developed theories of partial differential equations can be used; (iii) Standard discretization methods such as finite differences, finite elements, finite volumes, and spectral methods can be employed; and (iv) The model does not have be discretized on a mesh associated with the given data set and indeed, mesh adaptation can be used to improve accuracy and efficiency. As mentioned early, we shall define the diffusion matrix $\mathbb{D}$ using the input image and explore properties of the eigenvalue problem. One interesting property is that for many input images, the first few eigenfunctions of the model are close to being piecewise constant, which are very useful for image segmentation. However, this also means that these eigenfunctions change abruptly between objects and their efficient numerical approximation requires mesh adaptation. In this work, we shall use an anisotropic mesh adaptation strategy developed by the authors [@Huang-Wang-13] for differential eigenvalue problems. Another property of (\[eq:HW-eigen\]) is that eigenfunctions associated with small eigenvalues possess coarse, global features of the input image whereas eigenfunctions associated with larger eigenvalues carry more detailed, localized features. The decomposition of features agrees with the view of Shi and Malik [@shi_normalized_2000] on the hierarchical structure of image segmentation but in a slightly different sense since all eigenfunctions come from low level brightness knowledge.
The paper is organized as follows. In Section \[SEC:eigen\], we give a detailed description of the eigenvalue problem based on an anisotropic diffusion operator and discuss its relations to some commonly used discrete spectral clustering models and diffusion filters and some other models in image segmentation. Section \[SEC:implement\] is devoted to the description of the finite element implementation of the model and an anisotropic mesh adaptation strategy. In Section \[SEC:numerics\], we present a number of applications in image segmentation and edge detection and demonstrate several properties of the model. Some explanations to the piecewise constant property of eigenfunctions are given in Section \[SEC:piecewise\]. Concluding remarks are given in Section \[SEC:conclusion\].
Description of the eigenvalue problem {#SEC:eigen}
=====================================
Eigenvalue problem of an anisotropic diffusion operator
-------------------------------------------------------
We shall use the eigenvalue problem (\[eq:HW-eigen\]) subject to a Dirichlet or Neumann boundary condition for image segmentation.
We are inspired by the physics of anisotropic heat transport process (e.g., see [@Gunter-Yu-Kruger-Lackner-05; @Sharma-Hammett-07]), treating the dynamics of image diffusion as the transport of energy (pixel values) and viewing the eigenvalue problem as the steady state of the dynamic process. Denote the principal diffusion direction by $v$ (a unit direction field) and its perpendicular unit direction by $v^{\perp}$. Let the conductivity coefficients along these directions be $\chi_{\parallel}$ and $\chi_{\perp}$. ($v$, $\chi_{\parallel}$, and $\chi_{\perp}$ will be defined below.) Then the diffusion matrix can be written as $$\mathbb{D}=\chi_{\parallel}vv^{T}+\chi_{\perp}v^{\perp}(v^{\perp})^{T} .
\label{D-1}$$ When $\chi_{\parallel}$ and $\chi_{\perp}$ do not depend on $u$, the diffusion operator in (\[eq:HW-eigen\]) is simply a linear symmetric second order elliptic operator. The anisotropy of the diffusion tensor $\mathbb{D}$ depends on the choice of the conductivity coefficients. For example, if $\chi_{\parallel}\gg\chi_{\perp}$, the diffusion is preferred along the direction of $v$. Moreover, if $\chi_{\parallel}=\chi_{\perp}$, the diffusion is isotropic, having no preferred diffusion direction.
To define $\mathbb{D}$, we assume that an input image is given. Denote its gray level by $u_0$. In image segmentation, pixels with similar values of gray level will be grouped and the interfaces between those groups provide object boundaries. Since those interfaces are orthogonal to $\nabla u_0$, it is natural to choose the principal diffusion direction as $v=\nabla u_{0}/|\nabla u_{0}|$. With this choice, we can rewrite (\[D-1\]) into $$\mathbb{D} = \frac{\chi_{\parallel}}{|\nabla u_{0}|^{2}}
\begin{bmatrix}|\partial_x u_{0}|^{2}+\mu |\partial_y u_{0}|^{2} & (1-\mu)\left|\partial_x u_{0}\partial_y u_{0}\right|\\
(1-\mu)\left|\partial_x u_{0} \partial_y u_{0}\right| & |\partial_y u_{0}|^{2}+\mu |\partial_x u_{0}|^{2} \end{bmatrix}
\label{D-2}$$ where $\mu = \chi_{\perp}/\chi_{\parallel}$. We consider two choices of $\chi_{\parallel}$ and $\mu$. The first one is $$\chi_{\parallel}=g(|\nabla u_{0}|), \quad \mu = 1,
\label{D-3}$$ where $g(x)$ is a conductance function that governs the behavior of diffusion. This corresponds to linear isotropic diffusion. As in [@Perona-Malik-90], we require $g$ to satisfy $g(0)=1$, $g(x)\ge0$, and $g(x)\to 0$ as $x \to \infty$. For this choice, both $\chi_{\parallel}$ and $\chi_{\perp}$ becomes very small across the interfaces of the pixel groups and therefore, almost no diffusion is allowed along the normal and tangential directions of the interfaces.
The second choice is $$\chi_{\parallel}=g(|\nabla u_{0}|), \quad \mu=1+|\nabla u_{0}|^{2}.
\label{D-4}$$ This choice results in an anisotropic diffusion process. Like the first case, almost no diffusion is allowed across the interfaces of the pixel groups but, depending on the choice of $g$, some degree of diffusion is allowed on the tangential direction of the interfaces. We shall show later that with a properly chosen $g$ the eigenfunctions of (\[eq:HW-eigen\]) capture certain “grouping” features of the input image $u_{0}$ very well. This phenomenon has already been observed and explored in many applications such as shape analysis [@Reuter-09; @Reuter-06], image segmentation and data clustering [@Grady-01; @shi_normalized_2000; @Shi-Malik-2001; @Wu-Leahy-1993], and high dimensional data analysis and machine learning [@Belkin_towards_2005; @Nadler_diffusion_2005; @Nadler_diffusion_2006; @Luxburg-2007]. In these applications, all eigenvalue problems are formulated on a discrete graph using the graph spectral theory, which is different from what is considered here, i.e., eigenvalue problems of differential operators. The application of the latter to image segmentation is much less known. We shall discuss the connection of these discrete eigenvalue problems with continuous ones in the next subsection.
It is noted that the gray level function $u_0$ is defined only at pixels. Even we can view $u_{0}$ as the “ground truth” function (assuming there is one function whose discrete sample is the input image), it may not be smooth and the gradient cannot be defined in the classical sense. Following [@Alvarez-Lions-92; @Catte-Lions-92], we may treat $u_{0}$ as a properly regularized approximation of the “true image” so that the solution to the eigenvalue problem (\[eq:HW-eigen\]) exists. In the following, we simply take $u_{0}$ as the linear interpolation of the sampled pixel values (essentially an implicit regularization from the numerical scheme). More sophisticated regularization methods can also be employed.
We only deal with gray level images in this work. The approach can be extended to color or texture images when a diffusion matrix can be defined appropriately based on all channels. In our computation, we use both Dirichlet and Neumann boundary conditions, with the latter being more common in image processing.
Relation to discrete spectral clustering models {#SEC:relation-1}
-----------------------------------------------
The eigenvalue problem (\[eq:HW-eigen\]) is closely related to a family of discrete spectral clustering models, with the earliest one being the normalized cut method proposed by Shi and Malik [@shi_normalized_2000]. To describe it, we define the degree of dissimilarity (called $cut$) between any two disjoint sets $A,B$ of a weighted undirected graph $G=(V,E)$ (where $V$ and $E$ denote the sets of the nodes and edges of the graph) as the total weight of the edges connecting nodes in the two sets, i.e., $$cut(A,B)=\sum_{p\in A,\; q\in B}w(p,q).$$ Wu and Leahy [@Wu-Leahy-1993] proposed to find $k$-subgraphs by minimizing the maximum cut across the subgroups and use them for a segmentation of an image. However, this approach usually favors small sets of isolated nodes in the graph. To address this problem, Shi and Malik [@shi_normalized_2000] used the normalized cut defined as $$Ncut(A,B)=\frac{cut(A,B)}{assoc(A,A\cup B)}+\frac{cut(A,B)}{assoc(B,A\cup B)} ,$$ where $assoc(A,A\cup B)=\sum_{p\in A,\; q\in A\cup B}w(p,q)$. They sought the minimum of the functional $Ncut(A,B)$ recursively to obtain a $k$-partition of the image. The edge weight $w(p,q)$ is chosen as $$w(p,q)=\begin{cases}
e^{-|u_{q}-u_{p}|^2/\sigma^2}, & q\in\mathcal{N}_{p},\\
0, & {\rm otherwise,}
\end{cases}$$ where $\mathcal{N}_{p}$ is a neighborhood of pixel $p$ and $\sigma$ is a positive parameter. Shi and Malik showed that the above optimization problem is NP-hard but a binary solution to the normalized cut problem can be mapped to a binary solution to the algebraic eigenvalue problem (\[eq:Malik-Shi-eigen\]) with $D$ being a diagonal matrix with diagonal entries $d_{p}=\sum_{q}w(p,q)$ and $W$ being the weight matrix $(w(p,q))_{p,q}^{N\times N}$. Eigenvectors of this algebraic eigenvalue problem are generally not binary. They are used to approximate binary solutions of the normalized cut problem through certain partitioning.
To see the connection between the algebraic eigenvalue problem (\[eq:Malik-Shi-eigen\]) (and therefore, the normalized cut method) with the continuous eigenvalue problem (\[eq:HW-eigen\]), we consider an eigenvalue problem in the form of (\[eq:HW-eigen\]) with the diffusion matrix defined as $$\mathbb{D} = \begin{bmatrix}e^{-|\partial_{x}u_{0}|^2/\sigma^2} & 0\\
0 & e^{-|\partial_{y}u_{0}|^2/\sigma^2} \end{bmatrix}
\label{D-5}$$ A standard central finite difference discretization of this problem on a rectangular mesh gives rise to $$\frac{(c_{E_{i,j}}+c_{W_{i,j}}+c_{N_{i,j}}+c_{S_{i,j}})u_{i,j}-c_{E_{i,j}}u_{i+1,j}-c_{W_{i,j}}u_{i-1,j}-c_{N_{i,j}}u_{i,j+1}-c_{S_{i,j}}u_{i,j-1}}{h^{2}}=\lambda u_{i,j},
\label{eq:orthotropic-fe-scheme}$$ where $h$ is the grid spacing and the coefficients $c_{E_{i,j}},\; c_{W_{i,j}},\; c_{N_{i,j}},\; c_{S_{i,j}}$ are given as $$\begin{aligned}
c_{E_{i,j}} = e^{-|u_{i+1,j}-u_{i,j}|^2/\sigma^2},\quad
c_{W_{i,j}} = e^{-|u_{i-1,j}-u_{i,j}|^2/\sigma^2},\\
c_{N_{i,j}} = e^{-|u_{i,j+1}-u_{i,j}|^2/\sigma^2},\quad
c_{S_{i,j}} = e^{-|u_{i,j-1}-u_{i,j}|^2/\sigma^2}.\end{aligned}$$ It is easy to see that (\[eq:orthotropic-fe-scheme\]) is almost the same as (\[eq:Malik-Shi-eigen\]) with the neighborhood $\mathcal{N}_{i,j}$ of a pixel location $(i,j)$ being chosen to include the four closest pixel locations $\{(i+1,j),(i-1,j),(i,j+1),(i,j-1)\}$. The difference lies in that (\[eq:Malik-Shi-eigen\]) has a weight function on its right-hand side. Moreover, it can be shown that (\[eq:orthotropic-fe-scheme\]) gives [*exactly*]{} the algebraic eigenvalue problem for the average cut problem $${\rm min}\frac{cut(A,B)}{|A|}+\frac{cut(A,B)}{|B|} ,$$ where $|A|$ and $|B|$ denote the total numbers of nodes in $A$ and $B$, respectively. Notice that this problem is slightly different from the normalized cut problem and its solution is known as the Fiedler value. Furthermore, if we consider the following generalized eigenvalue problem (by multiplying the right-hand side of (\[eq:HW-eigen\]) with a mass-density function), $$-\nabla \cdot \left(\begin{bmatrix}e^{-|\partial_{x}u_{0}|^2/\sigma^2} & 0\\
0 & e^{-|\partial_{y}u_{0}|^2/\sigma^2} \end{bmatrix}
\nabla u\right)= (e^{-|\partial_x u_{0}|^2/\sigma^2} + e^{-|\partial_y u_{0}|^2/\sigma^2}) \lambda u,
\label{eq:pm-aniso-1}$$ we can obtain (\[eq:Malik-Shi-eigen\]) exactly with a proper central finite difference discretization.
The above analysis shows that either the average cut or normalized cut model can be approximated by a finite difference discretization of the continuous eigenvalue problem (\[eq:HW-eigen\]) with the diffusion matrix (\[D-5\]) which treats diffusion differently in the $x$ and $y$ directions. While (\[D-5\]) is anisotropic in general, it results in isotropic diffusion near oblique interfaces where $\partial_x u_0 \approx \partial_y u_0$ or $\partial_x u_0 \approx - \partial_y u_0$. This can be avoided with the diffusion matrix (\[D-2\]) which defines diffusion differently along the normal and tangential directions of group interfaces. In this sense, our method consisting of (\[eq:HW-eigen\]) with (\[D-2\]) can be regarded as an improved version of (\[eq:HW-eigen\]) with (\[D-5\]), and thus, an improved continuous generalization of the normalized cut or the average cut method.
It should be pointed out that there is a fundamental difference between discrete spectral clustering methods and those based on continuous eigenvalue problems. The former are defined and operated directly on a graph or data set and their cost depends very much on the size of the graph or data. On the other hand, methods based on continuous eigenvalue problems treat an image as a sampled function and are defined by a discretization of some differential operators. They have the advantage that many standard discretization methods such as finite difference, finite element, finite volume, and spectral methods can be used. Another advantage is that they do not have to be operated directly on the graph or the data set. As shown in [@Huang-Wang-13], continuous eigenvalue problems can be solved efficiently on adaptive, and especially anisotropic adaptive, meshes (also see Section \[SEC:numerics\]).
It is worth pointing out that the graph Laplacian can be connected to a continuous diffusion operator by defining the latter on a manifold and proving it to be the limit of the discrete Laplacian. The interested reader is referred to the work of [@Belkin_towards_2005; @Nadler_diffusion_2005; @Nadler_diffusion_2006; @Singer-06; @Luxburg-2007].
Relation to diffusion models
----------------------------
The eigenvalue problem (\[eq:HW-eigen\]) is related to several diffusion models used in image processing. They can be cast in the form $$\frac{\partial u}{\partial t}=\nabla \cdot \left(\mathbb{D}\nabla u\right)
\label{eq:linear-diffusion}$$ with various definitions of the diffusion matrix. For example, the Perona-Malik nonlinear filter [@Perona-Malik-90] is in this form with $\mathbb{D} = g(|\nabla u|) I$, where $g$ is the same function in (\[D-3\]) and $I$ is the identity matrix. The above equation with $\mathbb{D}$ defined in (\[D-2\]) with $\mu=1$ and $\chi_{\parallel}=g(|\nabla u_{0}|)$ gives rise to a linear diffusion process that has similar effects as the affine Gaussian smoothing process [@Nitzberg-Shiota-92]. The diffusion matrix we use in this paper in most cases is in the form (\[D-2\]) with $\mu$ and $\chi_{\parallel}$ defined in (\[D-4\]). A similar but not equivalent process was studied as a structure adaptive filter by Yang et al. [@Yang-Burger-96]. The diffusion matrix (\[D-2\]) can be made $u$-dependent by choosing $\mu$ and $\chi_{\parallel}$ as functions of $\nabla u$. Weickert [@Weickert-1996] considered a nonlinear anisotropic diffusion model with a diffusion matrix in a similar form as (\[D-2\]) but with $\nabla u_0$ being replaced by the gradient of a smoothed gray level function $u_\sigma$ and with $\chi_{\parallel} = g(|\nabla u_\sigma|)$ and $\mu = 1/g(|\nabla u_\sigma|)$.
Interestingly, Perona and Malik [@Perona-Malik-90] considered $$\begin{aligned}
\frac{\partial u}{\partial t} = \nabla \cdot \left(\begin{bmatrix}g(|\partial_{x}u |) & 0\\
0 & g(|\partial_{y} u|) \end{bmatrix}\nabla u\right)
\label{eq:pm-linear-diffusion}\end{aligned}$$ as an easy-to-compute variant to the Perona-Malik diffusion model (with $\mathbb{D} = g(|\nabla u|) I$). Zhang and Hancock in [@Zhang-Hancock-08] considered $$\begin{aligned}
\frac{\partial u}{\partial t} = -\mathcal{L}(u_0) u,
\label{eq:Zhang-Hancock}\end{aligned}$$ where $\mathcal{L}$ is the graph Laplacian defined on the input image $u_0$ and image pixels are treated as the nodes of a graph. The weight between two nodes $i,j$ is defined as $$w_{i,j}=\begin{cases}
e^{-(u_{0}(i)-u_{0}(j))^{2}/\sigma^{2}}, & \quad \text{ for }\|i-j\|\le r\\
0, & \quad \text{otherwise}
\end{cases}$$ where $r$ is a prescribed positive integer and $\sigma$ is a positive parameter. As in Section \[SEC:relation-1\], it can be shown that this model can be regarded as a discrete form of a linear anisotropic diffusion model. It has been reported in [@Buades-Chien-08; @Nitzberg-Shiota-92; @Yang-Burger-96; @Zhang-Hancock-08] that the image denoising effect with this type of linear diffusion model is comparable to or in some cases better than nonlinear evolution models.
Numerical Implementation {#SEC:implement}
========================
The eigenvalue problem (\[eq:HW-eigen\]) is discretized using the standard linear finite element method with a triangular mesh for $\Omega$. The finite element method preserves the symmetry of the underlying continuous problem and can readily be implemented with (anisotropic) mesh adaptation. As will be seen in Section \[SEC:numerics\], the eigenfunctions of (\[eq:HW-eigen\]) can have very strong anisotropic behavior, and (anisotropic) mesh adaptation is essential to improving the efficiency of their numerical approximation.
While both Dirichlet and Neumann boundary conditions are considered in our computation, to be specific we consider only a Dirichlet boundary condition in the following. The case with a Neumann boundary condition can be discussed similarly.
We assume that a triangular mesh $\mathcal{T}_{h}$ is given for $\Omega$. Denote the number of the elements of $\mathcal{T}_h$ by $N$ and the linear finite element space associated with $\mathcal{T}_h$ by $V^{h}\subset H_{0}^{1}\left(\Omega\right)$. Then the finite element approximation to the eigenvalue problem (\[eq:HW-eigen\]) subject to a Dirichlet boundary condition is to find $0 \not\equiv u^h \in V^{h}$ and $\lambda^h \in \mathbb{R}$ such that $$\int_{\Omega}(\nabla v^h)^t \mathbb{D}\nabla u^h =\lambda^h \int_{\Omega} u^h v^h,\qquad\forall v^h\in V^{h}.
\label{eq:fem-1}$$ This equation can be written into a matrix form as $$A {\bf u} = \lambda^h M {\bf u},$$ where $A$ and $M$ are the stiffness and mass matrices, respectively, and ${\bf u}$ is the vector formed by the nodal values of the eigenfunction at the interior mesh nodes.
An error bound for the linear finite element approximation of the eigenvalues is given by a classical result of Raviart and Thomas [@Raviart-Thomas-83]. It states that for any given integer $k$ ($1\le k\le N$), $$0\le\frac{\lambda_{j}^{h}-\lambda_{j}}{\lambda_{j}^{h}}\le C(k)\sup_{v\in E_{k},\|v\|=1}\| v-\Pi_{h}v\|_E^{2},
\qquad1\le j\le k$$ where $\lambda_j$ and $\lambda_j^h$ are the eigenvalues (ordered in an increasing order) of the continuous and discrete problems, respectively, $E_{k}$ is the linear space spanned by the first $k$ eigenfunctions of the continuous problem, $\Pi_{h}$ is the projection operator from $L^{2}(\Omega)$ to the finite element space $V^{h}$, and $\| \cdot \|_E$ is the energy norm, namely, $$\| v-\Pi_{h}v\|_E^{2} = \int_\Omega \nabla (v-\Pi_{h}v)^t \mathbb{D} \nabla (v-\Pi_{h}v).$$ It is easy to show (e.g., see [@Huang-Wang-13]) that the project error can be bounded by the error of the interpolation associated with the underlying finite element space, with the latter depending directly on the mesh. When the eigenfunctions change abruptly over the domain and exhibit strong anisotropic behavior, anisotropic mesh adaptation is necessary to reduce the error or improve the computational efficiency (e.g. see [@Boff-10; @Huang-Wang-13]).
An anisotropic mesh adaptation method was proposed for eigenvalue problems by the authors [@Huang-Wang-13], following the so-called $\mathbb{M}$-uniform mesh approach developed in [@Huang-05; @Huang-06; @Huang-Russell-11] for the numerical solution of PDEs. Anisotropic mesh adaptation provides one advantage over isotropic one in that, in addition to the size, the orientation of triangles is also adapted to be aligned with the geometry of the solution locally. In the context of image processing, this mesh alignment will help better capture the geometry of edges than with isotropic meshes. The $\mathbb{M}$-uniform mesh approach of anisotropic mesh adaptation views and generates anisotropic adaptive meshes as uniform ones in the metric specified by a metric tensor $\mathbb{M} = \mathbb{M}(x,y)$. Putting it in a simplified scenario, we may consider a uniform mesh defined on the surface of the gray level $u$ and obtain an anisotropic adaptive mesh by projecting the uniform mesh into the physical domain. In the actual computation, instead of using the surface of $u$ we employ a manifold associated with a metric tensor defined based on the Hessian of the eigenfunctions. An optimal choice of the metric tensor (corresponding to the energy norm) is given [@Huang-Wang-13] as $$\mathbb{M}_{K}=\det\left(H_{K}\right)^{-1/4}\max_{(x,y)\in K}\|H_{K}\mathbb{D}(x,y)\|^{1/2}\left(\frac{1}
{|K|}\|H_{K}^{-1}H\|_{L^{2}(K)}^{2}\right)^{1/2}H_{K},\qquad\forall K\in\mathcal{T}_{h}$$ where $K$ denotes a triangle element of the mesh, $H$ is the intersection of the recovered Hessian matrices of the computed first $k$ eigenfunctions, and $H_{K}$ is the average of $H$ over $K$. A least squares fitting method is used for Hessian recovery. That is, a quadratic polynomial is constructed locally for each node via least squares fitting to neighboring nodal function values and then an approximate Hessian at the node is obtained by differentiating the polynomial. The recovered Hessian is regularized with a prescribed small positive constant which is taken to be $0.01$ in our computation.
An outline of the computational procedure of the anisotropic adaptive mesh finite element approximation for the eigenvalue problem (\[eq:HW-eigen\]) is given in Algorithm \[alg:aniso\]. In Step 5, BAMG (Bidimensional Anisotropic Mesh Generator) developed by Hecht [@Hecht-Bamg-98] is used to generate the new mesh based on the computed metric tensor defined on the current mesh. The resultant algebraic eigenvalue problems are solved using the Matlab eigenvalue solver [eigs]{} for large sparse matrices. Note that the algorithm is iterative. Ten iterations are used in our computation, which was found to be enough to produce an adaptive mesh with good quality (see [@Huang-05] for mesh quality measures).
1\. Initialize a background mesh.
2\. Compute the stiffness and mass matrices on the mesh.
3\. Solve the algebraic eigenvalue problem for the first $k$ eigenpairs.
4\. Use the eigenvectors obtained in Step 3 to compute the metric tensor.
5\. Use the metric tensor to generate a new mesh (anisotropic, adaptive) and go to Step 2.
Numerical results {#SEC:numerics}
=================
In this section, all input images are of size $256\times256$ and normalized so that the gray values are between 0 and 1. The domain of input images is set to be $[0,1]\times[0,1]$. All eigenfunctions are computed with a homogeneous Neumann boundary condition unless otherwise specified. When we count the indices of eigenfunctions, we ignore the first trivial constant eigenfunction and start the indexing from the second one.
Properties of eigenfunctions
----------------------------
### Almost piecewise constant eigenfunctions
A remarkable feature of the eigenvalue problem (\[eq:HW-eigen\]) with the diffusion matrix (\[D-2\]) is that for certain input images, the first few eigenfunctions are close to being piecewise constant. In Fig. \[fig:synth1\], we display a synthetic image containing 4 objects and the first 7 eigenfunctions. The gaps between objects are 4 pixel wide. To make the problem more interesting, the gray level is made to vary within each object (so the gray value of the input image is not piecewise-constant). We use the anisotropic diffusion tensor $\mathbb{D}$ defined in (\[D-2\]) and (\[D-4\]) with $$g(x)=\frac{1}{(1+x^{2})^{\alpha}},
\label{D-6}$$ where $\alpha$ is a positive parameter. Through numerical experiment (cf. Section \[SEC:4.1.6\]), we observe that the larger $\alpha$ is, the closer to being piecewise constant the eigenfunctions are. In the same time, the eigenvalue problem (\[eq:HW-eigen\]) is also harder to solve numerically since the eigenfunctions change more abruptly between the objects. We use $\alpha = 1.5$ in the computation for Fig. \[fig:synth1\]. The computed eigenfunctions are normalized such that they have the range of $[0,255]$ and can be rendered as gray level images. The results are obtained with an adaptive mesh of 65902 vertices and re-interpolated to a $256\times256$ mesh for rendering.
The histograms of the first 3 eigenfunctions together with the plot of the first 10 eigenvalues are shown in Fig. \[fig:synth7hist\]. It is clear that the first 3 eigenfunctions are almost piecewise constant. In fact, the fourth, fifth, and sixth are also almost piece constant whereas the seventh is clearly not. (Their histograms are not shown here to save space but this can be seen in Fig. \[fig:synth1\].)
Fig.\[fig:nzsynth1x\] shows the results obtained an image with a mild level of noise. The computation is done with the same condition as for Fig. \[fig:synth1\] except that the input image is different. We can see that the first few eigenfunctions are also piecewise constant and thus the phenomenon is relatively robust to noise.
![The input synthetic image and the first 7 eigenfunctions (excluding the trivial constant eigenfunction), from left to right, top to bottom. The results are obtained with the diffusion matrix defined in (\[D-2\]), (\[D-4\]), and (\[D-6\]) ($\alpha = 1.5$).[]{data-label="fig:synth1"}](synth7_set){width="10cm"}
![The first 10 eigenvalues and the histograms of the first 3 eigenfunctions in Fig.\[fig:synth1\]. The $x$ and $y$ axes of the histograms are the gray value and the number of pixels having the same gray value.[]{data-label="fig:synth7hist"}](synth7ev.pdf)
![The first 10 eigenvalues and the histograms of the first 3 eigenfunctions in Fig.\[fig:synth1\]. The $x$ and $y$ axes of the histograms are the gray value and the number of pixels having the same gray value.[]{data-label="fig:synth7hist"}](synth7hist_2.pdf)
![The first 10 eigenvalues and the histograms of the first 3 eigenfunctions in Fig.\[fig:synth1\]. The $x$ and $y$ axes of the histograms are the gray value and the number of pixels having the same gray value.[]{data-label="fig:synth7hist"}](synth7hist_3.pdf)
![The first 10 eigenvalues and the histograms of the first 3 eigenfunctions in Fig.\[fig:synth1\]. The $x$ and $y$ axes of the histograms are the gray value and the number of pixels having the same gray value.[]{data-label="fig:synth7hist"}](synth7hist_4.pdf)
![A noisy synthetic image and the first 7 eigenfunctions, left to right, top to bottom. \[fig:nzsynth1x\]](nzsynth1x_set){width="10cm"}
### Eigenvalue problem (\[eq:HW-eigen\]) versus Laplace-Beltrami operator
Eigenfunctions of the Laplace-Beltrami operator (on surfaces) have been studied for image segmentation [@Shah-00; @Sochen-Kimmel-Malladi-98] and shape analysis [@Reuter-09; @Reuter-06]. Thus, it is natural to compare the performance of the Laplace-Beltrami operator and that of the eigenvalue problem (\[eq:HW-eigen\]). For this purpose, we choose a surface such that the Laplace-Beltrami operator has the same diffusion matrix as that defined in (\[D-2\]), (\[D-4\]), and (\[D-6\]) and takes the form as $$- \nabla \cdot \left(\frac{1}{\sqrt{1+|\nabla u|^{2}}}\begin{bmatrix}1+|\partial_y u_{0}|^{2} & -|\partial_x u_{0}\partial_y u_{0}|\\
-|\partial_x u_{0}\partial_y u_{0}| & 1+|\partial_x u_{0}|^{2}
\end{bmatrix}\nabla u\right) = \lambda \sqrt{1+|\nabla u_0|^{2}}\; u .
\label{LB-1}$$ The main difference between this eigenvalue problem with (\[eq:HW-eigen\]) is that there is a weight function on the right-hand side of (\[LB-1\]), and in our model the parameter $\alpha$ in (\[D-6\]) is typically greater than 1.
The eigenfunctions of the Laplace-Beltrami operator obtained with a clean input image of Fig. \[fig:nzsynth1x\] are shown in Fig. \[fig:LB-eigenfunctions\]. From these figures one can see that the eigenfunctions of the Laplace-Beltrami operator are far less close to being piecewise constant, and thus, less suitable for image segmentation.
![The clean input image of Fig. \[fig:nzsynth1x\] and the first 7 eigenfunctions of its associated Laplace-Beltrami operator.[]{data-label="fig:LB-eigenfunctions"}](LBsynth1x_set){width="10cm"}
### Open or closed edges
We continue to study the piecewise constant property of eigenfunctions of (\[eq:HW-eigen\]). Interestingly, this property seems related to whether the edges of the input image form a closed curve. We examine the two input images in Fig. \[fig:openarc\], one containing a few open arcs and the other having a closed curve that makes a jump in the gray level. The first eigenfunction for the open-arc image changes gradually where that for the second image is close to being piecewise constant.
![From left to right, input image with open arcs, the corresponding first eigenfunction, input image with connected arcs, the corresponding first eigenfunction.[]{data-label="fig:openarc"}](openarc){width="10cm"}
### Anisotropic mesh adaptation
For the purpose of image segmentation, we would like the eigenfunctions to be as close to being piecewise constant as possible. This would mean that they change abruptly in narrow regions between objects. As a consequence, their numerical approximation can be difficult, and (anisotropic) mesh adaptation is then necessary in lieu of accuracy and efficiency. The reader is referred to [@Huang-Wang-13] for the detailed studies of convergence and advantages of using anisotropic mesh adaptation in finite element approximation of anisotropic eigenvalue problems with anisotropic diffusion operators. Here, we demonstrate the advantage of using an anisotropic adaptive mesh over a uniform one for the eigenvalue problem (\[eq:HW-eigen\]) with the diffusion matrix defined in (\[D-2\]), (\[D-4\]), and (\[D-6\]) and subject to the homogeneous Dirichlet boundary condition. The input image is taken as the Stanford bunny; see Fig. \[fig:bunny41\]. The figure also shows the eigenfunctions obtained on an adaptive mesh and uniform meshes of several sizes. It can be seen that the eigenfunctions obtained with the adaptive mesh have very sharp boundaries, which are comparable to those obtained with a uniform mesh of more than ten times of vertices.
![Top row: the image of the Stanford bunny and the first 3 eigenfunctions computed with an anisotropic adaptive mesh with 45383 vertices; Bottom row: the first eigenfunction on a uniform mesh with 93732, 276044, 550394 vertices and on an adaptive mesh with 45383 vertices, respectively. All eigenfunctions are computed with the same diffusion matrix defined in (\[D-2\]), (\[D-4\]), and (\[D-6\]) ($\alpha = 1.5$) and subject to the homogeneous Dirichlet boundary condition.[]{data-label="fig:bunny41"}](bunny_set_cmp){width="10cm"}
### Anisotropic and less anisotropic diffusion
Next, we compare the performance of the diffusion matrix (\[D-2\]) (with (\[D-4\]), (\[D-6\]), and $\alpha = 1.5$) and that of a less anisotropic diffusion matrix (cf. (\[eq:pm-linear-diffusion\]), with (\[D-6\]) and $\alpha = 1.5$) $$\mathbb{D} = \begin{bmatrix}g(|\partial_{x}u_0 |) & 0\\ 0 & g(|\partial_{y} u_0|) \end{bmatrix} .
\label{D-7}$$ The eigenfunctions of (\[eq:HW-eigen\]) with those diffusion matrices with the Stanford bunny as the input image are shown in Fig. \[fig:ani-vs-iso\]. For (\[D-7\]), we compute the eigenfunction on both a uniform mesh of size $256\times256$ and an adaptive mesh of 46974 vertices. The computation with (\[D-2\]) is done with an adaptive mesh of 45562 vertices. The most perceptible difference in the results is that the right ear of the bunny (not as bright as other parts) almost disappears in the first eigenfunction with the less anisotropic diffusion matrix. This can be recovered if the conductance is changed from $\alpha = 1.5$ to $\alpha = 1.0$, but in this case, the eigenfunction becomes farther from being piecewise-constant. The image associated with the first eigenfunction for (\[D-2\]) seems sharper than that with (\[D-7\]).
![The first eigenfunction of (\[eq:HW-eigen\]) with the Stanford bunny and with the diffusion matrix (\[D-2\]) and a less anisotropic one (\[D-7\]). From left to right, (\[D-2\]) on an adaptive mesh of 45562 vertices, (\[D-7\]) on a uniform mesh of size $256\times256$, and (\[D-7\]) on an adaptive mesh of 46974 vertices.[]{data-label="fig:ani-vs-iso"}](bunny_aniVSiso){width="10cm"}
### Effects of the conductance $g$ {#SEC:4.1.6}
We now examine the effects of the conductance and consider four cases: $g_1$ ((\[D-6\]) with $\alpha = 1.0$), $g_2$ ((\[D-6\]) with $\alpha = 1.5$), $g_3$ ((\[D-6\]) with $\alpha = 3.0$), and $$g_{4}(x)=\begin{cases}
(1-(x/\sigma)^{2})^{2}/2, & \text{ for }|x|\le\sigma\\
\epsilon, & \text{ for }|x|>\sigma
\end{cases}$$ where $\sigma$ and $\epsilon$ are positive parameters. The last function is called Tukey’s biweight function and considered in [@Black-Sapiro-01] as a more robust choice of the edge-stopping function in the Perona-Malik diffusion. We show the results with (\[D-2\]) on the Stanford bunny in Fig. \[fig:g-choice\]. We take $\sigma=9$ and $\epsilon=10^{-6}$ for Tukey’s biweight function. Increasing the power $\alpha$ in $g(x)$ defined in (\[D-6\]) will make eigenfunctions steeper in the regions between different objects and thus, closer to being piecewise constant. Tukey’s biweight function gives a sharp result but the body and legs are indistinguishable.
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](g_curves.pdf)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyef_2_1.png)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyef_2_1dot5.png)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyef_2_2.png)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyef_tukey.png)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyefhist_2_1.pdf)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyefhist_2_1dot5.pdf)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyefhist_2_2.pdf)
![Top row: the graphs of $g_1, g_2, g_3, g_4$. Middle row: the first eigenfunctions on the bunny image for $g_1, g_2, g_3, g_4$, respectively. Bottom row: the histograms of the corresponding first eigenfunctions.[]{data-label="fig:g-choice"}](bunnyefhist_tukey.pdf)
Applications in Edge Detection and Image Segmentation
-----------------------------------------------------
Eigenfunctions can serve as a low level image feature extraction device to facilitate image segmentation or object edge detection. Generally speaking, eigenfunctions associated with small eigenvalues contain “global” segmentation features of an image while eigenfunctions associated with larger eigenvalues carry more information on the detail. Once the eigenfunctions are obtained, one can use numerous well developed edge detection or data clustering techniques to extract edge or segmentation information. We point out that spectral clustering methods also follow this paradigm. In this section, we focus on the feature extraction step and employ only simple, well known techniques such as thresholding by hand, $k$-means clustering, or Canny edge detector in the partitioning step. More sophisticated schemes can be easily integrated to automatically detect edges or get the segmentations.
We point out that boundary conditions have an interesting effect on the eigenfunctions. A homogeneous Dirichlet boundary condition forces the eigenfunctions to be zero on the boundary and may wipe out some structures there (and therefore, emphasize objects inside the domain). It essentially plays the role of defining “seeds” that indicates background pixels on the image border. The idea of using user-defined seeds or intervene cues has been widely used in graph based image segmentation methods [@Grady-01], [@Rother-Blake-04], [@Shi-Yu-04], [@Malik-Martin-04]. The PDE eigenvalue problem (\[eq:HW-eigen\]) can also be solved with more sophisticated boundary conditions that are defined either on the image border or inside the image. On the other hand, a homogeneous Neumann boundary condition tends to keep those structures. Since mostly we are interested in objects inside the domain, we consider here a homogeneous Dirichlet boundary condition. The diffusion matrix defined in (\[D-2\]), (\[D-4\]), and (\[D-6\]) ($\alpha = 1.5$) is used.
In Fig. \[fig:Lenna\_ef1\], we show the first eigenfunctions obtained with Dirichlet and Neumann boundary conditions with Lenna as the input image. For the edge detection for Lenna, it is natural to extract the “big picture” from the first eigenfunction and get the edge information from it. We show the edges obtained by thresholding a few level lines in the top row of Fig. \[fig:lenna-contour\]. Since any level line with value $s$ is the boundary of the level set $L_{s}=\{(x,y):I(x,y)\ge s\}$ of an image $I$, and $L_{s}$ is non-increasing with respect to $s$, the level line is “shrinking” from the boundary of a wider shape to empty as $s$ increases from 0 to 255. Some intermediate steps give salient boundaries of the interior figure. However, to make the “shrinking” automatically stop at the correct edge, other clues potentially from mid or high level knowledge in addition to the low level brightness info should be integrated in the edge detection step. We also use the MATLAB function [imcontour]{} to get major contours, and apply $k$-means clustering to the eigenfunctions with $k=2,3,4,5$, shown in the second row of Fig. \[fig:lenna-contour\].
![From left to right, Lenna, first eigenfunctions obtained with Dirichlet and Neumann boundary conditions, respectively.[]{data-label="fig:Lenna_ef1"}](lenna){width="10cm"}
![Top row: contour drawing by MATLAB (with no level parameters specified), level line 50, 240, 249; bottom row: segmentation with $k$-means, $k=2,3,4,5$.[]{data-label="fig:lenna-contour"}](lenna_gallery){width="12cm"}
We next compute for an image with more textures from [@MartinFTM01] (Fig. \[fig:tiger-gallery\]). This is a more difficult image for segmentation or edge detection due to many open boundary arcs and ill-defined boundaries. We display the the first eigenfunction and the $k$-means clustering results in Fig. \[fig:tiger-gallery\]. The $k$-means clustering does not capture the object as well as in the previous example. Better separation of the object and the background can be obtained if additional information is integrated into the clustering strategy. For instance, the edges detected by the Canny detector (which uses the gradient magnitude of the image) on the first eigenfunction clearly give the location of the tiger. Thus, the use of the gradient map of the first eigenfunction in the clustering process yields more accurate object boundaries. For comparison, we also show the edges detected from the input image with the Canny detector.
![Top row: the input image, the edges of the input image with the Canny detector, Level lines with value 50, 240, 254. bottom row: the first eigenfunction, the edges of the first eigenfunction with the Canny detector, $k$-means clustering with $k=2,3,4$.[]{data-label="fig:tiger-gallery"}](tiger_gallery){width="14cm"}
Another way to extract “simple” features is to change the conductance $g$ (e.g., by increasing $\alpha$ in (\[D-6\])) to make the eigenfunctions closer to being piecewise constant. This makes eigenfunctions more clustered but wipes out some detail of the image too. To avoid this difficulty, we can employ a number of eigenfunctions and use the projection of the input image into the space spanned by the eigenfunctions to construct a composite image. A much better result obtained in this way with 64 eigenfunctions is shown in Fig. \[fig:tigerfin\].
![From left to right, the first eigenfunction with $\alpha = 2$ in (\[D-6\]), the projection of the tiger image into the space spanned by the first 64 eigenfunctions, the contour of the projection image, the $k$-means clustering with $k=2$ of the projection image.[]{data-label="fig:tigerfin"}](tiger_eigen_1.png)
![From left to right, the first eigenfunction with $\alpha = 2$ in (\[D-6\]), the projection of the tiger image into the space spanned by the first 64 eigenfunctions, the contour of the projection image, the $k$-means clustering with $k=2$ of the projection image.[]{data-label="fig:tigerfin"}](tiger_fin_1.png)
![From left to right, the first eigenfunction with $\alpha = 2$ in (\[D-6\]), the projection of the tiger image into the space spanned by the first 64 eigenfunctions, the contour of the projection image, the $k$-means clustering with $k=2$ of the projection image.[]{data-label="fig:tigerfin"}](tigerfin_contour.pdf)
![From left to right, the first eigenfunction with $\alpha = 2$ in (\[D-6\]), the projection of the tiger image into the space spanned by the first 64 eigenfunctions, the contour of the projection image, the $k$-means clustering with $k=2$ of the projection image.[]{data-label="fig:tigerfin"}](tigerfin_kmeans2.pdf)
It should be pointed out that not always the first few eigenfunctions cary most useful information of the input image. Indeed, Fig. \[fig:sports-gallery\] shows that the first eigenfunction carries very little information. Since the eigenfunctions form an orthogonal set in $L^{2}$, we can project the input image onto the computed eigenfunctions. The coefficients are shown in Fig. \[fig:sports-components\]. We can see that the coefficients for the first two eigenfunctions are very small compared with those for the several following eigenfunctions. It is reasonable to use the eigenfunctions with the greatest magnitudes of the coefficients. These major eigenfunctions will provide most useful information; see Fig. \[fig:sports-gallery\].
![Top row: from left to right, the input image and the first 6 eigenfunctions. Bottom row: from left to right, the edges on the input image (Canny), the edges on the 3rd and 4th eigenfunctions (Canny), the $k$-means clustering results with $k=3$ for the 3rd and the 4th eigenfunctions, respectively; Level line of value 205 of the 3rd eigenfunction, level line of value 150 of the 4th eigenfunction.[]{data-label="fig:sports-gallery"}](sports-gallery){width="14cm"}
![The coefficients of the input image projected onto the first 64 eigenfunctions in Fig. \[fig:sports-gallery\].[]{data-label="fig:sports-components"}](sports-comp)
The piecewise constant property of eigenfunctions {#SEC:piecewise}
=================================================
As we have seen in Section \[SEC:numerics\], the eigenfunctions of problem (\[eq:HW-eigen\]) are localized in sub-regions of the input image and the first few of them are close to being piecewise constant for most input images except for two types of images. The first type of images is those containing regions of which part of their boundaries is not clearly defined (such as open arcs that are common in natural images). In this case, the first eigenfunction is no longer piecewise-constant although the function values can still be well clustered. The other type is input images for which the gray level changes gradually and its gradient is bounded (i.e., the image contrast is mild). In this case, the diffusion operator simply behaves like the Laplace operator and has smooth eigenfunctions. For other types of images, the gray level has an abrupt change across the edges of objects, which causes the conductance $g(|\nabla u_0|)$ to become nearly zero on the boundaries between the objects. As a consequence, the first few eigenfunctions are close to being constant within each object. This property forms the basis for the use of the eigenvalue problem (\[eq:HW-eigen\]) (and its eigenfunctions) in image segmentation and edge detection. In this section, we attempt to explain this property from the physical, mathematical, and graph spectral points of view. We hope that the analysis, although not rigorous, provides some insight of the phenomenon.
From the physical point of view, when the conductance $g(|\nabla u_0|)$ becomes nearly zero across the boundaries between the objects, the diffusion flux will be nearly zero and each object can be viewed as a separated region from other objects. As a consequence, the eigenvalue problem can be viewed as a problem defined on multiple separated subdomains, subject to homogeneous Neumann boundary conditions (a.k.a. insulated boundary conditions) on the boundary of the whole image and the internal boundaries between the objects. Then, it is easy to see that the eigenfunctions corresponding to the eigenvalue 0 include constant and piecewise constant (taking a different constant value on each object) functions. This may explain why piecewise constant eigenfunctions have been observed for most input images. On the other hand, for images with mild contrast or open arc object edges, the portion of the domain associated any object is no longer totally separated from other objects and thus the eigenvalue problem may not have piecewise constant eigenfunctions.
![A piecewise smooth function representing an input image with two objects.[]{data-label="fig:a-fun"}](funf)
Mathematically, we consider a 1D example with an input image with two objects. The gray level of the image is sketched in Fig. \[fig:a-fun\]. The edge is located at the origin, and the segmentation of the image is a 2-partition of $[-L,0]$ and $[0,L]$. The 1D version of the eigenvalue problem (\[eq:HW-eigen\]) reads as $$- \frac{d}{d x}\left(g(|u_0'(x) |)\frac{d u}{d x} \right) =\lambda u ,
\label{eq:example1}$$ subject to the Neumann boundary conditions $u'(-L)=u'(L)=0$. We take the conductance function as in (\[D-6\]) with $\alpha = 2$. Although $u_0$ is not differentiable, we could imagine that $u_0$ were replaced by a smoothed function which has a very large or an infinite derivative at the origin. Then, (\[eq:example1\]) is degenerate since $g(|u_0'(x)|$ vanishes at $x=0$. As a consequence, its eigenfunctions can be non-smooth. Generally speaking, studying the eigenvalue problem of a degenerate elliptic operator is a difficult task, and this is also beyond the scope of the current work. Instead of performing a rigorous analysis, we consider a simple approximation to $g(|u_0'(x)|)$, $$g_{\epsilon}(x)=\begin{cases}
g(|u_0'(x)|), & \text{ for }-L\le x<-\epsilon{\rm \ or\ }\epsilon<x\le L\\
0, & \text{ for }-\epsilon \le x\le \epsilon
\end{cases}$$ where $\epsilon$ is a small positive number. The corresponding approximate eigenvalue problem is $$- \frac{d}{d x}\left(g_\epsilon(|u_0'(x) |)\frac{d u}{d x} \right) =\lambda u .
\label{eq:exampe1-approx}$$ The variational formulation of this eigenvalue problem is given by $$\min_{u\in H^1(-L,L)}\int_{-L}^{L}g_{\epsilon}(|u_0'(x)|)(u')^{2}, \quad \text{ subject to } \int_{-L}^{L}u^{2}=1 .
\label{eq:ex1-variation-prob}$$ Once again, the eigenvalue problem (\[eq:exampe1-approx\]) and the variational problem (\[eq:ex1-variation-prob\]) are degenerate and should be allowed to admit non-smooth solutions.
The first eigenvalue of (\[eq:exampe1-approx\]) is 0, and a trivial eigenfunction associated with this eigenvalue is a constant function. To get other eigenfunctions associated with 0, we consider functions that are orthogonal to constant eigenfunctions, i.e., we append to the optimization problem (\[eq:ex1-variation-prob\]) with the constraint $$\int_{-L}^{L}u=0.
\label{eq:ex1-constraints}$$ It can be verified that an eigenfunction is $$u_{\epsilon}(x)=\begin{cases}
-c, & \text{ for }x\in [-L,-\epsilon)\\
\frac{c x}{\epsilon}, & \text{ for } x\in [-\epsilon, \epsilon]\\
c, & \text{ for } x\in (\epsilon, L]
\end{cases}
\label{1D-solution}$$ where $c = (2 (L-2\epsilon/3))^{-1/2}$. This function is piecewise constant for most part of the domain except the small region $[-\epsilon, \epsilon]$. Since the original problem (\[eq:example1\]) can be viewed to some extent as the limit of (\[eq:exampe1-approx\]) as $\epsilon \to 0$, the above analysis may explain why some of the eigenfunctions of (\[eq:example1\]) behave like piecewise constant functions.
The piecewise constant property can also be understood in the context of the graph spectral theory. We first state a result from [@Mohar-Alavi-91; @Luxburg-2007].
Assume that $G$ is a undirected graph with $k$ connected components and the edge weights between those components are zero. If the nonnegative weights matrix $W$ and the diagonal matrix $D$ are defined as in Section \[SEC:relation-1\], then the multiplicity of the eigenvalue 0 of the matrix $D-W$ equals the number of the connected components in the graph. Moreover, the eigenspace of eigenvalue 0 is spanned by the indicator vectors $1_{A_{1}},\cdots,1_{A_{k}}$ of those components ${A_{1}},\cdots, {A_{k}}$. \[prop-1\]
This proposition shows that the eigenvalue zero of the algebraic eigenvalue problem (\[eq:Malik-Shi-eigen\]) could admit multiple eigenvectors as indicators of the components. As shown in Section \[SEC:relation-1\], (\[eq:Malik-Shi-eigen\]) can be derived from a finite difference discretization of the continuous eigenvalue problem (\[eq:HW-eigen\]) (with a proper choice of $\mathbb{D}$). Thus, the indicators of the components can also be regarded as discrete approximations of some continuous eigenfunctions. This implies that the latter must behave like piecewise constant functions. Interestingly, Szlam and Bresson [@Szlam-Bresson-10] recently proved that global binary minimizers exist for a graph based problem called Cheeger Cut where the minimum of the cut is not necessarily zero.
In the continuous setting, a properly designed conductance $g(|\nabla u_0|)$ can act like cutting the input image $u_0$ into subregions along the boundary of the subregions and forcing the eigenvalue problem to be solved on each subregion. In a simplified case, we can have the following continuous analogue of Proposition \[prop-1\]. The proof is straightforward.
Suppose $\Omega\subset \mathbb{R}^2 $ is a bounded Lipschitz domain, $u_0 \in SBV(\Omega)$ (the collect of special functions of bounded variation) and the discontinuity set $K$ of $u_0$ is a finite union of $C^{1}$ closed curves, $g(\cdot)$ is a bounded positive continuous function. We define $g(|\nabla u_0|)=0$ for $(x,y)\in K$, and $g(|\nabla u_0|)$ takes its usual meaning for $(x,y)\in\Omega\backslash K$. For any function $u\in SBV(\Omega)$, assuming $\Gamma$ is the discontinuity set of $u$, we define the Rayleigh quotient on $u$ as $$R(u)=\frac{\widetilde{\int}_{\Omega}(\nabla u)^{T}g(|\nabla u_0|)\nabla u}{\int_{\Omega}u^{2}} ,$$ where $$\widetilde{\int}_{\Omega}(\nabla u)^{T}g(|\nabla u_0|)\nabla u=\begin{cases}
\int_{\Omega\backslash K}g(\nabla u_0|)|\nabla u|^{2}, & \text{ for }\Gamma\subseteq K\\
\infty, & \text{ for } \Gamma\nsubseteq K .
\end{cases}$$ Then, the minimum of $R(u)$ is zero and any piecewise constant function in $SBV(\Omega)$ with discontinuity set in $K$ is a minimizer. \[prop-2\]
The eigenvalue problem related to the above variational problem can be formally written as
$$-\nabla \cdot \left(g(|\nabla u_0|)\nabla u\right)=\lambda u.
\label{eq:general-eigen-prob}$$
The equation should be properly defined for all $u_0$, $u$ possibly in the space of $BV$. This is a degenerate elliptic problem which could admit discontinuous solutions, and it seems to be far from being fully understood. In the following proposition, we suggest a definition of weak solution in a simplified case. The property indicates that problem (\[eq:general-eigen-prob\]) is quite different from a classical elliptic eigenvalue problem if it has a solution in $BV$ that takes a non-zero constant value on an open set. The proof is not difficult and thus omitted.
Suppose $\Omega$ is a bounded Lipschitz domain in $R^{2}$, $u_0\in SBV(\Omega)$ and the discontinuity set $K$ of $u_0$ is a finite union of $C^{1}$ closed curves, $g(\cdot)$ is a bounded positive continuous function. We define $g(|\nabla u_0|)=0$ for $(x,y)\in K$, and $g(|\nabla u_0|)$ takes its usual meaning for $(x,y)\in\Omega\backslash K$. We define $u\in SBV(\Omega)$ to be a weak eigenfunction of (\[eq:general-eigen-prob\]) satisfying a homogeneous Dirichlet boundary condition if $$\int_{\Omega}(\nabla u)^{T}g(|\nabla u_0|)\nabla\phi=\int_{\Omega}\lambda u\phi,
\quad \forall \phi \in C_{0}^{1}(\Omega)$$ where, assuming that $\Gamma$ is the discontinuity set of $u$, the integral on the left side is defined by $$\int_{\Omega}(\nabla u)^{T}g(|\nabla u_0|)\phi=\begin{cases}
\int_{\Omega\backslash K}(\nabla u)^{T}g(\nabla u_0|)\phi, & \Gamma\subseteq K\\
\infty, & {\rm \Gamma\nsubseteq K.}
\end{cases}$$ If a weak eigenfunction $u\in SBV(\Omega)$ exists and takes a non-zero constant value on a ball $B_{\epsilon}(x_{0}, y_0)\subset\Omega$, then the corresponding eigenvalue $\lambda$ is zero. \[prop-3\]
If (\[eq:general-eigen-prob\]) indeed admits non-zero piecewise-constant eigenfunctions, one can see an interesting connection between (\[eq:general-eigen-prob\]) (for simplicity we assume a homogeneous Dirichlet boundary condition is used) and Grady’s Random Walk image segmentation model [@Grady-01] where multiple combinatorial Dirichlet problems are solved for a $k$-region segmentation with predefined seeds indicating segmentation labels. Using a similar argument in Section 2.2, one can show that the numerical implementation of the method is equivalent to solving a set of Laplace problems which are subject to a Neumann boundary condition on the image border and Dirichlet boundary conditions on the seeds and are discretized on a uniform mesh for potentials $u^{i}$, $i=1,\cdots,k$. These boundary problems read as $$\begin{aligned}
\nabla \cdot \left(\begin{bmatrix}g(|\partial_{x}u_{0}|) & 0\\
0 & g(|\partial_{y}u_{0}|)
\end{bmatrix}\nabla u^{i}\right) & = & 0,\ {\rm in}\ \Omega\backslash S
\label{eq:Grady-model}\\
\frac{\partial u^{i}}{\partial n} & = & 0,\ {\rm on}\ \partial\Omega \nonumber \\
u^{i} & = & 1,\ {\rm in}\ S_{i} \nonumber \\
u^{i} & = & 0,\ {\rm in}\ S\backslash S_{i}
\nonumber \end{aligned}$$ where $S_{i}$ is the set of seeds for label $i$ and $S$ is the set of all seeds. This problem with a proper choice of $g$ also gives a solution that has well clustered function values, a phenomenon called “histogram concentration” in [@Buades-Chien-08].
Note that when $\lambda=0$, the following equation, which is an anisotropic generalization of (\[eq:general-eigen-prob\]), $$-\nabla \cdot
\left(\begin{bmatrix}g(|\partial_{x}u_{0}|) & 0\\
0 & g(|\partial_{y}u_{0}|)
\end{bmatrix}\nabla u^{i}\right) = \lambda u^i,$$ becomes exactly the Laplace equation in Grady’s model which is the Euler-Lagrange equation of the energy $$\int_{\Omega}(\nabla u)^{T}\begin{bmatrix}g(|\partial_{x}u_{0}|) & 0\\
0 & g(|\partial_{y}u_{0}|)
\end{bmatrix}\nabla u .
\label{eq:bv-energy}$$ While the proper definition of functional (\[eq:bv-energy\]) for general $u$, $u_{0}$ possibly in $BV$ is missing, we can still define it for a simpler case as in Proposition 5.2. Then, there is no unique minimizer of this energy, and, as stated in Proposition 5.2, any minimizer of the above energy in $SBV$ yields the minimum value 0 in the ideal case (with proper $g$ and $u_{0}$ as in Proposition 5.2). While the eigenvalue method considers the minimizer of the above energy on the admissible set $\left\{ u:u|_{\partial\Omega}=0,\int u^{2}\ dx=1\right\} $, the Random Walk method considers the minimizer satisfying boundary conditions in (\[eq:Grady-model\]). Both gives piecewise-constant minimizers that can be used for image segmentation.
Concluding remarks {#SEC:conclusion}
==================
We have introduced an eigenvalue problem of an anisotropic differential operator as a tool for image segmentation. It is a continuous and anisotropic generalization of some commonly used, discrete spectral clustering models for image segmentation. The continuous formulation of the eigenvalue problem allows for accurate and efficient numerical implementation, which is crucial in locating the boundaries in an image. An important observation from numerical experiment is that non-trivial, almost piecewise constant eigenfunctions associated with very small eigenvalues exist, and this phenomenon seems to be an inherent property of the model. These eigenfunctions can be used as the basis for image segmentation and edge detection. The mathematical theory behind this is still unknown and will be an interesting topic for future research.
We have implemented our model with a finite element method and shown that anisotropic mesh adaptation is essential to the accuracy and efficiency for the numerical solution of the model. Numerical tests on segmentation of synthetic, natural or texture images based on computed eigenfunctions have been conducted. It has been shown that the adaptive mesh implementation of the model can lead to a significant gain in efficiency. Moreover, numerical results also show that the anisotropic nature of the model can enhance some nontrivial regions of eigenfunctions which may not be captured by a less anisotropic or an isotropic model.
[10]{}
L. Alvarez, P.-L. Lions, and J.-M. Morel. Image selective smoothing and edge detection by nonlinear diffusion. [II]{}. , 29(3):845–866, 1992.
M. Belkin and P. Niyogi. Towards a theoretical foundation for laplacian-based manifold methods. In [*Learning theory*]{}, pages 486–500. Springer, 2005.
M. Black, G. Sapiro, D. Marimont, and D. Heeger. Robust anisotropic diffusion. , 7(3):421–432, 1998.
D. Boffi. Finite element approximation of eigenvalue problems. , 19:1–120, 2010.
A. Buades, A. Chien, J. M. Morel, and S. J. Osher. Topology preserving linear filtering applied to medical imaging. , 1:26–50, 2008.
F. Catt[é]{}, P.-L. Lions, J.-M. Morel, and T. Coll. Image selective smoothing and edge detection by nonlinear diffusion. , 29(1):182–193, 1992.
L. Grady. Random walks for image segmentation. , 28(11):1768–1783, 2006.
S. Gunter, Q. Yu, J. Kruger, and K. Lackner. Modelling of heat transport in magnetised plasmas using non-aligned coordinates. , 209:354–370, 2005.
F. Hecht. Bamg: bidimensional anisotropic mesh generator. , 1998.
W. Huang. Metric tensors for anisotropic mesh generation. , 204(2):633–665, 2005.
W. Huang. Mathematical principles of anisotropic mesh adaptation. , 204(1):276–310, 2006.
W. Huang and R. D. Russell. , volume 174 of [*Applied Mathematical Sciences*]{}. Springer, New York, 2011.
D. Martin, C. Fowlkes, and J. Malik. Learning to detect natural image boundaries using local brightness, color, and texture cues. , 11(13):293–327, 2001.
D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In [*Proc. 8th Int’l Conf. Computer Vision*]{}, volume 2, pages 416–423, July 2001.
B. Mohar and Y. Alavi. The laplacian spectrum of graphs. , 2:871–898, 1991.
B. Nadler, S. Lafon, R. Coifman, and I. Kevrekidis. Diffusion maps, spectral clustering and eigenfunctions of fokker-planck operators. In [*[NIPS]{}*]{}, 2005.
B. Nadler, S. Lafon, R. R. Coifman, and I. G. Kevrekidis. Diffusion maps, spectral clustering and reaction coordinates of dynamical systems. , 21(1):113–127, 2001.
M. Nitzberg and T. Shiota. Nonlinear image filtering with edge and corner enhancement. , 14(8):826–833, 1992.
P. Perona and J. Malik. Scale-space and edge detection using anisotropic diffusion. , 12(7):629–639, 1990.
P.-A. Raviart and J.-M. Thomas. . Collection Mathématiques Appliquées pour la Maîtrise. \[Collection of Applied Mathematics for the Master’s Degree\]. Masson, Paris, 1983.
M. Reuter, S. Biasotti, D. Giorgi, G. Patan[è]{}, and M. Spagnuolo. Discrete laplace-beltrami operators for shape analysis and segmentation. , 33(3):381–390, 2009.
M. Reuter, F.-E. Wolter, and N. Peinecke. Laplace-beltrami spectra as “shape-dna” of surfaces and solids. , 38(4):342–366, 2006.
C. Rother, V. Kolmogorov, and A. Blake. Grabcut: Interactive foreground extraction using iterated graph cuts. , 23(3):309–314, 2004.
J. Shah. Riemannian drums, anisotropic curve evolution, and segmentation. , 11(2):142–153, 2000.
P. Sharma and G. Hammett. Preserving monotonicity in anisotropic diffusion. , 227:123–142, 2007.
J. Shi and J. Malik. Normalized cuts and image segmentation. , 22(8):888–905, 2000.
J. Shi and M. Malik. A random walks view of spectral segmentation. , 1(1):1, 2001.
A. Singer. From graph to manifold [L]{}aplacian: the convergence rate. , 21(1):128–134, 2006.
N. Sochen, R. Kimmel, and R. Malladi. A general framework for low level vision. , 7(3):310–318, 1998.
A. D. Szlam and X. Bresson. Total variation, cheeger cuts. , pages 1039–1046, 2010.
U. Von Luxburg. A tutorial on spectral clustering. , 17(4):295–416, 2007.
J. Wang and W. Huang. A study on anisotropic mesh adaptation for finite element approximation of eigenvalue problems with anisotropic diffusion operators. .
J. Weickert. Theoretical foundations of anisotropic diffusion in image processing. , Suppl. 11:221–236, 1996.
Z. Wu and R. Leahy. An optimal graph theoretic approach to data clustering: Theory and its application to image segmentation. , 15(11):1101–1113, 1993.
G.-Z. Yang, P. Burger, D. Firmin, and S. Underwood. Structure adaptive anisotropic image filtering. , 14(2):135–145, 1996.
S. Yu and J. Shi. Segmentation given partial grouping constraints. , 26(2):173–183, 2004.
F. Zhang and E. Hancock. Graph spectral image smoothing using the heat kernel. , 41(11):3328–3342, 2008.
[^1]: Department of Mathematics, the University of Kansas, Lawrence, KS 66045, U.S.A. ([*jwang@math.ku.edu*]{}).
[^2]: Department of Mathematics, the University of Kansas, Lawrence, KS 66045, U.S.A. ([*whuang@ku.edu*]{}).
[^3]: This work was supported in part by the NSF under Grant DMS-1115118.
|
---
abstract: 'The ionization dynamics of a Xenon cluster with 40 atoms is analyzed under a pump probe scenario of laser pulses where an infrared laser pulse of 50 fs length follows with a well defined time delay a VUV pulse of the same length and peak intensity. The mechanism of resonant energy absorption due to the coincidence of the IR laser frequency with the frequency of collective motion of quasi free electrons in the cluster is mapped out by varying the time delay between the pulses.'
author:
- Christian Siedschlag
- 'Jan M. Rost'
title: '**Surface plasma resonance in small rare gas clusters by mixing IR and VUV laser pulses**'
---
In recent years, much work has been devoted to the ionization mechanisms of clusters in few-cycle, intense laser fields (i.e. pulse lengths of the order of 100 fs and intensities $I=10^{13} \ldots 10^{16} \mathrm{W/cm^2}$): from the case of plasmon excitation when exposing metal clusters to relatively weak fields [@SuRe00] over [*enhanced ionization*]{} akin of molecular ionization for small rare gas clusters in intense fields [@SiRo02] to collective excitation of a plasma resonance in clusters of intermediate [@SaRo03] to large sizes [@jung04], ultimately leading to ionic charge states of 40+ and higher [@ditmire], thus potentially providing a new source for the generation of x-rays, energetic ions or electrons and, via nuclear fusion, even neutrons [@jortner]. A new parameter regime for laser-cluster interaction has been proven to become accesible with the first experiment using VUV-FEL light of 98nm wavelength for the ionization of rare gas clusters [@Waal02], soon followed by the first proposals for an explanation of the unexpectedly high charge states seen in this experiment [@SaGr03; @SiRo04].
While XUV-cluster interaction is still the subject of an ongoing debate, there seems to be a more or less common understanding regarding the qualitative picture of IR laser-cluster interaction: during the rising part of the laser pulse, a few electrons are ionized [^1] leaving the cluster with a net positive charge which leads to an expansion typically on the same time scale as the duration of the laser pulse. Hence, effects which depend on the internuclear distances can be resolved by varying the pulse length [@kolal99] and/or applying pump-probe techniques [@dopal01; @Zaal04]. The resolution of an optimum time delay $\Delta t$ and the contrast of the signal in a pump-probe experiment increases if $\Delta t \gg T$, the length of each pulse. Since $\Delta t \approx t_{c}$, the critical expansion time of the cluster at which maximum absorption of energy from the cluster pulse is possible, long times $t_{c}$ are desirable which implies large clusters consisting of heavy atoms (slower Coulomb explosion). Also, the large number of quasi-free electrons temporarily trapped in the cluster, lead to a good contrast for the optimized versus non-optimized signal [@Zaal04].
The critical time $t_{c}$ originates from a critical radius $R_{c} = R(t_{c})$ of the cluster, usually larger than the equilibrium radius $R_{0}$, where energy absorption is most efficient. For larger clusters (resonant mechanism) this radius is determined by the surface plasma frequency approximately given by $$\begin{aligned}
\label{plasmon}
\Omega_{t}=\sqrt{\frac{N_{t}Z_{t}}{R_{t}^3}}= \frac{\omega_{pl}}{\sqrt{3}},\end{aligned}$$ where $N_{t}$ is the number of atoms/ions in the cluster, $Z_{t}$ is their average charge, $R_{t}$ is the cluster radius and $\omega_{pl}$ the bulk plasma frequency. The indices $t$ indicate a slow dependence on time (at this point we want to emphasize that, at least as long as the cluster is neutral, the surface [*plasma*]{} resonance is mathematically completely equivalent to the surface [*plasmon*]{} resonance and can be derived along the same pathway. Nevertheless, we prefer to call it a plasma resonance since there is in principle a physical difference between a metal cluster being excited perturbatively and a rare gas cluster turned into a nanoplasma by a nonperturbative laser field). If $\Omega_t \approx \omega$, the laser frequency, then the cloud of electrons which are trapped inside the cluster behaves like a (damped) harmonic oscillator driven to resonance [@SaRo03], leading to efficient energy absorption and ionization.
The most important prerequisite for this mechanism is a significant amount of trapped (quasi-free) electrons before $R_{c}$ is reached, which can be achieved in two ways: either the laser field strength is small enough to leave enough electrons inside the cluster before $R=R_{c}$; this possibility is limited, however, by the fact that the inner ionization process will eventually not start if the field strength is too small, so that no quasi-free electrons will be created in the first place. For metal clusters, this problem does obviously not occur; on the other hand, one has to ionize the cluster to a certain degree in order to start the expansion process, and the range of intensities and pulse lengths which can start the Coulomb explosion while at the same time keeping the valence electron cloud intact is quite small [@calv00]. On the other hand, the force that keeps the electrons inside the cluster is generated by the space charge of the ions. Hence, going to larger clusters while leaving the average ion charge constant, will make it more and more difficult for electrons to leave the cluster, so that the number of trapped electrons will increase with the cluster size. This is the reason why the resonance absorption is much clearer seen with IR pulses for large clusters (compare [@dopal01] with [@Zaal04]).
To summarize, resonant absorption in intense IR fields occurs (i) for metal clusters in weak fields ($I \lesssim 10^{13}
\mathrm{W/cm^2}$); the size of the clusters then only plays a role in so far as it will change the expansion speed, which has to be accounted for by changing the pulse lengths accordingly; or (ii) for clusters with $N \gtrsim 10^{2}$ and intensities of $I\lesssim 10^{14} \mathrm{W/cm^2}$, where a substantial fraction of the quasi-free electrons which are created by the laser is kept inside the cluster, so that a collective oscillation can develop.
In the case of small rare gas clusters in strong IR fields, none of the above scenarions applies. Rather, ionization is dominated by [ *charge enhanced ionization*]{}, known already from diatomic molecules, where the shape of the interatomic barrier leads to an optimal distance between the cluster nuclei which results in an efficient interplay between inner and outer ionization [@enhanced]: the neighbouring charges must be close enough to start an [*ionization avalanche*]{} [@SiRo02] once the first electrons are created; on the other hand, they must not be too close in order to decrease the space charge which prevents the electrons from escaping the cluster. Obviously, the whole process relies on the ionization of the first electrons relatively early in the pulse; once the field strength drops significantly below the field strength $F_{th}$ required to field-ionize a single cluster atom ($F_{th}=E_b^2/4$, where $E_b$ is the first atomic ionization energy), the avalanche will not be started and the cluster will survive the radiation relatively undamaged. For Xe clusters, for example, $F_{th}=0.0493$ a.u., which corresponds to an intensity of $I_{th}=8.53\cdot 10^{13}
\mathrm{W/cm^2}$.
As has been shown in the Hamburg experiment, the threshold for an ionization avalanche is considerably lower when using VUV instead of IR light. With intensities of the order of $10^{12} \ \mathrm{W/cm^2}$ and a photon energy of 12.7 eV, complete breakup of Xe clusters and unexpectedly high ionic charges have been observed [@Waal02]. These findings can be explained [@SiRo04] by using standard atomic photoabsorption rates but taking into account the effective (inner) ionization threshold which is lowered by the surrounding charges in a cluster (see Fig \[fig1\]). Due to this mechanism and due to the fact that the quiver amplitude is two orders of magnitude smaller than for IR radiation, a VUV pulse is much more efficient in creating quasi-free electrons than an IR pulse of the same peak intensity.
This opens up an elegant way to study the dynamics of collectively excited electrons and hence the resonance absorption mechanism in a [*small*]{} cluster by combining a VUV pump pulse with a time delayed IR probe pulse: The VUV pulse generates a large number of quasi-free electrons. At same time the cluster gets only moderately charged and a slow expansion sets in mainly driven by the hydrodynamic pressure of the quasi-free electrons. Hence, one can observe with a time delayed probe pulse very cleanly the optimum condition for energy absorption by the quasi-free electrons as a function of cluster size starting at a size as small as $N=40$ as we will demonstrate with the following pump-probe scenario: a Xe$_{40}$ cluster is first irradiated by a 50 fs VUV pulse ($\omega=12.7$ eV, $I=7.9\cdot10^{12} \
\mathrm{W/cm^2}$) Then, after a variable time delay $\Delta t$, we apply a second pulse of the same length and intensity, but now with a wavelength of 780 nm. The simulation has been done using the quasiclassical model introduced in [@SiRo02].
\[\]\[\]\[1\][atom]{} \[\]\[\]\[1\][cluster]{} \[\]\[\]\[1\][x \[a.u.\]]{} \[\]\[\]\[1\][potential energy \[arb. units\]]{} ![Comparison of the inner ionization process for a single atom and a model cluster of three atoms. The effective barrier is lowered in a cluster due to the charged environment.[]{data-label="fig1"}](atom_cluster.xfig.eps "fig:")
\[\]\[\]\[0.9\][$\Delta t=0$ a.u.]{} \[\]\[\]\[0.9\][$\Delta t=1000$ a.u.]{} \[\]\[\]\[0.9\][$\Delta t=3500$ a.u.]{} \[\]\[\]\[0.9\][$\Delta t=8000 $ a.u.]{} \[t\]\[\]\[1\][charge state]{} \[b\]\[\]\[1\][counts \[arb. units\]]{} ![Single ion charge spectra after the interaction of Xe$_{40}$ with a VUV pump pulse followed by an IR probe pulse as a function of the delay $\Delta t$ between the two pulses. The highest (maximum and average) charge states are created for $\Delta t$=3500 a.u.[]{data-label="fig2"}](charge_spectra.xfig.eps "fig:")
The charge spectra after the interaction of the cluster with the two pulses is shown in Fig. \[fig2\] for various times delay $\Delta t$. Two features stand out: first of all, ions with charges of $5+$ or $6+$ are produced already with these comparatively low intensities. In fact, the ionization efficiency is comparable to the results from our calculation in [@SiRo04], where only a single VUV pulse was applied, but with an intensity one order of magnitude higher than in the present case (note that, with an average charge per atom of 2.5 for $\Delta t$=3500 a.u., the space charge of the Xe$_{40}$ cluster from the present work is approximately equal to the space charge of the Xe$_{80}$ cluster from [@SiRo04], where an average charge per atom of only 1.5 was achieved, so that the two cases can really be compared). This shows that by combining VUV and IR pulses, significantly higher charge states can be achieved than by applying an IR or a VUV pulse alone. Second, we see that the ionization is most efficient for a time delay of $\Delta t$=3500 a.u.\
\[t\]\[\]\[0.8\][delay \[a.u.\]]{} \[b\]\[\]\[0.8\][abs. energy \[a.u.\]]{} \[b\]\[\]\[0.8\][aver. charge/atom]{} \[\]\[\]\[1\][a)]{} \[\]\[\]\[1\][b)]{} ![Absorbed energy \[a)\] and average charge per atom \[b)\] as a function of time delay between the VUV and the IR pulse.[]{data-label="fig3"}](energy_charge.xfig.eps "fig:")
The absorbed energy and the average charge state per atom after the cluster has disintegrated are shown in [Fig. \[fig3\]]{} as a function of the time delay $\Delta t$. The curve for the energy absorption as well as the one for the charge shows a maximum $\Delta t \approx$ 3500 a.u. between the pump VUV and the probe IR pulse. If our physical picture is correct, this maximum should be due to the existence of a collective resonance with frequency $\Omega_{t}$ of the quasi-free electrons that were created by the VUV pulse. This resonance is most efficiently excited by the IR pulse when the laser frequency $\omega=\Omega_{t}$. We will now proceed to give evidence for this hypothesis.\
In principle there are two ways to check numerically whether the electron cloud is at resonance with the laser field: first, treating the electrons and ions of the cluster as a homogenic positively and negatively charged sphere, respectively, one ends up with the resonance condition $\Omega_{t}=\omega$. However, the definition [Eq. (\[plasmon\])]{} of $\Omega_{t}$ is by no means unique, since one has to define a cluster “volume” which itself is time dependent (through the increase of the cluster radius) and so are the charge of the ions and the number of electrons. The resonance condition can be determined more reliably by calculating the phase difference between the oscillation of the electronic center of mass (ECM) and the driving laser field $F(t) = F_{t}\cos\omega t$ [@SaRo03]: if one assumes a collective oscillation with a damping constant $\gamma$, the time-dependent dipole amplitude for the ECM reads $$\begin{aligned}
X(t)=A_t \cos(\omega-\phi_t)\end{aligned}$$ with $$\begin{aligned}
\label{amplitude}
A_t&=F_{t}/\sqrt{(\Omega_t^2-\omega^2)^2+(2\Gamma_t\omega)^2}\\
\phi_t&=\arctan(2\Gamma_t\omega/(\Omega_t^2-\omega^2)).\end{aligned}$$ For $\phi_t=\pi/2$ the system is at resonance and the laser cycle averaged energy absorption $$\begin{aligned}
\langle dE/dt\rangle = \frac 1T\int_{0}^{T}\frac{dX_{t}}{dt}F(t)\,dt\propto \sin\phi_{t}\end{aligned}$$ is at its maximum. Note, however, that the amplitude $X_{t}$ does not necessarily increase considerably at resonance due to strong damping (see [Eq. (\[amplitude\])]{}). Hence, we take the condition $\phi_t=\pi/2$ as the [*definition*]{} for the plasmon resonance and calculate $\phi_{t}$ from the phase lag between the driving field $F(t)$ and the dipole oscillation $dX_{t}/dt$ by extracting the maximum of the time correlation $c(\delta t)$ between the two signals, where $$\begin{aligned}
c(\delta t)=\int_{t_1}^{t_2} F(t) X(t+\delta t) \, dt
\label{corr_eq}\end{aligned}$$ (note that $\delta t \neq \Delta t$!). We chose the limits of integration in [Eq. (\[corr\_eq\])]{} to be 10 IR cycles before and after the maximum of the probe laser; the radial evolution $R_{t}$ of the cluster is sufficiently slow so that the phase lag only changes by a small amount during that time. The outcome of this calculation is shown in Fig. \[fig4\]. Indeed the phase lag is equal to $\pi/2$ for $\Delta t\approx
3800 a.u.$. This proves that the maxima in Fig. \[fig3\] are due to a resonance of the collective electron oscillation with the driving laser field.
Hence, the VUV pump combined with a IR probe pulse can map out the internal collective cluster dynamics very clearly and may be the only possibility to resolve this dynamics for small clusters. The reason is simply that the VUV pulse produces a large number of quasi-free electrons which can participate in collective electron motion. On the other hand the VUV pulse itself does not couple to this collective motion. Hence, only the probe pulse probes literally the collective electron dyamics. Using two IR pulses for pump and probe make this distinction difficult: The pump pulse must be very short in order not to “probe” itself the collective electron dynamics. On the other hand, if it is very short it will not produce a significant number of quasi-free electron which could move collectively since they are produced most efficiently through resonant coupling. In the next phase of the VUV-FEL at Hamburg, a pump-probe facility as “used” in this theoretical investigation will be available for experiments. Furthermore a similar experiment, but much simpler than the one at DESY as far as the experimental set is concerned, is planned in Saclay [@hubertus]; it is planned to use an 800 nm femtosecond laser, together with its 9th harmonic (delivering photons with $E\approx$ 14 eV) as the IR and VUV pulse, respectively; the intensities will be a bit lower than the ones considered here, but this should only result in a quantitative difference. The present paper shows that indeed, interesting and unique experiments can be done with such a time-delayed combination of pulses.\
This work is part of the research program of FOM (Fundamental Research on Matter), which is subsidized by the NWO (Netherlands Organization for the Advancement of Research).
\[t\]\[\]\[1\][delay \[a.u.\]]{} \[b\]\[\]\[1\][phase lag]{} \[\]\[\]\[1\][$\frac{\pi}{2}$]{} \[\]\[\]\[1\][$\pi$]{} ![Phase lag between the center-of-mass oscillation of the electron cloud and the driving laser field as a function of time delay between the VUV and the IR pulse.[]{data-label="fig4"}](phaselag "fig:")
[99]{} E. Suraud and P. G. Reinhard, Phys. Rev. Lett. [**85**]{}, 2296 (2000). T. Seideman, M.Y. Ivanov and P.B. Corkum, Phys. Rev. Lett. [**75**]{}, 2819 (1995). Ch. Siedschlag and J.M. Rost, Phys. Rev. Lett. [**89**]{}, 173401 (2002); Phys. Rev. A [**67**]{}, 013404 (2003). U. Saalmann and J.M. Rost, Phys. Rev. Lett. [**91**]{}, 223401 (2003). C. Jungreuthmayer, M. Geissler, J. Zanghellini and T. Brabec, Phys. Rev. Lett. [**92**]{}, 133401 (2004). T. Ditmire, T. Donnelly, A.M. Rubenchik, R.W. Falcone and M.D. Perry, Phys. Rev. A [**53**]{}, 3379 (1996). I. Last and J. Jortner, J. Phys. Chem. A [**106**]{}, 10872 (2002). H. Wabnitz [*et. al.*]{}, Nature [**420**]{}, 482 (2002). R. Santra and C. H. Greene, Phys. Rev. Lett. [**91**]{}, 233401 (2003). Ch. Siedschlag and J.M. Rost, Phys. Rev. Lett. [**93**]{}, 043402 (2004). L. K[ö]{}ller, M. Schumacher, J. K[ö]{}hn, S. Teuber, J. Tiggesb[ä]{}umker, and K. H. Meiwes-Broer, Phys. Rev. Lett. [**82**]{}, 3783 (1999). T. Döppner, Th. Diederich, J. Tiggesbäumker, and K.ÐH. MeiwesÐBroer, Eur. Phys. J. D [**16**]{}, 13 (2001). S. Zamith, T. Martchenko, Y. Ni, S. A. Aseyev, H. G. Muller, and M. J. J. Vrakking, Phys. Rev. [**A 70**]{}, 011201(R) (2004). F. Calvayrac, P.-G. Reinhard, E. Suraud and C.A. Ullrich, Phys. Rep. [**337**]{}, 493 (2000) H. Wabnitz, private communication.
[^1]: For rare gas clusters, the concepts of [*inner*]{} and [*outer*]{} ionization, the first being the ionization of an electron out of an atomic orbital into the cluster environment, the latter the process of an electron leaving the cluster as a whole, have proven to be very useful. If one is dealing with the valence electrons of a metal cluster, the inner ionization step is skipped, since these electrons can already move freely throughout the cluster. From the second shell onwards, metal clusters should not behave different from rare gas clusters
|
---
abstract: 'I present a detailed derivation of wideband optical pulses interacting with a Raman transition in the kind of scheme currently used to generate the ultra broadband light fields needed to create ultrashort pulses. In contrast to the usual approach using separate field envelopes for the pump, Stokes, and anti-Stokes spectral lines, I use a [*single*]{} field envelope. This requires the inclusion of few-cycle corrections to the pulse propagation. The single-field model makes fewer approximations and is mathematically (and hence computationally) simpler, although it does require greater computational resources to implement. The single-field theory reduces to the traditional multi-field one using appropriate approximations.'
author:
- Paul Kinsler
title: 'Wideband pulse propagation: a detailed calculation including Raman processes'
---
=10000
[*This report should be read along with the paper Phys. Rev. [**72**]{}, 033804 (2005) “Wideband pulse propagation: single-field and multi-field approaches to Raman interactions” by P. Kinsler, G.H.C. New [@Kinsler-N-2005pra] for proper context. This document is primarily intended as a complete (as possible) record of the calculational steps that were necessarily abbreviated (or omitted) from that published work. It is an edited version of a longer document from which on-going work has been excised; and, as a “work in progress”, despite my best efforts, may contain occasional mistakes. Please contact me if you have any comments, corrections or queries.* ]{}
\[\*\] I worked at this institution while doing the bulk of this calculation. My main project was with Jon Marangos & Prof. GHC New on ultrabroadband multifrequency Raman generation, and I was funded with money from the EPSRC.
WWW: QOLS Group <http://www.qols.ph.ic.ac.uk/>\
WWW: Physics Dept. <http://www.ph.ic.ac.uk/>\
WWW: Imperial College <http://www.ic.ac.uk/>\
Email: Paul Kinsler <Dr.Paul.Kinsler@physics.org>\
Email: G.H.C. New <g.new@ic.ac.uk>
Introduction {#s-intro}
============
An important aim of current wideband Raman experiments is to try to efficiently generate few-cycle pulses [@Harris-S-1998prl; @Sokolov-WYYH-2001prl; @Hakuta-SKL-2000prl; @Sali-MTHM-2004ol]. If driven strongly enough, the two-photon Raman transition modulates the incoming field by adding sidebands separated by the transition frequency. Wideband fields are generated as these sidebands generate sidebands of their own (and so on), thus generating a wide comb of frequency components separated by the transition frequency. If a scheme can be implemented that adjusts the phases of each component appropriately, then few- or single- cycle optical pulses can be obtained (see e.g. [@Sokolov-WYYH-2001prl]). Standard theoretical treatments of this process split the field into fields components centred on the teeth of this comb. The approach has the advantage that the components can be modeled reasonably well with slowly varying envelopes, but of course it has the disadvantage of needing to keep track of a large number of components.
In experiments like those of Sali et.al. [@Sali-MTHM-2004ol; @Sali-KNMHTM-2005pra], the Raman transition is driven near-resonantly by a pair of intense pump pulses about 100fs long; compared to the transition frequency of about 130THz, the spectra of each pump pulse (and hence the generated sidebands) are relatively narrow. This means that a multi-component model is still not unreasonable, even if numerical considerations might demand that the arrays used to store these spectra overlap in frequency space. However, if we were to move to shorter pump pulses, or to a single (much shorter) pump pulse with enough bandwidth to efficiently excite the transition, we would reach the regime where the “teeth” from the spectral comb significantly overlap. At this point, not only would we be forced to move from an SVEA (Slowly Varying Envelope Approximation) solution of the field propagation to a more accurate Generalized Few-cycle Envelope Approximation (GFEA) [@Kinsler-N-2003pra; @Kinsler-FCPP], but the utility of multiple field components becomes questionable. In this regime it can be advantageous to treat the field as a single unit, rather than splitting it into pieces. Note that this approach still differs from solutions of Maxwell’s equations such as FDTD (finite difference time domain)[@Joseph-T-1997itap] or PSSD (pseudospectral spatial domain)[@Tyyrell-KN-2005jmo], because our single-field is based on a second-order wave equation, and uses a convenient choice of carrier function to define a field envelope.
Following these considerations, we now derive a [*single-field*]{} model for Raman generation, and, apart from that notable detail, follow an analogous path to that of Hickman, Paisner, and Bischel (HPB) [@Hickman-PB-1986pra]. In the model, we find that the coupling constants retain an oscillatory behaviour at the transition frequency, and that it is this that impresses the sideband modulation on the propagating field. Since the field is not only wideband, but contains significant sideband components, we need to propagate this (no longer slowly varying) field envelope using the GFEA. The necessity of allowing for these can be demonstrated by converting the single-field model into a multi-field counterpart – without the envelope-gradient corrections, we will not get a correct multi-field model.
Summary of the theory and the numerical implementation {#ss-intro-theorysummary}
------------------------------------------------------
We model the wideband Raman generation process in the following way. We specify the field frequencies ($\omega_i$) of interest, which are usually at integer spacings of the transition frequency ($\omega_A$) from the main pump laser frequency ($\omega_0$). Each of these field components is described usin a standard envelope theory (i.e. as $A_i(t)$) with a time-history, allowing us to simulate pulses as they travel through the Raman medium. The Raman medium is modelled as a two-level atom using a (its) Bloch vector $(u,v,w)$, and this Bloch vector is driven by each combination of spectrally adjacent field components ($\sim \sum A_i A_{-1}$). Each of the field components $A_i$ is driven by the atomic polarization ($\sim v$) in combination with its pair of adjacent field components (i.e. $A_{i-1},
A_{i+1}$).
Each field component evolves as $$\begin{aligned}
\partial_z A_j
&=&
\frac{\sigma \omega_j \alpha_{12}}
{ \epsilon_0 c_0}
\left\{
-
\left[
v'
-
\imath u'
\right]
A_{j+1}
\exp\left( +\imath (k'_{j+1} -k'_j) z - \imath \Delta t \right)
+
\left[
v'
+
\imath u'
\right]
A_{j-1}
\exp\left( +\imath (k'_{j-1} -k'_j) z + \imath \Delta t \right)
\right\}
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~
-
\left( k_j - k_0 \right) A_j\end{aligned}$$
The transition evolves as ($\rho_{12}' \equiv (u',v') $) $$\begin{aligned}
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta + 2 \imath g' \sum_j A_j^* A_j
\right)
\rho_{12}'
+
8 \imath f' \sum_j A_{j} A_{j-1}^*
. w
. e^{+\imath \left( k_j-k_{j-1} \right) z - \imath \Delta t } \end{aligned}$$
$$\begin{aligned}
\partial_t w
&=&
- \gamma_1 \left( w - w_i \right)
+ 4 \imath f' \sum_j
\left[
A_j^*A_{j+1}
\rho_{12}'
. e^{+\imath \left( k_{j+1} -k_j \right) z - \imath \Delta t }
-
A_j^*A_{j-1}
\rho_{12}'^*
. e^{+\imath \left( k_{j-1} -k_j \right) z + \imath \Delta t }
\right]\end{aligned}$$
Here $\sigma$ is the number density of the atoms or molecules; $\gamma_1$, $\gamma_2$ are the population and polarization decay rates for the transition; $f'$ is the field-transition coupling constant; $g'$ is the stark shift coefficient; $\omega_j, k_j$ are the frequencies and wavevectors for the field components, $\Delta$ gives a rotating frame for $\rho_{12} \rightarrow \rho_{12}'$
Additionally, a Cauchy dispersion is applied to the propagation of the field components.
Because each of the field components has a time-history, this translates to a spectral width. In typical cases with roughly nanosecond pulse lengths, the bandwidth of each component will be tiny compared to the transition frequency, so there will be large uneventful gaps in the total spectrum. In contrast, if the pulse lengths drop to (say) roughly 100 femtoseconds, the bandwidths of the field components will form a noticeable fraction of the total spectrum
A comment on Cauchy dispersion {#ss-intro-comments}
------------------------------
As regards the mismatch terms in the Raman model, Geoff New has remarked (email, 20040121) that there’s a key point about the $\gamma_n$’s that is not made properly in most of the McDonald et al publications; since it is usually said rather enigmatically that the $\gamma$’s are “parameterized” by $\gamma_1$. The point is that if one assumes a Cauchy-type law for the refractive index variation, all the gamma’s are linked by a recurrence relation, and so you only need to specify one of them, from which all the others will follow. The point is made properly (to GN’s knowledge) only in ref [@McDonald-NCLL-1998jmo].
Single-field wideband Raman {#s-singlefieldraman}
===========================
[*Note: This single field derivation does mean some of the approximations as to the “slowness” of the field variation seem somewhat stringent. However, since a conversion to a multi-field model is possible, it would seem the field variation constraints are less stringent than they would first appear.* ]{}
Coupled wavefunction equations {#ss-single-coupled-wnf}
------------------------------
I start by considering the wave function $\psi$ of a single molecule (e.g. H$_2$) and the electric field $E$, and write the time-dependent wave function by expanding it in terms of the eigenfunctions in the field-free (i.e. $E=0$) case. This means I can get the expansion coefficients by solving for an effective Schröedinger equation that contains a two-photon Rabi frequency created by means of an interaction term based on a field-dependent dipole moment. I assume a dispersionless medium and write all equations in terms of position $z$ and retarded times $t=t_{lab}-z/c$. Here I follow the method of HPB [@Hickman-PB-1986pra], but use only a single $E$ field rather than multiple components. Note that HPB use [*Gaussian*]{} units, so there can appear to be inconsistencies when comparing my formulae (in S.I.) to theirs.
I denote the known molecular eigenfunctions of the unperturbed Hamiltonian $H_0$ as $\left| n \right>$, and their corresponding energies $\hbar W_n$. I want to obtain the solution to $$\begin{aligned}
\left( H_0 + V \right) \psi
&=&
\imath \hbar
\frac{\partial \psi}
{\partial t}
\label{eqn-hamltonian-def}
\\
\textrm{for} ~~~ ~~~
V &=&
-d E
\label{eqn-perturbation-def}
\\
\psi &=&
\sum_n
c_n e^{-\imath W_n t} \left| n \right>
,
\label{eqn-psi-def}\end{aligned}$$ where $d$ is the electronic dipole moment operator and the $c_n$ are a set of complex probability amplitudes.
A standard derivation for the equations of motion of the $c_i$ co-efficients proceeds as – $$\begin{aligned}
\imath \hbar
\frac{\partial}
{\partial t}
\sum_i c_i e^{-\imath W_i t}
\left| i \right>
&=&
\left( H_0 + V \right)
\sum_j c_j e^{-\imath W_j t}
\left| j \right>
\\
\imath \hbar
\sum_i
\left\{
-\imath W_i
c_i
e^{-\imath W_i t}
+
e^{-\imath W_i t}
\frac{\partial c_i}
{\partial t}
\right\}
\left| i \right>
&=&
\left( H_0 + V \right)
\sum_j c_j e^{-\imath W_j t}
\left| j \right>
\\
\imath \hbar
\left< n \right|
\sum_i
\left\{
-\imath W_i
c_i
e^{-\imath W_i t}
+
e^{-\imath W_i t}
\frac{\partial c_i}
{\partial t}
\right\}
\left| i \right>
&=&
\left< n \right|
\left( H_0 + V \right)
\sum_j c_j e^{-\imath W_j t}
\left| j \right>
\\
\imath \hbar
\left\{
-\imath W_n
c_n
e^{-\imath W_n t}
+
e^{-\imath W_n t}
\frac{\partial c_n}
{\partial t}
\right\}
&=&
\hbar
c_n
W_n
e^{-\imath W_n t}
+
\sum_j
c_j
e^{-\imath W_j t}
\left< n \right|
-d.E
\left| j \right>
\\
\imath \hbar
\left\{
-\imath W_n
c_n
+
\frac{\partial c_n}
{\partial t}
\right\}
&=&
\hbar
W_n
c_n
-
\sum_j
c_j
e^{-\imath W_j t + \imath W_n t}
\left< n \right|
d.E
\left| j \right>
\\
\imath \hbar
\frac{\partial c_n}
{\partial t}
&=&
-
\sum_j
c_j
e^{-\imath \left( W_j - W_n \right) t}
\left< n \right|
d.E
\left| j \right>
.\end{aligned}$$
We now use perturbation theory, & $d_{nm} = \left< n \right| \hat{d} \left| n \right>$, following two independent (but related) strands.
### CASE (i): Electric field {#sss-single-Efield}
This strand leaves the perturbing potential as a function of electric field $E$, and does not replace it with a carrier-envelope description. Although apparently the simplest strategy, it is generally impractical as it imposes constraints on the field and other model parameters that are too restrictive to be useful.
$$\begin{aligned}
\imath \hbar \frac{d c_n}{dt}
&=&
-
\sum_i
c_i
d_{ni} E
\exp \left[ -\imath \left( W_i - W_n \right) t \right]
\\
\textrm{... assume } c_n, d_{ni} &&
\textrm{vary only slowly, so I can integrate just the exponentials }
\\
c_i
&=&
-
\frac{1}{ \imath \hbar}
\sum_j
c_j
d_{ij} E
\frac{
\exp \left[ -\imath \left( W_j - W_i \right) t \right]
}
{-\imath \left( W_j-W_i \right) }
\\
&=&
\frac{1}{ \hbar}
\sum_j
c_j
d_{ij} E
\frac{
\exp \left[ -\imath \left( W_j - W_i \right) t \right]
}
{W_i-W_j}\end{aligned}$$
Now substitute the $c_i$ solution into the $d c_n / dt$ equations, and introduce the shorthand notation $W_{ij} = W_i - W_j$, $$\begin{aligned}
\imath \hbar \frac{d c_n}{dt}
&=&
-
\sum_i
\left[
\frac{1}{ \hbar}
\sum_j
c_j
d_{ij} E
\frac{
\exp \left[ -\imath W_{ji} t \right]
}
{W_i-W_j}
\right]
d_{ni} E
\exp \left[ -\imath W_{in} t \right]
\\
&=&
-\frac{1}{ \hbar}
\sum_i
\sum_j
c_j
d_{ij} E
\frac{
\exp \left[ -\imath W_{ji} t \right]
}
{W_{ij} }
d_{ni} E
\exp \left[ -\imath W_{in} t \right]
\\
&=&
-
E^2
\sum_j
\frac{1}{\hbar}
\sum_i
d_{ni} d_{ij} c_j
\frac{
\exp \left[ -\imath W_{jn} t \right]
}
{W_i-W_j}
\\
&=&
-
\sum_j
c_j
\alpha_{nj} E^2
,
\label{eqn-single-E-DcnDt}\end{aligned}$$ where $$\begin{aligned}
\alpha_{nj}
&=&
\frac{1}{\hbar}
\exp \left[ -\imath W_{jn} t \right]
\sum_i
\frac{ d_{ni} d_{ij} }
{W_i-W_j}
\label{eqn-single-E-alpha}
\\
\textrm{UNITS:} ~~~~
\left[ \alpha_{nj} \right]
&\equiv& J^{-1} s^{-1} . Cm . Cm . \left( s^{-1} \right)^{-1}
~~~~ = C^2 m^2 J^{-1}
~~~~ = \left( J V^{-1} \right)^2 m^2 J^{-1}
~~~~ = J m^2 V^{-2}
\\
\textrm{UNITS:} ~~~~
\left[ \frac{\alpha_{nj}}{\hbar} E^2\right]
&\equiv&
J m^2 V^{-2} . J^{-1} s^{-1} . \left( V m^{-1} \right)^2
~~~~ = J . J^{-1} \times V^{-2} . V^2 \times m^2 . m^{-2} \times s^{-1}
~~~~ = s^{-1}\end{aligned}$$
Since we are only interested in the end states $j=1, 2$, between which the Raman transition occurs, we only need calculate $c_1, c_2$; however we still retain the sum over all $i$ intermediates states, as they affect the coupling between $1$ and $2$. The diagonal couplings $\{ \alpha_{nj}, n = j
\}$ are real; but the off-diagonal couplings $\{ \alpha_{nj}, n \neq j \}$ undergo complex oscillations according to the difference in their energy levels.
Their [*frequency dependence*]{} is discussed after the following subsection; HPB’s corresponding parameters, which [*do not oscillate*]{}, were assumed to be frequency independent.
### CASE (ii): Electric Field Envelope {#sss-single-Eenvelope}
This strand replaces the electric field $E$ with a carrier-envelope description, but, unlike HPB, I use only a single carrier-envelope component rather than a set indexed by some integer $j$. This is necessary, because in the previous strand I ended up with coupling constants $\alpha_{nj}$ with strong frequency dependences.
I introduce the envelope and carrier [@Gabor-1946jiee] for the field: $$\begin{aligned}
E &=&
\left[
A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}
\right]
\label{eqn-single-EfromA}
\\
\textrm{so that } ~~~~ ~~~~
\imath \hbar \frac{d c_n}{dt}
&=&
-
\sum_l
c_l
d_{nl}
\exp \left[ -\imath \left( W_l - W_n \right) t \right]
\left[
A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}
\right]
\\
\textrm{now use~}l\rightarrow i; \textrm{~and assume ~} &c_n, d_{ni}&
\textrm{vary only slowly, so I can integrate just the exponentials }
\\
c_i
&=&
-\frac{1}{ \imath \hbar}
\sum_j
c_j
d_{ij}
\left[
A
\frac{
\exp \left[ -\imath \left( W_j - W_i + \omega_0\right) t \right]
}
{-\imath \left( W_j-W_i+\omega_0 \right) }
+
A^*
\frac{
\exp \left[ -\imath \left( W_j - W_i - \omega_0 \right) t \right]
}
{-\imath \left( W_j-W_i-\omega_0 \right) }
\right]
\\
&=&
\frac{1}{ \hbar}
\sum_j
c_j
d_{ij}
\left[
A
\frac{
\exp \left[ -\imath \left( W_j - W_i + \omega_0\right) t \right]
}
{W_i-W_j-\omega_0}
+
A^*
\frac{
\exp \left[ -\imath \left( W_j - W_i - \omega_0 \right) t \right]
}
{W_i-W_j+\omega_0}
\right]\end{aligned}$$
Note the swap of $\left(W_j-W_i\right)$ to $-\left( W_i-W_j \right)$ in the denominators. Now substitute the $c_i$ solution into the $d c_n / dt$ equations (using $l\rightarrow j$), and introduce the shorthand notation $W_{ij} = W_i - W_j$, $$\begin{aligned}
\imath \hbar \frac{d c_n}{dt}
&=&
-
\sum_i
\left\{
\frac{1}{ \hbar}
\sum_j
c_j
d_{ij}
\left[
A
\frac{
\exp \left[ -\imath \left( W_{ji} + \omega_0\right) t \right]
}
{W_{ij}-\omega_0}
+
A^*
\frac{
\exp \left[ -\imath \left( W_{ji} - \omega_0 \right) t \right]
}
{W_{ij}+\omega_0}
\right]
\right\}
d_{ni} E
\exp \left[ -\imath W_{in} t \right]
\\
&=&
-\frac{1}{ \hbar}
\sum_i
\sum_j
c_j
d_{ij}
\left[
A
\frac{
\exp \left[ -\imath \left( W_{ji} + \omega_0\right) t \right]
}
{W_{ij}-\omega_0}
+
A^*
\frac{
\exp \left[ -\imath \left( W_{ji} - \omega_0 \right) t \right]
}
{W_{ij}+\omega_0}
\right]
d_{ni}
\left[
A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}
\right]
\exp \left[ -\imath W_{in} t \right]
\\
&=&
-\frac{1}{ \hbar}
\sum_i
\sum_j
c_j
d_{ni}
d_{ij}
\left[
A
\frac{
\exp \left[ -\imath \left( W_{jn} + \omega_0\right) t \right]
}
{W_{ij}-\omega_0}
+
A^*
\frac{
\exp \left[ -\imath \left( W_{jn} - \omega_0 \right) t \right]
}
{W_{ij}+\omega_0}
\right]
\left[
A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}
\right]
\\
&=&
-\frac{1}{ \hbar}
\sum_i
\sum_j
c_j
d_{ni}
d_{ij}
\left\{
A^2
\frac{
\exp \left[ -\imath \left( W_{jn} + 2 \omega_0\right) t \right]
}
{W_{ij}-\omega_0}
+
A A^*
\frac{
\exp \left[ -\imath W_{jn} t \right]
}
{W_{ij}-\omega_0}
\right.
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~
\left.
+
A^* A
\frac{
\exp \left[ -\imath W_{jn} t \right]
}
{W_{ij}+\omega_0}
+
{A^*}^2
\frac{
\exp \left[ -\imath \left( W_{jn} - 2 \omega_0 \right) t \right]
}
{W_{ij}+\omega_0}
\right\}
\\
&\approx&
-\frac{1}{ \hbar}
\sum_i
\sum_j
c_j
d_{ni}
d_{ij}
A A^*
\exp \left[ -\imath W_{jn} t \right]
\left\{
\frac{1}
{W_{ij}-\omega_0}
+
\frac{1}
{W_{ij}+\omega_0}
\right\}
;
~~ \textrm{by discarding the $2 \omega_0$ terms;}
\\
&=&
-\frac{1}{ \hbar}
\sum_i
\sum_j
c_j
d_{ni}
d_{ij}
A A^*
\exp \left[ -\imath W_{jn} t \right]
\frac{W_{ij}+\omega_0 ~~ + ~~ {W_{ij}-\omega_0}}
{ \left( W_{ij}-\omega_0 \right) \left( W_{ij}+\omega_0 \right) }
\\
&=&
-\frac{1}{ \hbar}
\sum_i
\sum_j
c_j
d_{ni}
d_{ij}
A A^*
\exp \left[ -\imath W_{jn} t \right]
\frac{2 W_{ij} }
{W_{ij}^2-\omega_0^2}
\\
&=&
-
A A^*
\sum_j
c_j
\sum_i
\frac{1}{ \hbar}
d_{ni}
d_{ij}
\exp \left[ -\imath W_{jn} t \right]
\frac{2 W_{ij} }
{W_{ij}^2-\omega_0^2}
\\
&=&
-
\sum_j
c_j
\alpha_{nj}'
~~ . 2 \left| A \right|^2
\label{eqn-single-A-DcnDt}\end{aligned}$$ where $$\begin{aligned}
\alpha_{nj}'
&=&
\frac{1}{ \hbar}
\exp \left[ -\imath W_{jn} t \right]
\sum_i
d_{ni}
d_{ij}
\frac{W_{ij} }
{W_{ij}^2-\omega_0^2}
\label{eqn-single-A-alpha}
~~~~ ~~~~ = ~~
\frac{1}{ \hbar}
\exp \left[ +\imath W_{nj} t \right]
\sum_i
d_{ni}
d_{ij}
\frac{W_{ij} }
{W_{ij}^2-\omega_0^2}\end{aligned}$$
These redefined $\alpha_{nj}'$ parameters still oscillate, as they must because unlike in the HPB derivation, there is no frequency difference between field components to cancel with the Raman transition frequency. The coupling also varies with frequency, which is discussed next.
Raman coupling parameters: approximations
-----------------------------------------
I now discuss two particular (and vital) approximations applied to the Raman coupling parameters.
First, note that (as in HPB), I will take the indices $1$ and $2$ to correspond to the two states involved in the (Raman) transition of interest; these will be the $0$ and $1$ vibrational (or perhaps rotational) levels of the electronic ground state. Indices $3$ and above will correspond to (quoting HPB) “translational motion on higher electronic states”.
[*Note: I can see that assigning these indices to higher electronic states will conveniently keep the energy separations for transitions to greater than that of the $1 \leftrightarrow 2$ transitions, but it’s not so clear to me why I can ignore all the higher vibrational (or rotational) states.*]{}
Since I am interested only in the Raman transition, I specialise the above equations for the coefficients $c_n$, calculating $c_1$ and $c_2$ only, and assuming that the $d_{12} = \left< 1 \right| d \left| \right>$ dipole moment is zero. This means we will only be including transitions between indices $1$ and $2$ that [*go via one of the higher states*]{} $j \ge 3$, since we still allow $d_{1j}, ~d_{2j} \neq 0 ~~ \forall j \ge 3$. Further, I solve for the coefficients for the higher states in terms of $c_1$ and $c_2$, in an adiabatic approximation justified when $c_1$ and $c_2$ vary only slowly compared to the exponential terms.
[*Note: Separate from the oscillations that occur in my coupling parameters (but not in HPB), there is the issue of [*frequency dependence*]{} which applies to both HPB and my parameters.*]{}
For both HPB (their eqn.(15)), and my field-carrier based derivation (eqn.(\[eqn-single-A-alpha\])), the presence of the field carrier in the denominator is helpful. Since it can reasonably be assumed to be much greater than the inter(Raman)-level energy differences, the fractional terms will be correspondingly small, and so the $\alpha_{nj}$ ($\alpha_{nj}'$) parameters can be assumed to be independent of frequency (or at least nearly so).
This is not in any way obviously (or even partially) true for the no-carrier $E$ based (eqn.(\[eqn-single-E-alpha\]), $\alpha_{nj}$) parameters, which depend only on the difference in Raman levels – at best you might perhaps hope that the denominators were approximate multiples of each other.
### Near-zero difference
Here I assume that the forward and backward transitions (between levels 2 and 1) have nearly the same amplitude. It is not self-evidently true, but HPB must have made an equivalent assumption. In any case, since $\alpha_{12}'^* - \alpha_{21}'$ is the difference of similar terms, it will be smaller by at worst a factor of two ($(1+\delta-1) / (1+\delta+1) \sim \delta/2$) $$\begin{aligned}
\alpha_{12}'^* - \alpha_{21}'
&=&
\frac{1}{ \hbar}
\exp \left[ +\imath W_{21} t \right]
\sum_i
d_{1i}^*
d_{i2}^*
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
~~~~
-
\frac{1}{ \hbar}
\exp \left[ -\imath W_{12} t \right]
\sum_i
d_{2i}
d_{i1}
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
\\
&=&
\frac{1}{ \hbar}
\sum_i
d_{1i}^*
d_{i2}^*
\left[
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
-
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
\right]
\exp \left[ +\imath W_{21} t \right]
\\
&\approx&
0\end{aligned}$$
This approximation allows me to equate $\alpha_{21}'$ to $\alpha_{12}'^*$, and hence change $- \alpha_{21}' c_1^* c_1
+ \alpha_{12}'^* c_2 c_2^* \longrightarrow
\alpha_{21}' \left( c_2^* c_2
- c_1^* c_1 \right)$ in the $\rho_{12}$ equation below. This is a vital step in getting to a simple form equivalent to the Bloch equations.
### Sum is double
This follows from the above assumption (as per HPB) that the forward and backward transitions (between levels 2 and 1) have nearly the same amplitude. This approximation allows me to replace $\alpha_{21}'$ with $\alpha_{12}'$ in the $w$ equation below, which simplifies my notation and makes the coupling term match that in the $\rho_{12}$ equation, important in getting to true Bloch equations.
$$\begin{aligned}
\alpha_{12}' + \alpha_{21}'^*
&=&
\frac{1}{ \hbar}
\exp \left[ -\imath W_{21} t \right]
\sum_i
d_{1i}
d_{i2}
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
~~~~
+
\frac{1}{ \hbar}
\exp \left[ +\imath W_{12} t \right]
\sum_i
d_{2i}^*
d_{i1}^*
\frac{ W_{i1} }
{W_{i2}^2-\omega_0^2}
\\
&=&
\frac{1}{ \hbar}
\sum_i
d_{1i}
d_{i2}
\left[
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
+
\frac{ W_{i1} }
{W_{i2}^2-\omega_0^2}
\right]
\exp \left[ -\imath W_{21} t \right]
\\
&\approx&
\bar{\alpha}_{12}
e^{-\imath W_{21} t}
\label{eqn-alpha-bar}\end{aligned}$$
### Sum is cosine
I do not reduce the exponential sum to a (trig) cosine function, for potential use later when matching and/or differencing exponentials. However, this sum is does not occur, so the $\bar{a}_{12}$ is never used. $$\begin{aligned}
\alpha_{12}' + \alpha_{21}'
&=&
\frac{1}{ \hbar}
\exp \left[ -\imath W_{21} t \right]
\sum_i
d_{1i}
d_{i2}
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
~~~~
+
\frac{1}{ \hbar}
\exp \left[ -\imath W_{12} t \right]
\sum_i
d_{2i}
d_{i1}
\frac{ W_{i1} }
{W_{i2}^2-\omega_0^2}
\\
&=&
\frac{1}{ \hbar}
\sum_i
d_{1i}
d_{i2}
\left[
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
\exp \left[ -\imath W_{21} t \right]
+
\frac{ W_{i1} }
{W_{i2}^2-\omega_0^2}
\exp \left[ +\imath W_{21} t \right]
\right]
\\
&\approx&
\bar{a}_{12} \left[ e^{-\imath W_{21} t} + e^{+\imath W_{21} t} \right]
\label{eqn-alpha-bar-B}\end{aligned}$$
### Stark Shifts
Here I calculate the Stark shift term $$\begin{aligned}
\alpha_{11}' - \alpha_{22}'^*
&=&
\frac{1}{ \hbar}
\exp \left[ -\imath W_{11} t \right]
\sum_i
d_{1i}
d_{i1}
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
~~~~
-
\frac{1}{ \hbar}
\exp \left[ +\imath W_{22} t \right]
\sum_i
d_{2i}^*
d_{i2}^*
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
\\
&=&
\frac{1}{ \hbar}
\sum_i
\left[
d_{i1}^*d_{i1}
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
-
d_{2i}^* d_{2i}
\frac{ W_{i2} }
{W_{i2}^2-\omega_0^2}
\right]
\label{eqn-alpha-stark}\end{aligned}$$
Simplify assuming $W_{21} \ll W_{i1}, W_{i2}$, while assuming $W_{i1}/\left(W_{i1}^2-\omega_0^2 \right) \sim 1$; or, more accurately, that $$\begin{aligned}
1
&\gg&
\frac{W_{21}^n}{ W_{i1}^n}
\frac{ W_{i1}^2 }
{W_{i1}^2-\omega_0^2}
~~~~ ~~~~ \textrm{for } ~~~~ n ~ = ~~1, ~2
,\end{aligned}$$
and keeping only the terms first order in $W_{21}/W_{i1}, W_{21}/W_{i2}$, etc $$\begin{aligned}
\alpha_{11}' - \alpha_{22}'^*
&=&
\frac{1}{ \hbar}
\sum_i
\left\{
\left| d_{1i} \right|^2
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
-
\left| d_{2i} \right|^2
\frac{ W_{i1} + W_{21} }
{W_{i1}^2+ 2W_{i1} W_{21} + W_{21}^2 -\omega_0^2}
\right\}
\\
&=&
\frac{1}{ \hbar}
\sum_i
\left\{
\left| d_{1i} \right|^2
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
-
\left| d_{2i} \right|^2
\frac{ W_{i1} + W_{21} }
{
\left( W_{i1}^2 -\omega_0^2 \right)
}
\frac{ 1}
{
\left[
1
+ \left( 2W_{21} / W_{i1} + W_{21}^2/ W_{i1}^2 \right)
W_{i1}^2 / \left( W_{i1}^2 -\omega_0^2 \right)
\right]
}
\right\}
\\
&=&
\frac{1}{ \hbar}
\sum_i
\left\{
\left| d_{1i} \right|^2
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
-
\left| d_{2i} \right|^2
\frac{ W_{i1} + W_{21} }
{ W_{i1}^2 -\omega_0^2 }
\left[
1
- \left( 2 \frac{W_{21}}{ W_{i1} }
+ \frac{W_{21}^2}{ W_{i1}^2}
\right)
\frac{W_{i1}^2}{ \left( W_{i1}^2 -\omega_0^2 \right)}
+ ...
\right]
\right\}
\\
&\approx&
\frac{1}{ \hbar}
\sum_i
\left\{
\left| d_{1i} \right|^2
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
-
\left| d_{2i} \right|^2
\frac{ W_{i1} + W_{21} }
{W_{i1}^2 -\omega_0^2 }
\left[
1
- 2 \frac{W_{21}}{ W_{i1} }
\frac{W_{i1}^2 }{ \left( W_{i1}^2 -\omega_0^2 \right)}
\right]
\right\}
\\
&\approx&
\frac{1}{ \hbar}
\sum_i
\left\{
\left[
\left| d_{1i} \right|^2
-
\left| d_{2i} \right|^2
-
\left| d_{2i} \right|^2
\frac{W_{21}}{ W_{i1}}
\right]
\frac{ W_{i1} }
{W_{i1}^2-\omega_0^2}
-
\left[
2 \left| d_{2i} \right|^2
\frac{W_{21}}{ W_{i1}}
\right]
\frac{ W_{i1}^3 }
{ \left( W_{i1}^2 -\omega_0^2 \right)^2 }
\right\}
\label{eqn-alpha-stark-approx}\end{aligned}$$
Ignoring the $W_{21}$ transition frequency terms, this is just the difference in the energy shifts of the two levels with field intensity. This is a purely [*real*]{} quantity, with no imaginary part.
Two photon Bloch equations
--------------------------
From now on, I use $\mathscr{I}$ (a script “I”) to represent $E^2$ if following on from \[sss-single-Efield\] and eqns.(\[eqn-single-E-DcnDt\], \[eqn-single-E-alpha\]); or to represent $2\left| A \right|^2$ if following on from \[sss-single-Eenvelope\] and eqns.(\[eqn-single-A-DcnDt\], \[eqn-single-A-alpha\]). Since it is the envelope-carrier description of the field $E$ which is most useful (in \[sss-single-Eenvelope\]), for most purposes $\mathscr{I} = 2\left| A \right|^2$ holds; remember that it is difficult to maintain the accuracy of the approximations relied on above for the $E^2$ picture (in \[sss-single-Efield\]). I also drop the prime on the $\alpha_{nj}'$ parameters used in the (Case Iii)) “$2\left| A \right|^2$” electric field envelope model.
Since we are only interested in $c_1, c_2$, and because we keep only stationary or slowly varying terms, we can write equations for $c_1, c_2$ as – $$\begin{aligned}
\imath \hbar
\frac{d c_1}{dt}
&=&
- \alpha_{11} \mathscr{I} c_1
- \alpha_{12} \mathscr{I} c_2
\\
\Longrightarrow
~~~~ ~~~~
\imath \hbar
\frac{d c_1^*}{dt}
&=&
+ \alpha_{11}^* \mathscr{I} c_1^*
+ \alpha_{12}^* \mathscr{I} c_2^*
\\
\imath \hbar
\frac{d c_2}{dt}
&=&
- \alpha_{21} \mathscr{I} c_1
- \alpha_{22} \mathscr{I} c_2
\\
\Longrightarrow
~~~~ ~~~~
\imath \hbar
\frac{d c_2^*}{dt}
&=&
+ \alpha_{21}^* \mathscr{I} c_1^*
+ \alpha_{22}^* \mathscr{I} c_2^*
\\
\textrm{In matrix form: }
~~~~ ~~~~
\frac{d}{dt}
\left[
\begin {array}{c}
c_1
\\
c_2
\end {array}
\right]
&=&
-\frac{1}{\imath \hbar}
\left[
\begin {array}{cc}
\alpha_{11} \mathscr{I} & \alpha_{12} \mathscr{I} \\
\noalign{\medskip}
\alpha_{21} \mathscr{I} & \alpha_{22} \mathscr{I} \\
\end {array}
\right]
~
\left[
\begin {array}{c}
c_1
\\
c_2
\end {array}
\right]
\label{eqn-single-dcdt}
,\end{aligned}$$
Alternative formulations: Kien et.al. (KLKOHS), Hickman et.al. (HPB)
--------------------------------------------------------------------
This interlude compares my matrix equation for $c_1, c_2$ to those from Kien et.al.[@Kien-LKOHS-1999pra] (KLKOHS) and Hickman et.al. [@Hickman-PB-1986pra] (HPB). This is useful as a fixed point of refernce beween the approaches, enabling quick conversions between the parameter variables.
$$\begin{aligned}
\textrm{cf. (KLKOHS) eqn.(13): } ~~~~ ~~~~
\frac{d}{dt}
\left[
\begin {array}{c}
c_a
\\
c_b
\end {array}
\right]
&=&
-\frac{1}
{\imath \hbar}
\left[
\begin {array}{cc}
\hbar \Omega_{aa} & \hbar \Omega_{ab} \\
\noalign{\medskip}
\hbar \Omega_{ba} & \hbar \Omega_{bb} - \delta \\
\end {array}
\right]
~
\left[
\begin {array}{c}
c_a
\\
c_b
\end {array}
\right]
,
\label{eqn-KLKOHS-dcdt}
\\
\textrm{where $B_q$ are KLKOHS's $E_q$ envelopes:}
~~~~ ~~~~
E
&=&
\frac{1}{2}
\left[
B_q e^{\imath \Xi} + B_q^* e^{-\imath \Xi}
\right]
\\
\hbar \Omega_{aa}
&=&
\frac{\hbar }{2}
\sum_q
a_q B_q B_q^*
\\
\hbar \Omega_{ab}
&=&
\frac{\hbar }{2}
\sum_q
d_q B_q B_q^*
~~~~ ~~~~
=
\hbar \Omega_{ba}^*
\\
\hbar \Omega_{bb}
&=&
\frac{\hbar }{2}
\sum_q
b_q B_q B_{q+1}^* \end{aligned}$$
$$\begin{aligned}
\textrm{cf. (HPB) eqn.(13): } ~~~~ ~~~~
\frac{d}{dt}
\left[
\begin {array}{c}
c_1
\\
c_2
\end {array}
\right]
&=&
-\frac{1}{\imath \hbar}
\left[
\begin {array}{cc}
-H_{11} & -H_{12} \\
\noalign{\medskip}
-H_{21} & -H_{22} \\
\end {array}
\right]
~
\left[
\begin {array}{c}
c_1
\\
c_2
\end {array}
\right]
\label{eqn-HPB-dcdt}
,
\\
\textrm{where $\alpha_{Hij}$ are HPB's $\alpha_{ij}$ parameters:}
~~~~ ~~~~
H_{11}
&=&
-\frac{1}{4}
\sum_j \alpha_{H11} (\omega_j) V_j V_j^*
\\
H_{12}
&=&
-\frac{1}{4}
\sum_j \alpha_{H12} (\omega_j) V_j V_{j-1}^*
~~~~ ~~~~ =
H_{21}^*
\\
H_{22}
&=&
-\frac{1}{4}
\sum_j \alpha_{H22} (\omega_j) V_j V_j^*\end{aligned}$$
Thus comparing the Rabi-like parts of my eqn (\[eqn-single-dcdt\]) with that of KLKOHS (my eqn (\[eqn-KLKOHS-dcdt\])) and that of HPB (my eqn (\[eqn-HPB-dcdt\])), gives $$\begin{aligned}
\alpha_{12} \mathscr{I} = \alpha_{12} . 2 A^*A = \hbar f' . 2 A^*A
~~~~ ~~~~
&=&
\frac{\hbar}{2}
\sum_q
d_q B_q B_{q+1}^*
=
\hbar
\sum_q
d_q
.
\frac{1}{2}
B_q B_{q+1}^*
~~~~ ~~~~
=
-\frac{1}{4}
\sum_j
\alpha_{H12}(\omega_j) V_j V_{j-1}^*\end{aligned}$$
This uses KLKOHS: $ E = \frac{1}{2} \left( B_1 + B_1^* + B_2 + B_2^* \right)$\
$\rightarrow E^2 = \frac{1}{4} \left( B_1^2 + B_1^{*2} + B_2^2 + B_2^{2*}
+ 2 B_1 B_1^* + 2 B_2 B_2^* + 2 B_1 B_2^* + 2 B_1^* B_2
\right)$\
$\rightarrow E_{1-2}^2 \sim \frac{1}{4} 2 B_1 B_2^* = \frac{1}{2} B_1 B_2^*$\
which tells us the size of the field contribution to the KLKOHS Rabi-like term.
This uses HPB: which tells us the size of the field contribution to the HPB Rabi-like term. (cf KLKOHS). So $\Omega_{aa}=\alpha_{11}
\mathscr{I} / \hbar$, $\Omega_{ab}=\alpha_{12} \mathscr{I} / \hbar$, $\Omega_{bb}= \alpha_{22} \mathscr{I} / \hbar$ (temporarily ignoring their detuning $\delta$); and looking ahead to the definition of $f=
\bar{\alpha}_{12}/2\hbar = \left( \alpha_{12}+\alpha_{21} \right) /2\hbar$ gives us $\Omega_{ab}=f \mathscr{I}$, noting $\mathscr{I}=2 A_q A_{q+1}^*$, compared to $\mathscr{I}=\frac{1}{2}E_q E_{q+1}^*$. Note that KLKOHS have defined $\Omega_{ab}=\frac{1}{2} \sum_q d_q E_q E_{q+1}^*$, so that $f=d_q$.
Two photon Bloch equations (cont)
---------------------------------
An important difference between KLKOHS & HPB, and my equations is that my $\alpha_{ij}$ coupling parameters retain a “slow” time dependence at the Raman frequency $W_{12}$. I now turn my equations (\[eqn-single-dcdt\]) into Bloch equations by first working out $d/dt$ of $c_1^* c_1$, $c_2^* c_2$, and $c_1^* c_2$, $$\begin{aligned}
\imath \hbar
\frac{d c_1^* c_1}{dt}
=
\imath \hbar
c_1^* \frac{d c_1}{dt}
+
\imath \hbar
c_1 \frac{d c_1^*}{dt}
&=&
- \alpha_{11} \mathscr{I} c_1^* c_1
- \alpha_{12} \mathscr{I} c_1^* c_2
+ \alpha_{11}^* \mathscr{I} c_1 c_1^*
+ \alpha_{12}^* \mathscr{I} c_1 c_2^*
\\
&=&
+ \alpha_{12}^* \mathscr{I} c_1 c_2^*
- \alpha_{12} \mathscr{I} c_1^* c_2
- \left[ \alpha_{11} - \alpha_{11}^* \right] \mathscr{I} c_1^* c_1
\\
&=&
+ \mathscr{I} \left( \alpha_{12}^* c_1 c_2^* - \alpha_{12} c_1^* c_2 \right)
; ~~~~ ~~~~ ~~~~ ~~~~ \textrm{since} ~~ \alpha_{11}=\alpha_{11}^*
\\
\imath \hbar
\frac{d c_2^* c_2}{dt}
=
\imath \hbar
c_2^* \frac{d c_2}{dt}
+
\imath \hbar
c_2 \frac{d c_2^*}{dt}
&=&
- \alpha_{21} \mathscr{I} c_2^* c_1
- \alpha_{22} \mathscr{I} c_2^* c_2
+ \alpha_{21}^* \mathscr{I} c_2 c_1^*
+ \alpha_{22}^* \mathscr{I} c_2 c_2^*
\\
&=&
- \alpha_{21} \mathscr{I} c_2^* c_1
+ \alpha_{21}^* \mathscr{I} c_2 c_1^*
- \left[ \alpha_{22} - \alpha_{22}^* \right] \mathscr{I} c_2^* c_2
\\
&=&
- \mathscr{I} \left( \alpha_{21} c_2^* c_1 - \alpha_{21}^* c_2 c_1^* \right)
; ~~~~ ~~~~ ~~~~ ~~~~ \textrm{since} ~~ \alpha_{22}=\alpha_{22}^*
\\
\imath \hbar
\frac{d c_1 c_2^*}{dt}
=
\imath \hbar
c_1 \frac{d c_2^*}{dt}
+
\imath \hbar
c_2^* \frac{d c_1}{dt}
&=&
+ \alpha_{21}^* \mathscr{I} c_1 c_1^*
+ \alpha_{22}^* \mathscr{I} c_1 c_2^*
- \alpha_{11} \mathscr{I} c_2^* c_1
- \alpha_{12} \mathscr{I} c_2^* c_2
\\
&=&
-\left( \alpha_{11} - \alpha_{22}^* \right) \mathscr{I} c_2^* c_1
+ \alpha_{21}^* \mathscr{I} c_1 c_1^*
- \alpha_{12} \mathscr{I} c_2^* c_2
\\
&=&
- \left( \alpha_{11} - \alpha_{22}^* \right) \mathscr{I} c_1 c_2^*
- \alpha_{12} \mathscr{I} \left[ c_2 c_2^* - c_1^* c_1 \right]
; ~~~~ ~~~~ \textrm{since} ~~ \alpha_{21}^* \approx \alpha_{12}\end{aligned}$$
Use $\rho_{12} = c_1 c_2^*$ and $w=c_2 c_2^*-c_1^* c_1$, $$\begin{aligned}
\imath \hbar
\frac{d w}{dt}
&=&
- \mathscr{I} \left( \alpha_{21} c_2^* c_1 - \alpha_{21}^* c_2 c_1^* \right)
- \mathscr{I} \left( \alpha_{12}^* c_1 c_2^* - \alpha_{12} c_1^* c_2 \right)
\\
&=&
\mathscr{I}
\left(
- \alpha_{12}^* c_1 c_2^*
+ \alpha_{21}^* c_1^* c_2
- \alpha_{21} c_1 c_2^*
+ \alpha_{12} c_1^* c_2
\right)
\\
&=&
-
\mathscr{I} \left( \alpha_{12}^* + \alpha_{21} \right) c_1 c_2^*
+
\mathscr{I} \left( \alpha_{21}^* + \alpha_{12} \right) c_1^* c_2
\\
&=&
-
\mathscr{I} \left( \alpha_{12}^* + \alpha_{21} \right) \rho_{12}
+
\mathscr{I} \left( \alpha_{21}^* + \alpha_{12} \right) \rho_{12}^*
\\
\imath \hbar
\frac{d \rho_{12}}{dt}
&=&
-
\left( \alpha_{11} - \alpha_{22}^* \right) \mathscr{I}
c_1 c_2^*
-
\alpha_{12} \mathscr{I}
\left[ c_2 c_2^* - c_1^* c_1 \right]
\\
&=&
-
\left( \alpha_{11} - \alpha_{22}^* \right) \mathscr{I} \rho_{12}
-
\alpha_{12} \mathscr{I} w\end{aligned}$$
Hence $$\begin{aligned}
\frac{d \rho_{12}}{dt}
&=&
\imath
\frac{\left( \alpha_{11} - \alpha_{22}^* \right) }
{\hbar}
\mathscr{I} \rho_{12}
+ \imath
\frac{\alpha_{12} }{ \hbar}
\mathscr{I} w
\label{eqn-basic2pbloch-rho}
\\
\frac{dw}{dt}
&=&
+ \imath
\frac{\left( \alpha_{12}^* + \alpha_{21} \right) }
{ \hbar}
\mathscr{I}
\rho_{12}
-
\imath
\frac{\left( \alpha_{21}^* + \alpha_{12} \right) }
{ \hbar}
\mathscr{I}
\rho_{12}^*
,
\label{eqn-basic2pbloch-w}
\\
&=&
+ \imath
\frac{2 \alpha_{12}^* }
{ \hbar}
\mathscr{I}
\rho_{12}
-
\imath
\frac{ 2 \alpha_{12} }
{ \hbar}
\mathscr{I}
\rho_{12}^*
.\end{aligned}$$
Transformations of the Bloch Equations {#aa-single-transforms}
--------------------------------------
I define a new coupling parameter $f'$ (c.f. $\omega_B$), following the definition of $\bar{\alpha}_{12}$ in eqn.(\[eqn-alpha-bar\]), and $\omega_b = W_2 - W_1 = W_{21}$. I also define a $\omega_A$, which corresponds to the intensity dependent shift detuning shift proportional to $(\alpha_{22} - \alpha_{11})$. Thus, factoring any complex phase of $\bar{\alpha}_{12}$ into the angle $\delta'$, I get – $$\begin{aligned}
f' e^{-\imath \omega_b t}
&=&
\frac{\alpha_{12}}
{ \hbar}
~~~~ ~~~~
\simeq
\frac{\bar{\alpha}_{12}}
{ 2 \hbar}
e^{-\imath \omega_b t - \imath \delta'}
; ~~~~ ~~~~
\alpha_{12}
=
\frac{1}{\hbar}
e^{-\imath \omega_b t}
\sum_j d_{ij}d_{j2} \frac{-W_{j2}}{W_{j2}^2-\omega_0^2}
\label{eqn-blochcoupling}
\\
g'
&=&
\frac{\alpha_{11} - \alpha_{22}^*}
{ \hbar}
~~~~ ~~~~
\textrm{NB: this is a real quantity, see eqns.
(\ref{eqn-alpha-stark}, \ref{eqn-alpha-stark-approx})}
\label{eqn-blochstark}
\\
\omega_B
&=& \frac{\left( \alpha_{12} + \alpha_{21}^* \right) }
{ 2 \hbar}
\mathscr{I}
~~~~
= \frac{\bar{\alpha}_{12}}
{ 2 \hbar}
e^{-\imath \omega_b t - \imath \delta'}
\mathscr{I}
~~~~
=
f'
e^{-\imath \omega_b t - \imath \delta'}
\mathscr{I}
\label{eqn-blochcoupling-wB}
\\
\omega_A
&=&
\frac{\left( \alpha_{11} - \alpha_{22}^* \right) }
{\hbar}
\mathscr{I}
\label{eqn-blochcoupling-wA}
.\end{aligned}$$
The Bloch equations (\[eqn-basic2pbloch-rho\], \[eqn-basic2pbloch-w\]) can now be rewritten – $$\begin{aligned}
\frac{d \rho_{12}}{dt}
&=&
\imath g' \mathscr{I}
\rho_{12}
+
\imath f' \mathscr{I} w
e^{-\imath \omega_b t - \imath \delta'}
\\
\frac{dw}{dt}
&=&
+
2 \imath f'
\mathscr{I}
\left[
\rho_{12}
e^{+\imath \omega_b t + \imath \delta'}
-
\rho_{12}^*
e^{-\imath \omega_b t - \imath \delta'}
\right]
.\end{aligned}$$
In analogy to both the standard two level atom Bloch equations, and those in HPB (HPB 18), these equations can have losses introduced, in – $$\begin{aligned}
\frac{d \rho_{12}}{dt}
&=&
-
\gamma_2 \rho_{12}
+
\imath g' \mathscr{I}
\rho_{12}
+
\imath f' \mathscr{I} w
e^{-\imath \omega_b t - \imath \delta'}
\\
\frac{dw}{dt}
&=&
-
\gamma_1
\left(
w - w_i
\right)
+
2 \imath f'
\mathscr{I}
\left[
\rho_{12}
e^{+\imath \omega_b t + \imath \delta'}
-
\rho_{12}^*
e^{-\imath \omega_b t - \imath \delta'}
\right]
.\end{aligned}$$
In the standard atom-field case, the atom and field frequencies are similar, so an atom carrier could be chosen to match its evolution the field carrier frequency, leading naturally to a detuning term. However, in the Raman situation, the atomic frequency is far removed from the field frequency, so the frequency evolution that appears in the definition of $\psi$ (eqn.(\[eqn-psi-def\])) is sufficient.
Here I define two things (a) new $u, v$ variables that represent the density matrix element $\rho_{12}$, and (b) allow for a “detuning” rotation in $\rho_{12}$. This detuning rotation looks rather like a carrier+envelope representation for $\rho_{12}$, but without a “+c.c.” since $\rho_{12}$ is a complex quantity. There is no need for any kind of carrier+envelope transformation for the inversion $w$.
So, $$\begin{aligned}
\rho_{12}
= \frac{u}{2} + \imath \frac{v}{2}
&=&
\rho_{12}' \exp \left( -\imath \Delta t -\imath \delta' \right)
= \frac{u'}{2} + \imath \frac{v'}{2}
\label{eqn-blochcoupling-rho12-uv}\end{aligned}$$
I now adapt the Bloch equations to allow for the rotation in $\rho_{12}$, and introduce a detuned transition frequency $\omega_b'=\omega_b-\Delta$, $$\begin{aligned}
\partial_t
\left(
\rho_{12}' e^{-\imath \Delta t -\imath \delta'}
\right)
&=&
-
\imath \Delta \rho_{12}
e^{-\imath \Delta t-\imath \delta'}
+
e^{-\imath \Delta t-\imath \delta'}
\partial_t
\rho_{12}'
\\
&=&
-
\gamma_2
\rho_{12}' e^{-\imath \Delta t-\imath \delta'}
+
\imath g' \mathscr{I}
\rho_{12}' e^{-\imath \Delta t-\imath \delta'}
+
\imath f' \mathscr{I} w e^{-\imath \omega_b t-\imath \delta'}
\\
\partial_t w
&=&
-
\gamma_1 \left( w - w_i \right)
+
2 \imath f' \mathscr{I}
\left(
\rho_{12}'
e^{-\imath \Delta t-\imath \delta'}
e^{+\imath \omega_b t+\imath \delta'}
-
\rho_{12}'^*
e^{+\imath \Delta t+\imath \delta'}
e^{-\imath \omega_b t-\imath \delta'}
\right)
\\
\Longrightarrow ~~~~ ~~~~
e^{-\imath \Delta t-\imath \delta'}
\partial_t
\rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta
\right)
\rho_{12}' e^{-\imath \Delta t-\imath \delta'}
+
\imath g' \mathscr{I}
\rho_{12}' e^{-\imath \Delta t-\imath \delta'}
+
\imath f' \mathscr{I} w e^{-\imath \omega_b t-\imath \delta'}
\\
\partial_t w
&=&
- \gamma_1 \left( w - w_i \right)
+
2 \imath f' \mathscr{I}
\left(
\rho_{12}'
e^{+\imath \omega_b' t}
-
\rho_{12}'^*
e^{-\imath \omega_b' t}
\right)
\\
\Longrightarrow ~~~~ ~~~~
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta
\right)
\rho_{12}'
+
\imath g' \mathscr{I}
\rho_{12}'
+
\imath f' \mathscr{I}
e^{\imath \omega_b' t}
\\
\partial_t w
&=&
- \gamma_1 \left( w - w_i \right)
+
2 \imath f' \mathscr{I}
\left(
\rho_{12}'
e^{+\imath \omega_b' t}
-
\rho_{12}'^*
e^{-\imath \omega_b' t}
\right)\end{aligned}$$
Notice that we have made the fixed complex phase (the $\delta'$) vanish from the equations; this is not dependent on the presence of a finite “detuning” $\Delta$. In what follows, $\Delta=0$, and there are two cases depending on the chosen meaning for $\mathscr{I}$; however note that [*both give the same result*]{}.
CASE (i): $\mathscr{I} = E^2$: continue by applying the carrier+envelope for the $E$ field $E = A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}$.
CASE (ii): $\mathscr{I} = 2A^*A$: we can jump straight to eqn.(\[eqn-rbpostRWA-w0-rho\], \[eqn-rbpostRWA-w0-w\]) (i.e. $(D)$), since the RWA was made when deriving the couplings $\alpha_{nj}$.
The CASE (i) $\mathscr{I} = E^2$ starting point is – $$\begin{aligned}
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta + \imath g' \mathscr{I}
\right)
\rho_{12}'
~~~~
+
\imath f' \mathscr{I} w
e^{-\imath \omega_b' t}
\\
\partial_t w
&=&
- \gamma_1 \left( w - w_i \right)
~~~~
+ 2 \imath f' \mathscr{I}
\left(
\rho_{12}'
e^{+\imath \omega_b' t}
-
\rho_{12}'^*
e^{-\imath \omega_b' t}
\right)
,
\\
(B) \Longrightarrow ~~~~ ~~~~
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta + \imath g' E^2
\right)
\rho_{12}'
~~~~
\imath f' E^2
e^{-\imath \omega_b' t}
.
w
\\
\partial_t w
&=&
- \gamma_1 \left( w - w_i \right)
~~~~
+ 2 \imath f' E^2
\left(
\rho_{12}'
e^{+\imath \omega_b' t }
-
\rho_{12}'^*
e^{-\imath \omega_b' t }
\right)
,
\\
(C) \Longrightarrow ~~~~ ~~~~
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta + \imath g' E^2
\right)
\rho_{12}'
~~~~
+
\imath f'
\left[
A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}
\right]^2
w
e^{-\imath \omega_b' t}
\\
\partial_t w
&=&
-
\gamma_1 \left( w - w_i \right)
~~~~
+
2 \imath f'
\left[
A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}
\right]^2
\left(
\rho_{12}
e^{+\imath \omega_b' t}
-
\rho_{12}^*
e^{-\imath \omega_b' t}
\right)
,
~~~~
\\
(\textrm{apply a RWA about}~ \omega_0) ~~~~ ~~~~ ~~~~ ~~~~
&:::&
\left[
A e^{-\imath \omega_0 t} + A^* e^{+\imath \omega_0 t}
\right]^2
\nonumber
\\
&& ~~~~ ~~~~ =
\left[
A^2 e^{-2 \imath \omega_0 t}
+ 2 A^* A
+ {A^*}^2 e^{+2 \imath \omega_0 t}
\right]
\\
\Longrightarrow ~~~~ ~~~~
&& ~~~~ ~~~~ \approx
2 A^* A
\\
CASE(ii) ~\&~ (D) \Longrightarrow ~~~~ ~~~~
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta + 2 \imath g' A^* A
\right)
\rho_{12}'
~~~~
+
2 \imath f' A^* A w
e^{-\imath \omega_b' t}
\label{eqn-rbpostRWA-w0-rho}
\\
\partial_t w
&=&
-
\gamma_1 \left( w - w_i \right)
~~~~
+
4\imath f'
A^* A
\left(
\rho_{12}
e^{+\imath \omega_b' t }
-
\rho_{12}^*
e^{-\imath \omega_b' t }
\right)
\label{eqn-rbpostRWA-w0-w}
,
\\
(E) ~~~~ ~~~~
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta + 2 \imath g' A^* A
\right)
\rho_{12}'
+
2 \imath f' A^*A w .
e^{-\imath \omega_b' t}
\label{eqn-rbpostRWA-wb-rho}
\\
\partial_t w
&=&
-
\gamma_1 \left( w - w_i \right)
+
4 \imath f'
A^* A
\left(
\rho_{12}'
e^{+\imath \omega_b' t}
-
\rho_{12}'^*
e^{-\imath \omega_b' t}
\right)
\label{eqn-rbpostRWA-wb-w}
,
\\
(F: \textrm{split ~} \rho_{12},) \Longrightarrow ~~~~ ~~~~
\partial_t u'
&=&
-
\gamma_2 u'
-
\left( \Delta + 2 g' A^* A \right) v'
~~~~
+
4 f' A^* A w . \sin \left( \omega_b' t \right)
\label{eqn-rbpostRWA-last-u}
\\
\partial_t v'
&=&
-
\gamma_2 v'
+
\left( \Delta + 2 g' A^* A \right) u'
~~~~
+
4 f' A^* A w . \cos \left( \omega_b' t \right)
\label{eqn-rbpostRWA-last-v}
\\
\partial_t w
&=&
- \gamma_1 \left( w' - w_i \right)
~~~~
- 4 f' A^* A u' . \sin \left( \omega_b' t \right)
~~~~
- 4 f' A^* A v' . \cos \left( \omega_b' t \right)
,
\label{eqn-rbpostRWA-last-w}\end{aligned}$$
Note $2\rho_{12} = u+\imath v$, since $\rho_{12}=c_1 c_2^*$, as per eqn.(\[eqn-blochcoupling-rho12-uv\]); this means the $2$ in the $u,v$ equations becomes $4$, whereas the $4$ in the $w$ equation isn’t doubled to $8$. Note (again) that the Stark shift parameter $g'$ is real valued.
### Rotations
Looking at eqns.(\[eqn-rbpostRWA-last-u\], \[eqn-rbpostRWA-last-v\], \[eqn-rbpostRWA-last-w\]), we can see three rotation angles: $\theta_{uv} = \Delta$, $\theta_{uw} = 4 f' A^* A \sin \left( \omega_b t - \delta' \right) $, and $\theta_{vw} = 4 f' A^* A \cos \left( \omega_b t - \delta' \right)$, which apply to the coordinate pairs $\left( u', v' \right) $, $\left( u', w\right) $, and $\left( v', w\right) $ respectively. For extra generality, I will allow a complex valued $f' = f_r + \imath f_i$, but note that previous definitions make $f'$ real. )
In vector notation, ignoring the losses, we can construct a torque vector $\vec{\Omega}$ (unchecked signs), $$\begin{aligned}
\partial_t u' &=& 0.u' - \theta_{uv}.v' + \theta_{uw}.w
\\
\partial_t v' &=& \Delta.u' + 0.v' - \theta_{vw}.w
\\
\partial_t w &=& -\theta_{uw}.u' + \theta_{vw}.v' + 0.w
\\
\frac{d}{dt} \left[ u', v', w \right]
&=&
\left[ -4 f'_r A^* A,
~~ -4 f'_i A^* A,
~~ \Delta
\right]
\\
&=&
\left[ \theta_{vw}, ~ \theta_{uw}, ~ \theta_{uv}
\right]
\times
\left[ u', v', w \right]
\\
\Longrightarrow ~~~~ ~~~~
\frac{d}{dt} \vec{U} &=& \vec{\Omega} \times \vec{U}\end{aligned}$$
I should now be able to turn this into a (unitary) rotation matrix for the $\vec{U}$ vector.
The polarization driving the field {#ss-single-driving}
----------------------------------
There is a distinction between the atomic polarization $P$, and the effect on the field of that atomic polarization. This is beause we are dealing with a nonlinear interaction. I denote the (effective) Raman polarization $\mathscr{P}$, and this quantity is the one that appears in the wave equation. Allen and Eberly [*“Optical resonance and two level atoms”*]{} [@AllenEberly-ORTLA] have, for the standard (non-Raman) case (skipping the integral), an eqn.(AE 4.2) – $$\begin{aligned}
(AE ~ 4.2) ~~~~ ~~~~ ~~~~ ~~~~
P(t,z) &=&
\mathscr{N} d
\left[
u \cos(\omega t - Kz) - v \sin(\omega t - Kz)
\right]\end{aligned}$$
HPB, after summing the electric field components (which absorbs a $1/2$), have $$\begin{aligned}
(HPB ~ 21) ~~~~ ~~~~ ~~~~ ~~~~
P &=&
\frac{1}{4}
e^{-\imath \theta}
\left(
u + \imath v
\right)
\sigma
\alpha_{12}
\sum_j
\left[
V_j e^{ \imath \omega_{j-1} t}
+
V_j^* e^{ -\imath \omega_{j+1} t}
\right]
~~~~
+ c.c.
\\
&=&
\frac{1}{2}
e^{-\imath \theta}
\left(
u + \imath v
\right)
\sigma
\alpha_{12}
e^{ -\imath \omega_b t}
\frac{1}{2}
\sum_j
\left[
V_j e^{ \imath \omega_j t}
+
V_j^* e^{ -\imath \omega_j t}
\right]
~~~~
+ c.c.
\\
&=&
\frac{1}{2}
e^{-\imath \theta}
\left(
u + \imath v
\right)
\sigma
\alpha_{12}
e^{ -\imath \omega_b t}
E
~~~~
+ c.c.\end{aligned}$$
I now write down the the polarization envelope $B$, in my variables, and based on the same carrier as the electric field. I also introduce a complex factor equivalent to HPB’s $e^{ -\imath \theta}$, for closer matching of the expressions. $$\begin{aligned}
\mathscr{P} ~~
=
\mathscr{B} e^{-\imath \Xi}
+
\mathscr{B}^* e^{+\imath \Xi}
&=&
\rho_{12}
e^{-\imath \theta }
\sigma
\alpha_{12}
\left(
A e^{-\imath \Xi}
+
A^* e^{+\imath \Xi}
\right)
~~~~
+ c.c.
\\
&=&
\zeta \rho'_{12}
e^{-\imath \theta}
\sigma .
\bar{\alpha}_{12}
e^{-\imath \omega_b' t }
\left(
A e^{-\imath \Xi}
+
A^* e^{+\imath \Xi}
\right)
~~~~
+ c.c.
\\
&=&
\zeta
\rho'_{12}
e^{-\imath \theta}
\sigma .
\bar{\alpha}_{12}'
e^{-\imath \omega_b' t }
\left(
A e^{-\imath \Xi}
+
A^* e^{+\imath \Xi}
\right)
+
\zeta
{\rho'_{12}}^*
e^{+\imath \theta}
\sigma .
\bar{\alpha}_{12}'
e^{+\imath \omega_b' t }
\left(
A^* e^{+\imath \Xi}
+
A e^{-\imath \Xi}
\right)
\\
&=&
\zeta
\rho'_{12}
e^{-\imath \theta}
\sigma .
\bar{\alpha}_{12}'
e^{+\imath \omega_b' t }
A e^{-\imath \Xi}
~~~~
+
\zeta
{\rho'_{12}}^*
e^{+\imath \theta}
\sigma .
\bar{\alpha}_{12}'
e^{-\imath \omega_b' t }
A e^{-\imath \Xi}
\nonumber
\\
&& +
\zeta
\rho'_{12}
e^{-\imath \theta}
\sigma .
\bar{\alpha}_{12}'
e^{+\imath \omega_b' t }
A^* e^{+\imath \Xi}
~~~~
+
\zeta
{\rho'_{12}}^*
e^{+\imath \theta}
\sigma .
\bar{\alpha}_{12}'
e^{-\imath \omega_b' t }
A^* e^{+\imath \Xi}
\\
&=&
\zeta
\sigma .
\bar{\alpha}_{12}'
\left\{
\rho'_{12}
e^{+\imath \omega_b' t -\imath \theta}
+
{\rho'_{12}}^*
e^{-\imath \omega_b' t +\imath \theta}
\right\}
A e^{-\imath \Xi}
\nonumber
\\
&& +
\zeta
\sigma .
\bar{\alpha}_{12}'
\left\{
\rho'_{12}
e^{+\imath \omega_b' t -\imath \theta}
+
{\rho'_{12}}^*
e^{-\imath \omega_b' t +\imath \theta}
\right\}
A^* e^{+\imath \Xi}
\\
&=&
\zeta
\sigma .
\bar{\alpha}_{12}'
X(t)
A e^{-\imath \Xi}
+
\zeta
\sigma .
\bar{\alpha}_{12}'
X(t)
A^* e^{+\imath \Xi}
\\
\textrm{ where the real valued is } ~~~~
X(t)
&=&
\left\{
\rho'_{12}
e^{+\imath \omega_b' t -\imath \theta}
+
{\rho'_{12}}^*
e^{-\imath \omega_b' t +\imath \theta}
\right\}
\label{eqn-single-Xrho}
\\
&=&
\left\{
\frac{1}{2}
\left( u' + \imath v' \right)
e^{+\imath \omega_b' t -\imath \theta}
+
\frac{1}{2}
\left( u' - \imath v' \right)
e^{-\imath \omega_b' t +\imath \theta}
\right\}
\\
&=&
\left\{
u' \cos \left( \omega_b' t - \theta \right)
-
v' \sin \left( \omega_b' t - \theta \right)
\right\}
\label{eqn-single-X}
\\
\textrm{and} ~~~~ ~~~~
\mathscr{B}(t)
&=&
\zeta
\sigma \bar{\alpha}_{12}'
X(t) A(t)\end{aligned}$$
From the post-$\omega_0$ RWA at eqns.(\[eqn-rbpostRWA-last-u\], \[eqn-rbpostRWA-last-v\], \[eqn-rbpostRWA-last-w\]); above; and an un-scaled eqn. ([FCPP](../_pulse-fewcyc/fewcyc.dvi) 3.48), – $$\begin{aligned}
\partial_t u'
&=&
-\gamma_2 u'
-\left( \Delta + 2 g' A^* A \right) v'
~~~~
+ 4 f' A^* A w \sin( \omega_b' t )
\label{eqn-single-Aprop-du}
\\
\partial_t v'
&=&
-\gamma_2 v
+ \left( \Delta + 2 g' A^* A \right) u'
~~~~
+ 4 f' A^* A w \cos( \omega_b' t )
\label{eqn-single-Aprop-dv}
\\
\partial_t w
&=&
- \gamma_1 \left( w' - w_i \right)
~~~~
- 4 f' A^* A . \sin(\omega_b' t ) u'
~~~~
- 4 f' A^* A . \cos(\omega_b' t ) v'
,
\label{eqn-single-Aprop-dw}
\\
\partial_z A(t)
&=&
\frac{2 \imath \pi \omega_0}{c_0 n_0}
\frac{ \mathscr{B}(t) }{4 \pi \epsilon_0}
,
~~~~ ~~~~
\textrm{ NB: units } [A]/m=\frac{s^{-1}}{m.s^{-1}} [\mathscr{B}]
~~\rightarrow [A] = [\chi] [B] [A]
~~\rightarrow [\chi] = [B^{-1}] = 1
\\
&=&
\frac{\imath \omega_0}{2 c_0 n_0 \epsilon_0}
\mathscr{B}(t)
\\
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12} \omega_0}{2 c_0 n_0 \epsilon_0 }
~.~
\left[
u' \cos \left( \omega_b' t - \theta \right)
-
v' \sin \left( \omega_b' t - \theta \right)
\right]
~.~
A(t)
\\
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12} \omega_0}{2 c_0 n_0 \epsilon_0 }
~.~
A(t) X(t)
\label{eqn-single-Apropagate}\end{aligned}$$
Comparing the prefactor of eqn.(\[eqn-single-Apropagate\]) to that of the corrected (HPB 22), $$\begin{aligned}
\textrm{(HPB 22)} ~~~~ ~~~~
\frac{\partial V_j}{\partial z}
&=&
\frac{\sigma \alpha_{12}}{4 \epsilon_0 c}
\omega_j
\left[
e^{-\imath \theta} \left(u-\imath v\right) V_{j+1}
-
e^{+\imath \theta} \left(u+\imath v\right) V_{j-1}
\right]\end{aligned}$$ we can see that the only apparent differences are a factor of $n_0$, and in that HPB have a $e^{\pm \imath \theta}$ term that I omit. By imagining the cosine term split up into $+$ and $-$ frequency exponentials, it is easy to see how the relations between $V_j$ and $V_{j\pm 1}$ will arise. Note the appearance (in HPB) of a carrier-dependent $\omega_i$ term, which in a standard multi-field approach would lead to different prefactors on the different $\partial_\xi A_i$ equations.
### Simulation “photon” variables
Scale the field $E$ into square-root intensity “photon” variables, from $I = 2 n_0 c_0 \epsilon_0 E^2$, so that $$\begin{aligned}
A_p(t)
&=&
\sqrt{2 c_0 n_0 \epsilon_0 } ~ . ~ A(t)
\label{eqn-aphotonscale}
\\
f_p &=&
\frac{f'}{2 c_0 n_0 \epsilon_0};
~~~~ ~~~~
g_p ~=~
\frac{g'}{2 c_0 n_0 \epsilon_0};
\label{eqn-fphotonscale}\end{aligned}$$
Hence $f' A^*A \longrightarrow f_p A_p^* A_p$, so that $$\begin{aligned}
\partial_t u'
&=&
-
\gamma_2 u'
-
\left( \Delta + 2 g_p A^* A \right) v'
~~~~
+
4 f_p A_p^* A_p w \sin(\omega_b' t)
\\
\partial_t v'
&=&
-
\gamma_2 v
+
\left( \Delta + 2 g_p A^* A \right) u'
~~~~
+
4 f_p A_p^* A_p w \cos(\omega_b' t)
\\
\partial_t w
&=&
- \gamma_1 \left( w' - w_i \right)
~~~~
- 4 f_p A_p^* A_p u' \sin(\omega_b' t)
~~~~
- 4 f_p A_p^* A_p v' \cos(\omega_b' t)
\\
\partial_z A_p(t)
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12} \omega_0}
{ 2 c_0 n_0 \epsilon_0}
A_p(t)
X(t)
\label{eqn-single-Appropagate2}
\\
&=&
\imath
\zeta \sigma \omega_0
\frac{ 4 \hbar c_n n_0 \epsilon_0 f_p}
{ 2 c_0 n_0 \epsilon_0}
A_p(t)
X(t)
\\
&=&
\imath
\frac{\zeta}{2}
\left( 4 \sigma \hbar \right)
\omega_0 f_p
A_p(t)
X(t)
~~~~ ~~~~
=
\imath
\frac{\zeta}{2}
\omega_0 R f_p
A_p(t)
X(t); ~~~~
R = 4 \sigma \hbar
\label{eqn-single-Apropphoton}\end{aligned}$$
since $$\begin{aligned}
\bar{\alpha}_{12} = 2 \hbar f'
~~~~ \rightarrow
\frac{\bar{\alpha}_{12}}{2 c_0 n_0 \epsilon_0} = 2 \hbar f_p
~~~~ \rightarrow
\bar{\alpha}_{12} = 4 \hbar c_0 n_0 \epsilon_0 f_p\end{aligned}$$
Scalings:\
$\hbar = 1.05\times10^{-34}$Js $\rightarrow 1.05\times10^{-10}$nJ.fs\
$\sigma = X$m$^{-3} \rightarrow X\times10^{-18} \mu$m$^{-3}$\
$2 c_0 n_0 \epsilon_0 = 5.31 \times 10^{-3}$m.s$^{-1}$ . J.m$^{-1}$.V$^{-2}
\rightarrow 5.31 \times 10^{-3}$ . J.s$^{-1}$.V$^{-2}
\rightarrow 5.31 \times 10^{-9}$ . nJ.fs$^{-1}$.V$^{-2}$
Scalings:\
$[c \epsilon_0 ] = m.s^{-1} ~.~ J.m^{-1}.V^{-2} ~~~ = J.s^{-1}.C^2.J^{-2}
~~~ = C^2. J^{-1}.s^{-1} ~~~ = C^2 ~.~ 10^{-9} nJ^{-1} ~.~ 10^{-15} fs^{-1}
~~~ = 10^{-24} . C^2 . nJ^{-1} . fs^{-1} $\
$[\epsilon_0 ] = J.m^{-1}.V^{-2} ~~~ = J.m^{-1}.C^2.J^{-2}
~~~ = C^2 . J^{-1} . m^{-1} = C^2 ~.~ 10^{-9} nJ^{-1} ~.~ 10^{-6} \mu m^{-1}
~~~ = 10^{-15} . C^2 . nJ^{-1} . \mu m^{-1} $\
### Comments on the field propagation
We can see that the derivative of $A$ is proportional to $\imath A$, hence the (time-domain) evolution just rotates each point $A(t)$ differently without changing its amplitude. This might seem to imply that we will never get a shorter pulse than we put in; but note that the $A$ is an [*envelope*]{}, and the field is $A + A^*$, so that in principle the phase of $A$ might be such that its amplitudes cancel in certain $t$ regions but not others, leading to a shorter pulse (and the magnetic field $H$ also, since the fields are plane polarized). However, it seems unlikely that this will happen (barring some miraculous coincidence) from purely Raman effects.
If I can predict the output spectral phases though, a structure with a suitably designed dispersion [*might*]{} be able to impose the desired phases. It would only be necessary to get the dispersion right at the comb points of the spectrum. See Shverdin et.al. [@Shverdin-WYYH-2004pre], who do a four-wave mixing optimization procedure in their experiment to match their phases appropriately.
The steady state and gain co-efficient {#ss-steadystate}
--------------------------------------
Starting from the eqns.(\[eqn-single-Aprop-du\], \[eqn-single-Aprop-dv\], \[eqn-single-Aprop-dw\]), assume $v'$ is steady state (NB $\gamma_2 = 1/T_2$), so (with $w=-1$) – $$\begin{aligned}
\partial_t \approx 0
~~~~ ~~~~
&=&
-\gamma_2 v' + 4 f' A^* A (-1) \cos(\omega_b t)
\\
\Longrightarrow ~~~~ ~~~~
v_0' &=&
\frac{4 f'}{\gamma_2} ~.~ A^* A . \cos(\omega_b t)
~~~~ ~~~~ =
4 f' T_2 ~.~ A^* A . \cos(\omega_b t)\end{aligned}$$
I now use the field propagation equation (\[eqn-single-Apropphoton\]), with the $A$ field envelope on each side scaled into “photon” variables. Inserting the steady state of $v'$ calculated immediately above, we have – $$\begin{aligned}
\partial_z A_p
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12} \omega_0}
{2 c_0 n_0 \epsilon_0}
A_p v_0' \sin(\omega_b t)
\\
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12} \omega_0}
{ 2 c_0 n_0 \epsilon_0}
A_p
~.~
4 f' T_2 ~.~ A^* A . \cos(\omega_b t) . \sin(\omega_b t)
\\
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12} \omega_0}
{2 c_0 n_0 \epsilon_0}
A_p
~.~
4 T_2
\frac{\bar{\alpha}_{12}}
{2 \hbar}
\frac{A_p^* A_p}
{2 c_0 n_0 \epsilon_0}
\left[ \cos(\omega_b t) . \sin(\omega_b t) \right]
\\
&=&
\imath
\frac{ \zeta \sigma \omega_0 T_2 \bar{\alpha}_{12}^2}
{ 2 c_0^2 n_0^2 \epsilon_0^2 \hbar }
A_p
A_p^* A_p .
\left[ \cos(\omega_b t) . \sin(\omega_b t) \right]
\\
&=&
\imath G_p
A_p
A_p^* A_p .
\left[ \cos(\omega_b t) . \sin(\omega_b t) \right]
,
\\
G_p &=&
\frac{ \zeta }{2}
\frac{ \sigma \omega_0 T_2 \bar{\alpha}_{12}^2}
{ c_0^2 n_0^2 \epsilon_0^2 \hbar }
\label{eqn-ssgain-Gp}
.\end{aligned}$$
Apparently, therefore, this $G_p$ corresponds to the usual “gain co-efficient” $g$: $$\begin{aligned}
\textrm{cf (SMN) } ~~~~ ~~~~
g' &=&
\frac{ \sigma \omega_0 T_2 \alpha_{12}^2 }
{ c^2 \epsilon_0^2 \hbar }
\label{eqn-ssgain-gp-smn}
,\end{aligned}$$
and we can assume that they are identical but for a factor of $2 n_0^2 /
\zeta$. I do not know whether SMN silently assumes $n_0=1$, or factors $n_0$ into $c$. In any case $n_0=1$ is usually accurate enough in gases.
However, one hidden complication with my above equation is that $A_p$ includes both the center “pump” field and the Raman sideband we are amplifying – thus to properly check the gain co-efficient, we should expand it into its components, as is done below.
### Pump and sideband calculation
$A_p$ includes both fundamental $A_1$ and its Raman sideband $A_2=A_2'
e^{-\imath \omega_b t}$. For the moment, I leave the calculation in a rather abbreviated state – probably it should be shifted to the Multi-field section following later. So $$\begin{aligned}
A_p A_p^* A_p
&=&
\left[ A_1 + A_2 \right]
\left[ A_1^* + A_2^* \right]
\left[ A_1 + A_2 \right]
\\
&=&
\left[
A_1 A_1^* + A_1 A_2^* + A_2 A_1^* + A_2 A_2^*
\right]
\left[ A_1 + A_2 \right]
\\
&=&
A_1 A_1^* A_1 + A_1 A_2^* A_1 + A_2 A_1^* A_1 + A_2 A_2^* A_1
+
A_1 A_1^* A_2 + A_1 A_2^* A_2 + A_2 A_1^* A_2 + A_2 A_2^* A_2
\\
\textrm{1st order terms in $A_2$ only:}
~~~~ ~~~~
&\simeq&
A_1^2 A_2^* + 2 A_2 A_1^* A_1
\\
\textrm{drop counter rotating $A_1^2 A_2^*$: }
~~~~ ~~~~
&\simeq&
2 A_2 A_1^* A_1
.\end{aligned}$$
Thus the sideband envelope $A_2'$ evolves as $$\begin{aligned}
\partial_z A_2'
&=&
\imath G_p
A_2'
.
2 A_1^* A_1 . \cos(\omega_b t)
\\
&=&
\imath G_p
A_2'
.
I_1 . \cos(\omega_b t)
\\
&=&
\imath G'_p
A_2'
.
I_1 . \cos(\omega_b t)
,
\\
\textrm{where}
~~~~ ~~~~
G'_p
&=&
\frac{ \zeta }{2}
\frac{ \sigma \omega_0 T_2 \bar{\alpha}_{12}^2}
{ c_0^2 n_0^2 \epsilon_0^2 \hbar }
~~~~ ~~~~
=
G_p\end{aligned}$$
since $I_m = \frac{1}{2} A_m^* A_m$. I don’t convert $A_2'$ into $I_2$ because it occurs equally on both side of the equation.
Multi-field variant of single-field Raman theory {#s-multifield}
================================================
The single-field Raman model above can be converted into a traditional multi-field model as developed in e.g. HPB [@Hickman-PB-1986pra] or Syed, McDonald and New [@Syed-MN-2000josab] by replacing the field envelope with a sum of multiple envelopes using carrier exponentials spaced at the Raman frequency. When doing this, I will only get the correct multi-field form if few-cycle (either SEWA or GFEA) corrections to the field evolution part of the theory are applied to the effective polarization caused by the Raman transition.
The idea is to replace the single field envelope with a sum of multiple envelopes spaced at the Raman frequency, which are best placed to represent the comb of frequencies generated by the Raman interaction. Note that it will not necessarily be identical to HPB and/or SMN, because the field equations are derived from a propagation equation using a $\omega_0, \beta_0$ carrier, but it will be very closely related.
Starting from eqns.(\[eqn-rbpostRWA-wb-rho\],\[eqn-rbpostRWA-wb-w\]), Since the single-field evolution equation (eqn.(\[eqn-single-Apropagate\])) uses an envelope $A$ that is based on a carrier (see eqn.(\[eqn-single-EfromA\])), the single-field envelope $A$ is replaced with $A_j$’s at frequency $\omega_j = \omega_0 + j \omega_b$ with wavevector $k_j =
k(\omega_j)$; thus $\omega'_j = \omega_j - \omega_0$, $k'_j = k_j - k(\omega_0) = k_j - k_0$; also $\beta \leftrightarrow k$. The single-field envelope in terms of the new $A_j$’s is $$\begin{aligned}
A = \sum_j A_j \exp\left[ -\imath \left( \omega_j t - k_j z \right) \right]\end{aligned}$$
Polarization ($ \rho_{12}'$) {#ss-multifield-rho}
----------------------------
First I will handle the polarization ($ \rho_{12}'$) equation (from eqn.(\[eqn-rbpostRWA-w0-rho\])) (watch for any $\omega_b' = \omega_b - \Delta$ confusion, and note $2\rho_{12}'=u'+\imath v'$) – $$\begin{aligned}
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta + 2 \imath g' A^* A
\right)
\rho_{12}'
+
2 \imath f' A^*A w
e^{-\imath \omega_b' t }
\\
(A1) ~~ \Longrightarrow ~~
\partial_t \rho_{12}'
&=&
\left(
-\gamma_2 + \imath \Delta
\right)
\rho_{12}'
+
2 \imath g' \rho_{12}'
\sum_j \sum_k A_j^*A_k
e^{+\imath \left( \omega_j - \omega_k \right) t}
e^{+\imath \left( -k_j + k_k \right) z }
\nonumber
\\
&& ~~~~ ~~~~
+
2 \imath f' \sum_j \sum_k A_j^*A_k w
e^{+\imath \left( \omega_j - \omega_k - \omega_b' \right) t}
e^{+\imath \left( -k_j + k_k \right) z }
\\
&=&
\left(
-\gamma_2 + \imath \Delta
\right)
\rho_{12}'
+
2 \imath g' \rho_{12}'
\sum_j \sum_k A_j^*A_k
e^{+\imath \left( j - k \right) \omega_b t}
e^{+\imath \left( -k_j + k_k \right) z }
\nonumber
\\
&& ~~~~ ~~~~
+
2 \imath f' \sum_j \sum_k A_j^*A_k w
e^{+\imath \left( j - 1 - k \right) \omega_b t + \imath \Delta t}
e^{+\imath \left( k_k-k_j \right) z }
\\(RWA) ~~~~ ~~~~
&\approx&
\left(
-\gamma_2 + \imath \Delta
\right)
\rho_{12}'
+
2 \imath g' \rho_{12}'
\sum_j A_j^*A_j w
+
4 \imath f' \sum_j A_j^*A_{j-1}
. w
. e^{+\imath \Delta t }
. e^{+\imath \left( k_{j+1}-k_j \right) z }
\\
&\approx&
\left(
-\gamma_2 + \imath \Delta + 2 \imath g' \sum_j A_j^* A_j
\right)
\rho_{12}'
+
4 \imath f' \sum_j A_{j} A_{j+1}^*
. w
. e^{+\imath \Delta t }
. e^{+\imath \left( k_j-k_{j-1} \right) z }
\\
&\approx&
\left(
-\gamma_2 + \imath \Delta + 2 \imath g' \sum_j A_j^* A_j
\right)
\rho_{12}'
+
4 \imath f' w .
\mathbb{R}e
\left[
\sum_j A_{j} A_{j+1}^*
. e^{+\imath \Delta t }
. e^{+\imath \left( k_j-k_{j-1} \right) z }
\right]
\nonumber
\\
&& ~~~~ ~~~~
+
4 \imath^2 f' w .
\mathbb{I}m
\left[
\sum_j A_{j} A_{j+1}^*
. e^{+\imath \Delta t }
. e^{+\imath \left( k_j-k_{j-1} \right) z }
\right]
\\
\left(\textrm{split}~ \rho_{12}'\right)
~~~~
~~~~
\partial_t u
&=&
-\gamma_2 u
-
\left(
\Delta + 2 g' \sum_j A_j^* A_j
\right)
v
-
8 f' w .
\mathbb{I}m
\left[
\sum_j A_{j} A_{j+1}^*
. e^{+\imath \Delta t }
. e^{+\imath \left( k_j-k_{j-1} \right) z }
\right]
\\
\partial_t v
&=&
-\gamma_2 v
+
\left(
\Delta + 2 g' \sum_j A_j^* A_j
\right)
u
+
8 f' w .
\mathbb{R}e
\left[
\sum_j A_{j} A_{j+1}^*
. e^{+\imath \Delta t }
. e^{+\imath \left( k_j-k_{j-1} \right) z }
\right]\end{aligned}$$
Where the factor $2 \imath f'$ turns into $4 \imath f'$ because the double summation gives two identical terms that only occur once in the single summation. When split into equations for $u'$ and $v'$, the corresponding factor becomes $8 \imath f'$. This equation for $\rho_{12}'$ is equivalent to (SMN 2)[@Syed-MN-2000josab], except I have just $k_j-k_{j-1}$ whereas they have $\Delta_j = k_j - k_{j-1} - k_0 + k_{-1}$; however note they have the reverse sign in their carrier wave, so the only physical difference is the $k_0 - k_{-1}$ part; also my definitions of the coupling differs slightly.
Inversion ($w$) {#ss-multifield-w}
---------------
And now the inversion ($w$) equation (also from eqn.(\[eqn-rbpostRWA-w0-rho\])) (watch for any $\omega_b$ vs $\omega_b'$ confusion) – $$\begin{aligned}
\partial_t w
&=&
- \gamma_1 \left( w - w_i \right)
+ 4 \imath f'
A^* A
\left[
\rho_{12}' e^{+\imath \omega_b' t}
-
\rho_{12}'^* e^{-\imath \omega_b' t}
\right]
\\
(A2) ~~
\partial_t w
&=&
- \gamma_1 \left( w - w_i \right)
+ 4 \imath f'
\sum_j \sum_k A_j^*A_k
. e^{+\imath \left( \omega_j - \omega_k \right) t}
. e^{+\imath \left( -k_j + k_k \right) z}
\left[
\rho_{12}' e^{+\imath \omega_b' t}
-
\rho_{12}'^* e^{-\imath \omega_b' t}
\right]
\\
&=&
- \gamma_1 \left( w - w_i \right)
+ 4 \imath f'
\sum_j \sum_k
\left[
A_j^*A_k
\rho_{12}'
e^{+\imath \left( j -k +1 \right) \omega_b t - \imath \Delta t }
-
A_j^*A_k
\rho_{12}'^*
e^{+\imath \left( j -k -1 \right) \omega_b t + \imath \Delta t}
\right]
. e^{+\imath \left( k_k -k_j \right) z}
\\(RWA) ~~
&\approx&
- \gamma_1 \left( w - w_i \right)
+ 8 \imath f' \sum_j
\left[
A_j^*A_{j+1}
\rho_{12}'
. e^{-\imath \Delta t }
. e^{+\imath \left( k_{j+1} -k_j \right) z }
-
A_j^*A_{j-1}
\rho_{12}'^*
. e^{+\imath \Delta t }
. e^{+\imath \left( k_{j-1} -k_j \right) z }
\right]
\\
&=&
- \gamma_1 \left( w - w_i \right)
+ 4 \imath f' \sum_j
\left[
A_j^*A_{j+1}
\left( u' + \imath v' \right)
. e^{-\imath \Delta t }
. e^{+\imath \left( k_{j+1} -k_j \right) z }
\right.
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~
\left.
-
A_{j+1}^*A_{j}
\left( u' - \imath v' \right)
. e^{+\imath \Delta t }
. e^{+\imath \left( k_{j} -k_{j+1} \right) z }
\right]
\\
&=&
- \gamma_1 \left( w - w_i \right)
+
4 \imath u' f' \sum_j
\left[
A_j^*A_{j+1}
. e^{-\imath \Delta t }
. e^{+\imath \left( k_{j+1} -k_j \right) z }
- c.c.
\right]
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~
-
4 v' f' \sum_j
\left[
A_j^*A_{j+1}
. e^{-\imath \Delta t }
. e^{+\imath \left( k_{j+1} -k_j \right) z }
+ c.c.
\right]
\\
&=&
- \gamma_1 \left( w - w_i \right)
-
8 u' f' \sum_j
\mathbb{I}\textrm{m} \left[
A_j^*A_{j+1}
. e^{-\imath \Delta t }
. e^{+\imath \left( k_{j+1} -k_j \right) z }
\right]
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~
-
8 v' f' \sum_j
\mathbb{R}\textrm{e} \left[
A_j^*A_{j+1}
. e^{-\imath \Delta t }
. e^{+\imath \left( k_{j+1} -k_j \right) z }
\right]\end{aligned}$$
Note the RWA’s above (for both polarization and inversion equations) discard modulations at frequency of multiples of $\omega_b$. Quite a lot of physics has been removed by these RWA approximations, although it is very reasonable except in the very wideband limit. For example, the effect of next-nearest neighbour field components acting on the transition have been ignored, as have all more distant field-field interactions. In the next-nearest neighbour case, the dropped terms would impose a rapid $\omega_b$ oscillation onto the polarization $\rho_{12}$, which would in turn tend to impose sidebands at $\pm \omega_b$ onto each field component. It is reasonable to ignore such sidebands in the narrowband limit studied by most users of a multi-field Raman theory; but, in principle one might extend a multi-field theory to include them by inventing a scheme to apply the sidebands to the field component they are (near) resonant with.
Fields ($A_j$) {#ss-multifield-A}
--------------
Note that the field evolution equation already has a carrier of $\exp \left[ -\imath \left( \omega_0 t - k_0 z \right) \right]$ factored out of it. Thus I use $A'$ from $A = A' \exp \left[ -\imath \left( \omega_0 t - k_0 z \right) \right]$, not $A$. Finally, I need to insert the GFEA few-cycle correction to the polarization term, because my ($j\ne 0$) sub-envelopes $A_j$ have an $\imath j \omega_b t$ time dependence that cannot be neglected.
From eqns.(\[eqn-single-Apropagate\], \[eqn-single-Xrho\]), I get $$\begin{aligned}
\partial_z A'(t)
&=&
\imath
\left[
1 + \frac{\imath \partial_t }{\omega_0}
\right]
\frac{\zeta \sigma \omega_0 \bar{\alpha}_{12}'}
{2 \epsilon_0 c_0}
A'(t) X(t)
\label{eqn-single-Apropagate2-copy}
\\
\partial_z
\sum_j A_j \exp\left[ -\imath \left( \omega'_j t - k'_j z \right) \right]
&=&
\imath
\left[
1 + \frac{\imath \partial_t}{\omega_0}
\right]
\frac{\zeta \sigma \omega_0 \bar{\alpha}_{12}'}
{2 \epsilon_0 c_0}
\left[
\rho_{12}' e^{ +\imath \omega_b' t }
+
\rho_{12}'^* e^{ -\imath \omega_b' t }
\right]
\nonumber
\\
&&
~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ \times
\sum_j A_j \exp\left[ -\imath \left( \omega'_j t - k'_j z \right) \right]
\\
(\textrm{match} ~ \omega_j ~ \textrm{terms}) ~~~~
\left[
\imath
k'_j
A_j
+
\partial_z A_j
\right]
e^{-\imath j \omega_b t }
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12}'}
{2 \epsilon_0 c_0}
\left\{
A_{j+1} \rho_{12}'
\exp\left[ +\imath (k'_{j+1} -k'_j) z - \imath \Delta t \right]
\right.
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~
\left.
+
A_{j-1} \rho_{12}'^*
\exp\left( +\imath (k'_{j-1} -k'_j) z + \imath \Delta t \right)
\right\}
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ \times ~~~~
\left[
\omega_0 + \imath \partial_t
\right]
e^{-\imath j \omega_b t }\end{aligned}$$
Then, using $$\begin{aligned}
\left[
\omega_0 + \imath \partial_t
\right]
e^{-\imath j \omega_b t }
&\longrightarrow&
\omega_0
e^{-\imath j \omega_b t }
-
\imath^2
j \omega_b
e^{-\imath j \omega_b t }
~~~~
\longrightarrow
~~~~
\left[
\omega_0 + j \omega_b
\right]
e^{-\imath j \omega_b t }
~~~~
\longrightarrow
~~~~
\omega_j
e^{-\imath j \omega_b t }\end{aligned}$$
So $$\begin{aligned}
\left[
\imath
k'_j
A_j
+
\partial_z A_j
\right]
e^{-\imath j \omega_b t }
&=&
\imath
\frac{\zeta \sigma \bar{\alpha}_{12}'}
{2 \epsilon_0 c_0}
\left\{
A_{j+1} \rho_{12}'
\exp\left[ +\imath (k'_{j+1} -k'_j) z - \imath \Delta t \right]
\right.
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~ ~~~~ ~~~~
\left.
+
A_{j-1} \rho_{12}'^*
\exp\left( +\imath (k'_{j-1} -k'_j) z + \imath \Delta t \right)
\right\}
~~~~ \times ~~~~
\omega_j
e^{-\imath j \omega_b t }
\\
\imath
k'_j
A_j
+
\partial_z A_j
&=&
\imath
\frac{\zeta \sigma \omega_j \bar{\alpha}_{12}'}
{4 \epsilon_0 c_0}
u
\left[
A_{j+1}
\exp\left[ +\imath (k'_{j+1} -k'_j) z - \imath \Delta t \right]
+
A_{j-1}
\exp\left[ +\imath (k'_{j-1} -k'_j) z + \imath \Delta t \right]
\right]
\nonumber
\\
&& ~~
-
\frac{\zeta \sigma \omega_j \alpha_{12}}
{4 \epsilon_0 c_0}
v
\left[
A_{j+1}
\exp\left( +\imath (k'_{j+1} -k'_j) z - \imath \Delta t \right)
-
A_{j-1}
\exp\left( +\imath (k'_{j-1} -k'_j) z + \imath \Delta t \right)
\right]
~~~~ ~~~~
\\
\partial_z A_j
&=&
\frac{\zeta \sigma \omega_j \alpha_{12}}
{4 \epsilon_0 c_0}
\left\{
-
\left[
v
-
\imath u
\right]
A_{j+1}
\exp\left( +\imath (k'_{j+1} -k'_j) z - \imath \Delta t \right)
+
\left[
v
+
\imath u
\right]
A_{j-1}
\exp\left( +\imath (k'_{j-1} -k'_j) z + \imath \Delta t \right)
\right\}
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~
-
\imath \left( k_j - k_0 \right) A_j\end{aligned}$$
This is in agreement with both HPB[@Hickman-PB-1986pra] barring the opposite sign on the RHS – similar agreement occurs with SMN[@Syed-MN-2000josab] once I identify $q = (v+\imath u)$. Note that generally $\Delta=0$, as it just controls a frame rotation for $\rho_{12}$.
Note that we can assume, quite reasonably, that the multi-field envelopes $A_j$ will be better behaved than the single-field envelope $A$. However, we have made [*more*]{} approximations, notably by RWA’ing away all the off-resonant cross terms driving the atomic transition so a multi-field approach is not always better. These off-resonant terms are at $2\omega_0 \pm \omega_b$ – see just prior to the starting point above of eqns.(\[eqn-rbpostRWA-w0-rho\], \[eqn-rbpostRWA-w0-w\]).
In photon variables, the above field propagation equation is (using $R=4 \hbar \sigma$, also note $\zeta=2$ to conserve energy) $$\begin{aligned}
\partial_z A_{p,j}
&=&
\frac{\zeta}{4}
R
\omega_j f_p
\left\{
-
\left[
v
-
\imath u
\right]
A_{p,j+1}
\exp\left( +\imath (k'_{j+1} -k'_j) z - \imath \Delta t \right)
+
\left[
v
+
\imath u
\right]
A_{p,j-1}
\exp\left( +\imath (k'_{j-1} -k'_j) z + \imath \Delta t \right)
\right\}
\nonumber
\\
&& ~~~~ ~~~~ ~~~~ ~~~~
-
\left( k_j - k_0 \right) A_{p,j}\end{aligned}$$
Comparisons {#s-comparisons}
===========
See [*“Wideband pulse propagation: single-field and multi-field approaches to Raman interactions”*]{} by Kinsler and New [@Kinsler-N-2005pra].
Summary {#s-summary}
=======
I describe how to model a multi-frequency field such as that seen in a wideband Raman generation experiment using a single field envelope rather than a set of envelopes, one at each Stokes or anti-Stokes frequency. This requires that the field be propagated taking into account wideband effects, as described by either the SEWA theory of Brabec and Krausz [@Brabec-K-1997prl], or the more general GFEA of Kinsler and New [@Kinsler-N-2003pra].
The usefulness of this single-field approach is not restricted to the Raman interaction described in this paper. It would be equally valuable for a near-degenerate optical parametric oscillator, or indeed any system where any two or more field components contain spectra that start to overlap as the pump or probe pulses get shorter.
It is important to note that it will usually only be more efficient to use a single-field simulation if pump pulses are very short, and effects like the next-nearest neighbour field interactions, neglected in the multi-field theory, need to be included, or if the extra computational overhead is not inconvenient. This is because in a single-field simulation, a very fine time-resolution is necessary to model the polarization and field oscillations closely enough to get good numerical convergence. However, this situation improves when the Raman transition has a smaller frequency compared to the pump pulse frequencies. One useful side effect of the fine time resolution is that it naturally gives a wide spectral bandwidth, so that many Stokes and anti-Stokes lines are modeled quite naturally. Further, our single-field model could be invaluable in modeling a short pulse pump-probe experiment where the probe frequency does not match any of the Stokes or anti-Stokes spectral lines generated by the pump pulse(s). A multi-field simulation would then need arrays for both the pump and probe Raman ‘ladders’’ of Stokes/anti-Stokes lines, and the role of next nearest neighbour interactions (ignored in the multi-field model) could well become more significant.
In summary, the advantages of our single-field approach are twofold. First, it includes more physics than the multi-field approach, even compared to a multi-field approach enhanced by adding GFEA corrections to the propagation of the field components. Secondly, it deals effortlessly with the complications of overlapping spectra in the multi-field case.
[19]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, (), <http://link.aps.org/abstract/PRA/v72/e033804>.
, ****, ().
, , , , , ****, ().
, , , , ****, ().
, , , , , ****, ().
, , , , , , , ****, (), <http://link.aps.org/abstract/PRA/v72/e013813>.
, ****, (), <http://link.aps.org/abstract/PRA/v67/e023813>.
, (), <http://arXiv.org/physics/0212014>.
, ****, ().
, , , ****, (), [http://journalsonline.tandf.co.uk/openurl.asp?genre=article&i%
d=doi:10.1080/09500340512331334086](http://journalsonline.tandf.co.uk/openurl.asp?genre=article&i%
d=doi:10.1080/09500340512331334086).
, , , ****, (), <http://link.aps.org/abstract/PRA/v33/p1788>.
, , , , , ****, ().
, ****, ().
, , , , , ****, (), <http://link.aps.org/abstract/PRA/v60/p1562>.
, ** (, ).
, , , , , ****, ().
, , , ****, ().
, ****, ().
, , , , , ****, ().
|
---
abstract: 'Variational inference approximates the posterior distribution of a probabilistic model with a parameterized density by maximizing a lower bound for the model evidence. Modern solutions fit a flexible approximation with stochastic gradient descent, using Monte Carlo approximation for the gradients. This enables variational inference for arbitrary differentiable probabilistic models, and consequently makes variational inference feasible for probabilistic programming languages. In this work we develop more efficient inference algorithms for the task by considering importance sampling estimates for the gradients. We show how the gradient with respect to the approximation parameters can often be evaluated efficiently without needing to re-compute gradients of the model itself, and then proceed to derive practical algorithms that use importance sampled estimates to speed up computation. We present importance sampled stochastic gradient descent that outperforms standard stochastic gradient descent by a clear margin for a range of models, and provide a justifiable variant of stochastic average gradients for variational inference.'
author:
- |
[**Joseph Sakaya**]{}\
Helsinki Institute for Information Technology HIIT\
Department of Computer Science\
University of Helsinki, Finland\
\
Helsinki Institute for Information Technology HIIT\
Department of Computer Science\
University of Helsinki, Finland\
bibliography:
- 'paper.bib'
title: Importance Sampled Stochastic Optimization for Variational Inference
---
INTRODUCTION
============
Variational inference considers parametric approximations for posterior densities of probabilistic models. Following @Jordan1999 the classical variational approximation algorithms are based on coordinate descent algorithms for which individual steps of the algorithm are often carried out analytically. This limits the use of variational approximation to models with conjugate priors (or simple extensions of those) and restricts the family of potential approximating distributions based on analytic tractability.
Recent advances in variational approximation have lead to a phase transition; instead of closed-form updates, the approximation is nowadays often fit using generic gradient descent techniques instead – for a good overview see, [*e.g.*]{}, @gal. The key behind these advances is in using Monte Carlo approximation to estimate the gradient of the objective function that is an integral over the approximating distribution. This can be done in two alternative ways. The *reparameterization estimate* [@dsvi; @vae; @salimans] allows expressing the gradient of the objective function using the gradients of the model itself, whereas the *score function estimate* [@bbvi] is based on gradients of the approximation. Given the new family of algorithms we can apply variational approximation for a considerably wider range of probabilistic models, enabling for example use of variational inference as the inference backend in probabilistic programming languages [@advi; @pymc3; @edward].
The main research efforts in variational inference are nowadays geared towards making the approach applicable to a still wider family of models, by constructing even more flexible approximations [@rezende15; @ranganath16; @kingma16] or by generalizing the gradient estimators [@grep; @naesseth]. The question of how exactly the resulting optimization problem is solved has largely remained unattended to – practically all authors are satisfied with standard stochastic gradient descent (SGD) as the underlying optimizer, although some effort has been put into improving convergence by reducing the variance of the gradient estimate near the optimum [@roeder17].
We turn our attention to the optimizer itself, looking into ways of speeding up the computation of gradient-based variational approximation. Practically all of the computational effort during learning goes into evaluating the gradient of the model (or the approximation if using the score-function estimate). Our contribution is in reducing the number of times we need to evaluate the gradient of the model during the optimization process, based on an importance sampling scheme specifically designed for optimization problems where the gradients are computed using Monte Carlo approximation.
The key observation is that the gradient of the objective function with respect to the parameters of the approximation consists of two parts. One part is the gradient of the model itself, evaluated at parameter values drawn from the approximation, whereas the other part is the gradient of the transformation used in the reparameterization estimate. We show that the gradient required for optimization can be computed for the newly updated approximation without re-computing the first part, which is computationally heavier. Instead, we can re-use existing computation by appropriately modifying and re-weighting the available terms.
We show how to formulate this idea in a justified manner, by constructing an importance sampling estimate for the gradient. Importance sampling is typically used for cases where one cannot sample from the distribution of interest but instead has to resort to sampling from a related proposal distribution. In our case we could sample from the distribution of interest – the current approximation – but choose not to, since by using an earlier approximation as a proposal we can avoid costly computation. The idea is conceptually similar to the way @ep reuses samples from previous iterations in expectation propagation.
Since the advances in our case are related to the computation of the gradient itself, the idea can readily be combined with several optimization algorithms. In this work, we derive practical algorithms extending standard SGD and stochastic average gradients [@sage]. We demonstrate them in learning a variational approximation for several probabilistic models, showing how they improve the convergence speed in a model-independent manner.
In the following we first give a brief overview of the state-of-the-art in gradient-based variational approximation, covering both the gradient estimates and stochastic optimization algorithms in Section 2. We then proceed to describe the importance sampling estimate for the gradient in Section 3, followed by practical optimization algorithms outlined in Section 4. Empirical experiments and illustrations are provided in Section 5.
BACKGROUND
==========
VARIATIONAL APPROXIMATION
-------------------------
Variational inference refers to approximating the posterior distribution $p(z|x)$ of a probabilistic model $p(x,z)$ using a distribution $q_\lambda(z)$ parameterized by $\lambda$. Usually this is achieved by maximizing a lower bound $\mathcal{L}(\lambda)$ for the evidence (also called the marginal likelihood) $p(x)$: $$\mathcal{L}(\lambda) = \int q_\lambda(z) \log \frac{p(x, z)} {q_\lambda(z)}\ \delta z \le p(x).
\label{eq:ELBO}$$ Traditionally, the problem has been made tractable by assuming a factorized mean-field approximation and models with conjugate priors, resulting in closed-form coordinate ascent algorithms specific for individual models – for a full derivation and examples, see, [*e.g.*]{}, @blei16.
In recent years several novel types of algorithms applicable for a wider range of models have been proposed [@dsvi; @vae; @salimans; @bbvi], based on direct gradient-based optimization of the lower bound . The core idea behind these algorithms is in using Monte Carlo estimates for the loss and its gradient $$\nabla_{\lambda} \mathcal{L}(\lambda) =
\nabla_\lambda \mathbb{E}_{q_\lambda(z)}[\log p(x,z) - \log q_\lambda(z)].
\label{eq:ELBOgrad}$$ Given such estimates, the inference problem can be solved by standard gradient descent algorithms. This enables inference for non-conjugate likelihoods and for complex models for which closed-form updates would be hard to derive, making variational inference a feasible inference strategy for probabilistic programming languages [@advi; @edward; @pymc3]. In the following, we briefly describe two alternative strategies of estimating the gradient. In Section \[sec:method\] we will then show how the proposed importance sampling technique is applied for both cases.
### REPARAMETERIZATION ESTIMATE
The reparameterization estimate for (and consequently ) is based on representing the approximation $q_\lambda(z)$ using a differentiable transformation $z = f(\epsilon,
\lambda)$ of an underlying standard distribution $\phi(\epsilon)$ that does not have any free parameters. The core idea of how this enables computing the gradient was developed simultaneously by @vae [@salimans] and @dsvi with many of the mathematical details visible already in the early work by @opper. Plugging the transformation $f(\cdot)$ into gives $$\mathcal{L}(\lambda) = \int \phi(\epsilon) \log \frac{p(x,
f(\epsilon,\lambda))|\operatorname*{det}_{J_f}(\epsilon,
\lambda)|}{\phi(\epsilon)}\ \delta \epsilon,$$ where the integral is over the standard distribution that does not depend on $z$, and $|\operatorname*{det}_{J_f}(\epsilon, \lambda)|$ is the absolute value of the determinant of the Jacobian of $f(\epsilon, \lambda)$. Consequently, it can be replaced by a stochastic approximation $$\mathcal{L}(\lambda) = \frac{1}{M} \sum_{m=1}^M \log \frac{p(x,
f(\epsilon_m,\lambda))|\operatorname*{det}_{J_f}(\epsilon_m,
\lambda)|}{\phi(\epsilon_m)},$$ where $\epsilon_m$ is drawn from $\phi(\epsilon)$. We can now easily compute the gradients using the chain rule, by first differentiating $\log p(x, f(\epsilon, \lambda))$ w.r.t $z$ and then $z = f(\epsilon,
\lambda)$ w.r.t $\lambda$, resulting in
\_ () \_[m=1]{}\^M . \[eq:MCELBO\]
The combination of the standard distribution $\phi(\epsilon)$ and the transformation $f(\epsilon, \lambda)$ defines the approximation family. For example, $\phi(\epsilon) = \mathcal{N}(0, I)$ and $f(\epsilon, \lambda=\{\mu, L\}) = \mu + L \epsilon$ defines arbitrary Gaussian approximations [@opper; @dsvi], where $L$ is the Cholesky factor of the covariance. To create richer approximations we can concatenate multiple transformations; @advi combines the transformation above with a rich family of univariate transformations designed for different kinds of parameter constraints. For example, by using $f(\epsilon, \lambda=\{\mu, L\}) = \operatorname*{Softplus}(\mu + L \epsilon)$ we can approximate parameters constrained for positive values. Alternatively, we can directly reparameterize other common distributions such as Gamma or Dirichlet – see @naesseth [@grep] for details.
### SCORE FUNCTION ESTIMATE
An alternative estimate for can be derived based on manipulation of the log-derivatives; the use of the estimate for simulation of models was originally presented by @score and its use for variational approximation by @bbvi. The estimate for $\nabla_\lambda \mathcal{L}(\lambda) $ is provided by $$\mathbb{E}_{q_\lambda(z)}[(\log p(x,z) - \log q_\lambda(z)) \nabla_\lambda \log q_\lambda(z)],$$ which again leads into a straightforward Monte Carlo approximation of $\nabla_\lambda \mathcal{L}(\lambda)$ as $$\frac 1M \sum_{m = 1}^M(\log p(x,z_m) - \log q_\lambda(z_m)) \nabla_\lambda \log q_\lambda(z_m),
\label{eq:score}$$ where $z_m \sim {q_\lambda(z)}$. The notable property of this technique is that it does not require derivatives of the model (that is, $\log p(x, z)$) itself, but instead relies solely on derivatives of the approximation. This is both a pro and a con; the model does not need to be differentiable, but at the same time the estimate is not using the valuable information the model gradient provides. This is shown to result in considerably higher variance compared to the reparameterization estimate, often by orders of magnitude. Variance reduction techniques [@bbvi] help, but for differentiable models the reparameterization technique is typically considerably more efficient [@naesseth; @grep].
STOCHASTIC GRADIENT OPTIMIZATION
--------------------------------
Given estimates for the gradient computed with either method, the optimization problem[^1] is solved by standard gradient descent techniques. In practice all of the automatic variational inference papers have resorted to stochastic gradient descent (SGD) on mini-batches, adaptively tuning the step lengths with the state-of-the-art techniques.
In recent years several more advanced stochastic optimization algorithms have been proposed, such as stochastic average gradients (SAG) [@sage], stochastic variance reduced gradients (SVRG) [@svrg], and SAGA that combines elements of both [@saga]. However, to our knowledge these techniques have not been successfully adapted for automatic variational inference. In Section \[sec:isag\] we will present a new variant of SAG that works also when the gradients are estimated as Monte Carlo approximations, and therefore briefly describe below the basic idea behind stochastic average gradients.
SAG performs gradient updates based on an estimate for the full batch gradient, obtained by summing up gradients stored for individual data points (or for mini-batches to save memory). Whenever a data point is seen again during the optimization the stored gradient for that point is replaced by the gradient evaluated at the current parameter values. The full gradient estimate hence consists of individual gradients estimated for different parameter values; the most recently computed gradients are accurate but the ones computed long time ago may correspond to vastly different parameter values. This introduces bias [@saga], but especially towards the convergence the variance of the estimated full batch gradient is considerably smaller than that of the latest mini-batch, speeding up convergence.
METHOD {#sec:method}
======
All gradient-based optimization algorithms follow the same basic pattern of computing a gradient for a mini-batch of samples and updating the parameters. The computational effort required goes almost solely into evaluating the gradient of the loss. To speed up the optimization, we next present a technique that allows computationally lighter evaluation of the gradient in scenarios where the gradient is computed using a Monte Carlo approximation. The presentation here is based on the reparameterization estimate that benefits more off this treatment – as will become evident later – but for completeness we discuss also the score function estimate in Section \[sec:score\].
The Monte Carlo approximation for estimating the gradient depends on the data $x$ and a set of $M$ parameters $z_m$ drawn from the current approximation. As highlighted in , the actual computation factorizes into $\nabla_z p(x, z_m)$ that depends only on $x$ and $z_m$ and into $\nabla_\lambda f(\epsilon, \lambda)$ and $\nabla_{\lambda}|\text{det}_{J_f}(\epsilon_m, \lambda)|$ that depend only on $\epsilon_m$ and $\lambda$. The former part is typically considerably more computationally expensive. The observation that the slower part does not directly depend on the parameters hints that it should be possible to avoid re-computing the term even if the approximation changes, and this indeed is the case as explained next.
Assume we have already estimated the gradient at some parameters $\lambda$, implying that we have also evaluated $\nabla_z p(x, z_m)$ for some set of $z_m$. The question now is how to estimate the gradient at parameters $\lambda' = \lambda + \delta$ that are (typically only slightly) different. It turns out this can be done using the well-known concept of *importance sampling* originally designed for approximating expectations when we cannot directly draw samples from the density of interest. In our case, however, we *could* draw samples directly from the new approximation, but *choose not to* since the gradient can be estimated also using the old approximation as a proposal distribution. That is, we are using importance sampling for an unusual reason but can still use all the standard tools.
Typically, we use importance sampling to find the expectation of a function $f(z)$ over the target distribution $p(z)$ if we are able to draw samples only from a proposal distribution $q(z)$. The expectation of $f(z)$ over $p(z)$ can then be approximated by The quantities $w_m = {p(z_m)}/ {q(z_m)}$ are the *importance weights* that correct the bias introduced by sampling from the wrong distribution $q(z)$ [@bishop]. The weights $w_m$ are non-negative and tend to zero when $p(z)$ is completely mismatched to $q(z)$, and $w_m > 1$ when the sample $z_m$ is more likely under the $p(z)$. The estimate above is unbiased, but has high – potentially infinite – variance when $q(z)$ and $p(z)$ are dissimilar. Next we show how importance sampling can be used for evaluating the reparameterization gradient .
To save computation we want to re-use the model gradients $\nabla_z
\log p(x, z_m)$ already available for certain values of $z_m$ and hence need to consider estimates that keep these values fixed. This means we need to find the $\epsilon_m'$ under the new approximation $q_{\lambda'}(z)$ that correspond to these values, by computing $\epsilon_m' = f^{-1}(z_m, \lambda')$. Given these values we can evaluate the necessary quantities to compute both the importance weights and the other terms ($\nabla_{\lambda} f(\epsilon_{m}',
\lambda')$ and $\nabla_{\lambda}|\text{det}_{J_f}(\epsilon_m,
\lambda)|$) required for evaluating the gradient itself.
The resulting importance sampling estimate for is
\^i\_ () \_[m=1]{}\^M w\_m, \[eq:impgrad\]
where the $i$ in $\nabla^i_{\lambda} \mathcal{L}(\lambda) $ refers to the importance-sampled estimate of the gradient. The computationally expensive part of the gradient $\nabla_{z} \log p(x, z_m)$ is already available and need not be computed. The rest of the terms are efficient to evaluate, and hence the whole gradient estimate is obtained in a fraction of a time compared to computing it from scratch. The importance weights are provided by $$w_m =
\frac{\phi(\epsilon_m')}{\phi(\epsilon_m)} \label{eq:weight}$$ and hence only require evaluating densities of the standard distribution underlying the approximation. The above description is summarized in Algorithm \[alg:impgrad\] and illustrated graphically in Figure \[fig:normal\].
It is also worth noting that if $f(\cdot)$ is constructed as a series of transformations, for example as element-wise transformations of reparameterized normal distribution as done by [@advi], then parts of the term $\nabla_{\lambda}|\text{det}_{J_f}(\epsilon'_m,
\lambda')|$ can (and naturally should) also be re-used. This is, however, of secondary importance compared to the computational saving of re-using $\nabla_z \log p(x, z_m)$.
A practical challenge with importance sampling is that for high-dimensional densities the weights $w_m$ easily tend to zero. Variational approximation is, however, often conducted for approximations that factorize over the parameters of the approximation as $q(\lambda) = \prod_{s} q(\lambda_{S_s})$, where $\{S_s\}$ is a partitioning of the parameter vector. The importance sampling estimate can – and should – be done for each factor separately, since the gradient $\nabla_z \log p(x, z_m)$ can be computed for each approximation factor independently even if $\log p(x, z_m)$ itself does not factorize. We show in Section \[sec:dimensionality\] that the technique helps at least until factors consisting of roughly ten parameters.
$\epsilon_m' \leftarrow
f^{-1}(z_m, \lambda')$ $w_m \leftarrow
\frac{\phi(\epsilon'_m)}{\phi(\epsilon_m)}$ Calculate $\nabla^i_{\lambda} \mathcal{L}(\lambda)$ using
REPARAMETERIZATION GRADIENT EXAMPLE
-----------------------------------
To further clarify the derivation above, we next illustrate the procedure for the common scenario of Gaussian reparameterization combining $\phi(\epsilon) = \mathcal{N}(0, I)$ with $z = \mu + L
\epsilon$, denoting $\lambda = \{\mu, L\}$. Given a set of $z_m$ drawn from $q(\lambda)$ we want to estimate the gradient evaluated at $\lambda'$ using .
First we compute $\epsilon_m' = f^{-1}(z_m, \lambda') = L^{-1}(z_m -
\mu)$ and evaluate $\phi(\epsilon_m')$ under the standard normal distribution. The weight $w_m$ can then readily be evaluated using , computing $\phi(\epsilon_m)$ as well if it has not already been evaluated because of being used for another importance sampling estimate.
To compute the gradient we need to re-compute $\nabla_{\lambda} f(\epsilon_m', \lambda')$ and $\nabla_{\lambda}|\text{det}_{J_f}(\epsilon'_m, \lambda')|$. For $\mu$ these terms are simply identity and zero, whereas for $L$ we get $z_m$ and $\nabla_{L} \log |L'| \equiv \Delta_{L'}$. The exact form of $\Delta_{L'}$ depends on the assumptions made for $L$; see @dsvi for details. The final importance sampled estimate for the gradient is then $$\begin{aligned}
\nabla_{\mu}^{i} \mathcal{L}(\mu', L') &\approx
\frac{1}{M} \sum_{m=1}^M w_m \nabla_{z} \log p(x, z_m)\\
\nabla_{L}^{i} \mathcal{L}(\mu', L') &\approx \frac{1}{M}
\sum_{m=1}^M w_m \left [ \nabla_{z} \log p(x, z_m)
\epsilon'_m + \Delta_{L'} \right ]. \notag\end{aligned}$$
An important observation here is that the importance sampling procedure does not merely re-weight the terms $\nabla_z \log p(x,
z_m)$, but in addition the transformation that converts them into the $\lambda$ space changes because of the new values of $\epsilon_m'$. These values depend on the approximation parameters in a non-linear fashion and hence the gradient itself is a non-linear transformation of the gradient evaluated at $\lambda$ (for a graphical illustration, see Figure \[fig:normalcontour\]). This is crucially important for development of the practical optimization algorithms in Section \[sec:algorithms\]; if the transformation was linear then the importance sampling estimate would not necessarily provide improvement over careful adaptation of element-wise learning rates.
SCORE FUNCTION ESTIMATE {#sec:score}
-----------------------
Above we discussed the importance sampling estimate from the perspective of the reparameterization estimate. For completeness we also show how it can be applied for the score function estimate, and discuss why it is less useful there.
The Monte Carlo approximation for the score function is obtained directly by drawing samples from the approximation $q(\lambda)$. To approximate the gradient for $q(\lambda')$ we merely use the standard importance sampling equation to obtain the approximation $$\frac{1}{M} \sum_{m=1}^M w_m \left[\log p(x, z_m) - \log
q_\lambda(z_m)\right] \nabla_{\lambda} \log q_\lambda(z_m),$$ where $w_m = \frac{q(z|\lambda')}{q(z|\lambda)}$. This is still an unbiased estimate, but the computational saving is typically smaller than in the reparameterization case. We do not need to evaluate $\log
p(x, z_m)$ since the samples $z_m$ are kept constant, but all other terms need to be computed again and evaluating the gradient of the approximation is not cheap. This estimate is only useful when the evaluation of the log probability utterly dominates the total computation.
ALGORITHMS {#sec:algorithms}
==========
In the following we describe example optimization algorithms based on the importance sampling idea. The details are provided for a straightforward variant of SGD and for a generalization of stochastic average gradients, but other related algorithms could be instantiated as well.
IMPORTANCE SAMPLED SGD
----------------------
Stochastic gradient descent estimates the gradient based on a mini-batch and then takes a step along the gradient direction, typically using adaptive learning rates such as those by @adam [@adagrad].
The importance sampled SGD (I-SGD; Algorithm \[alg:isgd\]) follows otherwise the same pattern, but for each mini-batch we conduct several gradient steps instead of just one. For the first one we evaluate the gradient directly using . After updating the approximation we apply Algorithm \[alg:impgrad\] to obtain an importance-sampled estimate for the gradient evaluated at the new parameter values, and proceed to take another gradient step using that estimate. For each step we use a proper estimate for the mini-batch gradient that can, after the first evaluation, be computed in a fraction of a time. After taking a few steps we then proceed to analyze a new mini-batch, again needing to compute the gradient from scratch since now $x$ has changed.
After passing through the whole data we have evaluated $\log p(x, z)$ and its gradient once for every data point, just as in standard SGD. However, we have taken considerably more gradient steps, possibly by a factor of ten. Alternatively, we can think of it as performing more updates given a constant number of model gradient evaluations.
A practical detail concerns the choice of how many steps to take for each mini-batch. This choice is governed by two aspects. On one hand we should not use the importance sampled estimate if the approximation has changed too much since computing the $\nabla_z \log p(x, z_m)$ terms, recognized typically as $w_m$ tending to zero. On the other hand, we should not take too many steps even if the approximation does not change dramatically, since the gradient is still based on just a single mini-batch.
The empirical experiments in this paper are run with a simple heuristic that randomly determines whether to take another step with the current mini-batch or to proceed to the next one. This introduces a single tuning parameter $t$ that controls the expected number of steps per mini-batch. The algorithm is robust for this choice; we obtain practical speedups with values ranging from $t = 0.5$ to $t=0.9$. Finally, importance-sampling could in principle result in very large gradients if $w_m \gg 1$ for some $m$; we never encountered this in practice, but a safe choice is to proceed to the next batch if that happens.
For a practical illustration of the algorithm, see Figure \[fig:normalcontour\] that approximates the posterior over the mean and precision of a normal model. Here $t=0.9$ and hence we take on average $9$ importance-sampled gradient steps for each mini-batch. The I-SGD algorithm reaches the optimum in roughly as many steps as conventional SGD but achieves it almost ten times faster.
IMPORTANCE SAMPLED SAG {#sec:isag}
----------------------
Stochastic average gradients [@sage] stores the batch gradient and iteratively updates it for the samples in a given mini-batch. In the following we derive a variant of SAG (Algorithm \[alg:isag\]) that uses importance sampling to both re-weight and update the gradients for the historical mini-batches using , helping to detect and avoid using stale gradients whose parameter values have changed so much since computing them.
When visiting a new mini-batch we compute the gradient using . For all previously visited mini-batches we compute the importance weights and modify the gradient according to . The whole gradient is formed by summing up the terms for all mini-batches. It is important to note that the importance sampling changes the weight of the gradient, decreasing it towards zero for the mini-batches evaluated under clearly different parameter settings, and transforms the gradient to better match one that would have been calculated under the current approximation.
This algorithm provides a justified version of SAG for automatic variational inference. The computational cost is higher than for standard SAG since we need to evaluate the importance weights and compute the terms related to the gradient of the transformation for all past mini-batches. There is, however, no additional memory overhead and the amount of evaluations for the gradient of the model itself is the same. This overhead for importance sampling the gradients for other batches is not negligible, but usually still small enough that the resulting algorithm outperforms a naive implementation of SAG because of vastly more accurate gradient estimates, as shown in Section \[sec:isage\]. In case updating the past gradients becomes too costly, a simple remedy is to use only the latest $K$ mini-batches for some reasonable choice of $K$.
EXPERIMENTS
===========
In this section we first demonstrate how the behavior of importance sampling depends on the dimensionality of the approximation. We then empirically compared both I-SGD and I-SAG for real variational inference tasks on a range of alternative models and settings.
DIMENSIONALITY OF THE APPROXIMATION {#sec:dimensionality}
-----------------------------------
Figure \[fig:weight-decay\] studies importance weights of approximations factorized at different granularities on a 100-dimensional diagonal multivariate Gaussian. An important observation is that even if importance sampling itself fails for factors of high dimensionality, the I-SGD algorithm degrades gracefully. For low-dimensional factors, up to at least 5-10 dimensions, we can safely take 5-10 steps with each mini-batch while still having accurate gradient estimates. When the dimensionality of individual factors reaches around $25$ the weights tend to zero already after a single gradient step, but the algorithm does not break down. It merely needs to proceed immediately to the next step, reverting back to standard SGD.
[.28]{}
[.28]{}
[.28]{}
[.28]{}
[.28]{}
[.28]{}
IMPORTANCE SAMPLED SGD {#sec:impsamsgd}
----------------------
We show in Figure \[fig:experiments\] how I-SGD consistently bests SGD for a variety of different models, especially for the reparameterization estimate. For these experiments we use fully factorized mean-field approximation,
We first apply the reparameterization estimate for three probabilistic models: (a) a diagonal multivariate Gaussian with a standard normal prior on the mean and Gamma priors on the precisions, trained on $N =
50000$ data points of $D = 500$ dimensions; (b) a Bayesian linear regression model with standard normal priors on the weights and a Gamma prior on the precision with $N = 50000$ and $D = 500$; and (c) a Gaussian mixture model with the usual conjugate priors, $N = 10000$, $D = 2$ and $K = 25$ clusters. We used fully factorized variational approximations for all models, with $\operatorname*{Softplus}$ transforms for the positively constrained precision parameters and the stick breaking transform for the mixture weights of the last model. For all three choices the I-SGD algorithm with $t=0.9$ converges to the same optimal solution as SGD, but does so in roughly an order of magnitude faster. For fair comparison the size of the mini-batch, the initial learning rate were chosen for each method to work well for SGD, forcing I-SGD to use the same choices. For both algorithms, we used M = 1 sample to estimate the gradients and Adam [@adam] to adaptively control the learning rate during optimization.
We then compare I-SGD and SGD using the score function estimates on a Poisson likelihood with a single Gamma prior on the rate (d), and Bayesian linear regression models with large (e) and small (f) mini-batch. The variational approximations used were the same as the priors. For the sub-plots (d) and (e) evaluating the log-probability takes long compared to evaluating the gradient of the approximation because of a large mini-batch size, and hence I-SGD is faster. With a smaller mini-batch size (f) the advantage is lost because evaluating the gradient of the approximation starts to dominate. We used $M =
100$ samples and did not consider variance reduction techniques for simplicity.
IMPORTANCE SAMPLED SAG {#sec:isage}
----------------------
Figure \[fig:sag\] compares the I-SAG algorithm (Algorithm \[alg:isag\]) against naive implementation of SAG [@sage]. Both algorithms are initialized by passing once through the data with I-SGD, to provide the initial estimate for the full batch gradient.
While SAG eventually converges to the right solution, the progress is slow and erratic due to stale mini-batch gradients being accumulated into the full gradient. I-SAG fixes the issue by not only down-weighting gradients corresponding to mini-batches visited several updates ago, but also by transforming the gradients to match the current approximation. The additional computation required for adapting the gradients for other mini-batches results in a computational overhead of, here, roughly 30% per iteration, but the improved accuracy of the batch gradient estimate is more than enough to overcome this.
Stochastic running average (SRA) provides another baseline that down-weights older mini-batches exponentially. Similar to I-SAG, it avoids using mini-batches with badly outdated gradient estimates, by using a simple weighting scheme without transforming the gradients. It outperforms SAG, but converges more slowly than I-SAG. Hence, I-SAG is stable implementation of SAG for variational inference, outperforming the alternative of running averages often considered as a remedy for the issues of SAG.
DISCUSSION
==========
Automatic variational inference using automatically differentiated gradients has in recent years become a feasible technique for inference for a wide class of probabilistic models, extending the scope of variational approximations beyond simple conjugate models towards practical probabilistic inference engines. While standard computational platforms and advances in convex optimization are readily applicable for gradient-based variational inference, the need to use Monte Carlo approximation to estimate the gradients necessarily induces a computational overhead – with very few samples the gradients are noisy whereas the cost grows linearly as a function of the samples.
Our work addressed this central element, discussing ways to speed up the gradient-based inference of variational approximations. By highlighting how the gradient computation separates into two steps we derived an importance-sampling estimate for the gradient that often only needs to evaluate the computationally cheaper part to provide the estimate. Skipping the computationally costly evaluation of the gradient of the model itself as often as possible lead to a practical speedup that is independent of other improvements provided by more advanced optimization algorithms [@svrg; @saga]. Our method relies on the inverse transformation being unique and efficient to compute. This might not be the case for complex structured approximations or approximations parameterized by neural networks; we leave more efficient extensions for such cases as future work.
We demonstrated the core idea in creating a more efficient stochastic gradient descent algorithm for both reparameterization [@dsvi; @advi] and score function [@bbvi] estimates used for variational inference. In addition, we formulated a theoretically justified variant of stochastic average gradients [@sage] applicable for variational inference. The idea, however, extends well beyond these special cases. For example, the rejection sampling variational inference [@naesseth] can be readily combined with our importance sampling strategy and is expected to result in a similar speedup.
Our main focus was in practically applicable algorithms, with much of the theoretical analysis left for future work. Two particular directions are immediately apparent: (a) The decision of when to use importance sampling estimates and (b) the behavior for approximations that do not factorize into reasonably small factors. In this work we showed how simple randomized procedure for determining whether to re-compute the gradient for a new mini-batch results in practical and robust algorithm, but more theoretically justified decisions such as inspecting for example the variance of the importance sampling estimate could be considered. The proposed algorithms are efficient for approximating factors of dimensionality up to roughly ten; for factors of higher dimensionality the algorithms revert back to the standard variants since all gradients need to be computed from scratch for every iteration.
### Acknowledgements {#acknowledgements .unnumbered}
This work was financed by the Academy of Finland (decision number 266969) and by the *Scalable Probabilistic Analytics* project of Tekes, the Finnish funding agency for innovation.
[^1]: Here cast as minimization of negative evidence, to maintain consistent terminology with gradient descent literature
|
---
abstract: 'We report a determination of the antiproton magnetic moment, measured in a three-body system, independent of previous experiments. We present results from a systematic study of the hyperfine (HF) structure of antiprotonic helium where we have achieved a precision more than a factor of 10 better than our first measurement. A comparison between the experimental results and three-body quantum electrodynamic (QED) calculations leads to a new value for the antiproton magnetic moment $\mu^{{\ensuremath{\overline{\rm{p}}}}}_{s} \, = \, -2.7862 (83) \mu_{N}$, which agrees with the magnetic moment of the proton within 2.9 $\times$ 10$^{-3}$.'
address:
- 'Stefan Meyer Institute for Subatomic Physics, Austrian Academy of Sciences, Boltzmanngasse 3, A-1090 Vienna, Austria.'
- 'Department of Physics, University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan.'
- 'KFKI Research Institute for Particle and Nuclear Physics, H-1525 Budapest, PO Box 49, Hungary.'
- 'Max-Planck-Institut für Quantenoptik, Hans-Kopfermann-Strasse 1, D-85748 Garching, Germany.'
- 'Institute of Nuclear Research of the Hungarian Academy of Sciences, H-4001 Debrecen, PO Box 51, Hungary'
author:
- 'T. Pask'
- 'D. Barna'
- 'A. Dax'
- 'R. S. Hayano'
- 'M. Hori'
- 'D. Horváth'
- 'S. Friedreich'
- 'B. Juhász'
- 'O. Massiczek'
- 'N. Ono'
- 'A. Sótér'
- 'E. Widmann'
bibliography:
- 'ps205.bib'
- 'hbar.bib'
- 'EBW-new.bib'
- 'DH-new.bib'
- 'RSH-new.bib'
title: 'Antiproton magnetic moment determined from the HFS of ${\ensuremath{\overline{\rm{p}}\rm{He}^+}}$'
---
Antiprotonic helium ,Microwave spectroscopy ,Hyperfine structure ,CPT invariance 36.10.-k ,32.10.Fn ,33.40.+f
Antiprotonic helium (${\ensuremath{\overline{\rm{p}}\rm{He}^+}}$) is a three body system consisting of an antiproton, a helium nucleus and an electron (${\ensuremath{\overline{\rm{p}}}}$, He$^{++}$, e$^-$) [@Yamazaki:93; @Yamazaki:02; @Hayano:2007]. It is formed by stopping antiprotons in a helium medium. Because of its mass, the ${\ensuremath{\overline{\rm{p}}}}$ occupies states of principle quantum number $n$ $\sim$ 38 with the highest probability, while the e$^-$ remains in the ground state. The vast proportion of newly formed ${\ensuremath{\overline{\rm{p}}\rm{He}^+}}$ atoms proceed rapidly (within nanoseconds) to an ionised state by Auger excitation of the electron. The antiproton then annihilates almost instantaneously (within picoseconds) with one of the nucleons in the helium nucleus due to the overlap of their wavefunctions. However, a small proportion ($\sim$ 3%) occupy circular states $n$ $\sim$ $l$, where $l$ is the total angular momentum quantum number. Auger decay is then supressed by the large ionisation energy ($\sim~25$ eV) and degeneracy is lifted due to the presence of the electron [@Condo:64]. These states become relatively long lived (metastable) because the only decay channel available to them is the radiative one, and they cascade from ($n$, $l$) $\rightarrow$ ($n-1$, $l-1$) with typical energy level spacings $\sim$ 2 eV and lifetimes of $\sim$ 1.5 $\mu$s.
Shortly after its discovery, in 1991 [@Iwasaki:91], laser spectroscopy measurements were performed on various levels of the ${\ensuremath{\overline{\rm{p}}\rm{He}^+}}$ cascade [@Morita:94; @Torii:99; @Hori:01; @Hori:03]. Compared with three-body Quantum Electrodynamic (QED) calculations these measurements were used to determine the antiproton-to-electron mass ratio ($m_{{\ensuremath{\overline{\rm{p}}}}}/m_{\rm{e}}$). Over the years, improvements to the experimental system have increased the precision to 2 ppb [@Hori:06], which is one of the best tests of CPT invariance in the baryon sector. In 1997 the hyperfine structure was first revealed by laser spectroscopy [@Widmann:97] and, in 2002, measured via a laser-microwave-laser technique [@Widmann:02].
This paper concludes a systematic study, commenced in 2006 [@Pask:2008], where the error has been reduced by more than a factor of 10 over the first measurement [@Widmann:02] and a new value for the antiproton spin magnetic moment $\mu^{{\ensuremath{\overline{\rm{p}}}}}_{s}$ has been determined by comparison with QED calculations.
![(a) Schematic view of the level splitting of ${\ensuremath{\overline{\rm{p}}\rm{He}^+}}$ for the ($n$, $l$) $\rightarrow$ ($n$ $-$ 1, $l$ + 1) electric dipole transitions. The laser transitions $f^{+}$ and $f^{-}$, from the parent to daughter states, are indicated by straight lines and the microwave transitions, between the quadruplets of the parent, by wavy ones. For this experiment $(n,L)$ = (37, 35) and $(n',L')$ = (38, 34). (b) Laser resonance profile demonstrating the two sharp peaks and HF laser splitting $\Delta f$ $\equiv$ $f^- \, - \, f^+$. Although there are four SHF laser transitions only the HF ones can be resolved in this experiment.[]{data-label="fig:HF_trans"}](HF_Trans.eps)
The *hyperfine (HF)* splitting [@Yamazaki:02] of ${\ensuremath{\overline{\rm{p}}\rm{He}^+}}$ arises from the coupling of the e$^-$ spin $\vec{S}_{e}$ with the ${\ensuremath{\overline{\rm{p}}}}$ orbital angular momentum $\vert \vec{L} \vert$ $\sim$ 35$\hbar$ and results in a doublet structure of the order $\nu_{\rm{HF}}$ = 10 - 15 GHz. The interaction between the antiproton spin $\vec{S}_{{\ensuremath{\overline{\rm{p}}}}}$ with $\vec{S}_{e}$ and $\vec{L}$ causes a *superhyperfine (SHF)* splitting of size $\nu_{\rm{SHF}}$ = 150 - 300 MHz. A schematic of the energy level structure is presented in Fig. \[fig:HF\_trans\]a.
The theoretical framework for the level splitting has been developed by two separate groups [@Bakalov:98; @Korobov:01; @Yamanaka:01; @Kino:03APAC] which all use the same Hamiltonian, first derived by Bakalov and Korobov [@Bakalov:98], but different variational methods to extract the energy eigen values.
The HF doublet is characterised by the quantum number $\vec{F}$ = $\vec{L}$ + $\vec{S}_{e}$ and the SHF quadruplet by $\vec{J}$ = $\vec{F}$ + $\vec{S}_{{\ensuremath{\overline{\rm{p}}}}}$ = $\vec{L}$ + $\vec{S}_{e}$ + $\vec{S}_{{\ensuremath{\overline{\rm{p}}}}}$. An electron spin flip transition can be induced by an oscillating magnetic field, resulting in two M1 transitions: $\nu^{+}_{\rm{HF}}: \, J^{++} \, = F^{+} \, + \, \frac{1}{2} \, \longleftrightarrow \, J^{-+} \, = F^{-} \, + \, \frac{1}{2}$ and $\nu^{-}_{\rm{HF}}: \, J^{+-} \, = F^{+} \, - \, \frac{1}{2} \, \longleftrightarrow \, J^{--} \, = F^{-} \, - \, \frac{1}{2}$
Bakalov and Widmann [@Bakalov:07] indicate the sensitivity of certain states on $\mu^{{\ensuremath{\overline{\rm{p}}}}}_{s}$ and the precision required to improve its value over the most precise measurement [@Kreissl:88]. Some of these states are not practical due to limitations in laser capability. Others, like the ($n$, $L$) = (39, 35) state, which are within laser capabilities, have a HF laser splitting $\Delta f$ = 0.5 GHz, of the same order as the Doppler broadening $\Delta f_{D}$ = 420 MHz. The previously measured (37, 35) state remains the best candidate for a precision study because there is an easily stimulated laser transition between the ($n$, $L$) = (37, 35) and (38, 34) states with $\Delta f$ = 1.75 GHz. It is therefore possible to individually resolve the $F^{\pm}$ states, demonstrated in Fig. \[fig:HF\_trans\]b, since $\Delta f_{D}$ = 320 MHz. The parent state is also relatively highly populated, containing some 0.3% of the antiprotons stopped in the target [@Hori:02].
The laser spectroscopy experiments employ a technique by which the annihilation decay products are detected [@Yamazaki:02]. A sharp prompt peak is first observed, where the majority of states annihilate within picoseconds of formation, then an exponential tail, where the metastable states cascade more slowly towards the nucleus. This tail constitutes the background.
A narrow band pulsed laser is scanned over the region of an expected transition between a radiative decay dominated parent state and an Auger decay dominated daughter. Because the daughter state is relatively short lived ($\sim$ 10 ns), resonance is indicated by a sharp peak against the background at the time of the pulse. The ratio of the peak area to this background (peak-to-total) indicates the size of the population transferred.
The HF splitting measurement method is illustrated in Fig. \[fig:subfig:timing\] where the first laser pulse remained fixed to the $f^+$ transition between the (37, 35) and (38, 34) states, creating a population asymmetry. A microwave pulse, if on resonance with either $\nu^{+}_{\rm{HF}}$ or $\nu^{-}_{\rm{HF}}$, then transferred the population from $J^{-+}$ or $J^{--}$ to refill the $F^+$ state. A second laser pulse was tuned to the same $f^+$ transition and fired with a delay $T$ = 200 - 500 ns from the first, which measured the population transfer. Plotting the peak-to-total of the second laser induced annihilation peak as a function of microwave frequency yields the two HF transitions as distinct peaks [@Pask:2008].
![Microwave frequency profiles averaged from scans at a common pressure $p$ = 150 mbar and laser delays $T$ = 200 ns, 350 ns and 500 ns. The broadening is due to the Fourier transform of the rectangular microwave pulse of length $T$.[]{data-label="fig:All"}](All_scans_at_150mb.eps)
The experiment was carried out at CERN’s Antiproton Decelerator (AD) which provided a pulsed beam of (1 - 3) $\times$ $10^{7}$ antiprotons with length $\sim$ 200 ns and energy 5.3 MeV. Every 90 - 120 s, such a pulse was stopped in a helium gas target at a temperature of 6.1 K and pressures $p$ = 150 - 500 mbar (number density 1.7-6.2 $\times$ 10$^{20}$ cm$^{-3}$). The antiproton annihilation products passed through one of two Lucite plates either side of the target where their Cherenkov photons were detected by photomultipliers (PMT) [@CherHori:03] and the signal was displayed on a digital oscilloscope. The PMTs were gated off for the ${\ensuremath{\overline{\rm{p}}}}$ arrival so that only the 3% metastable tail was recorded.
$p$ (mbar) $T$ (ns) $P$ (W) ${\ensuremath{\overline{\rm{p}}}}$ shots $\nu^{+}_{\rm{HF}}$ (GHz) $\Gamma_+$ (MHz) $\nu^{-}_{\rm{HF}}$ (GHz) $\Gamma_-$ (MHz)
------------------ ---------- --------- ------------------------------------------ --------------------------- ------------------ --------------------------- ------------------
150 200 15 1070 12.896 45(12) 3.84(8) 12.924 30(14) 4.70(3)
150 350 5 1028 12.896 709(78) 2.76(8) 12.924 579(59) 2.21(4)
150 500 3 2236 12.896 688(48) 1.68(6) 12.924 470(41) 1.48(4)
250 [@Pask:2008] 350 5 2938 12.896 651(35) 2.24(2) 12.924 431(35) 2.41(3)
250 [@Pask:2008] 500 3 230 12.896 53(12) 2.55(7) 12.924 446(65) 1.65(5)
500 350 5 1844 12.896 525(80) 2.06(5) 12.924 446(99) 2.01(7)
The pulse-amplified continuous wave (CW) laser system [@Hori:06; @Hori:09p] was constructed by splitting a CW laser of wavelength $\lambda$ $\sim$ 726.1 nm into two seed beams. These were each pulse amplified by a NG:Yag laser and three Bethune dye cells, the second delayed by a time $T$ after the first. The pump beams were stretched so that the pulse lengths of the two lasers were 18 ns and 13 ns [@Pask:2008], slightly longer than the Auger lifetime of the daughter state, to achieve a high depopulation efficiency.
The microwave apparatus was similar to that described in Sakaguchi *et al.* [@Sakaguchi:04] and a schematic is displayed in Fig. \[fig:subfig:target\]. The microwave pulse was synthesised by a vector network analyser (Anritsu 37225B) referenced to a 10 MHz satellite signal (HP 58503B) and amplified by a travelling wave tube amplifier (TMD PTC6358). A waveguide carried the pulse to a custom made stainless steel cylindrical cavity, with central frequency $\sim$ 12.91 GHz, which provided the desired shape for the field (TM$_{110}$ mode) at the target. Steel meshes (92% transparency) covered both ends of the cavity so that antiprotions and the two laser beams could enter the target from opposite directions. The cavity was overcoupled to the waveguide to achive a broad frequency range $\sim$ 100 MHz. A mu-metal shell surrounded the target region to protect from external magnetic fields. Indeed the field measured in three dimensions within the target was $B$ $<$ 0.03 G.
Previously, different choke positions of a triple-stub-tuner were used to match the impedance of the waveguide to that of the cavity for a range of frequencies [@Sakaguchi:04]. This time, a constant microwave power $P$ was produced at the target by firing a predetermined signal strength down the unmatched waveguide. Most of the signal was reflected and dumped to a 50 $\mathrm{\Omega}$ terminator by a three-way circulator. This removed standing waves from the system and allowed the relatively small amount of power absorbed by the cavity to be controlled to within 1 dB over the frequency range. The power was monitored by an undercoupled pickup antenna situated opposite the waveguide.
Table \[tab:List\] shows a summary of all data measured for this experiment. The line shape was determined in Pask *et al.* [@Pask:2008] to be $$\label{eq:linewidth}
X(\omega) = \frac{|2b|^2}{|2b|^2+(\omega_0-\omega)^2}\sin^2 \bigg\{ \frac{1}{2} \Big[|2b|^2+(\omega_0-\omega)^2 \Big]^{\frac{1}{2}}T \bigg\}.$$ Data measured with the same $p$ and $T$ were fitted simultaneously with common parameters for height, width and central frequency. Two data sets were systematically examined: 1) Microwave power dependence, and 2) Pressure dependence
1\) The ac Stark effect shifts the E1 transitions by less than one part in $10^9$ [@Hori:06]. Its equivalent, the ac Zeeman shift of the M1 transitions is far weaker and therefore far too small to be resolved. A power dependence measurement was nevertheless examined for a complete understanding of the systematics. At a constant pressure $p$ = 150 mbar, resonance profiles were measured with various laser delays $T$ = 200 ns, 350 ns and 500 ns at microwave powers carefully chosen to achieve a $\pi$-pulse [@Pask:2008p], $P$ = 15 W, 5 W and 3 W, respectively. For illustrative purposes the average of these scans is shown in Fig. \[fig:All\] where the the dominating broadening effect was due to the Fourier transform of the rectangular microwave pulse of length $T$ [@Pask:2009p]. Figure \[fig:subfig:power\_points\] shows that no power dependent trend was observed, therefore all data measured at common target densities could be averaged.
2\) In 2006 Korenman predicted a collisional shift of $\Delta$ $\equiv$ d$\nu/$d$p$ = 0.3 kHz/mbar [@Korenman:2006] and, for the first time, a resolution has been achieved to examine this. High statistic microwave resonance profiles were measured at $p$ = 150 mbar and 500 mbar. Previous measurements had been made at $p$ = 250 mbar [@Pask:2008].
The results are displayed in Fig. \[fig:subfig:density\_points\] and clearly show that, if any such density dependence existed, the trend may have the same magnitude but opposite sign. Extrapolating the points to zero density can be performed with lines of average gradient $\Delta$ = $-$0.26 $\pm$ 0.2 kHz/mbar. Such a large error neither confirms nor precludes conclusively the existence of a shift. The two fits in Fig. \[fig:subfig:density\_points\] are shown with the same gradient because no density shift was observed for the difference between the transitions $\Delta \nu_{\rm{HF}}$ = $\nu^{-}_{\rm{HF}}$ $-$ $\nu^{+}_{\rm{HF}}$, see Fig. \[fig:Diff\], in accordance with predictions.
A more recent calculation [@Korenman:2009p], based on measurements of the collisional broadening [@Pask:2009p], predicts $\Delta$ = $-$0.048 kHz/mbar which is less than the experimental precision. As a negative slope contradicts the predictions, the average of each measurement was taken but the error used was that of the extrapolation to zero density from the fit.
![The difference for $\Delta \nu_{\rm{HF}}$ between experiment and theory [@Korobov:01; @Kino:03APAC] as a function of target pressure. The experimental values are shown as squares ($\blacksquare$). The point shown at $x$ = 0 and represented by a circle ([$\bullet$]{}) represents the average of the total data.[]{data-label="fig:Diff"}](Exp_The_All_Diff.eps)
Other systematic effects that influence the measurement include external magnetic fields, precision of the microwave frequency source, shot-to-shot microwave power fluctuations and variances in the laser position and fluence from day to day. However these effects have been determined to be far smaller than the shot-to-shot fluctuations of the antiproton beam. Data was measured over a long period to reduce these drift effects and variations in the ${\ensuremath{\overline{\rm{p}}}}$ intensity have been reduced by normalising the second laser induced annihilation peak with the first (proportional to the number of antiprotons captured). Despite these considerations the reduced chi-squared $\chi_{\rm{red}}$ of the fit was $\chi_{\rm{red}} \, \sim \, 3$. To adjust for this the error bars were inflated by $\sqrt{\chi_{\rm{red}}} \, \sim \, 1.7$.
Bakalov calculated that a broadening of $\nu^{\pm}_{\rm{HF}}$ due to an external magnetic field occurs at a rate of $\Gamma_{\pm}$ $\sim$ 5.6 MHz/G [@Bakalov:2009p]. The similarity between the Fourier transform of the microwave pulse and the spectral line widths [@Pask:2009p] confirms that the target region was well shielded during the experiment. Due to referencing to a 10 MHz GPS receiver, the precision of the frequency source is several orders of magnitude less than the resolution of this experiment. Thefore the statistical errors are much greater than the systematic.
$\nu^{+}_{\rm{HF}}$ (GHz) $\nu^{-}_{\rm{HF}}$ (GHz) $\Delta \nu_{\rm{HF}}$ (MHz)
----------------------- --------------------------- --------------------------- ------------------------------
This work 12.896 641(63) 12.924 461(63) 27.825(33)
2002 [@Widmann:02] 12.895 96(34) 12.924 67(29) 28.71(44)
Korobov [@Korobov:01] 12.896 3(13) 12.924 2(13) 27.896(33)
Kino [@Kino:03APAC] 12.896 0(13) 12.923 9(13) 27.889(33)
The individual transition frequencies have a negligible dependence on $\mu^{{\ensuremath{\overline{\rm{p}}}}}_{s}$. However $\Delta \nu_{\rm{HF}}$ is directly proportional to this value. The predicted density shift for $\Delta \nu_{\rm{HF}}$ is far smaller, $\Delta$ = 0.003 Hz/mbar [@Korenman:2009p] than the precision of this experiment. If this is the case, the total splitting can be calculated from the difference between each pair of transitions that were measured at common densities $\Delta \nu_{\rm{HF}}$ = $\sum^N_i (\nu^{-}_{\rm{HF}_i}$ $-$ $\nu^{+}_{\rm{HF}_i})/N$, rather than the difference between the sum of each transition measurement $\Delta \nu_{\rm{HF}}$ = $(\sum^N_i\nu^{-}_{\rm{HF}_i}$ $-$ $\sum^N_i\nu^{+}_{\rm{HF}_i})/N$, where $i$ is the index of a measurement and $N$ = 3 is the total number of density dependent measurements. Figure \[fig:Diff\] displays $\Delta \nu_{HF}$ as a function of target pressure compared to the two most recent theories.
Fitting a first order polynomial, results in a gradient almost half that of its associated error, $\Delta$ = 0.24 $\pm$ 0.37 kHz/mbar, so the above holds and the data can be averaged to obtain a final value of $\Delta \nu_{\rm{HF}}$. Table \[tab:Results\] presents the data for the recent and previous experiments compared to the two most up to date theories.
This work demonstrates the completion of a systematic experimental study on the HF splitting of the (37, 35) state of ${\ensuremath{\overline{\rm{p}}\rm{He}^+}}$. The experimental error of $\nu^{\pm}_{\rm{HF}}$ has been reduced by a factor 20 less than that of the theoretical calculations and, although $\Delta_{\rm{exp-th}}$ = 300 - 600 kHz, it is well within the estimated theoretical error 1.3 MHz. The experimental precision for $\Delta \nu_{\rm{HF}}$ has reached that of theory and has been improved by more than a factor of 10 over the first measurement [@Widmann:02]. There is a two sigma agreement between theory and experiment.
The sensitivity $S$ of $\Delta \nu_{\rm{HF}}$ on $\mu^{{\ensuremath{\overline{\rm{p}}}}}_{s}$ for the (37, 35) state is $S$ $\equiv$ d$E$/d$\mu^{{\ensuremath{\overline{\rm{p}}}}}_{s}$ = 10.1 MHz/$\mu_{N}$ [@Bakalov:07], where $\mu_{N}$ is the nuclear magneton. Thus the magnetic moment can be determined to be:
$$\mu^{{\ensuremath{\overline{\rm{p}}}}}_{s} \, = \, -2.7862 (83) \mu_{N},$$
where the uncertainty has been calculated by adding $\Delta_{\rm{exp-th}}$ with the errors of theory and experiment in quadrature, resulting in a one sigma error, slightly less than the value determined by Kreissl *et al.* [@Kreissl:88], while the deviation from the magnitude of the proton spin magnetic moment, $\mu^p_s \, = \, 2.792847351(28)$, is similar but opposite in sign.
The absolute values for the magnetic moments of the proton and antiproton are in agreement within
$$\frac{\mu^p_s \, - \mid \mu^{{\ensuremath{\overline{\rm{p}}}}}_{s} \mid}{\mu^p_s} \, = \, (2.4 \, \pm \, 2.9) \, \times \, 10^{-3}.$$
The limit of experimental precision has been reached for the (37, 35) state. The study of other states with larger $\Delta \nu_{\rm{HF}}$ could potentially increase the precision but the system cannot generally be improved due mainly to uncontrollable fluctuations of the AD beam. Preparations are underway to measure the HF splitting of ${\ensuremath{\overline{\rm{p}}}}^3 \rm{He}^+$ which, because of the additional helion spin, provides a more thorough test of the theory but yields no further information on the magnetic moment. The theorists are currently working on an $\alpha^6$ calculation but a significant change in $\Delta \nu_{\rm{HF}}$ is not expected.\
The authors would like to acknowledge V. Korobov (Joint Institute for Nuclear Research, Russia), D. Bakalov (INRNE, Bulgaria) for many helpful discussions. We thank two undergraduate students: P. Somkuti and K. Umlaub who contributed to this project. We are also grateful to the AD operators for providing the antiproton beam. This work was supported by Monbukagakusho (grant no. 15002005), by the Hungarian National Research Foundation (NK67974 and K72172), the EURYI Award of the European Science Foundation and the Deutsche Forschungsgemeinschaft (DFG), the Munich-Centre for Advanced Photonics (MAP) Cluster of DFG and the Austrian Federal Ministry of Science and Research.
|
---
abstract: 'It is necessary for a mobile robot to be able to efficiently plan a path from its starting, or current, location to a desired goal location. This is a trivial task when the environment is static. However, the operational environment of the robot is rarely static, and it often has many moving obstacles. The robot may encounter one, or many, of these unknown and unpredictable moving obstacles. The robot will need to decide how to proceed when one of these obstacles is obstructing it’s path. A method of dynamic replanning using RRT\* is presented. The robot will modify it’s current plan when an unknown random moving obstacle obstructs the path. Various experimental results show the effectiveness of the proposed method.'
author:
-
-
bibliography:
- 'Thesis\_bibliography.bib'
title: |
Dynamic Path Planning and Replanning\
for Mobile Robots using RRT\*
---
Introduction
============
Path planning has been one of the most researched problems in the area of robotics. The primary goal of any path planning algorithm is to provide a collision free path from a start state to an end state within the configuration space of the robot. Probabilistic planning algorithms, such as the Probabilistic Roadmap Method (PRM) [@Kavraki_96] and the Rapidly-exploring Random Tree (RRT) [@Lavalle98], provide a quick solution at the expense of optimality. Since its introduction the RRT algorithm has been one of the most popular probabilistic planning algorithms. The RRT is a fast, simple algorithm that incrementally generates a tree in the configuration space until the goal is found.
The RRT has a significant limitation in finding an asymptotically optimal path, and has been shown to never converge to an asymptotically optimal solution [@Karaman_2] [@Karaman_1]. There is extensive research on the subject of improving the performance of the RRT. Simple improvements such as the Bi-Directional RRT and the Rapidly-exploring Random Forest (RRF) improve the search coverage and speed at which a single-query solution is found. The Anytime RRT [@Ferguson_1] provides a significant improvement in cost-based planning. The RRT\* algorithm provides a significant improvement in the optimality of the RRT and has been shown to provide an asymptotically sub-optimal solution [@Karaman_2].
Since the introduction of the RRT\* algorithm, research has expanded to discover new ways to improve upon the algorithm. Research includes adding heuristics [@Perez_1] [@Gammell_1] or bounds [@Salzman_1] to the algorithm in order to maintain the convergence of the algorithm but reduce the execution time. Additional research attempts to guide the algorithm through intelligent sampling [@Islam_1], or guided sampling through an artificial potential field [@Qureshi_1].
In many scenarios the operational environment is rarely static. The path from a single query will often be obstructed during execution. For that reason the topic of replanning is very important to robotic path planning. It is not feasible to discard an entire search tree and start over. One method is to store waypoints and regrow trees called the ERRT [@Bruce_2002]. Another method (DRRT) is to place the root of the tree at the goal location, so that only a small number of branches may be lost or invalidated when replanning [@Ferguson_2]. The Multipartite RRT maintains a set of subtrees that may be pruned and reconnected, along with previous states to guide regrowth. It is essentially a combination of DRRT and ERRT [@Kuffner_1]. More recently the RRT\* algorithm has been incorporated into replanning. RRT^X^ is an algorithm that uses RRT\* to continuously update the path during execution [@Otte]. The RRT^X^ is able to compensate for instantaneous changes in the static environment which is outside the scope of this work.
The contribution of this paper is the method using the RRT\* algorithm for replanning in a dynamic environment with random, unpredictable moving obstacles. Also included is the comparison of RRT and RRT\* algorithms in a complex 2-D environment.
The remainder of this paper is organized as follows: Section II provides an overview of the RRT\* algorithm. Section III will present the replanning method using RRT\* in a dynamic environment. Section IV contains the results from all simulations. Section V presents the conclusions and future work.
Robot Path Planning using the RRT\*
===================================
The RRT\* algorithm provides a significant improvement in the quality of the paths discovered in the configuration space over it’s predecessor the RRT. The quality of the path is determined by the cost associated with moving from the start location to the end location. While RRT\* does produce higher quality paths, the algorithm does have a longer execution time. The longer execution time of RRT\* is due to the algorithm making many additional calls to the local planner in order to continuously improve the discovered paths. RRT\* operates in a very similar way as RRT. The algorithm builds a tree using random samples from the configuration space of the robot and connects new samples to the tree as they are discovered. There are two primary differences between RRT and RRT\*. The first difference is in the method that new edges are added to the tree. The second difference is an added step to change the tree in order to reduce path cost using the newly added vertex. Each of these differences contributes to the improvement of discovered paths over time and is reason RRT\* will converge to an asymptotically sub-optimal solution.
When a random vertex is added to the tree, the RRT will select the nearest neighbor as the parent for this new vertex and edge. RRT\* will select the best neighbor as the parent for the new vertex. While finding the nearest neighbor, RRT\* considers all the nodes within a neighborhood of the random sample. RRT\* will then examine the cost associated with connecting to each of these nodes. The node yielding the lowest cost to reach the random sample will be selected as the parent, and the vertex and edge are added accordingly.
[latex@errorgobble]{}
$T \leftarrow $InitializeTree() $T \leftarrow $InsertNode($\emptyset$,$q_{init}$,$T$)
The RRT\* algorithm begins in the same way as the RRT. However, when selecting the nearest neighbor the algorithm also selects the set of nodes, $Q_{near}$, in the tree that are in the neighborhood of the random sample $q_{rand}$. Line 6 of Algorithm 3 is the first major difference between RRT\* and the RRT. Instead of selecting the nearest neighbor to the random sample, the $ChooseParent()$ function will select the best parent from the neighborhood of nodes.
[latex@errorgobble]{}
$q_{min} \leftarrow q_{nearest}$ $c_{min} \leftarrow $Cost($q_{nearest}$) + c($q_{rand}$) $q_{min}$
Algorithm 4 describes the $ChooseParent()$ function. This function maintains the node with the lowest total cost for reaching $q_{rand}$. At line 1 of Algorithm 4 the nearest neighbor, $q_{nearest}$, is considered the minimum cost neighbor, or $q_{min}$. On line 2 the cost associated with reaching the new random sample $q_{rand}$ by using $q_{nearest}$ as the parent is stored as the current best cost, or $c_{min}$. The algorithm then searches the set of nodes in the neighborhood of $q_{rand}$. The $Steer()$ function on line 4 of Algorithm 4 will return a path from the nearby node, $q_{near}$ to $q_{rand}$. If this path is obstacle free and has a lower cost than the current minimum cost, then the nearby node becomes the best neighbor, $q_{min}$ and this cost becomes the best cost $c_{min}$ (lines 7-9 of Algorithm 4). When all nearby nodes have been examined the function returns the best neighbor. The new random node is inserted into the tree using $q_{min}$ as the parent. The next step is the second major difference between the RRT\* and the RRT algorithms. Line 8 of Algorithm 3 calls the $Rewire()$ function.
[latex@errorgobble]{}
$T$
The $Rewire()$ function, described in Algorithm 5, changes the tree structure based on the newly inserted node $q_{rand}$. This function again uses the nearby neighborhood of nodes, $Q_{near}$, as candidates for rewiring. The $Rewire()$ function uses the $Steer()$ function to get the path, except this time the path will start from the new node, $q_{rand}$ and go to the nearby node $q_{near}$. If this path is obstacle free and the total cost of this path is lower than the current cost to reach $q_{near}$ (line 3 of Algorithm 5). Then the new node $q_{rand}$ is a better parent than the current parent of $q_{near}$. The tree is then rewired to remove the edge to the current parent of $q_{near}$, and add an edge to make $q_{rand}$ the parent of $q_{near}$. This is done using the $ReConnect()$ function on line 4 of Algorithm 5.
The functions $ChooseParent()$ and $Rewire()$ change the structure of the search tree when compared to the RRT algorithm. The tree generated by the RRT has branches that move in all directions. The tree generated by the RRT\* algorithm rarely has branches that move back in the direction of the parent. The $ChooseParent()$ function ensures edges are created and always moving away from the start location. The $Rewire()$ function changes the internal structure of the tree to ensure internal vertices do not add unnecessary steps on any discovered path. The $ChooseParent()$ and $Rewire()$ functions guarantee the paths discovered will be asymptotically sub-optimal because these functions are always minimizing the costs to reach each node within the tree.
Dynamic Replanning
==================
Overview
--------
A real world environment is not static, and it is full of moving obstacles. These obstacles are often moving in unpredictable directions, which makes planning tasks to avoid them difficult. When a moving obstacle is known and is following a known trajectory, the configuration space can be modified to account for this trajectory. When the obstacle is unknown, the robot will need to be able to dynamically determine a course of action in order to avoid a collision. In this section a method of dynamic replanning is proposed in order to avoid a random obstacle when it is detected by the robot.
Simulation Environment
----------------------
For the following simulations the environment remains very similar. The robot is given a configuration space from which to build a tree using RRT\* and determine the best path to reach the goal configuration from the start configuration. In all the experiments below, the robot was allowed a tree of varying sizes to evaluate the performance with different node densities. During the simulation a few random moving obstacles are added to the environment, described in the next section. These obstacles represent a region of the configuration space that would be a collision if the robot were to enter.
### Path Execution
After the initial planning process, the robot begins to execute the optimal path found by the search tree. The robot traverses the optimal path by selecting the next node and required velocity vector to reach it, see lines 3 and 4 of Algorithm 6. This process is described in Algorithm 7 below. When the vertex is reached the robot changes the velocity vector to move toward the next node. This process continues until the robot reaches the goal node. If the robot encounters a random moving obstacle that is obstructing the path a replan event occurs.
[latex@errorgobble]{}
SetObsDestination($numObs$) SetObsVelocities($numObs$) SetRobotDestination() SetRobotVelocity()
Random Moving Obstacles
-----------------------
The random moving obstacles force the robot to dynamically plan around the obstacle using RRT\*. In order for the obstacles to move about the environment, a graph is created to provide the paths between the static obstacles, and the vertices are the intersections of these paths. Upon initialization of the simulation the obstacles are placed at random vertices. The vertices are chosen such that the robot will have a chance to move before encountering a random obstacle. When the simulation begins the moving obstacles choose a random adjacent vertex and begins moving toward that vertex, see lines 1 and 2 of Algorithm 6. When the vertex is reached a new random vertex is chosen and the obstacle moves in the new direction, line 6 of Algorithm 6.
### Random Obstacle Detection
Robots operating in a real world scenario will have sensors, such as a LIDAR, to detect both static and dynamic obstacles. Sensors are not included in this simulation. Instead a detection range is placed on the robot. The simulation controls whether or not a moving obstacle is within the detection range of the robot (lines 7 and 8 of Algorithm 7). If a moving obstacle is within range the $Steer()$ function is used, by the simulation, to determine if any static obstacles are blocking the robot’s line of sight to the moving obstacle.
[latex@errorgobble]{}
$robotLocation \leftarrow robotLocation + robotVelocity$
The obstacle must be observed for a minimum of two time steps in order to determine the direction that the obstacle is moving. Once the direction is observed the robot can determine if the moving obstacle is blocking the path or not, line 11 of Algorithm 7. If the robot decides that the path is blocked, the replanning event begins.
Path Replanning
---------------
Path replanning begins with the determination of whether or not the moving obstacle is blocking the path, described in the next section. Algorithm 8, below, lists all the steps executed during the replanning process. The next step is to find the location along the optimal path that is beyond the obstacle. Next, the tree generated by RRT\* is modified and expanded in order to find a path around the obstacle. Finally, the best path around the obstacle is chosen and the execution of this sub-path begins. Each of these steps is described in the following sub-sections.
[latex@errorgobble]{}
InvalidateNodes() GetReplanGoalLocation() SetReplanSamplingLimits() Rewire($T$, $Q_{all}$, $NULL$, $q_{robot}$) RRT\*($q_{robot}$) SetReplanPath()
### Path Obstruction
The method for determining if the moving obstacle is blocking the path is a series of trigonometric functions using a direction vector from the robot to the moving obstacle and a comparison between the robot velocity vector and the moving obstacle velocity vector. Since the configuration space is 2-Dimensional, the inverse tangent can be used to find the angles of the vectors. To obtain the direction vector to the moving obstacle the equation is: $$angle_{direction} = atan2((Y_{obs} - Y_{robot}), (X_{obs} - X_{robot})).$$ Where $ (X_{robot},Y_{robot})$ is the position of the robot and $ (X_{obs},Y_{obs})$ is the position of the obstacle. This will return an angle in degrees over the range $ (-180, 180)$. Similarly the angle of the robot’s velocity vector can be obtained using the following equation: $$angle_{V_{robot}} = atan2(V_{j}, V_{i}).$$ Where $V_{i}$ and $V_{j}$ are the X and Y components of the robot velocity vector. Using the angles from (1) and (2) the difference can be taken to see if they are similar. If the absolute value of the difference between the two angles is less than some threshold, then the robot is moving toward the moving obstacle. Note, the angle difference is normalized to be in the range$ (-180, 180)$ before the absolute value is taken. This is done for all angle comparisons: $$|angle_{direction} - angle_{V_{robot}}| < angle_{thresh}.$$ If the robot is moving in the direction of the random obstacle the velocity vectors are examined. Substituting the obstacle velocity into (2) above the angle of the obstacle velocity can be obtained. Next, the differences between the velocity vectors is found: $$angle_{V_{diff}} = |angle_{V_{robot}} - angle_{V_{obs}}| .$$ There are three possibilities from this point. If the angle difference between the velocity vectors is less than the angle threshold, then the robot and the obstacle are moving in a similar direction. Second, if the angle difference between the velocity vectors is greather than $ 180 - angle_{thresh}$, then the robot and the obstacle are approximately moving toward each other. Last, if the angle difference falls outside of these ranges the moving obstacle and the robot are moving in different directions.
For the first case: The robot will simply follow the obstacle, until the obstacle changes direction, or the path takes the robot away from the obstacle. The robot will then choose from one of the other conditions. For the second case: The robot quickly activates a replan event to get out of the way. The random obstacle may move out of the way on its own, but there is no way of predicting that will occur. Finally, if the robot and the obstacle are moving in different directions, the robot ignores the obstacle unless it gets too close. This third condition catches the event that the robot moves out from a corner and a random obstacle is detected very close by. This event is best summed up with the following example: When two people approach a hallway intersection they will run into each other if they continue on their current course, it is only when they see each other that they can adjust to avoid a collision.
### Select Replan Goal Location
The second step in the replan event, line 2 in Algorithm 8, is to find a location that will navigate the robot around the random obstacle. First, any nodes that are in collision with a random moving obstacle are invalidated, not deleted. The only exception is the goal location of the optimal path. After this step an assumption had to be made to simplify and speed up the rest of the replanning process. The assumption is that the robot is currently following the best path in order to reach the goal location and should return to this path after the moving obstacle is avoided. Using this assumption only the nodes along the optimal path are examined. Nodes that are farther from the robot than the random obstacle are candidate nodes. The node on the optimal path that is immediately following the node that is closest to the obstacle, without colliding, will be the replan goal location.
### Modify Search Tree
The third step is to modify the original search tree in order to find a way around the moving obstacle. First a node is added to the tree at the robot’s current location. Using the distance to the replan goal node as a metric, a sampling area is established, line 3 in Algorithm 8. Then, using $Rewire()$, every node within the sampling area is rewired such that the robot’s current location becomes the parent of that node. New nodes are then sampled within this area and added to the tree using RRT\*. Since there are already many nodes in the tree only a small number will need to be added. However, the number of nearest neighbors used during the $ChooseParent()$ function and the $Rewire()$ function is increased. This increase allows each new node to direct the existing tree toward the replan goal location.
### Sub-path Selection and Execution
When the search tree modification is complete the best path to the replan goal location is found. Path execution will begin again as it did at the beginning of the simulation. When the robot reaches the replan goal location, the execution of the original optimal path resumes. If the robot encounters another moving obstacle and determines the path is obstructed again. The replanning is repeated, however the replan goal location will always be a node on the original optimal path.
Results
=======
The Simulation Environment
--------------------------
The environment for all of the experiments is a complex 2-Dimensional environment that will also serve as the configuration space for the robot. The environment is complex due to the number of obstacles and several narrow passages. There are also several sub-optimal paths where the algorithm may get stuck. For each experiment the path cost is measured in Euclidean distance. The environment is also intended to mimic a potential real world situation where there would be streets or sidewalks and open areas such as parks and plazas. In the RRT\* results presented below, the algorithm is allowed a maximum of 5000 nodes. The optimal path length is 98.48 units.
RRT\* Results
-------------
The RRT\* evaluation was conducted in the following way: There is not a growth factor for extending the tree and the tree is goal oriented. The expecation as the tree grows will be long branches. These branches are often inefficient, the RRT\* $Rewire()$ function will remove these long inefficient branches as the algorithm executes in favor of shorter, lower cost branches. The algorithm also has a maximum number of nearest neighbors, or neighborhood size that is configurable to the algorithm. This implementation of RRT\* has a maximum number of nearest neighbors equal to 1% of the total number of nodes.
Fig. \[F.movietree2:c\] shows the result of the search tree using the RRT\* algorithm. The best path length found is 103.96 units.
Dynamic Replanning Results
--------------------------
The simulation results shown below demonstrate the robot’s ability to plan a path around the moving obstacle and reach the goal. Using RRT\* during the replanning step allows an efficient path to be found to avoid the moving obstacle and continue on the original optimal path. Since the obstacles move randomly, it is possible for the robot to execute the optimal path and never be obstucted by a moving obstacle. Only examples where the robot did encounter these obstacles are shown.
The first set of results use a search tree containing 2000 nodes. Upon completion of the search tree the moving obstacles are placed randomly within the configuration space, and the simulation begins. Fig. \[F.movietree1:a\] shows the search tree found by the robot upon reaching 2000 nodes.
When executing this path the robot encounters two random obstacles near the center of the configuration space. One obstacle moves across the path and obstructs the robot. The robot triggers a replanning event at this time. Fig. \[F.movietree1:b\] shows when the robot encountered the moving obstacles and replanned to avoid them. A second obstacle is nearby and can be seen by the robot and must be considered when replanning. Following the replanning steps in Algorithm 8, the robot will select a goal location, then modify the search tree to avoid the obstruction.
Fig. \[F.movietree1:c\] shows the full modified search tree. There is an obvious empty region of the configuration space where the moving obstacles are located. The abscence of branches within this area shows the robot has considered this space as obstacle space, rather than free space. Note, the nodes in this region are not removed, they are considered invalid. If the robot should need to replan again, and this area is free of any moving obstacles, these nodes would be available.
When the robot reaches the replanning goal location the original path can resume. Fig. \[F.movietree1:d\] shows the completed path, along with the final positions of the random moving obstacles. The magenta line shows the path followed by the robot. The sections in blue are unexecuted portions of the original path.
The second set of results is similar to the first, with one exception. The random moving obstacle initial positions were selected in order to increase the probability the robot would encounter one, or many, while executing the path. Fig. \[F.movietree2:a\] shows the starting locations of the moving obstacles. The robot was obstructed three times during the execution of the path and successfully planned around the moving obstacle each time. Fig. \[F.movietree2:b\] shows the final positions of moving obstacles and the path followed by the robot.
The final set of results is a simulation with a search tree containing 5000 nodes and 3 obstacles moving at random in the configuration space. The three moving obstacles in this simulation were placed similarly to those in the second simulation. In this simulation the robot encounters a moving obstacle very early in the execution of the path. The robot finds a path around the moving obstacle and completes the path. Fig. \[F.movietree2:d\] shows the final positions of the obstacles and the executed path.
Conclusion and Future Work
==========================
The replanning method presented performs well and is a good first step toward a more robust method of replanning when unknown randomly moving obstacles obstruct the robot’s path. Future work will research a floating replan goal location to minimize the total remaining cost to the query goal. Minimizing the modifications to the original search tree is another area of improvement. This method has only been implemented in a 2-Dimensional configuration space. The algorithm must be expanded and modified to operate in higher dimension configuration spaces. The simulations up to this point have been with small numbers of moving obstacles, further research is needed to determine how the algorithm performs when there are many random moving obstacles. Additional research is pursuing multi-robot systems. Efficient path planning and replanning will benefit: cooperative sensing systems [@La_2; @La_4], formation control systems[@La_1; @La_3; @La_5; @La_7; @La_ICRA10] , and target tracking and observation [@La_6].
![Replanning results from the first set[]{data-label="F.movietree1"}](movie1.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree1:a\]
![Replanning results from the first set[]{data-label="F.movietree1"}](movie1-snapshot-sm.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree1:b\]
![Replanning results from the first set[]{data-label="F.movietree1"}](movie1-replanTree-sm.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree1:c\]
![Replanning results from the first set[]{data-label="F.movietree1"}](movie1-final-sm.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree1:d\]
![A sample RRT\* search tree containing 5000 nodes and replanning results from the second and third sets. []{data-label="F.movietree2"}](movie2-snapshot-sm.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree2:a\]
![A sample RRT\* search tree containing 5000 nodes and replanning results from the second and third sets. []{data-label="F.movietree2"}](movie2-final-sm.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree2:b\]
![A sample RRT\* search tree containing 5000 nodes and replanning results from the second and third sets. []{data-label="F.movietree2"}](RRT-star5000-103p960-sm.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree2:c\]
![A sample RRT\* search tree containing 5000 nodes and replanning results from the second and third sets. []{data-label="F.movietree2"}](movie3-final-sm.png "fig:"){width="\columnwidth" height="5cm"} \[F.movietree2:d\]
|
---
address: 'Physics Department, Duke University, Durham, North Carolina 27708-0305'
author:
- 'K. M. O’Hara, M. E. Gehm, S. R. Granade, S. Bali, and J. E. Thomas'
title: 'Stable, Strongly Attractive, Two-State Mixture of Lithium Fermions in an Optical Trap'
---
\#1[[$\backslash$\#1]{}]{}
Trapped, ultracold atomic vapors offer exciting new opportunities for fundamental studies of an interacting Fermi gas in which the temperature, density and interaction strength can be independently controlled. Recently, a degenerate gas of fermionic $^{40}{\rm K}$ has been produced by using a two-state mixture to enable s-wave scattering and evaporation in a magnetic trap [@Jin1]. By removing one species, the properties of the noninteracting degenerate gas were measured, demonstrating that the momentum distribution and the total energy obey Fermi-Dirac statistics [@Jin1]. However, the properties of interacting two-state fermionic vapors have not been explored experimentally.
Theoretical treatments of an interacting Fermi gas have focused extensively on $^{6}{\rm
Li}$ [@BCS; @Stringari; @HighTC; @Hydrodynamic; @Parametric; @Cooperpair; @BruunBurnett; @ConMatRev]. Certain two-state $^{6}{\rm Li}$ mixtures are predicted to be strongly attractive, i.e., they have anomalously large and negative scattering lengths [@Elastic] arising from a near-zero energy resonance in the triplet state [@s-wave]. It has been predicted that these strongly attractive mixtures can undergo a transition to a superfluid state at a relatively high transition temperature [@BCS; @HighTC]. In addition, the two-state effective interaction potential is widely tunable in a magnetic field, permitting systematic studies of fundamental phenomena such as collective oscillations for both the normal and superfluid phases [@Stringari; @Hydrodynamic; @Parametric], as well as new tests of superconductivity theory [@HighTC].
Unfortunately, magnetically trappable mixtures in $^{6}{\rm Li}$ with large s-wave scattering lengths are not stable, since there are correspondingly large spin-exchange and dipolar decay rates [@BCS; @Cooperpair; @Elastic]. Hence, the methods employed to study degenerate $^{40}{\rm K}$ are not applicable. For this reason, we developed an ultrastable ${\rm CO}_{2}$ laser trap to confine a stable mixture of the two lowest $^{6}{\rm Li}$ hyperfine states [@O'Hara]. However, attaining a large and negative scattering length in this mixture requires high magnetic fields $B \geq 800$ G to exploit either a Feshbach resonance or the triplet scattering length [@Cooperpair; @Elastic].
In this Letter, we show that there exists another stable hyperfine state mixture in $^{6}{\rm Li}$ which has the following unique properties. First, we predict that the scattering length $a$ is large, negative, and widely tunable at [*low*]{} magnetic field $B$. By monitoring the rate of evaporation from the ${\rm CO}_{2}$ laser trap at a fixed well depth, we measure $|a|=540_{-100}^{+210}\,a_{0}$ at $B=8.3$ G. This result confirms for the first time that very large scattering lengths exist in $^{6}{\rm Li}$ mixtures. The predicted scattering length is $-490\,a_{0}$ at $B=8.3$ G, consistent with our observations, and is expected to increase to $-1615\,a_{0}$ as $B\rightarrow 0$. Second, we find that this system is stable against spin exchange collisions provided that $B\neq 0$. In addition, the dipolar decay rate is predicted to be very small [@Verhaar], consistent with our observations. Finally, in the experiments, a Raman $\pi$ pulse is employed to abruptly create an interacting mixture from a noninteracting one, a desirable feature for studies of many-body quantum dynamics [@Zoller].
Fig. \[fig:levels\] shows the hyperfine states for $^{6}{\rm Li}$ labelled $|1\rangle -|6\rangle$, in order of increasing energy in a magnetic field. At low field, the states $|1\rangle$ and $|2\rangle$ correspond to the $|F=1/2,\,m\rangle$ states, while states $|3\rangle$ through $|6\rangle$ correspond to states $|F=3/2,\,m\rangle$. At nonzero magnetic field, only the magnetic quantum number $m$ is conserved. The subject of this paper is the $|3\rangle - |1\rangle$ mixture.
Fig. \[fig:a31\] shows the scattering length $a_{31}$ for the $|3\rangle - |1\rangle$ mixture as a function of magnetic bias field $B$. We estimate $a_{31}(B)$ by using the asymptotic boundary condition (ABC) approximation [@Elastic]. This calculation incorporates the singlet and triplet scattering lengths [@s-wave], $a_{S}=45.5\pm 2.5\,a_{0}$ and $a_{T}=-2160\pm 250\,a_{0}$, and a boundary radius which we take to be $R=40\,a_{0}$ [@Elastic]. The scattering length varies from $-1620\,a_{0}$ ($\simeq 3a_{T}/4$ as $B\rightarrow 0$) to $-480\,a_{0}$ at $B=10$ G. The results of our approximate calculation for $B=0$ to $B=200$ G are confirmed within 10% by van Abeelen and Verhaar [@Verhaar] using a coupled channel calculation which includes the uncertainties in the potentials. At higher fields, near 800 G, we believe the scattering length exhibits a Feshbach resonance (not shown). Above this resonance, the scattering length approaches the triplet scattering length of $-2160\,a_{0}$.
The $|3\rangle - |1\rangle$ mixture is stable against spin-exchange collisions provided that a small bias magnetic field is applied. Spin-exchange inelastic collisions conserve the two-particle total magnetic quantum number $M_{T}$, where $M_{T}=-1$ for the $|\{3,1\}\rangle$ state. Note that $\{,\}$ denotes the antisymmetric two-particle spin state, as required for s-wave scattering which dominates at low temperatures. There are no lower-lying antisymmetric states with $M_{T}=-1$. Hence, exothermic collisions are precluded. The only other states with $M_{T}=-1$ are $|\{4,2\}\rangle$ and $|\{5,3\}\rangle$. Without an adequate bias magnetic field, transitions to these states lead to population in level $|4\rangle$. Then, exothermic $| \{3,4\}\rangle\rightarrow |\{3,2\}\rangle$ and $|\{4,1\}\rangle
\rightarrow |\{1,2\}\rangle$ collisions can occur. With an adequate bias magnetic field, the energy of states $|\{4,2\}\rangle$ and $|\{5,3\}\rangle$ can be increased relative to that of state $|\{3,1\}\rangle$ by more than the maximum relative kinetic energy, i.e., twice the well depth during evaporative cooling. By energy conservation, spin-exchange transfer is then suppressed. In this case, the inelastic rate is limited to magnetic dipole-dipole (dipolar) interactions which contain a rank 2 relative coordinate operator of even parity [@BCS]. Since parity is conserved and ${\rm
p-wave}\rightarrow {\rm p-wave}$ scattering is frozen out at low temperature, the dominant dipolar process is a small ${\rm
s}\rightarrow {\rm d}$ rate in which $|\{3,1\}\rangle\rightarrow |\{1,2\}\rangle$ [@Verhaar].
In the experiments, the ${\rm CO}_{2}$ laser trap is initially loaded from a magneto-optical trap (MOT)[@O'Hara]. At the end of the loading period, the MOT laser beams are tuned near resonance and the intensity is lowered to decrease the temperature. Then, optical pumping is used to empty the $F=3/2$ state to produce a 50-50 mixture of the $|1\rangle -|2\rangle$ states. These states are noninteracting at low magnetic field, i.e, the scattering amplitude vanishes as a result of an accidental cancellation [@Elastic]. With a ${\rm CO}_{2}$ laser trap depth of $330\,\mu$K, up to $4\times 10^{5}$ atoms are confined in the lowest-lying hyperfine states at an initial temperature between 100 and 200 $\mu$K. A bias magnetic field of 8.3 G is applied to split the two-particle energy states by $\simeq 16$ MHz. This is twice the maximum attainable energy at our largest well depth of $400\,\mu{\rm K}=8$ MHz. After a delay of 0.5 second relative to the loading phase, a pair of optical fields is pulsed on to induce a Raman $\pi$ pulse. This pulse transfers the population in state $|2\rangle$ to state $|3\rangle$ in two microseconds, initiating evaporative cooling in the resulting $|3\rangle
-|1\rangle$ mixture. The optical fields are detuned from resonance with the D2 transition by $\simeq 700$ MHz to suppress optical pumping. If the Raman pulse is not applied, the trapped atoms remain in the noninteracting $|1\rangle-|2\rangle$ mixture and exhibit a purely exponential decay with a time constant $\simeq 300$ seconds.
An acousto-optic modulator (A/O) in front of the ${\rm CO}_{2}$ laser controls the laser intensity, which is reduced to yield a shallow trap depth of $100\,\mu$K. By using a shallow well, we avoid the problem that the elastic cross section becomes independent of the scattering length at high energy, as described below. In addition, the shallow well greatly reduces the number of loaded atoms and makes the sample optically thin, simplifying calibration of the number of trapped atoms. To determine the trap parameters, the laser power is modulated and parametric resonances [@Hansch] are observed at drive frequencies of $2\nu$ for three different trap oscillation frequencies $\nu$: At $100\,\mu$K well depth, $\nu_{x}=2.4$ kHz $\nu_{y}=1.8$ kHz and $\nu_{z}=100$ Hz, where the trap laser beam propagates along ${\bf z}$. Using the measured total power as a constraint, we obtain the trap intensity $1/e^{2}$ radii, $w_{x}=50\,\mu$m and $w_{y}=67\,\mu$m, and the axial intensity $1/e^{2}$ length, $z_{f}\simeq 1.13$ mm, where $z_{f}$ is consistent with the expected Rayleigh length within 15%.
The number of atoms in the trap $N(t)$ is estimated using a calibrated photomultiplier. The detection system monitors the fluorescence induced by pulsed, retroreflected, $\sigma_{\pm}$ probe and repumper beams which are strongly saturating ($I/I_{sat}=26$ for the strongest transition). To simplify calibration, only the isotropic component of the fluorescence angular distribution is measured: The collecting lens is placed at the magic angle [@Magic] of $55^{\circ}$ ($P_{2}(\cos\theta )=0$) with respect to the propagation direction of the probe beams. The net efficiency of the detection system is determined using laser light of known power. The primary uncertainty in the calibration arises from the excited state population fraction, which we estimate lies between 1/4 and 1/2.
Fig. \[fig:number\] shows the number of trapped atoms $N(t)$ measured for the $|3\rangle -|1\rangle$ mixture at a well depth $U_{0}=100\,\mu$K and a bias field of 8.3 G as a function of time between 5 ms and 20 seconds after evaporation is initiated. For times beyond 50 seconds (not shown), the evaporation stagnates, and we observe an exponential decay of the cooled $|3\rangle -|1\rangle$ mixture with a time constant of 370 seconds over a period of a few hundred seconds. The error bars are the standard deviation of the mean of ten complete runs through the entire time sequence.
A model based on the s-wave Boltzmann equation [@Walraven] is used to predict $N(t)$ for comparison to the experiments. This equation is modified to include the density of states for a gaussian potential well [@O'Hara] and to include the energy dependence of the elastic cross section. Assuming a short range potential and a symmetric (s-wave) spatial state, the cross section takes the form $$\sigma (k)=\frac{8\pi a_{31}^{2}}{1+k^{2}a_{31}^{2}},
\label{eq:cross}$$ where $\hbar k$ is the relative momentum. For $k|a_{31}|<<1$, the cross section is maximized. When $k|a_{31}|>>1$, the cross section approaches the unitarity limit $8\pi /k^{2}$ which is independent of $a_{31}$. Note that $k|a_{31}|=1$ corresponds to a relative kinetic energy of $\epsilon =\hbar^{2}/(2\mu\,a_{31}^{2})$, where $\mu =M/2$ is the reduced mass. For $|a_{31}|=500\,a_{0}$, $\epsilon =115\,\mu$K.
For a two-state mixture of fermions, the effective cross section is reduced from that of Eq. \[eq:cross\] by a factor of 2 since pairs of colliding atoms are in an antisymmetric hyperfine state with a probability 1/2. This effective cross section is used in a Boltzmann collision integral for each state $i=1,3$. A decay term $-N_{i}(t)/\tau$ with $\tau =370$ sec is added to account for the measured trap lifetime. A detailed description of our coupled Boltzmann equation model will be published elsewhere.
The coupled s-wave Boltzmann equations for the two states are numerically integrated to determine $N(t)$ using the well parameters as fixed inputs. From the calibrated photomultiplier signal, assuming that 1/3 of the atoms are in the excited state, we obtain an initial total number $N_{0}=44,000$. For this case, the initial collision rate in Hz is estimated to be $1/(2\pi \tau_{c})\simeq N_{0}M\sigma_{0}\nu^{3}/(k_{B}T)$, where $\nu^{3}=\nu_{x}\nu_{y}\nu_{z}$, $\sigma_{0}=8\pi a_{31}^{2}$, and $M$ is the $^{6}{\rm Li}$ mass. Assuming $|a_{31}|=500\,a_{0}$, $\tau_{c} =30$ ms. Hence, for $t>0.3$ seconds, when on average 10 collisions have occurred, the sample should be thermalized as assumed in the theory.
The best fit to the data starting with 22,000 atoms in each state is shown as the solid curve in Fig. \[fig:number\]. The $\chi^{2}$ per degree of freedom for this fit is 1.4 and is found to be very sensitive to the initial temperature $T_{0}$ of the atoms in the optical trap. >From the fit, we find $T_{0}=46\mu$K, which is less than the well depth. We believe that this low temperature is a consequence of the MOT gradient magnet, which is turned off after the MOT laser beams. The effective well depth of the optical trap is therefore reduced until the gradient is fully off, allowing hotter atoms to escape before the Raman pulse is applied to create the $|3\rangle -|1\rangle$ mixture. The fit is most sensitive to data for $t>0.5$ second, where the thermal approximation is expected to be valid. From the fit, we obtain the scattering length $|a_{31}|=540\pm 25\, a_{0}$, which is within 10% of the predictions of Fig. \[fig:a31\]. The quoted error corresponds to a change of 1 in the total $\chi^{2}$.
We determine the systematic errors in $a_{31}$ due to the uncertainties in the calibration and in the population imbalance as follows. The data is fit for an initial number of atoms $N_{0}$ of 58,000 and 29,000, corresponding to an excited state fraction of 1/4 and 1/2. This yields $|a_{31}|=440\pm 20\,a_{0}$ and $|a_{31}|=750\pm 42\,a_{0}$, respectively. Note that for the larger scattering lengths, the cross section given by Eq. \[eq:cross\] approaches the unitarity limit and the error increases. We assume that the initial population imbalance for states $|3\rangle$ and $|1\rangle$ is comparable to that of states $|2\rangle$ and $|1\rangle$ in the optically pumped MOT. To estimate the latter population imbalance, we use state-selective Raman $\pi$ pulses to excite $|2\rangle\rightarrow |3\rangle$ or $|1\rangle\rightarrow |6\rangle$ transitions in the MOT. Probe-induced fluorescence signals from states $|3\rangle$ or $|6\rangle$ show that the initial $|1\rangle$ and $|2\rangle$ populations are equal within 10%. Note that residual population in state $|2\rangle$ is expected to be stable and weakly interacting, since we estimate $|a_{32}|<30\,a_{0}$ for $0\leq B\leq 50$ G using the ABC method, and $a_{12}\simeq 0$ [@Elastic]. Using the parameters for the fit shown in Fig. \[fig:number\], but changing the initial mixture from 50-50 to 60-40, we find a slight increase in the fitted scattering length from $540\,a_{0}$ to $563\,a_{0}$. Thus, the uncertainty in the calibration of the number of atoms produces the dominant uncertainty and $|a_{31}|=540^{+210}_{-100}\,a_{0}$.
To demonstrate that evaporative cooling is occuring, rather than just trap loss, we have also measured the final temperature of the mixture using release and recapture [@Chu] from the ${\rm CO}_{2}$ laser trap. We obtain $9.8\pm 1\,\mu$K, which is within 10 % of the final temperature of $8.7\,\mu$K predicted by the Boltzmann equation model. An excellent fit to the data is obtained for the final temperature, which describes a thermal distribution. However, the initial temperature is not so readily measured, as it is nonthermal before evaporation is initiated, and is rapidly changing during evaporation, unlike the final temperature, which stagnates.
Good fits to the evaporation data are obtained neglecting inelastic collisions, suggesting that the dipolar rate for the $|3\rangle -|1\rangle$ mixture is small, in contrast to the scattering length. A limit on the dipolar loss rate for the $|3\rangle -|1\rangle$ mixture can be estimated from the $\tau
=370$ second lifetime of the mixture after evaporation stagnates. For equal populations in both states, dipolar decay results in an initial loss rate $\dot{n}=-Gn^{2}/4$, where $G$ is the dipolar rate constant and $n$ is the total density. To obtain a high density, the trap is loaded at a well depth of $330\,\mu$K and the temperature of the atoms is reduced by evaporation to $T\simeq 30\pm 1\,\mu$K. The number of atoms remaining in each state after evaporation is estimated to be $N=6.5\pm 2.2\times
10^{4}$, where the uncertainty is in the calibration. We cannot rule out the possibility that one state is depleted on a long time scale, since we do not directly measure the individual state populations. However, we believe that, after evaporation stagnates in the deep well, a $|3\rangle-|1\rangle$ mixture remains, since subsequent reduction of the well depth yields final temperatures consistent with evaporative cooling. Note that the mixture ratio is not critical: An 80-20 mixture yields an initial loss rate $\dot{n}=-0.16\,Gn^{2}$, $\simeq 2/3$ that of a 50/50 mixture. For a fixed $330\,\mu$K trap depth, $\nu^{3}=2.6\pm
0.3\,{\rm kHz}^{3}$, and the phase space density for one state in the harmonic approximation is then $\rho_{ph}=N/(k_{B}T/h\nu
)^{3}=7\times 10^{-4}$. This corresponds to a maximum total density of $n=2\,\rho_{ph}/\lambda_{B}^{3}=6.4\times 10^{11}/{\rm
cm}^{3}$, where $\lambda_{B}\equiv h/\sqrt{2\pi M\,k_{B}T}$. Since the exponential decay time of the $|3\rangle -|1\rangle$ mixture is similar to that obtained in the noninteracting $|1\rangle
-|2\rangle$ mixture, we assume the loss is dominated by background gas collisions. Thus, we must have $Gn/4<<1/\tau$, which yields $G<<2\times 10^{-14}{\rm cm}^{3}/{\rm sec}$. This result is consistent with the value $G\simeq 2\times 10^{-15}\,{\rm cm}^{3}/{\rm sec}$ predicted for the dipolar rate constant at $30\,\mu$K by van Abeelen and Verhaar [@Verhaar].
Future experiments will employ continuous evaporation by slowly reducing the well depth [@Adams]. In this case, very large scattering lengths can be obtained at low temperatures and small well depths by using a reduced bias magnetic field $B$. By adiabatically recompressing the well, experiments can be carried out with the precooled atoms in a deep trap to obtain high density as well. In such experiments, the final low temperature limits the number of atoms in the high energy tail of the energy distribution, exponentially suppressing spin-exchange collisions for $B\neq 0$. For example, if a total of $3\times 10^{5}$ atoms were contained in our trap at a well depth of $400\,\mu$K, the Fermi temperature $T_{F}=7\,\mu$K and the Fermi density is $4\times 10^{13}/{\rm cm}^{3}$. At a temperature of $T=0.1\, T_{F}=0.7\,\mu$K, a bias field of $B=0.16$ G would split the two-particle hyperfine states by $k_{B}T_{F}+12\,k_{B}T$, suppressing the spin exchange rate by $\exp ( -12)$, and giving $a_{31}\simeq -1200\,a_{0}$. Alternatively, as shown in Fig. \[fig:a31\], large $a_{31}$ can be obtained at moderate $B\simeq 300$ G.
In conclusion, we have observed that an optically trapped $|3\rangle -|1\rangle$ mixture of $^{6}{\rm Li}$ atoms has a very large scattering length at low magnetic field. This mixture is stable against spin-exchange collisions provided that a small bias magnetic field is applied. The evaporation curves measured for this mixture are in good agreement with a model based on an s-wave Boltzmann equation which neglects inelastic processes. We have predicted that the scattering interactions are strongly attractive and widely tunable at low magnetic field. If the parameters described above for deep wells can be attained, the system will be close to the threshold for superfluidity [@BCS] and ideal for investigating frequency shifts and damping in collective oscillations [@Stringari; @Hydrodynamic]. Further, since s-wave interactions can be turned on and off in a few microseconds, this system is well suited for studies of many-body quantum dynamics.
This research has been supported by the Army Research Office and the National Science Foundation.
B. DeMarco, and D. S. Jin, Science [**285**]{}, 1703 (1999). H. T. C. Stoof, et al., Phys. Rev. Lett. [**76**]{}, 10 (1996); See also, M. Houbiers et al., Phys. Rev. A [**56**]{}, 4864 (1997). L. Vichi and S. Stringari, Phys. Rev. A [**60**]{}, 4734 (1999). R. Combescot, Phys. Rev Lett. [**83**]{}, 3766 (1999). G. M. Bruun, and C. W. Clark, Phys. Rev. Lett. [**83**]{}, 5415 (1999). , cond-mat/9906392. M. Houbiers and H. T. C. Stoof, Phys. Rev. A [**59**]{}, 1556 (1999). G. Bruun, et al., Eur. Phys. J. D [**7**]{}, 433 (1999). M. Houbiers and H. T. C. Stoof, cond-mat/9808171. M. Houbiers, et al., Phys. Rev. A [**57**]{}, R1497 (1998). E. R. I. Abraham, et al., Phys. Rev A [**55**]{}, R3299 (1997). K. M. O’Hara, et al., Phys. Rev. Lett. [**82**]{}, 4204 (1999). We are indebted to F. A. van Abeelen and B. J. Verhaar who calculated the inelastic $|\{3,1\}\rangle\rightarrow
|\{1,2\}\rangle$ dipolar rate and confirmed our calculations of the magnetic field dependence of $a_{31}$. P. Törmä and P. Zoller, Phys. Rev. Lett. [**85**]{}, 487 (2000). S. Friebel, et al., Phys. Rev. A [**57**]{}, R20 (1998). , ed. G. W. Drake, (AIP Press, New York, 1996), p. 176. O. J. Luiten, et al., Phys. Rev. A [**53**]{}, 381 (1996). S. Chu, et al., Phys. Rev. Lett. [**55**]{}, 48 (1985). C. S. Adams, et al., Phys. Rev. Lett. [**74**]{}, 3577 (1995).
|
---
abstract: 'Strong spin-orbit coupling (SOC) can result in ground states with non-trivial topological properties. The situation is even richer in magnetic systems where the magnetic ordering can potentially have strong influence over the electronic band structure. The class of AMnBi$_2$ (A = Sr, Ca) compounds are important in this context as they are known to host massive Dirac fermions with strongly anisotropic dispersion, which is believed to be due to the interplay between strong SOC and magnetic degrees of freedom. We report the optical conductivity of YbMnBi$_2$, a newly discovered member of this family and a proposed Weyl semi-metal (WSM) candidate with broken time reversal symmetry. Together with density functional theory (DFT) band structure calculations, we show that the complex conductivity can be interpreted as the sum of an intra-band Drude response and inter-band transitions. We argue that the canting of the magnetic moments that has been proposed to be essential for the realization of the WSM in an otherwise antiferromagnetically ordered system is not necessary to explain the optical conductivity. We believe our data is explained qualitatively by the uncanted magnetic structure with a small offset of the chemical potential from strict stochiometry. We find no definitive evidence of a bulk Weyl nodes. Instead we see signatures of a gapped Dirac dispersion, common in other members of AMnBi$_2$ family or compounds with similar 2D network of Bi atoms. We speculate that the evidence for a WSM seen in ARPES arises through a surface magnetic phase. Such an assumption reconciles all known experimental data.'
author:
- Dipanjan Chaudhuri
- Bing Cheng
- Alexander Yaresko
- 'Quinn D. Gibson'
- 'R. J. Cava'
- 'N. P. Armitage'
title: 'An optical investigation of the strong spin-orbital coupled magnetic semimetal YbMnBi$_2$'
---
Introduction
============
Correlated electron systems with strong SOC have been the subject of intensive research in recent years. The interplay of electronic correlations and SOC can result in emergent topological phases and has opened up a completely new direction in condensed matter physics. This interplay can be very different depending on the specifics of the electronic correlation. In weakly to moderately interacting electron systems, SOC can lead to non-trivial band topology as observed in conventional topological insulators [@moore2009topological], Dirac and Weyl semi-metals [@Balents; @turner2013beyond; @armitage2017weyl], axion insulators [@wan2012computational] and topological superconductors [@qi2011topological]. More recently, the effects of SOC on strongly correlated systems are being explored with the realization of new material systems with heavy 4d/5d transition metal compounds [@doi:10.1146/annurev-conmatphys-020911-125138]. The iridates deserve special mention in this category and have been instrumental in exploring much of this uncharted territory [@schaffer2015recent; @rau2015spin].
In addition to the emergence of topologically non-trivial ground states, the interplay between SOC and magnetic degrees of freedom themselves is also quite interesting. The family of AMnBi$_2$(A = Sr, Ca) compounds are particularly important in this context. Being structurally similar to iron based superconductors, they are referred to as manganese pnictides, which contain layers of Mn-Bi edge sharing tetrahedra and a Bi square net separated by a layer of A atoms [@guo2014coupling]. These compounds were expected from first principles DFT band calculations to host highly anisotropic Dirac dispersions with a finite gap at the Dirac point due to SOC [@lee2013anisotropic]. Such predictions were confirmed experimentally through magnetization, magneto-transport measurements [@park2011anisotropic; @wang2012two] and later using angle resolved photoemission spectroscopy (ARPES) [@feng2013strong]. What makes these compounds particularly interesting is the antiferromagnetic (AFM) ordering of the spin magnetic moment on the Mn$^{2+}$ atoms which has a $3d^5$ electronic configuration. The magnitude of the ordered moment is around $4\mu_B$ with an ordering temperature $\sim$300K as observed in several experiments [@lee2013anisotropic; @guo2014coupling]. Strong in-plane super exchange is responsible for the Néel-type AFM ordering observed in the $ab$ plane with a rather weak inter-layer magnetic coupling whose nature is dependent on the A atom [@guo2014coupling].
More recently, new members have been added to this family by substituting the rare earth elements with lanthanides such as Europium and Ytterbium. Whereas EuMnBi$_2$ appears to be very similar to its rare earth sister compounds [@PhysRevB.90.075109], YbMnBi$_2$ appears to have richer behavior and has been proposed as a prospective candidate for type II Weyl semi-metal (WSM) [@borisenko2015time]. Unlike most WSMs discovered so far that are non-centrosymmetric systems (TaAs, NbAs, NbP, TaP, SrSi$_2$, etc) with broken inversion symmetry, YbMnBi$_2$ has been proposed to be a potential WSM with broken time reversal symmetry (TRS). A magnetic WSM has continued to be elusive to date, with the possible exception of potential WSM phases in pyrochlore iridates [@sushkov2015optical] and magnetic heuslers under magnetic field [@hirschberger2016chiral]. In contrast to the other members of the AMnBi$_2$ family, ARPES has shown a Fermi surface that appears to be continuous with the hole-like lenses touching the electron-like pockets at what has been interpreted to be the Weyl nodes. It should be noted, that the band structure calculations predict that a TRS breaking WSM state can only be realized in this compound if a canting of the magnetic moment ($\sim$10$^\circ$) from the $c$ axis, resulting in an effective in-plane ferromagnetic component along (110) direction, is assumed in addition to the established AFM ordering. Although such a band structure is in reasonable agreement with the ARPES results, the ad-hoc assumption of canted antiferromagnetism has not been supported by neutron scattering experiments [@FuhrmanPrivateCommunication; @wang2016two]. Moreover, the magnetic space group found via neutron scattering that describes the AFM ordering in YbMnBi$_2$ has a symmetry breaking that does not allow canting of Mn magnetic moments away from the $c$ axis. From general considerations, the only way canting can occur is via an AFM transition that is not second order, or a lattice symmetry that is not tetragonal. Transport measurements reported in [@wang2016two] further supports a quasi two dimensional Dirac dispersion in this compound.
A number of interesting proposals have been made for the optical properties of WSMs [@hosur2013recent]. An ideal 3D Weyl state comprises at least a pair of non-degenerate band crossings in momentum space with the Fermi level at the nodal point. In 3D Dirac or Weyl systems, the joint density of states increases with frequency and is proportional to $\omega^2$ whereas the dipole optical matrix element goes as $\omega^{-1}$ as a result of which at low frequencies, the real optical conductivity is linear in $\omega$. In real systems however it is hard to achieve such fine tuning in the Fermi energy ($E_F$) and hence a cut-off is set by the Fermi level such that the real part of the conductivity is expected to be $$\sigma_1(\omega) = \dfrac{N G_0 \omega}{24 v_F}\Theta(\omega-2E_F) \cdot \label{eq:lin}$$ where $N$ is the number of massless fermion species in the Brillouin zone, $G_0 = 2e^2/h$ is the quantum conductance and $v_F$ is the Fermi velocity. Experimental evidence of such linear optical conductivity with zero intercepts has been reported in Eu$_2$Ir$_2$O$_7$ [@sushkov2015optical] and TaAs [@xu2016optical] among others [@armitage2017weyl]. A more detailed theoretical analysis of the inter-band optical response of TRS breaking WSMs [@PhysRevB.93.085442] however predicts that the single linear optical conductivity should be broken into two regions of quasi-linear conductivity with different slopes by a peculiar kink. The kink is a manifestation of the van Hove singularity, appearing at a frequency that corresponds to the energy difference between the bands crossing to form the Weyl nodes at the extrema between the two nodes. Such a feature has been experimentally observed in the inversion symmetry broken WSM TaAs [@xu2016optical].
In this paper, we report the complex optical conductivity of YbMnBi$_2$ obtained from Kramers-Kronig (KK) constrained variational dielectric function (VDF) fitting of the reflectivity [@chaudhuri2016optical]. A comparison of the experimental optical spectra with that calculated from the DFT band structure shows that the low energy features in the spectrum can be attributed to specific inter-band transitions and are not a signature of the van Hove singularity as proposed elsewehere [@chinotti2016electrodynamic]. The optical spectra has also been compared to the calculated band structures with both canted and uncanted AFM ordering which opens the possibility of reconciling the ARPES experiments with the neutron scattering and transport results. We believe our optical data is explained qualitatively by the uncanted magnetic structure with a small offset of the chemical potential from strict stochiometry.
Experimental and Computational Details
======================================
\
![\[fig:ref\] (a) Reflectivity of YbMnBi$_2$ measured using FTIR spectroscopy. (b) Raw reflectivity data at 300 K along with reflectivity curve as obtained from VDF fitting. (c) Resistivity of the sample measured using a 4-probe technique along with the Drude scattering rates.](ref_fitting_log.png "fig:"){width="48.00000%"}\
![\[fig:ref\] (a) Reflectivity of YbMnBi$_2$ measured using FTIR spectroscopy. (b) Raw reflectivity data at 300 K along with reflectivity curve as obtained from VDF fitting. (c) Resistivity of the sample measured using a 4-probe technique along with the Drude scattering rates.](res_sc_rt.png){width="53.00000%"}
Optical properties were measured on a cleaved (001) surface ($\sim 3$ mm $\times$ 3 mm) of a high quality YbMnBi$_2$ crystal using Fourier transform infrared (FTIR) spectroscopy. For details of the sample preparation please refer to Ref.[@borisenko2015time]. The as-grown (001) surface of another slightly smaller crystal was also measured and had identical results. Fourier transform spectroscopy in known to offer excellent signal to noise and frequency resolution. Reflectivity of YbMnBi$_2$ single crystals were measured using a commercial FTIR spectrometer (Bruker, Vertex 80V, Source: Globar/Hg Arc Lamp, Detectors: DLaTGS/MCT/Bolometer) across far and mid infrared spectral ranges spanning from 50 cm$^{-1}$ to 8500 cm$^{-1}$ (i.e. 1.5-2500 THz) for several temperatures between 5 K and 300 K. The reflection spectra at each temperature were referenced to that of elemental gold, deposited on the sample in situ by thermal evaporation. To extend the measurement across a broader spectral range, the infrared spectra was supplemented by visible reflection spectrum between 11150 cm$^{-1}$ and 29000 cm$^{-1}$ measured at room temperature using commercial spectrophotometry (Hitachi, U-3010) and referenced to an aluminum mirror. Additionally, the DC resistivity of the sample has been measured using conventional 4-probe measurements down to 2K.
Figure \[fig:ref\](a) shows the measured reflectivity in the infrared regime. The reflectivity tends to 1 at low frequencies and drops with a broad plasma edge-like feature above 4000 cm$^{-1}$ indicating metallic behavior. Besides the usual Drude response, two distinct features are observed in the reflectivity spectra that becomes particularly prominent at low temperatures. The lowest one is close to 200 cm$^{-1}$ where there is an abnormal increase in reflectivity and the second feature is a rather broad bump around 950 cm$^{-1}$.
Band structure calculations were performed using the linear muffin-tin orbital method [@And75] as implemented in the relativistic PY LMTO computer code. Some details of the implementation can be found in Ref. . Calculations were done assuming collinear AFM ordering of Mn moments which were aligned along the $c$ axis. Completely filled Yb$^{2+}$ 4$f^{14}$ states were treated as semi-core states. SOC was added to the LMTO Hamiltonian in the variational step. All theoretical results presented below were obtained within the local spin-density approximation (LSDA) with the Perdew-Wang parameterization [@PW92] for the exchange-correlation potential. Test calculations were also done using the PBE GGA potential. [@PBE96] They showed that the use of GGA slightly increases exchange splitting of Mn $d$ states but has only minor effect on Bi $p$ derived bands. Brillouin zone (BZ) integrations during the self-consistency loop were done on a 32$\times$32$\times$16 mesh using the improved tetrahedron method. [@BJA94]
Matrix elements for interband optical transitions were calculated in the dipole approximation. Then, the real part of the optical conductivity was calculated using the tetrahedron method and the imaginary part was obtained using the Kramers-Kronig relations.
We found that a very dense $k$ mesh should be used in order to achieve convergence of the conductivity below 0.3 eV. For instance, the in-plane conductivity spectrum calculated on the 32$\times$32$\times$16 mesh shows a broad peak centered at 0.18 eV. The peak shifts to $\sim$0.10 eV ($\sim$1000 cm$^{-1}$), becomes narrower, and its height doubles when a denser 80$\times$80$\times$48 mesh with almost 40000 symmetry inequivalent $k$ points in BZ is used. The use of even denser 128$\times$128$\times$48 ($\sim$10$^6$ $k$ points) mesh still leads to a small ($<$0.01 eV) shift of the peak position and 10% increase of its height. However, since calculations on such a dense mesh becomes extremely time consuming we present in the next section conductivity spectra calculated on the 80$\times$80$\times$48 mesh. It is worth noting, that above 0.5 eV convergence of the calculated conductivity spectrum is achieved already on the 32$\times$32$\times$16 mesh.
Analysis and Results
====================
\
[]{data-label="fig:sigma"}](sigma2.png "fig:"){width="48.00000%"}
The complex optical conductivity of the sample has been calculated from the reflectivity spectrum using KK constrained VDF fitting [@kuzmenko2005kramers]. This technique is an alternative to using conventional KK transformations to calculate the complex conductivity from a single measurement and is particularly useful in dealing with multiple reflection/transmission data from different disjoint frequency ranges. The method, implemented using the software RefFIT, involves an initial fitting of the reflectivity spectra to a standard Drude-Lorentz oscillator model with a few oscillators and then performing a KK constrained variational fitting with many oscillators of the difference spectra. For YbMnBi$_2$, the reflectivity spectra for each temperature has been fitted with one Drude and six Lorentz oscillators which has then been subjected to the VDF fitting analysis. Figure \[fig:ref\](b) shows the actual infrared and visible reflectivity data at 300K along with the output of the VDF fitting. In addition to reflectivity, the dc conductivity at each temperature as obtained from 4-probe resistivity measurements (see figure \[fig:ref\](c)) have also been used to constrain the VDF fitting subroutine. The real and imaginary parts of the complex conductivity thus obtained are shown in figure \[fig:sigma\](a) and \[fig:sigma\](b) respectively.
We model the low frequency response ($<$150 cm$^{-1}$) for all temperatures by a single Drude oscillator. Figure \[fig:ref\](c) shows the Drude scattering rate, $\Gamma_D$, as a function of temperature. Note that the temperature dependence of $\Gamma_D$ scales as resistivity within experimental error. At low temperatures, the resistivity of YbMnBi$_2$ has a strong temperature dependence showing the dominant role of inelastic scattering from either electron-electron or electron-phonon interactions. The similarity in temperature dependence of resistivity and $\Gamma_D$ thus points to the coherent nature of the quasiparticles within the Drude subsystem. The measured Drude plasma frequency at 5K is 1.74 eV.
The two most prominent features in the real part of the optical conductivity are the low energy peaks around 200 cm$^{-1}$ and 950 cm$^{-1}$. Such peaks in conductivity are ordinarily attributed to inter-band electronic transitions. To investigate this further, the Drude contribution (which mostly captures the intra-band optical conductivity) was subtracted. In general, strong electronic correlation can give low frequency intra-band contributions that would not generally be captured through a Drude term only. One possible way to isolate such effects is through the extended Drude model [@allen1977optical] analysis which can account for inelastic scattering due to electron-electron or electron-phonon interactions that can be frequency dependent even at low frequencies. However, the extended Drude model is only reliable much below the energy scales of inter-band transitions and thus it is not particularly useful in this context where the lowest possible inter-band transition could be as low as 200 cm$^{-1}$. We will therefore ignore such effects for now and assume that all the remaining high frequency optical conductivity is primarily from inter-band transitions and compare it to the inter-band optical conductivity calculated from band structure.
\
![(a) Band structure calculated along symmetry lines in tetragonal BZ. Bands crossing the Fermi level, which are discussed below, are plotted with blue (37,38), red (39,40), and green (41,42) lines. and (b) comparison of the calculated optical conductivity with experiment[]{data-label="fig:bs"}](S1_band.png "fig:"){width="48.00000%"}
Calculated inter-band optical conductivity (see figure \[fig:bs\](b)) is in rough agreement with the experimental observations at low frequencies and successfully reproduces a peak at 950 cm$^{-1}$ (green arrows) although the predicted amplitude is much larger. However, as one can see, the calculated conductivity is significantly higher than the experimental one almost all throughout the frequency range. The origin and significance of this peak will be discussed further below.
The more striking feature is perhaps the almost complete absence of a large peak close to 200 cm$^{-1}$ (red arrows) in the calculated optical conductivity spectra which might be indicative of some deeper inconsistencies. As the band structure in figure \[fig:bs\](a) does not include the canting of the Mn$^{2+}$ moments that is believed to be necessary for the magnetic WSM phase, a next obvious step is to include that in the DFT calculations.
\
\
![Band structure calculated along symmetry lines for a canting of (a) 5$^{\circ}$ and (b) 10$^{\circ}$ from the c-axis. The solid and dashed lines of the same colors represent the bands that were degenerate without canting. (c) Comparison of the calculated optical conductivity for different canting with experiment[]{data-label="fig:cant"}](S1_cant.png "fig:"){width="48.00000%"}
Figure \[fig:cant\](a) and \[fig:cant\](b) shows the band structure along (001) for a canting of 5$^{\circ}$ and 10$^{\circ}$ from the c-axis respectively. Two sets of Weyl points develop in the band structure for the canted system. The first set consists of two pair of Weyl nodes at momentum coordinates (0.193,0.193,0.12), (0.193,0.193,-0.09) and symmetry related points. In figure \[fig:cant\](a) and \[fig:cant\](b), along the $k$-space cut along the M-$\Gamma$ line, the Weyl point can be seen to be developing near 200 meV. The other four Weyl nodes are at (0.394, 0.045, 0.131) and symmetry related points. Details of these Weyl node structures have been discussed in reference [@borisenko2015time]. Note that for all points, there are certain non-idealities (both in distortions of the bands and their position from E$_F$) when compared to the simplest Weyl band structure proposed in [@PhysRevB.93.085442] which would make the observation of Weyl physics in the optical response challenging.
The inter-band conductivity for these three levels of canting, as well as the experimental data is plotted in figure \[fig:cant\](c). Note that a peak slightly above 200 cm$^{-1}$ gradually develops as one includes the effect of canting in the band structure. Nevertheless, one must be cautious in claiming this to be evidence for the WSM phase as there are noticeable alterations in the band structure unrelated to WSM physics that might bring about the observed changes in optical conductivity. The principal modification to the band structure from canting is the lifting of two fold degeneracy of the bands close to the Fermi level. This is the feature that gives rise to the Weyl nodes in the calculated band structure. However, this also opens up the possibility of having transitions between these previously degenerate bands as not all of them are above or below the Fermi level. The signatures of such inter-band transitions are not expected to be directly related to the existence of the Weyl nodes as one can see that they are generally far from the Fermi level.
If the important changes to the optical conductivity are indeed from the bands being rearranged around the Fermi level, somewhat similar changes may be observed by a shift of the chemical potential. Moreover, doping is rather common in semimetallic systems and thus exploring this possiblity is important. This was investigated by shifting the chemical potential above the calculated charge neutral Fermi level (see figure \[fig:fermi\](a)). An enhancement in the spectral weight under the calculated peak at 200 cm$^{-1}$ is seen upon shifting chemical potential by 20 meV above the charge neutral Fermi level (see figure \[fig:fermi\](b)). Thus it is possible that that the observed peak at 200 cm$^{-1}$ is simply indicative of a low energy inter-band transitions and not a signature of Weyl nodes as postulated elsewhere [@chinotti2016electrodynamic].
![Band structure along (001) with Fermi level at 0 meV (blue solid line) and 20 meV above (red dashed line). (b) Calculated optical conductivity for the corresponding Fermi levels.[]{data-label="fig:fermi"}](Fig05a_paqo80_dEf "fig:"){width="45.00000%"}\
![Band structure along (001) with Fermi level at 0 meV (blue solid line) and 20 meV above (red dashed line). (b) Calculated optical conductivity for the corresponding Fermi levels.[]{data-label="fig:fermi"}](S1_fermi.png "fig:"){width="48.00000%"}
To investigate the origin of these two peaks more closely, we decomposed the calculated conductivity into individual inter-band contributions. Figure \[fig:dec\](a) shows the result of that decomposition. These calculations provide more definitive proof to support that the peak in conductivity at 200 cm$^{-1}$ is indeed an inter-band transition from the set of degenerate bands that is mostly above the Fermi level (bands 39, 40) to the immediate higher one (bands 41, 42). This is why when the Fermi level is shifted to higher energies, the occupation in the lower degenerate set of bands (39, 40) and consequently the transition probability increases. Hence we observe the enhancement of the amplitude of the peak in conductivity upon shifting the Fermi level and is a probable cause of the experimentally observed peak.
![(a) Decomposition of the the total inter-band optical conductivity into individual band transitions. Transitions from bands (39, 40) to (41, 42) result in the peak at 200 cm$^{-1}$ whereas transitions from bands (37, 38) to (39, 40) is the origin of the peak at 950 cm$^{-1}$. (b) Similar decomposition assuming a 10$^{\circ}$ canted AFM order in the band structure.[]{data-label="fig:dec"}](S1_dec.png "fig:"){width="48.00000%"}\
![(a) Decomposition of the the total inter-band optical conductivity into individual band transitions. Transitions from bands (39, 40) to (41, 42) result in the peak at 200 cm$^{-1}$ whereas transitions from bands (37, 38) to (39, 40) is the origin of the peak at 950 cm$^{-1}$. (b) Similar decomposition assuming a 10$^{\circ}$ canted AFM order in the band structure.[]{data-label="fig:dec"}](S1_dec_canted.png "fig:"){width="48.00000%"}
It is interesting to note that in the context of the conductivity calculated from the band structure with canted AFM order, the origin of the peak at 200 cm$^{-1}$ is somewhat different than in the context of the uncanted structure, despite appearing at a similar energy scale. A decomposition into different inter-band contribution indicates that in this structure the peak is primarily from transitions between the previously degenerate bands, i.e., bands 37$\rightarrow$38 and 39$\rightarrow$40 (Band labels given in Fig. \[fig:bs\]). Although there is not much evidence from our experiment to choose one over the other, we do not favor this as the correct interpretation of our data as a canted order is not supported by neutron scattering experiments or theoretical considerations. However, this calculation further elucidates the fact that this peak is not a signature of Weyl node, even in a band structure that incorporates the effects of canted AFM order.
A possible explanation for the discrepancy in amplitude of the 200 cm$^{-1}$ peak would be electronic correlation effects which can potentially add to the intra-band conductivity but would not be captured through simple subtraction of the Drude contribution. As an alternative to this peak being an inter-band excitation, it may arise from the coupling of conduction electrons to magnetism. Neutron scattering experiments on YbMnBi$_2$ [@FuhrmanPrivateCommunication] show prominent spin wave excitations in the energy range $\sim$ 120 - 340 cm$^{-1}$ which matches the energy scale of the low energy peak observed in conductivity. The extent of this coupling is hard to estimate solely from optical conductivity measurements. One can expect such correlations to redistribute the Drude spectral weight giving an effective mass of the quasiparticles, $m^*$, different from the band mass, $m_b$. Such mass renormalizations can be extracted using extended Drude model analysis but as argued before, it would be unreliable in the present context because of the low lying inter-band transitions. However we can roughly characterize the renormalization effect by spectral weight analysis. We make use of the conductivity sum rule $$\dfrac{Z_0}{\pi^2}\int_0^{\omega_c} \sigma_1(\omega) d\omega = \omega_p^2 = \dfrac{4\pi n e^2}{m_b}$$ where $Z_0$ is vacuum impedance, $n$ is the carrier density and $\omega_c$ is chosen appropriately to include the bandwidth of interest. If we interpret both the Drude term and peak at 200 cm$^{-1}$ both as inter-band contributions, then the ratio of the total spectral weight to the Drude piece gives a measure of the renormalized mass. A naive estimate of this electron-magnon coupling constant from this exercise, $\gamma = m^*/m_b -1 = 0.4$. This is however a rather crude estimate as we know that there is a finite contribution to the conductivity at these frequencies from inter-band transitions and hence these numbers are only a rough guide. It would be interesting to estimate coupling constants from the neutron scattering data.
The peak at 950 cm$^{-1}$ is clearly from the inter-band transition between the bands of interest that would host the Weyl nodes in the canted structure i.e., right above (39,40) and below (37,38). However, because of the other low energy transition, it is nearly impossible to isolate any signature of the Weyl nodes if they exist in the optical conductivity spectrum. The peak derives from the quasi-2D Bi $p$-states (see Supplementary Materials) and is indicative of the linear dispersion. A sharp peak in conductivity is ordinarily expected to appear if a pair of occupied initial and empty final bands have nearly parallel dispersion in a large volume in $k$-space. Such an energy dispersion may occur if the Fermi level is inside a gap, which opens due to avoided crossing of two bands, and the size of the gap is constant in a large part of Brillouin zone. Hence it is reasonable to conclude that this peak is a signature of a Dirac dispersion with a small mass gap. This gap between the “lens” shaped Fermi surfaces from the (37,38) bands and the “boomerang” shaped Fermi surfaces from the (39,40) bands (see figure \[fig:dirac\](a) (inset)) extends along $k_z$ over the whole Brillouin zone as the derived Bi $p$ bands are quasi-2D. This is further supported by partial conductivity calculations where we have integrated over cylindrical regions in $k$-space centered around points shown in the inset in Fig. \[fig:dirac\](a). The radius of each cylinder was chosen to be $\approx 0.1\times\frac{2\pi}{a}$ and the results clearly indicate that the major contribution to this peak is from the D-point which is between the “lens” and the “boomerang” shaped Fermi surfaces where the Fermi level lies in the gap. Calculations also indicate that the magnitude of this gap does not depend on $k_z$ as expected. The contribution from B and X-points are also quite significant but as these points are in close proximity to the D-point, it is not surprising because the cylindrical volumes are not mutually exclusive. This is the reason why these spectra are not additive.
![(a) Decomposition of the the inter-band contribution to the optical conductivity from the transition between the bands (37,38) $\rightarrow$ (39,40) from different parts of the Fermi surface (fig. inset) (b) Linear fit of the inter-band transition peaked at 950cm$^{-1}$.[]{data-label="fig:dirac"}](dirac_dec.png "fig:"){width="48.00000%"}\
![(a) Decomposition of the the inter-band contribution to the optical conductivity from the transition between the bands (37,38) $\rightarrow$ (39,40) from different parts of the Fermi surface (fig. inset) (b) Linear fit of the inter-band transition peaked at 950cm$^{-1}$.[]{data-label="fig:dirac"}](vf.png "fig:"){width="48.00000%"}
We can isolate the low frequency features of this particular inter-band contribution from the conductivity measured experimentally by fitting the peak at 200 cm$^{-1}$ to a Lorentz oscillator and subtracting it from the previously calculated inter-band conductivity. The resultant shows a roughly linear regime between 300 cm$^{-1}$ and 950 cm$^{-1}$ that can be fitted to a straight line with zero intercept (see figure \[fig:dirac\](b)). With the assumption that this linearity is due to the Dirac dispersion, we can calculate the Fermi velocity using equation . There are four Dirac points in each Brillouin zone of YbMnBi$_2$ [@borisenko2015time] and thus the calculated Fermi velocity is (1.1$\pm$0.4)$\times$10$^5$ cm/s \[or 0.045$\pm$0.01 eV$\cdot$Å\]. This estimated Fermi velocity is within the range expected from ARPES which varies between 9 eV$\cdot$Å(perpendicular to the $k_xk_y$ plane) and 0.043 eV$\cdot$Å(along $k_z$). As discussed before, anisotropic Dirac cones are a hallmark of AMnBi$_2$ compounds and our observations here on YbMnBi$_2$ fits well into this description. The existence of a Dirac fermionic dispersion is also in good agreement with the reported magneto-transport experiments [@PhysRevB.93.085442].
We note that our interpretation of these peaks is at odds with another recent optical study [@chinotti2016electrodynamic] that has observed similar features in low energy conductivity but proposes that the peak around 200 cm$^{-1}$ derives from the van Hove singularity in a simplified WSM band structure. We do not favor this interpretation as realistic band structure calculations of YbMnBi$_2$ show it to exhibit certain non-idealities that will obscure the canonical behavior given in Ref. [@PhysRevB.93.085442]. Moreover, the Weyl state appears to be dependent on the existence of the magnetic structure for which there is no evidence. To reiterate, the the gross features of the optical response can be explained by the electronic structure of a uncanted state with a slightly shifted chemical potential.
Conclusion
==========
We have measured the reflectivity of YbMnBi$_2$ in the infrared regime and calculated the optical conductivity. The various features in the conductivity has been interpreted using DFT band structure calculations. We believe that the low energy peaks in conductivity at 200 cm$^{-1}$ and 950 cm$^{-1}$ are true inter-band transitions and the spectrum can be explained reasonably well without invoking a canted magnetic structure which is necessary to realize Weyl nodes in this system. The conductivity leading up to the peak at 950 cm$^{-1}$ is consistent with a linear Dirac dispersion with a small gap. Although canting of the Mn moments does improve the agreement of the calculated spectra with the experiments, similar effect can also be achieved by small shift in Fermi level. Precise control of the Fermi level in semimetals is rather difficult which when considering the lack evidence for canting, we believe makes the latter outcome more likely. However, in either case the high relative amplitude of the low energy peak is not apparent from band structure calculations but could be an manifestation of electronic correlation effects.
One must still reconcile the ARPES data that shows reasonably convincing evidence for Weyl physics [@borisenko2015time], with neutron scattering experiments [@FuhrmanPrivateCommunication; @wang2016two] that does not show evidence for canting and our optical data that does not need it. Moreover, we should reiterate that the crystal structure of this compound does not allow for a canted magnetic structure to develop through a second order phase transition. All of these considerations can be accommodated if we assume that the reduced symmetry of the surface allows a surface magnetic structure reconstruction that is consistent with the canted state and a surface Weyl phase. It is possible that the surface of this compound hosts a true WSM through the breaking of time-reversal symmetry. Detailed measurements of the surface electronic structure or magnetism would be very useful in this regard.
Acknowledgements
================
We would like to thank C. Broholm, S. Borisenko and W. Fuhrman for helpful discussions and E. S. Arinze and N. Drichko for assistance with visible range reflectivity measurements. Optical measurements at JHU were supported by the U.S. Department of Energy, Office of Basic Energy Sciences under Contract DE-FG02-08ER46544. Sample growth at Princeton was supported by the ARO MURI on topological insulators, Grant W911NF-12-0461.
[31]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} NP Armitage, EJ Mele, and Ashvin Vishwanath. Weyl and dirac semimetals in three dimensional solids. , 2017. @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [ ]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} in @noop [**]{} () @noop [****, ()]{} , ****, (). , , , ** (, , ). , ****, (). , , , ****, (). , , , ****, (). @noop [****, ()]{} @noop [****, ()]{}
Fermi surface map for the different band structures
===================================================
The Fermi surface for $k_z=0$ corresponding to the different band structure calculations are plotted in fig. \[fig:fs\].
\
\
\
\
\
\
Fat bands
=========
\
The origin of the peak in optical conductivity at 950 cm$^{-1}$, as discussed in the main text, is predominantly from the transition across the linearly dispersing bands. These massive Dirac like dispersion derives from the quasi-2D Bismuth $p$-states. So-called “fat bands” in Fig. \[fig:ocm\] clearly show that the strongly dispersing bands are almost exclusively composed of Bi $p$ states \[Fig. \[fig:ocm\] (a)\] whereas the relatively flat bands forming the electron pockets near the $Z$ point originate from Mn $d$ states \[Fig. \[fig:ocm\] (b)\].\
\
It is worth mentioning, that the wave functions of the predominantly Bi $p$ derived bands have also contributions of Bi $s$ and $d$ states which allows dipole transitions between these bands.
![Band plots showing orbital contributions from (a) Bi $p$ states and (b) Mn $d$ states.[]{data-label="fig:ocm"}](FigS2_FB_MnBi2.pdf){width="\textwidth"}
|
---
abstract: 'We use a discrete dislocation dynamics (DDD) approach to study the motion of a dislocation under strong stochastic forces that may cause bending and roughening of the dislocation line on scales that are comparable to the dislocation core radius. In such situations, which may be relevant in high entropy alloys (HEA) exhibiting strong atomic scale disorder, standard scaling arguments based upon a line tension approximation may be no longer adequate and corrections to scaling need to be considered. We first study the wandering of the dislocation under thermal Langevin forces. This leads to a linear stochastic differential equation which can be exactly solved. From the Fourier modes of the thermalized dislocation line we can directly deduce the scale dependent effective line tension. We then use this information to investigate the wandering of a dislocation in a crystal with spatial, time-independent (’quenched’) disorder. We establish the pinning length and show how this length can be used as a predictor of the flow stress. Implications for the determination of flow stresses in HEA from molecular dynamics simulations are discussed.'
address:
- 'Institute of Materials Simulation (WW8), Friedrich-Alexander University Erlangen-N[ü]{}rnberg (FAU), Dr.-Mack-Str. 77, 90762 F[ü]{}rth, Germany'
- 'Department of Engineering and Mechanics, Southwest Jiaotong University, Chengdu, People$''$s Republic of China'
author:
- 'Jian-Hui Zhai'
- Michael Zaiser
bibliography:
- 'references.bib'
title: 'Properties of dislocation lines in crystals with strong atomic-scale disorder'
---
,Discrete Dislocation Dynamics,Langevin force,Spatial force field ,High Entropy Alloys
Introduction
============
Discrete dislocation dynamics (DDD) as a mesoscale method to simulate plastic deformation by considering dislocation motion, reactions, and interactions with other defects has been developed over the past decades to simulate deformation of bulk materials [@Zbib1998_IJMS; @Verdier1998_MSMSE] and micropillars [@Schwarz1999_JAP; @Ghoniem1999_PRB; @Weygand2002_MSMSE]. Most of these simulations consider the evolution of dislocation systems in a deterministic setting, without explicitly accounting for stochastic influences that may be either due to thermal fluctuations (’annealed disorder’) or due to spatially random but temporally constant forces arising from small-scale (e.g. chemical) disorder (’quenched disorder’). Such random effects may be significant in novel materials with a large degree of atomic scale disorder such as HEA. They are the main focus of the present study.
Thermal effects can be considered by incorporating Langevin forces into the equations of motion [@Ronnpagel1993_PSS; @Mohles1996_CMS; @Hiratani2002_JEMT; @Hiratani2003_JNM]. R[ö]{}nnpagel et al. [@Ronnpagel1993_PSS] developed a stochastic model that considers the effects of temperature in a line tension model within Brownian dynamics scheme. They simulated dislocation glide in a stress field which was generated by localized obstacles and found that the effective attack frequency (inverse of waiting time) was indepedent of the choice of segment length and the activation enthalpy was dependent on applied stress and temperature. Mohles and R[ö]{}nnpagel [@Ronnpagel1993_PSS] studied dislocations in a field of obstacles or interacting with obstacle walls. They found that the activation volume for overcoming such obstacles was independent on the spacing of obstacle and the drag coefficient had no influence on waiting time. Their simulation system focused on two dimensional problems [@Ronnpagel1993_PSS; @Mohles1996_CMS]. Hiratani and Zbib [@Hiratani2002_JEMT; @Hiratani2003_JNM] extended the model proposed by R[ö]{}nnpagel et al. [@Ronnpagel1993_PSS] to three-dimensional simulations and used it to simulate dislocation glide through weak obstacles which were represented by stacking fault tetrahedra (SFTs). They found that dislocation motion was obstacle-controlled when the applied stress was below a critical resolved shear stress (CRSS), otherwise drag-controlled [@Hiratani2002_JEMT]; the dislocation line was found to exhibit a self-affine structure as manifested by non-trivial height-difference correlations of dislocation shapes [@Hiratani2003_JNM]. Langevin forces were implemented in different ways in these simulation schemes. Some researchers directly applied Langevin forces on dislocation nodes [@Ronnpagel1993_PSS; @Mohles1996_CMS], while others applied Langevin force on dislocation segments [@Hiratani2002_JEMT; @Hiratani2003_JNM]. In this paper we show that, to prevent artefacts, in a nodal dislocation dynamics scheme Langevin forces should be directly applied to dislocation nodes.
Solid solution strengthening is a common method to strengthen materials where foreign atoms interact with dislocations and impede their movement [@Argon2008_OUP] (’pinning’). This is one particular example of a class of problems studied extensively in statistical physics, namely the pinning of elastic manifolds by random fields, see [@Chauve2000_PRB]. Concepts of elastic manifold depinning and the associated statistical phenomena were applied to the athermal motion of dislocations by Zapperi and Zaiser [@Zapperi2001_MSEA] and Bako et. al. [@Bako2008_PRB], and to dislocation motion at finite temperature by Ioffe and Vinokur [@Ioffe1987_JPC] and Zaiser [@Zaiser2002_PM].
The behavior of dislocations in pinning fields created by the superposition of forces from multiple pinning centers depends crucially on the minimal wavelength of dislocation shape fluctuations. If this wavelength is comparable to the spacing of pinning centers along the dislocations, i.e., if the dislocation bends around the pinning centers individually, we speak of strong pinning, otherwise if the wavelength is larger than the pinning center spacing, we speak of weak or collective pinning. Fleischer [@Fleischer1961_AM; @Fleischer1963_AM] considered solute atoms as strong pinning centers. Labusch [@Labusch1970_PSS; @Labusch1972_AM] developed a statistical model to consider describe the interaction between dislocations and solute atoms where the solute atoms act as weak pinning centres atoms. The Fleischer model is appropriate to dilute solid solutions, where spacings between solute atoms are large, whereas the Labusch model is suitable for concentrated solid solutions and has been also applied to high entropy alloys (HEAs) [@Toda2015_AM; @Wu2016_AM; @Varvenne2016_AM; @Varvenne2017_AM]. Note that the concepts of ’strong’ vs ’weak’ pinning do not relate to the magnitude of the CRSS: In highly dilute solid solutions, the CRSS is low but pinning may be ’strong’ in the above defined technical sense, whereas in HEA the CRSS is high but pinning may be technically ’weak’.
Theoretical approaches to dislocation pinning, envisaged as elastic manifold pinning, often rely on a line tension approximation. This idea may be problematic in HEA where the atomic scale disorder may be strong and cause roughening of dislocations down to the nanometre scale, such that the minimal dislocation ’wavelength’ is no longer small compared to the dislocation core radius. Here we investigate systematically the corrections that need to be applied to standard pinning theories as a consequence. The remainder of this paper is organized as following. Section \[sec:2\] gives an overview of results from statistical physics concering the dynamics of elastic lines under the influence of random forces, and introduces several concepts that will be used in Sections to analyze the results of the dislocation dynamics simulations.
Section \[sec:3\] introduces discrete dislocation dynamics (DDD) under the influence of Langevin forces in \[sec:3.1\], the model is validated in \[sec:3.2\]. Results of simulations are reported in Section \[sec:4\], where spatio-temporal roughening under the influence of thermal forces is studied in \[sec:4.1\]. Comparison with analytical predictions from Section \[sec:2\] allows us to determine an effective, scale dependent line energy which is used in Section \[sec:4.2\] for studying dislocation glide under the influence of spatially random forces and applied stresses. Relations are given that allow to determine a characteristic pinning length from the zero-stress relaxed structure of the dislocation, and to relate this pinning length to the zero-temperature CRSS. Implications for determining CRSS values in HEA from molecular and ab-initio simulations are discussed and conclusions are given in Section \[sec:5\].
Scaling theory of elastic lines in random fields {#sec:2}
================================================
Thermal fluctuations
--------------------
The thermal equilibrium shape of an elastic line under the influence of thermal fluctuations can be deduced from simple thermodynamic arguments. We consider an elastic line of length $L$ and line energy ${\cal T}$ pinned at both ends. The initial line direction is identified with the $x$ direction of a Cartesian coordinate system. The fundamental modes of the line are given by $y_n(x) = A_n \sin(n \pi x/L)$ where $n$ is a positive integer number. The associated energy is $$E_n = \frac{{\cal T}}{2} \int \left(\frac{\partial y_n(x)}{\partial x} \right)^2 dx = n^2 \frac{{\cal T} \pi^2}{4 L} A_n^2.$$ From the equipartition theorem it then follows that $E_n = k_BT/2$, hence $$A_n^2 = \frac{2k_B T}{ {\cal T}L} \left(\frac{L}{\pi n}\right)^2,
\label{eq:intens}$$ i.e., the square mode amplitude (’intensity’) is expected to be proportional to temperature and inversely proportional to the square of the wave number. Conversely, from measurements of the mode amplitude the line energy ${\cal T}$ can be deduced as $${\cal T} = \frac{2 k_B T}{L} \frac{1}{(\pi n A_n)^{2}}
\label{eq:linetens}$$ For the case of dislocations, this provides an easy check to see whether a line tension approximation is warranted or whether corrections must be taken into account.
Defining the mode wave vector as $q = n \pi/L$ and noting that the phases of the different Fourier modes are independent random variables, we can deduce, in the limit $L \to \infty$ the scaling of the power spectrum of the line as $P(k) \propto q^{-a}$ where $a = 2$. Using general results for self affine manifolds [@Schmittbuhl1995_PRE], this implies that the line shape represents, in the limit $L \to \infty$, a self affine fractal with roughness exponent $\zeta = (a-1)/2 = 0.5$, i.e., it is equivalent to the graph of a random walk. We will investigate later to which extent this result is correct for dislocations.
To investigate the wandering dynamics, we consider an initially straight elastic line moving under the influence of thermal forces in an over-damped manner with a drag coefficient per unit length $B$. For a nearly straight line of line energy ${\cal T}$, the wandering dynamics under the influence of thermal Langevin forces is described by the annealed Edwards-Wilkinson equation $$B \frac{d y}{d t} = {\cal T}\frac{\partial^2 y}{\partial x^2} + f_T (x,t)
\label{eq:aedwilk}$$ where $f_T$ is a Gaussian random force with the correlation function $$\langle f_T(x,t) f_T(x',t') \rangle = 2 k_{\rm B} T B \delta(x-x') \delta (t-t').$$ The linearity of the governing equation allows for analytical solution in Fourier space, see e.g. [@Nattermann1992_PRA]. The evolution equation of the Fourier modes has the structure of an Ornstein-Uhlenbeck process; its solutions are Gaussian random variables with the correlation function $$\langle A(q,t) A(q',t') \rangle = \frac{k_B T}{ {\cal T}} \frac{1}{q^2} \left(1 - \exp\left[ - \frac{2 {\cal T}}{B} q^2 t\right ] \right) \delta(q - q')$$ or in terms of the wave number $$\langle A(n,t) A(n',t') \rangle = \frac{2 k_B T}{ {\cal T} L} \left(\frac{L}{\pi n}\right)^2 \left(1 -
\exp\left[ - \frac{2 \pi^2 {\cal T}}{BL^2} n^2 t\right ] \right) \delta_{nn'}$$
Spatially fluctuating forces
----------------------------
If the elastic line is subject to spatially fluctuating but temporally fixed forces, the evolution is given by the quenched Edwards-Wilkinson equation $$B \frac{d y}{d t} = {\cal T}\frac{\partial^2 y}{\partial x^2} + f_{\rm q} (x,y) + f_{\rm ext}
\label{eq:qedwilk}$$ where $f_{\rm ext}$ is an externally applied driving force per unit length and the random force $f_{\rm q}$ has the correlation function $$\langle f_{\rm q}(x,y) f_{\rm q}(x',y') \rangle = \hat{f}^2 \xi \delta(x-x') \phi(y-y').$$ where $\hat{f}$ is a characteristic force per unit length created by a random potential of correlation length $\xi$ and it is understood that we consider the line on scales well above the correlation length. The function $\phi(y)$ where $\phi(0) = 1, \int \phi dy = \xi$ describes short-range correlations of the random force in the $y$ direction of motion of the line. Owing to the $y$ dependence of the random force, is intrinsically nonlinear.
To understand the energy scales associated with , let us consider a bulge of length $L$ and width $\xi$ on an otherwise straight dislocation. The characteristic restoring force due to line tension is then easily estimated as $F_{\rm LT} \approx {\cal T} \xi/L$. The total random force acting on the bulge is a random variable with zero mean and standard deviation $F_{\rm R} = \hat{f} \sqrt{\xi L}$. Comparing the random and the restoring force defines a characteristic length, the so-called pinning length, given by $$L_{\rm p} = \left(\frac{\cal T}{\hat{f}}\right)^{2/3} \xi^{1/3}
\label{eq:lpin}$$ Below this length, the restoring force prevails, i.e., the line is bound to remain essentially straight. On scales above this length, the line will in the absence of an external driving force wander by an amount of the order of $\xi$ to reach the nearest energy minimum. For an initially straight line parallel to the $y$ axis this leads to a characteristic line shape where the line is straight but inclined on scales below $L_{\rm p}$, and exhibits irregular fluctuations with a standard deviation of the order of $\xi$ while keeping its original orientation on scales above $L_{\rm p}$. This allows us to obtain estimates of the pinning length directly from the shape of the relaxed line.
Averaging the pinning force over the pinning length defines a characteristic force per unit length. If the external force is smaller than this effective pinning force, the line is likely to be trapped into some metastable configuration. If the external force is large enough to make all metastable configurations disappear, the line will move indefinitely (depinning). The critical force for depinning can be estimated by comparing the external force to the pinning force, averaged over the pinning length. This gives $$f_{\rm ext,c}= \hat{f} \sqrt{\frac{\xi}{L_{\rm p}}} = \hat{f}^{4/3} {\cal T}^{-1/3} \xi^{1/3}.
\label{eq:fpin}$$ In the context of dislocation theory, this argument corresponds to the weak pinning limit. An interesting aspect is that, if the pinning length is known, we may relate the critical force directly to the pinning length by eliminating the characteristic force $\hat{f}$ from Eqs. (\[eq:lpin\]) and (\[eq:fpin\]) to obtain $$f_{\rm ext,c}= \frac{{\cal T}\xi^{1/3}}{L_{\rm p}^2}.
\label{eq:fpinL}$$ As we shall demonstrate, this relationship allows to obtain, in the weak pinning regime, rough estimates of the pinning force in a numerically very efficient manner.
Discrete Stochastic Dislocation Dynamics {#sec:3}
========================================
Description of model {#sec:3.1}
--------------------
In the present model which focuses on core effects, a non singular theory [@Cai2006_JMPS; @Arsenlis2007_MSMSE] is considered. After discretization of the dislocation into nodes connected by segments, the equation of motion of dislocation node $i$ is escribed by the following equation: $$\label{equ:1}
m_{0}L_{i} \bm{a}_{i}=-BL_{i} \bm{v}_{i} + \bm{F}_{i}^{\rm disloc} + \bm{F}_{i,T} + \bm{F}_{i}^{\rm a} + \bm{F}_{i}^{\rm q}.$$ Here $m_{0}\approx\rho b^{2}/2$ is the effective dislocation mass per unit length, , $\rho$ and $b$ are the material mass density and magnitude of Burgers vector, respectively; $L_{i}$ is an effective length associated with node $i$ due to dislocation segments connecting to it; $B$ is the drag coefficient per unit length; $ \bm{f}_{i}^{\rm disloc} $ includes the self force $\bm{f}_{i}^{\rm self}$, core force $\bm{f}_{i}^{\rm core} $ and the force from dislocation segment-segment interactions; $\bm{f}_{i,T}$ is a Langevin force which accounts for thermal fluctuations; $\bm{f}_{i}^{\rm a}$ is the force due to the externally applied stress and $\bm{f}_{i}^{\rm q}$ is the force due to a spatial force field that may be used to model the influence of atomic disorder.
The Langevin force is a stochastic force of magnitude (see \[app:1\]) $$\bm{F}_{i,T} =\Lambda_{i,j}\sqrt{\frac{2Bk_{B}TL_{i}}{\Delta t_{j}}}\bm{s}_{i},$$ where $\Lambda_{i,j}$ is a normal distributed Gaussian random number; $k_{B}$ is Boltzmann’s constant; $T$ is absolute temperature; $\Delta t_{j}$ is the increment of time at simulation step $j$, and the unit vector $\bm{s}_{i}$ denotes the effective glide direction of node $i$.
Temperature(K) Drag coefficient($\mu$Pa$\bm{\cdot}$s)
---------------- ----------------------------------------
300 30
600 60
900 90
1200 120
: Temperature dependence of drag coefficient, after [@Hiratani2002_JEMT].
\[table:1\]
[|c|c|c|c|c|>p[1.2cm]{}|c|c|c|]{} & & & &\
& (i) &(ii) &(i) &(ii) &(i) &(ii) &(i) &(ii)\
300 & 299 & 299 &52.7 &16.6 &52.3 &17.2 & $\sim$31.6&$\sim$36.5\
600 & 596 & 596 &74.5 &23.5 &72.9 &23.6 & $\sim$21.2& $\sim$16.7\
900 & 894 & 896 &91.2 &28.8 &92.5 &29.8 & $\sim$13.1& $\sim$15.9\
1200 & 1189& 1189&105.3 &33.3 &105.7 & 32.4& $\sim$9.0& $\sim$8.2\
\[table:2\]
Validation: thermal equilibrium properties {#sec:3.2}
------------------------------------------
To validate our model implementation, we consider an initially straight edge dislocation which we assume to run in $x$ direction and move in the $xy$ plane under the influence of thermal forces. After sufficient simulation time, once thermal equilibrium has been reached, the mean kinetic energy for each degree of freedom should be $k_{B}T/2$ and the distribution of nodal velocities along the local directions $\bm{s}_{i}$ should follow Maxwell’s distribution with a theoretical standard deviation (SD) of $\sqrt{k_{B}T/(m_{0}L_{i})}$. Two different simulations are considered in this part: (i) an edge dislocation line with fixed endpoints which is discretized into 1000 dislocation segments with segment length 20$b$; (ii) a similar edge dislocation line with non-equidistant nodes and $L_{i}$ changing from 20$b$ to 200$b$. The time step $\Delta t_{j}$ is chosen constant and equals 50fs. For larger time steps, the mean kinetic energy and velocity distribution can be used to determine whether the selection is suitable or not. A series of temperatures (300K, 600K, 900K and 1200K) are simulated to check the correctness of the Langevin force implementation. Material parameters are taken from [@Hiratani2002_JEMT] in order to compare with these simulations, in which Langevin force were applied on dislocation segments rather than nodes. The shear modulus and Poisson ratio are $\mu$=54.6 GPa and $\nu$=0.324, respectively; the magnitude of Burgers vector $b$=0.256nm and mass density is $\rho$=8900kg/m$^{3}$. The drag coefficient is assumed proportional to temperature and listed in Table \[table:1\].
Kinetic temperatures and velocity distributions for different Langevin temperatures are shown in Figs. \[fig:kinetic\_temperature\] and \[fig:velocity\_distribution\], respectively for a dislocation with equi-distant nodes. The principle of equipartition is observed in Fig. \[fig:kinetic\_temperature\] and the distribution of velocity about a specified node follows a Maxwell distribution with simulated SD close to the theoretical value $\sqrt{k_{B}T/(m_{0}L_{i})}$ as shown in Fig. \[fig:velocity\_distribution\]. Simulated temperature, theoretical value of standard deviation (TSD), simulated SD and time $t_{\rm equ}$ for thermal equilibration are shown in Table \[table:2\] (for dislocations with equi-distant nodes (case (i)) and for dislocations with variable node spacing (case (ii)), in the latter case, the SD and TSD data refer to the node with the largest associated effective length. As we can see, simulated temperatures are very close to the given temperatures (maximum relative error is within 1%) and the simulated SD of velocity distribution has matches the theoretical values. A non-equidistant node distribution scheme has no detrimental influence on the agreement between theory and simulation. $t_{\rm equ}$ is mainly influenced by drag coefficient and decreases with increasing drag coefficient. The influence of dislocation segment length is also studied. Simulated temperature, TSD, simulated SD and $t_{\rm equ}$ for different dislocation segment lengths in a simulation of type (i) are shown in Table \[table:3\] where $\Delta t$=50fs and $T$=1200K. The table shows that different dislocation segment lengths have no influence on simulated temperature and thermal equilibration time decreases with increased dislocation length, which is the opposite of the result obtained when Langevin forces are applied on dislocation segments [@Hiratani2002_JEMT].
Length($b$) simulated temperature(K) TSD(m/s) simulated SD(m/s) $t_{equ}$(ps)
------------- -------------------------- ---------- ------------------- ---------------
100 1188 47.1 47.0 $\sim$8.7
200 1190 33.3 32.3 $\sim$8.1
400 1190 23.5 23.7 $\sim$5.8
800 1185 16.6 15.4 $\sim$4.4
\[table:3\]
Simulation results {#sec:4}
==================
Spatio-temporal roughening of a dislocation line under influence of thermal forces {#sec:4.1}
----------------------------------------------------------------------------------
We simulate an edge dislocation of length $2 \times 10^4 b$ pinned at its endpoints. The line direction is identified with the $x$ direction of a Cartesian coordinate system, the glide plane is the $xy$ plane. The system is assumed to be of infinite extension in $z$ direction and periodic boundary conditions are imposed on a simulation cell of extension $2 \times 10^4 b$ in $x$ and $2 \times 10^4 b$ in $y$ direction. The dislocation is discretized into 1000 nodes with an inital spacing of $20b$.
In our investigation of the spatio-temporal behavior of the dislocation under the influence of thermal forces, we first consider a dislocation that has been evolving for sufficient time to establish thermal equilibrium. In this case the amplitudes of its Fourier modes fulfill Eq. .
![Top: Fourier mode amplitudes of a thermally equilibriated dislocation against wave number; dashed line: $A_n^2 \propto n^{-2}$ as expected for an elastic line according to ; bottom: effective line energy as determined from , full: guide to the eye.\[fig:eqdis\]](eqdis.pdf){width="3.3in"}
The amplitudes of the Fourier modes of the equilibrated dislocation are shown in Figure \[fig:eqdis\], top. We observe that, for short wavelengths (large wave numbers), characteristic deviations occur from the $A_n^2 \propto n^{-2}$ behavior expected according to Eq. \[fig:eqdis\]. Of course, in our simulations the self-interaction of a dislocation line is described by a non-local interaction kernel which is much more complex than a simple elastic line. So the question arises whether we can understand the deviations and quantify them.
To this end, we first use to determine a wave number dependent line energy. This is shown in Figure \[fig:eqdis\], bottom. We can see that, for long-wavelength fluctuations, the data can be well described by an approximately constant line energy. However, on short scales (large wave numbers) the effective line tension seems to be decreasing. To understand this behavior, we have, in Appendix A, calculated the Fourier transform of the non-singular dislocation self interaction kernel [@Cai2006_JMPS; @Arsenlis2007_MSMSE], both for an isolated dislocation and for a periodic array of rigidly coupled dislocations as implied in the present simulations by our boundary conditions.
![\[fig:rougheningT\]Top: time evolution of the structure function $R(d,t)$, simulation at $T$=300K; dashed line: $R \propto
(d/{\cal T}(d)^{0.5}$, average over 10 simulations; bottom: scaled structure function $w = R (k_{\rm B}T)^{-0.5} B^{0.25} {\cal T}^{0.25} t^{-0.25}$ as function of scaled displacement $u = d({\cal T} t/B)^{-0.5}$; data for different temperatures as shown in legend, data at each temperature for times $t=0.05ns, t= 0.5ns, t=3ns$.](rougheningT.pdf){width="3.3in"}
In a next step, we investigate the transient dynamics before the dislocation has reached thermal equilibrium, starting from an initially straight dislocation line configuration which, under the influence of thermal fluctuations, gradually develops a self affine shape $y(x)$. We statistically characterize this process by the structure factor $R(d,t)$ defined as: $$\label{equ:3}
R(d,t)=\left\langle \left\langle \left| y(x,t)-y(x+d,t)\right|\right\rangle_{x}\right\rangle _{N}$$ where $\left\langle \right\rangle_N $ means average over an ensemble of $N$ simulations and $\left\langle \right\rangle_x$ the average over all $x$ positions for which $y(x+d,t)$ can be computed. The time evolution of $R(d,t)$ is shown in Figure \[fig:rougheningT\] for $T=300$K. For a given time, the function $R(d,t)$ exhibits two regimes: At small $d$, thermal equilibrium has been reached, and the line has a self affine shape where the structure function is time independent and approximately follows the theoretically expected scaling for an elastic line, $R \propto (d k_{\rm B} T/{\cal T})^{\zeta}$ with $\zeta = 0.5$ [@Nattermann1992_PRA]. The apparent deviations from slope 0.5 on the double-logarithmic plot can be accounted for by noting that the effective line energy of the dislocation decreases on small scales. Setting ${\cal T} \propto \ln(d/b)$ produces a good representation of the data.
The overall behavior of $R(d,t)$ can be described by the equation $$R(d,t) = \left(\frac{d k_{\rm B} T}{{\cal T}}\right)^{\zeta} \Phi\left(\frac{\cal T}{B} \frac{t}{d^z}\right)$$ with the roughness exponent $\zeta = 0.5$ and the dynamic exponent $z=2$ for the elastic line under thermal forces. The scaling function $\Phi$ has the following properties: $\Phi(x) \to \Phi_{\infty}\; {\rm for}\; x \to \infty, \Phi(x) \propto x^{\zeta/z}\;
{\rm for}\; x \to 0$. This implies that, for sufficiently large $d$, the structure function approaches an approximately $d$ independent value $R_{\infty} \propto T^{\zeta} B^{-\zeta/z} {\cal T}^{\zeta(1-z)/z} t^{\zeta/z}$. With $\zeta = 0.5$, $z = 2$, and a linear temperature dependency $B \propto T$, we find that $R_{\infty} \propto T {0.25} t^{0.25}$. Figure \[fig:rtimetemp\] shows that the time- and temperature dependence of $R_{\infty}$ is in good agreement with this expectation.
![Top: time dependence of the asymptotic structure factor $R_{\infty}$ at a temperature of 1200 K, full line: slope 0.25 of the double-logarithmic plot; bottom: temperature dependence of $R_{\infty}$ at time $3ns$, full line: slope 0.25 of the double-logarithmic plot. \[fig:rtimetemp\]](Rtimetemp.pdf){width="3.3in"}
A generic representation of the roughening behavior at all times and all temperatures is obtained by introducing the new variables $u = d({\cal T} t/B)^{-1/z}$ and $w = R (k_{\rm B}T)^{-\zeta} B^{\zeta/z} {\cal T}^{\zeta(z-1)/z} t^{-\zeta/z}$ in which the structure factor has the universal form $$w = u^{\zeta} \Phi(u^{-z}).$$ This is illustrated in Figure \[fig:rougheningT\], bottom, which shows a compilation of structure factor curves pertaining to different temperatures and roughening times after re-scaling to the universal variables $u$ and $w$. The plot demonstrates the existence of an underlying generic scaling curve that represents the evolving shapes at all times and temperatures.
Roughening and pinning in a spatially random force field {#sec:4.2}
--------------------------------------------------------
In this section, we investigate the interaction of a dislocation with a spatial time-independent force field. Spatially random force fields can be used to model structural disorder on the atomic scale as prominent in random solid solutions and high-entropy alloys. Such disorder is, at low to intermediate temperatures, time independent (’quenched disorder’) and creates spatially fluctuating internal stresses and ensuing forces on the dislocations which depend on position and have zero average value. We use a simple model where a random force at the point $(x,y)$ is created by the superposition of forces from randomly located pinning points, each of which creates a Gaussian pinning potential: $$f(x,y)=\hat{f}\sum_{i=1}^{N}\frac{(x-x_{i})}{\xi}e^{-\frac{(x-x_{i})^{2}+(y-y_{i})^{2}}{2\xi^{2}}}.$$ Here $x_{i},y_{i}$ are the coordinates of pinning center $i$; $N$ is total number of randomly distributed pinning centers; $\xi$=$\sqrt{A/N}$ is the range of the pinning potential which also determines the correlation length of the random pinning force, and $\hat{f}$ defines the characteristic magnitude of the spatial force field.
![Top: determination of the pinning length, simulation with $\hat{f} = 0.04$N/m at zero applied stress; bottom: dependence of pinning length on amplitude $\hat{f}$ of the pinning field.[]{data-label="fig:pinlength"}](pinlength.pdf){width="3.3in"}
To quantify the action of the pinning force on the dislocation, we again consider the structure factor $R(d)$ which we evaluate for an ensemble of initially straight dislocations interacting with a pinning potential of range $\xi = 20 b$ and variable strength. During relaxation at zero applied stress, the dislocations develop local roughness as they adjust their line shape to minimize their energy in the random pinning field. The corresponding structure factor is shown in Figure \[fig:pinlength\]: On small scales one finds an increasing $R(d)$ function with an approximately linear dependency, $R \propto d^{\zeta}$ with $\zeta =1$, as indicative for an inclined dislocation line that has moved under the influence of random forces while remaining locally straight. On large scales the function $R(d)$ is constant as indicative of uncorrelated random displacements of the initially straight line in both directions. A crossover length $L_{\rm p}$ can be constructed as shown in Figure \[fig:pinlength\] by fitting two curves $R \propto d$ and $R = {\rm const.}$ to the two branches and identifying the crossover length (’pinning length’) with the intersection point.
Results are shown in Figure \[fig:pinlength\], top which shows the thus determined pinning length as a function of the strength of the pinning field. As expected according to , the pinning length decreases with increasing pinning field. However, this decrease is only in the regime of weak pinning fields quantitatively described by the theoretical slope of -2/3. At large fields, the pinning length saturates because $L_{\rm p}$ cannot become larger than the physical spacing $\xi$ of the pinning centers. This corresponds to a transition from weak to strong pinning.
We now turn to the behavior of the dislocation line under applied stress. An increasing applied stress causes the dislocation to move between metastable configurations until, at a critical stress, metastability is lost and the dislocation moves indefinitely (depinning). In terms of the self affine roughness of the dislocation line, roughening extends to larger and larger scales until, at the critical stress, the dislocation exhibits a self affine shape on all scales with roughness exponent $\zeta \approx 1$. This roughening is illustrated in Figure \[fig:roughquenched\] for two different amplitudes $\hat{f}$ of the pinning field.
![Evolution of the structure function $R(d,\tau)$ under increasing applied shear stress in a stationary pinning field; top: amplitude of the pinning field $\hat{f}$ = 0.128 N/m, bottom: $\hat{f}$ = 0.004 N/m; the resolved shear stress is given as a fraction of the critical resolved shear stress (crss), see Figure \[fig:crss\], top, for the respective crss values.\[fig:roughquenched\]](roughquenched.pdf){width="3.3in"}
We define the critical resolved shear stress (crss) as the minimum stress required to move the dislocation once across the periodic simulation cell of width $2000b$ in the (mean) direction of dislocation motion. To compare with the prediction for an elastic line, Eq. (\[eq:fpin\]), we need to specify a line energy: We identify the wave vector corresponding to the pinning length as $q_{\rm p} = 2 \pi/L_{\rm p}$ and obtain the corresponding line energy from Figure \[fig:eqdis\], thus assuming that pinning is controlled by an effective value of the line energy on the scale of the pinning length. This allows us to calculate a pinning force from Eq. (\[eq:fpin\]), from which the predicted crss derives by simply setting $f_{\rm c} = \tau_{\rm c}b$. Predicted crss values are compared to values deduced from the simulations in Figure \[fig:crss\], top. It can be seen that in the regime of weak pinning, here up to a force amplitude of about $\hat{f}=$0.02 N/m, corresponding to a pinning length of about 100$b$, the simulated crss data (black squares in Figure \[fig:crss\], top) follow to a good approximation the prediction for an elastic line (red circles in Figure \[fig:crss\], top). Deviations occur in the regime of strong pinning, corresponding to the regime where the pinning length deviates from the weak pinning result and tends towards a saturation: In this regime, we see a cross-over from an exponent $p=4/3$ in the $\tau_{\rm c} \propto \hat{f}^p$ relationship to an exponent of 1. The transition occurs at a pinning length of $L_{\rm p} \approx 5 \xi$, for larger pinning lengths the weak pinning relations are well fulfilled.
We next turn to the possibility of predicting the crss on the basis of the pinning length alone using Eq. (\[eq:fpin\]). As can be seen in Figure \[fig:crss\], bottom, this equation produces excellent results in the weak pinning regime $L_{\rm p} > 100 b = 5 \xi$.
![Top: critical resolved shear stress as function of amplitude $\hat{f}$ of the random force field, full squares: crss values determined from our simulations, open circles: prediction from Eq. (\[eq:fpin\]); bottom: critical resolved shear stress as function of pinning length (pinning length as in Figure \[fig:pinlength\]), full squares: crss values determined from our simulations, open circles: prediction from Eq. (\[eq:fpinL\]). \[fig:crss\]](CRSS.pdf){width="3.3in"}
This finding is important because it suggests that, in the weak pinning regime, estimates of the zero-temperature critical resolved shear stress can be obtained by studying the static properties of a dislocation at zero stress on scales of the pinning length only - scales that are accessible to molecular or even ab-initio simulation methods. Depinning, on the other hand, involves complex spatio-temporal processes and avalanche dynamics covering a wide range of spatial and temporal scales, which require mesoscopic methods such as discrete dislocation dynamics for their simulation.
Discussion and Conclusions {#sec:5}
==========================
We have analyzed the roughening of a dislocation line under the influence of thermal forces which has provided us with a method to determine, for the DDD model analyzed, a length-scale dependent effective line energy that controls the energy associated with perturbations of a straight line. We then studied the behavior of the simulated dislocation in a stationary random field created by pinning centers of spacing and interaction range $\xi$. Using scaling relations typical of elastic lines in the regime of weak pinning, we found that we can determine a pinning length $L_{\rm p}$ from analyzing the relaxed shape of a initially straight dislocation at zero applied stress. The main result of our investigation is that, in the weak pinning regime where $L_{\rm p} > \xi$, it is possible to obtain estimates of the depinning stress (the critical resolved shear stress needed to move the dislocation across the pinning field) without explicit knowledge of the pinning field strength, based upon the pinning length, the interaction range $\xi$ which relates to the density of the pinning obstacles, and the scale dependent dislocation line energy.
What is the use of this result? First of all we note that determining the pinning length can be done by static relaxation of the dislocation in a comparatively small periodic simulation cell. For the construction shown in Figure \[\] to work, it is sufficient that the simulation cell extends over about 5 pinning lengths in the direction parallel to the dislocation and about 5 $\xi$ in the perpendicular direction. Depinning, on the other hand, is preceded by a sequence of increasingly complex metastable configuurations where the dislocation develops self affine roughness on all scales and, as can be inferred from Figure \[\], accurate determination of the flow stress requires a simulation cell that extends over about $1000 \xi$ in the direction parallel and over $100\xi$ in the direction perpendicular to the dislocation. Hence, the size of the simulation cells needed for determination of $\tau_{\rm c}$ and of $L_{\rm p}$ differs by a factor of about 200. In addition, direct determination of the crss needs a search algorithm to find the critical stress at which, in such a simulation, metastability is lost, which necessitates a sequence of relaxation steps or a simulation of the time dependent dynamics under a very slowly ramping external stress.
In a conventional solution hardened alloy, e.g. a binary alloy system for which reliable inter-atomic potentials are available, it is nevertheless possible to perform medium-scale molecular dynamics simulations to directly determine the crss [@Rodary2004_PRB] from atomic simulation. This direct approach cuts out the need to parametrize a meso-scale model: In a concentrated solid solution the solute spacing and hence the correlation length $\xi$ is expected to be of the order of $\xi \sim 2b$ only, hence a system of size $100 \times 100 \times 1000 \xi^3$ as used in our study is accessible to large-scale MD simulations. In medium- or high-entropy alloy systems for which phenomenological potentials are available, similar simulations may be performed as demonstrated by [@Rao2017_AM] for the quarternary CoFeNiTi system. The here presented method may then allow for a computationally efficient screening of the space of compositions within such an alloy system in view of establishing optimal mechanical properties, by performing serial simulations of small systems and establishing the composition dependence of the pinning length from the respective relaxed dislocation configurations.
To show that our ideas are indeed applicable to high-entropy alloys we refer to experimental data from the literature. For fcc equi-atomic CoCrFeMnNi alloy a solute contribution to strength of about 300 MPa has been reported [@Otto2013_AM] while the shear modulus amounts to $\mu$=81 GPa and the Young’s modulus to 203 MPa [@Laplanche2015_JAC]. Using a line energy estimate of $0.3 \mu b^2$ and correlation length $\xi \approx b$ we find from Eq. (\[eq:fpin\]) a characteristic pinning length $L_{\rm p} > 9 \xi$ which indicates that this alloy is well within the weak pinning regime (see also the work of Varvenne et. al. [@Varvenne2016_AM] who apply weak-pinning type averaging methods to the same alloy). For bcc equi-atomic TaNbHfZrTi the solute contribution to strength is about 700MPa [@Senkov2011_JAC] which with an estimated shear modulus of $\mu$=40 GPa [@Senkov2011_JAC] and similar line energy and correlation length estimates implies a characteristic pinning length $L_{\rm p} \approx 5 \xi$, still at the borderline of the weak pinning regime. Finally, in a computational study of a quarternary CoFeNiTi alloy data determined by [@Rao2017_AM] inidcate a disorder contribution to the $T=0$ crss of screw dislocations of $0.02 \mu$ (difference between the crss values of the random alloy and the disorder-homogeneized alloy), which corresponds to a disorder-associated pinning length of $L_{\rm p} \approx 7 \xi$, again within the weak pinning regime. It is thus likely that weak pinning ideas can be applied to HEA or, more generally speaking, to compositionally complex random alloys, and that the method proposed here allows for a rapid screening of the composition space in such alloy systems where reliable potentials are available. We note that, in fcc HEA, an additional complication arises from the typically low stacking fault energy and resulting high degree of core splitting, which implies that the concept of solute pinning must be applied to split dislocations of even independently pinned particals. We will discuss the resulting modifications to the present considerations elsewhere but note that, as long as the splitting distance is larger than the pinning length, the analysis of weak pinning can be applied to the two partial dislocations separately.
Acknowledgements {#acknowledgements .unnumbered}
================
This work is funded by Deutsche Forschungsgemeinschaft (DFG) under grant 1Za-8/1. Discussions with Daniel Weygand and Volker Mohles are gratefully acknowledged.
Derivation of Langevin force on a dislocation {#app:1}
=============================================
Without external applied stress and spatial force field Eq. (\[equ:1\]) can be rewritten as $$\label{equ:a1}
m_{0}L_{i} \bm{a}_{i}=-BL_{i} \bm{v}_{i} + \bm{F}_{i}^{\mathit{disloc}} + \bm{F}_{i,T}$$ where $\bm{F}_{i}^{\rm{disloc}}=-U_{\bm r_{i}}$, $U$ is total potential energy of the dislocation network [@Bulatov2006_OUP] and $\bm r_{i}$ is the position of node $i$. $L_{i}$ is the effective segmeint length associated with node $i$ and can be expressed as: $$L_{i} = \sum\limits_{j=1}^N0.5L_{ij}$$ where $ j $ labels the $N$ nodes connected to node $ i $ by connecting segments of length $L_{ij}$. For a straight line not connected into a network, as considered here, $N =2$. Eq. (\[equ:a1\]) can be rewritten as following: $$\bm{a}_{i} + \gamma\bm{v}_{i} = \bm{\mathit{\Gamma}}_{i}^{\rm disloc} + \bm{\mathit{\Gamma}}_{i,T}$$ where $\gamma=B/m_{0}$ and $\bm{\mathit{\Gamma}}_{i}=\bm{\mathit{F}}_{i}/(m_{0}L_{i})$ denotes a force per unit mass.
The Langevin force describing a thermal white noise has the following statistical properties [@Risken1996_SPRINGER]: $$\begin{split}
&\left\langle \mathit{\Gamma}_{i,T}\left( t\right)\right\rangle = 0 \\
&\left\langle {\mathit{\Gamma}}_{i,T}(t){\mathit{\Gamma}}_{i,T}(t')\right\rangle = q\delta(t-t')
\end{split}
\label{equ:a4}$$ where the noise strength $q$ is given by $$q=2\gamma k_{B}T/(m_{0}L_{i})=2Bk_{B}T/(m_{0}^{2}L_{i})$$ where $k_{B}$ is Boltzmann constant and $T$ is absolute temperature. In a numerical implementation, we statistically integrate the Langevin force over the elementary time step, yielding $${\mathit{\Gamma}}_{i,T}(t_{j} \leq t<t_{j}+\Delta t_{j})=\Lambda_{i,j}\sqrt{\frac{2Bk_{B}T}{m_{0}^{2}L_{i}\Delta t_{j}}}$$ where $\Lambda_{i,j}$ is a Gaussian distributed random variable of zero mean and unit variance, and $\Delta t_{j}$ is the length of simulation time step $j$. Accordingly, the Langevin force on a dislocation segment of length $L_{i}$ has the following form: $$\label{equ:a7}
{\mathit{F}}_{i,T}( t_{j} \leq t<t_{j}+\Delta t_{j}) = \Lambda_{i,j}\sqrt{\frac{2Bk_{B}TL_{i}}{\Delta t_{j}}}$$ and the Langevin force per unit length is: $$\label{equ:a8}
{\mathit{f}}_{i,T}(t_{j} \leq t<t_{j}+\Delta t_{j})=\Lambda_{i,j}\sqrt{\frac{2Bk_{B}T}{L_{i}\Delta t_{j}}}$$
Following the suggestion of [@Mohles1996_TUB], the component of the Langevin force along the dislocation line direction is ignored to avoid numerical problems and the nodal Langevin force is thus given by the following expression: $$\bm{s}_{i} = \sum\limits_{j=1}^NL_{ij}\bm{s}_{ij}/|\sum\limits_{j=1}^NL_{ij}\bm{s}_{ij}|$$ where $\bm{s}_{ij} = \bm{n}_{ij}\times\bm{t}_{ij}$ is the glide direction of segment $i-j$, $\bm{n}_{ij}$ and $\bm{t}_{ij}$ are the unit normal vectors of glide plane and the unit line direction vector of dislocation segment $i-j$, respectively. Combining magnitude and direction gives the Langevin force and force per unit length as: $$\begin{split}
&\bm{\mathit{F}}_{i,T} = {\mathit{F}}_{i,T}\bm{s}_{i} \\
&\bm{\mathit{f}}_{i,T} = {\mathit{f}}_{i,T}\bm{s}_{i}
\end{split}
$$
References {#references .unnumbered}
==========
|
---
abstract: 'In this paper, the design and analysis of a new bandwidth-efficient signaling method over the bandlimited intensity-modulated direct-detection (IM/DD) channel is presented. The channel can be modeled as a bandlimited channel with nonnegative input and additive white Gaussian noise (AWGN). Due to the nonnegativity constraint, standard methods for coherent bandlimited channels cannot be applied here. Previously established techniques for the IM/DD channel require bandwidth twice the required bandwidth over the conventional coherent channel. We propose a method to transmit without intersymbol interference in a bandwidth no larger than the bit rate. This is done by combining Nyquist or root-Nyquist pulses with a constant bias and using higher-order modulation formats. In fact, we can transmit with a bandwidth equal to that of coherent transmission. A trade-off between the required average optical power and the bandwidth is investigated. Depending on the bandwidth required, the most power-efficient transmission is obtained by the parametric linear pulse, the so-called “better than Nyquist" pulse, or the root-raised cosine pulse.'
author:
- |
\
[^1]
bibliography:
- 'MyDatabase\_EA.bib'
title: Bandlimited Intensity Modulation
---
Intensity-modulated direct-detection (IM/DD), strictly bandlimited signaling.
Introduction
============
demand for high-speed data transmission systems has introduced new design paradigms for optical communications. The need for low-complexity and cost-effective systems has motivated the usage of affordable optical hardware (e.g., incoherent transmitters, optical intensity modulators, multimode fibers, direct-detection receivers) to design short-haul optical fiber links (e.g., fiber to the home and optical interconnects)[@4528765; @5875697] and diffuse indoor wireless optical links [@gfeller2005wireless; @kahn2002wireless; @hranilovic2005design]. These devices impose three important constraints on the signaling design. First, the transmitter only modulates information on the instantaneous intensity of an optical carrier, contrary to conventional coherent channels where the amplitude and phase of the carrier can be used to send information . In the receiver, only the optical intensity of the incoming signal will be detected [@kahn2002wireless]. Due to these limitations, the transmitted signal must be nonnegative. Such transmission is called intensity modulation with direct detection (IM/DD). Second, the peak and average optical power (i.e., the peak and average of the transmitted signal in the electrical domain) must be below a certain threshold for eye- and skin-safety concerns [@kahn2002wireless] and to avoid nonlinearities present in the devices [@westbergh200932; @inan2009impact]. In conventional channels, such constraints are usually imposed on the peak and average of the squared electrical signal. Third, the bandwidth is limited due to the impairments in the optoelectronic devices [@hranilovic2005design; @4132995] and other limitations (e.g., modal dispersion in short-haul optical fiber links [@5223628] and multipath distortion in diffuse indoor wireless optical links[@kahn2002wireless]). Consequently, the coherent modulation formats and pulse shaping methods designed for conventional electrical channels (i.e., with no nonnegativity constraint on the transmitted signal) cannot be directly applied to IM/DD channels.
Pulse shaping for the purpose of reducing intersymbol interference (ISI) in conventional channels has been previously investigated in , [@5055024; @1089824; @beaulieu2002better; @1369626; @chandan05; @192386]. Much research has been conducted on determining upper and lower bounds on the capacity of IM/DD channels considering power and bandwidth limitations [@farid2010capacity; @hranilovic2004capacity; @farid2009capacity; @lapidoth2009capacity; @you2002capacity; @you2002upper]. In [@kahn2002wireless; @16882; @524287; @780164; @490239; @HranilovicLattice; @karout2010power; @FSOIMDD], the performance of various modulation formats in IM/DD channels were studied using rectangular or other time-disjoint (i.e., infinite-bandwidth) pulses. Hranilovic in [@1577875] pioneered in investigating the problem of designing strictly bandlimited pulses for IM/DD channels with nonnegative pulse-amplitude modulation (PAM) schemes. He showed the existence of nonnegative bandlimited Nyquist pulses, which can be used for ISI-free transmission over IM/DD channels, and evaluated the performance of such pulses. He also showed that any nonnegative root-Nyquist pulse must be time limited (i.e., infinite bandwidth). Hence receivers with matched filters are not suitable for Hranilovic’s signaling method. He concluded that transmission is possible with a bandwidth twice the required bandwidth over the corresponding conventional electrical channels. This work was extended to other Nyquist pulses that can introduce a trade-off between bandwidth and average optical power in [@4132995; @bhandari2007squared].
In this paper, we present a new signaling method for bandlimited IM/DD channels, in which the transmitted signal becomes nonnegative by the addition of a constant direct-current (DC) bias. This method provides us with two benefits: (i) We can transmit ISI-free with a bandwidth equal to that of coherent conventional channels, while benefiting from the reduced complexity and cost of IM/DD system. (ii) We can implement the system using either Nyquist pulses with sampling receiver or root-Nyquist pulses with matched filter receiver. By being able to use a larger variety of pulses, the transmitted power can be reduced compared with known methods, which is advantageous in power-sensitive optical interconnects and indoor wireless optical links. We also evaluate the spectral efficiency and optical power efficiency of binary and 4-PAM formats with Nyquist and root-Nyquist pulses for achieving a specific noise-free eye opening or a specific symbol-error-rate (SER).
The remainder of the paper is organized as follows. Section \[sec:SYSTEM-MODEL\] presents the system model. In Section \[sec:Pulse-shaping\], we define the Nyquist pulses that have been used extensively for conventional bandlimited channels, as well as the ones that have been suggested for nonnegative bandlimited channels. In Section \[sec:RootPulse-shaping\], the root-Nyquist pulses used in this study are introduced. Section \[sec:Required-DC-bias\] discusses a method of computing the required DC bias for a general pulse. Section \[sec:RESULTS\] introduces the performance measures and analyzes the performance of the system under different scenarios. Finally, conclusions are drawn in Section \[sec:CONCLUSIONS\] on the performance of the system.
System Model\[sec:SYSTEM-MODEL\]
================================
In applications such as diffuse indoor wireless optical links and short-haul optical fiber communications, where inexpensive hardware is used, IM/DD is often employed. In such systems, the data is modulated on the optical intensity of the transmitted light using an optical intensity modulator such as a laser diode or a light-emitting diode. This optical intensity is proportional to the transmitted electrical signal. As a result, the transmitted electrical signal must be nonnegative. This is in contrast to conventional electrical channels, where the data is modulated on the amplitude and phase of the carrier . In the receiver, the direct-detection method is used in which the photodetector generates an output which is proportional to the incident received instantaneous power [@780164]. Another limitation, which is considered for safety purposes, is a constraint on the peak and average optical power, or equivalently, a constraint on the peak and average of the signal in the electrical domain [@4132995; @farid2010capacity; @hranilovic2004capacity; @lapidoth2009capacity; @kahn2002wireless]. In this study, we consider the IM/DD transmission system with a strict bandwidth limitation and general $M$-level modulation.
Fig. \[fig:system\] represents the system model for an IM/DD optical transmission system. It can be modeled as an electrical baseband transmission system with additive white Gaussian noise (AWGN) and a nonnegativity constraint on the channel input [@4132995; @kahn2002wireless; @kahn2002experimental; @gfeller2005wireless]. We consider an ergodic source with independent and identically distributed information symbols $a_{k}\in\mathcal{C}$, where $k\in\mathbb{Z}$ is the discrete time instant, and $\mathcal{C}$ is a finite set of constellation points. Based on these symbols, an electrical signal $I(t)$ is generated. The optical intensity modulator converts the electrical signal to an optical signal with optical carrier frequency $f_{\mathrm{c}}$ and random phase $\theta$, given by $O(t)=\sqrt{2x(t)} cos\left(2\pi f_{\mathrm{c}} t+\theta \right)$, where $x(t)$ is the intensity of the optical signal. This intensity is a linear function of $I(t)$ [@kahn2002wireless], given by $$x(t)=JI(t)=JA\left(\mu+\sum_{k=-\infty}^{\infty}a_{k}q(t-k{T_\mathrm{s}})\right),\label{eq:TxSignal}$$ where $J$ is the laser conversion factor, $A$ is a scaling factor that can be adjusted depending on the desired transmitted power, $\mu$ is the required DC bias, $q(t)$ is an arbitrary pulse, and ${T_\mathrm{s}}$ is the symbol duration.
Three requirements are placed on $x(t)$: it should be nonnegative, bandlimited, and ISI-free. The nonnegativity constraint, $x(t)\geq0$ for all $t\in\mathbb{R}$, is fulfilled by choosing $\mu$ in (\[eq:TxSignal\]) sufficiently large, see Sec. \[sec:Required-DC-bias\]. This DC bias is added equally to each symbol to maintain a strictly bandlimited signal $x(t)$, in contrast to works like [@karout2010power; @780164; @HranilovicLattice] in which the bias is allowed to vary with time. The bandwidth constraint is fulfilled by choosing the pulse $q(t)$ such that $$Q(\omega)=\intop_{-\infty}^{\infty}q(t)e^{-j\omega t}dt=0,\; |\omega|\geq 2\pi B,\label{eq:bandwidthLimitation}$$ where $Q(\omega)$ denotes the Fourier transform of $q(t)$. The condition of ISI-free transmission, finally, is fulfilled by either choosing $q(t)$ as a Nyquist pulse, see Sec. \[sec:Pulse-shaping\], when using a sampling receiver, or choosing $q(t)$ as a root-Nyquist pulse (also known as ${T_\mathrm{s}}$-orthogonal pulse), see Sec. \[sec:RootPulse-shaping\], when using a matched filter in the receiver. Fig. \[fig:TxSig\_RC\] illustrates an example of the transmitted intensity given by (\[eq:TxSignal\]) where $\mathcal{C}=\left\{ 0,1\right\} $.
Depending on the application, it is desirable to minimize the average optical power or the peak optical power [@4132995; @hranilovic2004capacity; @lapidoth2009capacity; @kahn2002wireless; @westbergh200932; @inan2009impact]. The average optical power is $$P_{\mathrm{opt}}=\frac{1}{{T_\mathrm{s}}}\intop_{0}^{{T_\mathrm{s}}}\mathbb{E}\left\{ x(t)\right\} dt,\nonumber$$ where $\mathbb{E}\left\{ \cdot\right\} $ denotes expectation, which for the definition of $x(t)$ in (\[eq:TxSignal\]) yields $$\begin{aligned}
P_{\mathrm{opt}} & =\frac{1}{{T_\mathrm{s}}}\intop_{0}^{{T_\mathrm{s}}}JA\left(\mu+\mathbb{E}\left\{ a_{k}\right\}\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})\right)dt\nonumber\\
&=JA\left(\mu+\mathbb{E}\left\{ a_{k}\right\}\overline{q}\right),\label{eq:OpticalPower}\end{aligned}$$ where $$\overline{q}=\frac{1}{{T_\mathrm{s}}}\intop_{-\infty}^{\infty}q(t)dt=\frac{Q(0)}{{T_\mathrm{s}}}.\label{eq:Overbar_q}$$ The peak optical power is$$P_\mathrm{max} = \max x(t)= JA \left(\mu+\max \sum_{k=-\infty}^{\infty} a_k q(t-k{T_\mathrm{s}}) \right)
\label{pmax}$$where the maximum is taken over all symbol sequences $\ldots, a_{-1}, a_0, a_1, a_2,\ldots$ and all times $t$.
The optical signal then propagates through the channel and is detected and converted to the electrical signal [@kahn2002wireless; @hranilovic2004capacity] $$y(t)=Rh(t)\otimes x(t)+n(t),\nonumber$$ where $R$ is the responsivity of the photodetector, $\otimes$ is the convolution operator, $h(t)$ is the channel impulse response, and $n(t)$ is the noise. In this study, the channel is considered to be flat in the bandwidth of interest, i.e., $h(t)=H(0)\delta(t)$. Without loss of generality, we assume that $R=J=1$ [@kahn2002wireless] and $H(0)=1$. Since the thermal noise of the receiver and the shot noise induced by ambient light are two major noise sources in this setup, which are independent from the signal, $n(t)$ can be modeled as a zero-mean AWGN with double-sided power spectral density $N_0/2$ [@kahn2002wireless; @audeh2002performance; @lapidoth2009capacity; @proakis2001digital]. Although the input signal to the channel $x(t)$ must be nonnegative, there is no such constraint on the received signal $y(t)$ [@farid2010capacity].
The received signal passes through a filter with impulse response $g(t)$, resulting in $$r(t)=y(t)\otimes g(t),\label{eq:FilteredSignal}$$ which is then sampled at the symbol rate. In this paper, two scenarios are considered for the receiver filter:
\(i) Similarly to [@1577875; @4132995], $y(t)$ can enter a sampling receiver, which in this paper is assumed to have a rectangular frequency response to limit the power of the noise in the receiver, and is given by $$G(\omega)=\begin{cases}
\begin{array}{cc}
G(0) & |\omega|<2\pi B\\
0 & |\omega|\geq 2 \pi B\end{array}.\end{cases}\label{eq:G(f)}$$ (ii) According to our proposed method, $y(t)$ can enter a matched filter receiver with frequency response $G(\omega)=\zeta Q^{*}(\omega)$ where [$\left( \cdot \right)^*$]{} is the complex conjugate and $\zeta$ is an arbitrary scaling factor. This type of filter will limit the power of the noise, and can also result in ISI-free transmission if the pulses are root-Nyquist (see Sec. \[sec:RootPulse-shaping\]).
The system model introduced in this section is a generalization of the one in [@4132995], which is obtained by considering $\mathcal{C}\subset\mathbb{R}^{+}$ and setting $\mu=0$ in (\[eq:TxSignal\]). If $\mu=0$, the pulse $q(t)$ should be nonnegative to guarantee a nonnegative signal $x(t)$. In our proposed system model, by introducing the bias $\mu$, the nonnegativity condition can be fulfilled for a wider selection of pulses $q(t)$ and constellation $\mathcal{C}\subset\mathbb{R}$.
Bandlimited Nyquist Pulses\[sec:Pulse-shaping\]
===============================================
In order to have ISI-free transmission with a sampling receiver, the pulse $q(t)$ must satisfy the Nyquist criterion [@5055024]. In other words, for any $k\in\mathbb{Z}$ , $$q(k{T_\mathrm{s}})=\begin{cases}
q(0), & k=0,\\
0, & k\neq0.\end{cases}\label{eq:NyquistInTime}$$ The most popular Nyquist pulses are the classical “sinc” pulse, defined as ${\mathop{\mathrm{sinc}}\nolimits}(x)= \sin(\pi x)/(\pi x)$, and the raised-cosine (RC) pulse [@proakis2001digital Sec. 9.2]. Many other Nyquist pulses have been proposed recently for the conventional channel; see [@alexandru11; @assimonis11] and references therein.
In this paper, we evaluate some of these pulses, defined in Table \[pulse-def\], for IM/DD transmission. Our selected pulses are the RC pulse, the so-called “better than Nyquist” (BTN) pulse [@beaulieu2002better], which in [@1369626] was referred to as the parametric exponential pulse, the parametric linear (PL) pulse of first order [@1369626], and one of the polynomial (Poly) pulses in [@chandan05]. Their bandwidth can be adjusted via the parameter $0 \le \alpha \le 1$ such that their lowpass bandwidth is $B=(1+\alpha)/(2{T_\mathrm{s}})$. Since these pulses may be negative, they must be used in a system with $\mu>0$. We denote these four pulses as *regular* Nyquist pulses.
Another option is to use *nonnegative* Nyquist pulses, which satisfy all the three aforementioned constraints. As a result, in (\[eq:TxSignal\]), $\mu=0$ and $q(t)\geq0$ for all $t\in\mathbb{R}$. In [@4132995], it has been shown that pulses that satisfy these requirements must be the square of a general Nyquist pulse. This will result in having pulses with bandwidth twice that of the original Nyquist pulses. Three pulses that satisfy these constraints were introduced in [@4132995], and we use them in our study for compatibility with previous works: squared sinc (S2), squared RC (SRC), and squared double-jump (SDJ), also defined in Table \[pulse-def\]. Their low-pass bandwidth is $B=1/{T_\mathrm{s}}$ for S2 and $B=(1+\alpha)/{T_\mathrm{s}}$ for SRC and SDJ, where $0\le\alpha\le 1$.
Figs. \[fig:TxSig\_RC\] and \[fig:TxSig\_SquareRC\] depict the normalized transmitted signal $x(t)/A$ using the RC and SRC pulses, respectively, assuming $\mathcal{C}= \left\{0,1\right\}$. The most important parameters of the pulses are summarized in Table \[tab:All\_Info\_Nyq\].
Bandlimited Root-Nyquist Pulses\[sec:RootPulse-shaping\]
========================================================
ISI-free transmission is achieved with the pulses in Sec. \[sec:Pulse-shaping\] as long as the input of the sampling unit satisfies the Nyquist criterion given in (\[eq:NyquistInTime\]). In addition to the method of using a Nyquist pulse in the transmitter and a rectangular filter in the receiver, other scenarios can be designed that generate Nyquist pulses at the input $r(t)$ of the sampling unit. In one of these methods, the transmitted pulse is a root-Nyquist pulse, and the receiver contains a filter matched to the transmitted pulse . Consequently, the output of the matched filter will be ISI-free if for any integer $k$ $$\intop_{-\infty}^{\infty}q(t)q(t-k{T_\mathrm{s}})dt=\begin{cases}
E_{{q}} & k=0\\
0 & k\neq0
\end{cases}
\label{eq:Root_Nyquist Criteria},$$ where ${E_q}=\intop_{-\infty}^{\infty}q^{2}(t)dt.$ Tables \[pulse-def\] and \[tab:All\_Info\_Nyq\] also includes two root-Nyquist pulses that have been previously used for conventional coherent channels, where again $0\leq \alpha \leq 1$. These are the root raised cosine (RRC) pulse and the first-order Xia pulse [@xia1997family]. Both have the lowpass bandwidth $B=(1+\alpha)/(2{T_\mathrm{s}})$.
Although the output of the matched filter for both the first order Xia pulse and the RRC pulse are similar ($r(t)$ consists of RC pulses in both cases), the RRC is symmetric in time, whereas the Xia pulse has more energy in the precursor (i.e., the part of the pulse before the peak) [@tan2004transmission]. Moreover, the maximum of Xia pulse does not happen at the origin. The important point with the Xia pulse is that it is both a Nyquist and a root-Nyquist pulse.
In contrast to Nyquist pulses, from which nonnegative Nyquist pulses can be generated by squaring the original pulse (see Sec. \[sec:Pulse-shaping\]), the square of a root-Nyquist pulse is not root-Nyquist anymore. Moreover, [@4132995] has proven that there is no nonnegative root-Nyquist pulse with strictly limited bandwidth.
Required DC Bias \[sec:Required-DC-bias\]
=========================================
Our goal is to find the lowest $\mu$ that guarantees the nonnegativity of $x(t)$. From (\[eq:TxSignal\]) and $x(t)\geq0$, the smallest required DC bias is $$\begin{aligned}
\mu & =-\min_{\forall a,-\infty< t < \infty}\sum_{k=-\infty}^{\infty}a_{k}q(t-k{T_\mathrm{s}}) \label{eq:DC1}\\
&=-\min_{\forall a,-\infty< t < \infty}\sum_{k=-\infty}^{\infty}\left[\left(a_{k}-L\right)q(t-k{T_\mathrm{s}})+Lq(t-k{T_\mathrm{s}})\right] \label{eq:DC2}\end{aligned}$$ where $L=(\hat{a}+\check{a})/2$, $\hat{a}=\max_{a\in \mathcal{C}}a$, and $\check{a}=\min_{a\in \mathcal{C}}a$. The notation $\forall a$ in and means that the minimization should be over all $a_k \in \mathcal{C}$ where $k = \ldots,-1,0,1,2,\ldots$ Going from (\[eq:DC1\]) to (\[eq:DC2\]), we created a factor ($a_{k}-L$) which is a function of $a_{k}$ and symmetric with respect to zero. As a result, the minimum of the first term in (\[eq:DC2\]) occurs if, for all $k$, either $a_{k}=\hat{a}$ and $q(t-k{T_\mathrm{s}})<0$ or $a_{k}=\check{a}$ and $q(t-k{T_\mathrm{s}})>0$. In both cases, due to the fact that the factor $\hat{a}-L=-(\check{a}-L)$, $$\begin{aligned}
\mu &=\max_{0\leq t< {T_\mathrm{s}}}\left[\left(\hat{a}-L\right)\sum_{k=-\infty}^{\infty}\left|q(t-k{T_\mathrm{s}})\right|-L\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})\right].\label{eq:DC}\end{aligned}$$ The reason why (\[eq:DC\]) is minimized over $0\leq t< {T_\mathrm{s}}$ is that $\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})$ and $\sum_{k=-\infty}^{\infty}|q(t-k{T_\mathrm{s}})|$ are periodic functions with period equal to ${T_\mathrm{s}}$. Since for all pulses defined in Sec. \[sec:Pulse-shaping\] and \[sec:RootPulse-shaping\], $q(t)$ rescales with ${T_\mathrm{s}}$ as $q(t)=v(t/{T_\mathrm{s}})$ for some function $v(t)$, then $\mu$ is independent of ${T_\mathrm{s}}$.
To simplify (\[eq:DC\]), Lemma \[LemFS\] and Corollary \[CorrAvval\] will be helpful, since they prove that the second term in (\[eq:DC\]) does not change over time.
\[LemFS\] For an arbitrary pulse $q(t)$, $$\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})=\frac{1}{{T_\mathrm{s}}}\sum_{n=-\infty}^{\infty}Q\left(\frac{2\pi n}{{T_\mathrm{s}}}\right)e^{\frac{j2\pi nt}{{T_\mathrm{s}}}}.$$
Since $f(t)=\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})$ is a periodic function with period ${T_\mathrm{s}}$, it can be expanded as a Fourier series. Its Fourier series coefficients are $$\begin{aligned}
C_{n} & =\frac{1}{{T_\mathrm{s}}}\intop_{-{T_\mathrm{s}}/2}^{{T_\mathrm{s}}/2}f(t)e^{-\frac{j2\pi nt}{{T_\mathrm{s}}}}dt\nonumber \\
&=\frac{1}{{T_\mathrm{s}}}\intop_{-{T_\mathrm{s}}/2}^{{T_\mathrm{s}}/2}\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})e^{-\frac{j2\pi nt}{{T_\mathrm{s}}}}dt.\label{eq:Fourier Coeffs}\end{aligned}$$ Since both $n$ and $k$ are integers, $e^{j2\pi nk}=1$. As a result, (\[eq:Fourier Coeffs\]) can be written as $$\begin{aligned}
C_{n} & =\frac{1}{{T_\mathrm{s}}}\intop_{-{T_\mathrm{s}}/2}^{{T_\mathrm{s}}/2}\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})e^{-\frac{j2\pi n}{{T_\mathrm{s}}}(t-k{T_\mathrm{s}})}dt\nonumber \\
& =\frac{1}{{T_\mathrm{s}}}\intop_{-\infty}^{\infty}q(t)e^{-\frac{j2\pi nt}{{T_\mathrm{s}}}}dt=\frac{1}{{T_\mathrm{s}}}Q\left(\frac{2\pi n}{{T_\mathrm{s}}}\right).\nonumber \end{aligned}$$ Hence, $$f(t)=\sum_{n=-\infty}^{\infty}C_{n}e^{\frac{j2\pi nt}{{T_\mathrm{s}}}}=\frac{1}{{T_\mathrm{s}}}\sum_{n=-\infty}^{\infty}Q\left(\frac{2\pi n}{{T_\mathrm{s}}}\right)e^{\frac{j2\pi nt}{{T_\mathrm{s}}}},\label{eq:bandlimited}$$ which proves the lemma.
The usefulness of this lemma follows from the fact that for bandlimited pulses $q(t)$, (\[eq:bandlimited\]) is reduced to a finite number of terms. As a special case, we have the following corollary.
If $q(t)$ is a bandlimited pulse defined in (\[eq:bandwidthLimitation\]), where $B{T_\mathrm{s}}\leq 1$, then (\[eq:bandlimited\]) can be written as $$f(t)=\sum_{k=-\infty}^{\infty}q(t-k{T_\mathrm{s}})=\frac{1}{{T_\mathrm{s}}}Q(0).\label{eq:coroll1}$$In other words, for such $q(t)$, this sum is not a function of time. \[CorrAvval\]
Since $B{T_\mathrm{s}}\leq 1$, the sum in (\[eq:bandlimited\]) has only one nonzero term (i.e., $Q(0)$ can be nonzero whereas $Q(2\pi n /{T_\mathrm{s}})=0$ for all $n\neq 0$ due to (\[eq:bandwidthLimitation\])).
As a result of Corollary \[CorrAvval\], (\[eq:DC\]) for the regular Nyquist pulses and root-Nyquist pulses considered in Sec. \[sec:Pulse-shaping\] and \[sec:RootPulse-shaping\] (but not SRC and SDJ) can be written as $$\mu=\left(\hat{a}-L\right)\max_{0\leq t< {T_\mathrm{s}}}\sum_{k=-\infty}^{\infty}\left|q(t-k{T_\mathrm{s}})\right|-L\frac{Q(0)}{{T_\mathrm{s}}},\label{eq:DC_New}$$ where $Q(0)=\overline{q}{T_\mathrm{s}}$ for all pulses, see (\[eq:Overbar\_q\]). It appears that solving the summation in (\[eq:DC\_New\]) is impossible analytically even for simple pulses.
For bandlimited pulses where $B{T_\mathrm{s}}\leq 1$, the transmitted signal (\[eq:TxSignal\]) is unchanged if all constellation points in $\mathcal{C}$ are shifted by a constant offset. \[CorrSevom\]
Since the chosen pulse has limited bandwidth given by (\[eq:bandwidthLimitation\]), using (\[eq:coroll1\]) given in Corollary \[CorrAvval\], the transmitted signal (\[eq:TxSignal\]) can be written as $$\begin{aligned}
x(t)=A\left(\mu+\sum_{k=-\infty}^{\infty}\left(a_{k}-L+L\right)q\left(t-k{T_\mathrm{s}}\right)\right)\nonumber\\ = A\left(\mu+\sum_{k=-\infty}^{\infty}\left(a_{k}-L\right)q\left(t-k{T_\mathrm{s}}\right)+L\frac{Q(0)}{{T_\mathrm{s}}}\right)\label{eq:alaki}.\end{aligned}$$ Substituting the required bias given by (\[eq:DC\_New\]), (\[eq:alaki\]) can be written as $$x(t)=A\Bigg(\left(\hat{a}-L\right)\max_{0\leq t< {T_\mathrm{s}}}\left[\sum_{i=-\infty}^{\infty}\left|q(t-i{T_\mathrm{s}})\right|\right]+\sum_{k=-\infty}^{\infty}\left(a_{k}-L\right)q\left(t-k{T_\mathrm{s}}\right)\Bigg).\label{eq:symmetric}$$ It can be seen that (\[eq:symmetric\]) only depends on symbols through $\hat{a} - L$ and $a_{k}-L$. Both terms are independent of the constellation offset.
Theorem \[CorrSevom\] shows that for narrow-band pulses defined in (\[eq:bandwidthLimitation\]), the constellation offset does not have an effect on the performance. This result which holds for intensity modulated channels (with nonnegative transmitted signal requirement) is in contrast to the standard result for conventional channels. For instance, binary phase-shift keying (BPSK) and on-off keying (OOK) are equivalent in this IM/DD system, whereas BPSK is 3 dB better over the conventional AWGN channel [@proakis2001digital Sec. 5]. Fig. \[fig:RequiredDC\] illustrates the required DC bias (\[eq:DC\_New\]) for various pulses considering any nonnegative $M$-PAM constellation ($\mathcal{C}=\left\{ 0,1,...,M-1\right\} $). In case of Nyquist pulses, due to the fact that by increasing $\alpha$, the ripples of the pulses decrease, the required DC bias decreases as well. It can be seen that the Poly and RC pulses always require more DC bias than other Nyquist pulses. The PL and BTN pulses require approximately the same DC bias. The BTN pulse requires slightly less DC bias in $0.250\le \alpha \le 0.256$, $0.333\le \alpha \le 0.363$, and $0.500 \le \alpha \le 0.610$, while the PL is better for all other roll-off factors in the range $0 < \alpha < 1$.
The RRC pulse has a different behavior. For $0< \alpha \le 0.420$, similar to Nyquist pulses, by increasing the roll-off factor, the required DC bias decreases, and is approximately equal to the required DC bias for BTN and PL. However, when $0.420\le \alpha < 1$, the required DC bias starts to fluctuate slightly around $\mu=0.25\hat{a}$ and the minimum happens for $\alpha =0.715$. The reason for this behavior is that in RRC, the peak is a function of $\alpha$, see Table \[pulse-def\]. As a result, by increasing the roll-off factor, there will be a compromise between the reduction in the sidelobe amplitude and the increase in peak amplitude. For small values of $\alpha$, the sidelobe reduction is more significant than the peak increase, and as a result, the required DC bias decreases. The Xia pulse always requires the largest DC bias. For $0< \alpha \le 0.730$, similar to other pulses, by increasing the roll-off factor, the required DC bias for Xia pulses decreases. However, when $0.730\le \alpha < 1$, the required DC bias starts to fluctuate slightly and starts to approach the required DC for RRC.
The expression for $\mu$ given in (\[eq:DC\]) illustrates the reason why the double-jump and sinc pulses are not considered in Sec. \[sec:Pulse-shaping\]. These pulses decay as $1/|t|$. As a result, the summation in (\[eq:DC\]) does not converge to a finite value. Hence, they require an infinite amount of DC bias to be nonnegative.
Analysis and Results\[sec:RESULTS\]
===================================
Received Sequence for Sampling Receiver\[subsec:RxSequence\_Nyq\]
-----------------------------------------------------------------
Considering the assumptions mentioned in Sec. \[sec:SYSTEM-MODEL\], the received signal (\[eq:FilteredSignal\]) is $$\begin{aligned}
r(t) &=\left(x(t)+n(t)\right)\otimes g(t) \nonumber\\
&=A\left(\mu+\sum_{k=-\infty}^{\infty}a_{k}q(t-k{T_\mathrm{s}})\right)\otimes g(t)+z(t) \nonumber\\
& =AG(0)\left[\mu+\sum_{k=-\infty}^{\infty}a_{k}q(t-k{T_\mathrm{s}})\right]+z(t),\label{eq:receivedSignal}\end{aligned}$$ where (\[eq:receivedSignal\]) holds since $g(t)$ has a flat frequency response given by (\[eq:G(f)\]) over the bandwidth of $q(t)$ given by (\[eq:bandwidthLimitation\]); Therefore, the convolution has no effect on $x(t)$. The noise at the output of the receiver filter, which is given by $z(t)=n(t)\otimes g(t)$, is zero mean additive white Gaussian with variance $\sigma_{z}^2=G(0)^2N_{0}B$.
Applying the Nyquist criterion given in (\[eq:NyquistInTime\]) to the sampled version of (\[eq:receivedSignal\]), we can write the $i$-th filtered sample as $$r(i{T_\mathrm{s}})=AG(0)\left[\mu+a_{i}q(0)\right]+z(i{T_\mathrm{s}}).\label{eq:RecSamplesEQ}$$ for any constellation $\mathcal{C}$. The received waveform $r(t)$, for several Nyquist pulses, is shown in Fig. \[eye-diagrams\], in the form of eye diagrams in a noise-free setting ($z(t) = 0$). As expected, the output samples $r(i {T_\mathrm{s}})$ are ISI-free.
Received Sequence for Matched Filter Receiver\[subsec:RxSequence\_MF\]
----------------------------------------------------------------------
Similar to Sec. \[subsec:RxSequence\_Nyq\], the received signal will be $$\begin{aligned}
\vspace{-3mm}
r(t) & =\left(x(t)+n(t)\right)\otimes g(t) \nonumber \\
& =A\left(\mu+\sum_{k=-\infty}^{\infty}a_{k}q(t-k{T_\mathrm{s}})\right)\otimes \zeta q(-t)+u(t) \nonumber \\
& =A\zeta \Big(\mu\intop_{-\infty}^{\infty}q(-t)dt+\sum_{k=-\infty}^{\infty}a_{k}\intop_{-\infty}^{\infty}q(\tau-k{T_\mathrm{s}})q(\tau-t)d\tau\Big)+u(t) \nonumber \\
& =A\zeta \left(\mu Q(0)+\sum_{k=-\infty}^{\infty}a_{k}\intop_{-\infty}^{\infty}q(\tau)q(\tau-t+k{T_\mathrm{s}})d\tau\right)+u(t)\label{eq:MF_Output}\end{aligned}$$ where $u(t)$ is zero mean additive white Gaussian noise with variance $\sigma_{u}^{2}=\zeta^2N_{0}{E_q}/2$. Applying the root-Nyquist criterion given in (\[eq:Root\_Nyquist Criteria\]) to the sampled version of (\[eq:MF\_Output\]), the $i$-th filtered sample will be, for any constellation $\mathcal{C}$, $$r(i{T_\mathrm{s}})=A\zeta \left(\mu Q(0)+a_{i}{E_q}\right)+u(i{T_\mathrm{s}}).\label{eq:Rec_MF}$$
Comparison Between Pulses
-------------------------
As mentioned in Sec. \[sec:SYSTEM-MODEL\], it may be desirable to minimize the average or peak optical power. The next theorem shows that these two criteria are equivalent for narrow-band pulses ($B{T_\mathrm{s}}< 1$) and symmetric constellations ($\mathbb{E}\{a_k\} = L$).
\[th:pmax\] If $B{T_\mathrm{s}}< 1$ and $\mathbb{E}\{a_k\} = L$, then $P_\mathrm{max} = 2 P_\mathrm{opt}$.
From and Corollary \[CorrAvval\], $$\begin{aligned}
P_{\mathrm{max}} &=JA\left(\mu+\max_{\forall a,-\infty< t < \infty}\sum_{k=-\infty}^{\infty}\left[\left(a_{k}-L\right)q(t-k{T_\mathrm{s}})+Lq(t-k{T_\mathrm{s}})\right]\right)\\
&=JA\left(\mu+\max_{\forall a,-\infty< t < \infty}\left[\sum_{k=-\infty}^{\infty}\left(a_{k}-L\right)q(t-k{T_\mathrm{s}})+\frac{LQ(0)}{{T_\mathrm{s}}}\right]\right).\end{aligned}$$ In analogy to , the maximum is $$\begin{aligned}
P_{\mathrm{max}} &=JA\left(\mu+(\hat{a}-L)\max_{0\leq t<{T_\mathrm{s}}} \sum_{k=-\infty}^{\infty}|q(t-k{T_\mathrm{s}})|+\frac{LQ(0)}{{T_\mathrm{s}}}\right) =JA\left(2\mu+2\frac{LQ(0)}{{T_\mathrm{s}}}\right)\end{aligned}$$ which compared with completes the proof.
To compare the optical power of various pulses, a criterion called optical power gain is used, which is defined as[@4132995] $$\Upsilon=10\log_{10}\left(\frac{P_{\mathrm{opt}}^{\mathrm{ref}}}{P_{\mathrm{opt}}}\right),\nonumber$$ where $P_{\mathrm{opt}}^{\mathrm{ref}}$ is the average optical power for a reference system. (According to Theorem \[th:pmax\], $\Upsilon$ would be the same if defined in terms of $P_\mathrm{max}$, for all pulses in our study except SRC and SDJ.) Similarly to [@1577875], this reference is chosen to be the S2 pulse with OOK modulation and sampling receiver, for which no bias is needed. Using (\[eq:OpticalPower\]), $P_{\mathrm{opt}}^{\mathrm{ref}}=A_{\mathrm{ref}}\mathbb{E}_{\mathrm{ref}}\left\{a_k\right\}$ and $$\Upsilon=10\log_{10}\left(\frac{A_{\mathrm{ref}}\mathbb{E}_{\mathrm{ref}}\left\{a_k\right\}}{A\left(\mu+\mathbb{E}\left\{a_k\right\}\overline{q}\right)}\right)\label{eq:OptGainGeneral}$$ where $A_{\mathrm{ref}}$ and $\mathbb{E}_{\mathrm{ref}}\left\{a_{k}\right\}$ are the scaling factor and the symbol average for the reference system, respectively. Defining $$\Delta a=\min_{a,a'\in\mathcal{C},a\neq a'}\left|a-a'\right|\label{eq:MinDistance}$$ as the minimum distance between any two constellation points $a$ and $a'$, $\mathbb{E}_{\mathrm{ref}}\left\{a_k\right\}=\Delta a_{\mathrm{ref}}/2$, where $\Delta a_{\mathrm{ref}}$ is the minimum distance for the reference system. The expressions in (\[eq:OptGainGeneral\]) and (\[eq:MinDistance\]) hold in general for all finite set of constellation points $\mathcal{C}$.
Initially, we compare the pulses in a noise-free setting. For any Nyquist pulse with a sampling receiver, the minimum eye opening after filtering is given by (\[eq:RecSamplesEQ\]) as $$\min_{ a,a'\in \mathcal{C}, a\neq a'} \left|AG(0)\left(\mu+aq(0)\right)-AG(0)\left(\mu+a'q(0)\right)\right|=AG(0)\Delta a q(0). \label{eq:MinEye_Sample}$$ As a result, to have equal eye opening we require ${A_{\mathrm{ref}}}/{A}={\Delta a q(0)}/{\Delta a_{\mathrm{ref}}}$, which substituted into yields $$\Upsilon=10\log_{10}\left(\frac{\Delta a q(0)}{\mu+\mathbb{E}\left\{a_k\right\}\overline{q}}\right).\label{eq:OptGainEye}$$
Fig. \[fig:OptPower\_Eye\] demonstrates the comparison of the optical power gain for various pulses defined in Sec. \[sec:Pulse-shaping\] for both OOK and 4-PAM formats, where the signals are scaled to have equal eye opening. The S2 pulse with OOK modulation, which is used as a baseline for comparison, is shown in the figure with an arrow. The results for SRC and SDJ have been derived before in [@4132995 Fig. 4], whereas the results for other pulses are novel, where $T_{\mathrm{b}}={T_\mathrm{s}}/\log_{2}M$ is the bit rate. OOK is chosen rather than BPSK for compatibility with [@4132995], although these binary formats are entirely equivalent for $BT_{\mathrm{b}}\leq 1$, as shown in Theorem \[CorrSevom\]. In these examples, we use $\Delta a = \Delta a_\mathrm{ref}$; however, rescaling the considered constellation $\mathcal{C}$ would not change the results, as it would affect the numerator and denominator of equally.
For the nonnegative pulses in Sec. \[sec:Pulse-shaping\] (i.e., SRC and SDJ) with OOK, where $\mu=0$, by increasing the bandwidth, the optical power gain, which depends on $\alpha$ through its dependence on $\overline{q}$, increases since $\overline{q}$ decreases. The results in Fig. \[fig:OptPower\_Eye\] are consistent with [@4132995 Fig. 4], where the same nonnegative pulses were presented. It can be seen that when the regular Nyquist pulses (RC, BTN, PL, and Poly) are used, and the nonnegativity constraint is satisfied by adding a DC bias, transmission is possible over a much narrower bandwidth. However, since the DC bias consumes energy and does not carry information, the optical power gain will be reduced.
There is a compromise between bandwidth and optical power gain, due to the fact that $\mu$ will be reduced by increasing the roll-off factor (see Fig. \[fig:RequiredDC\]), whereas the required bandwidth increases. The highest optical power gain for all pulses will be achieved when the roll-off factor $\alpha$ is one. The reason is that by increasing the roll-off factor, the required bias, which is the only parameter in (\[eq:OptGainEye\]) that depends on $\alpha$, decreases. The BTN and the PL pulses have approximately similar optical power gain, and the Poly and RC pulses have smaller gains, due to higher $\mu$, which is also visible in the eye diagrams of Fig. \[eye-diagrams\]. Comparing the binary and 4-PAM cases for the same $\alpha$ and $\Delta a$, we can see in Fig. \[fig:OptPower\_Eye\] that by using higher-order modulation formats, the optical power gain for all pulses decreases, since in (\[eq:OptGainEye\]), $\mathbb{E}\left\{a_{k}\right\}$ and $\mu$ will increase. For $0.5<BT_{\mathrm{b}}<1$, the optical power gain for the best 4-PAM system with nonnegative Nyquist pulses is up to 2.39 dB less than the gain of the best OOK system with regular Nyquist pulses.
For any root-Nyquist pulse with a matched filter receiver, the minimum eye opening after filtering is given by (\[eq:Rec\_MF\]) as $$\min_{a,a'\in \mathcal{C}, a\neq a'} \left|A \zeta\left(\mu Q(0)+a{E_q}\right)-A\zeta\left(\mu Q(0)+a'{E_q}\right)\right|=A\zeta\Delta a {E_q}.\label{eq:MinEye_MFR}$$ Since the eye openings in and depend on the receiver filter gains $G(0)$ or $\zeta$, pulses should be compared using the same receiver filter. In particular, it is not relevant to compare the sampling receiver with matched filters in this context, since the outcome would depend on the ratio $G(0)/\zeta$, which can be chosen arbitrarily. This is the reason why root-Nyquist pulses are not included in Fig. \[fig:OptPower\_Eye\]. It appears from Fig. \[fig:OptPower\_Eye\] that the studied pulses become more power-efficient when the bandwidth is increased. A higher bandwidth, however, for sampling receiver means that the receiver filter admits more noise, which reduces the receiver performance. In Fig. \[fig:OptPower\_BER\], we therefore compare the average optical power gain of Nyquist and root-Nyquist pulses, when the power is adjusted to yield a constant SER equal to $10^{-6}$. Since the amount of noise after the matched filter receiver does not depend on the bandwidth, we considered this fact as a potential advantage, and therefore included root-Nyquist pulses in the following analysis. Similarly to the previous case, the S2 pulse with OOK and sampling receiver is used as a baseline for comparison.
So far the analysis holds for a general $\mathcal{C}$. To find the optical power gain as a function of SER for the sampling receiver, we first apply a maximum likelihood detector to (\[eq:RecSamplesEQ\]), assuming a special case in which $\mathcal{C}$ is an $M$-PAM constellation, which yields the SER $$P_{\mathrm{err}}=2\frac{M-1}{M}Q\left(\frac{AG(0)\Delta a q(0)}{2\sqrt{G(0)^2N_{0}B}}\right)$$where $$Q(x)=\frac{1}{\sqrt{2\pi}}\intop_{x}^{\infty}\exp\left(\frac{-x^{2}}{2}\right)dx$$ is the Gaussian Q-function. As a result, $$A=\frac{2}{\Delta a q(0)}Q^{-1}\left(P_{\mathrm{err}}\frac{M}{2\left(M-1\right)}\right)\sqrt{N_{0}B}$$ and $$\frac{A_{\mathrm{ref}}}{A}=\frac{\Delta a q(0)}{\Delta a_{\mathrm{ref}}}\frac{Q^{-1}\left(P_{\mathrm{err}}\right)}{Q^{-1}\left(P_{\mathrm{err}}\frac{M}{2\left(M-1\right)}\right)}\sqrt{\frac{B_{\mathrm{ref}}}{B}},\label{eq:Ratio_Sampl}$$ where $B_{\mathrm{ref}}=1/T_\mathrm{b}$ is the bandwidth of the reference pulse. The optical power gain now follows from (\[eq:OptGainGeneral\]).
For the matched filter receiver, by applying the maximum likelihood detector to (\[eq:Rec\_MF\]), the SER will be $$\begin{aligned}
P_{\mathrm{err}}=2\frac{M-1}{M}Q\left(\frac{A\Delta a{E_q}\zeta}{2\sqrt{\frac{\zeta^2N_{0}{E_q}}{2}}}\right)\nonumber\\
\qquad=2\frac{M-1}{M}Q\left(A\Delta a\sqrt{\frac{{E_q}}{2N_{0}}}\right).\nonumber\end{aligned}$$ As a result, $$A=\frac{1}{\Delta a}Q^{-1}\left(P_{\mathrm{err}}\frac{M}{2\left(M-1\right)}\right)\sqrt{\frac{2N_{0}}{{E_q}}}$$and $$\frac{A_{\mathrm{ref}}}{A}=\frac{\Delta a}{\Delta a_{\mathrm{ref}}}\frac{\sqrt{2}Q^{-1}\left(P_{\mathrm{err}}\right)}{Q^{-1}\left(P_{\mathrm{err}}\frac{M}{2\left(M-1\right)}\right)}\sqrt{{E_q}B_{\mathrm{ref}}}.\label{eq:Ratio_MFRx}$$
In contrast to the case with equal eye openings (see Fig. \[fig:OptPower\_Eye\]), Nyquist and root-Nyquist pulses can be compared with each other when the SER is kept constant, since neither (\[eq:Ratio\_Sampl\]) nor (\[eq:Ratio\_MFRx\]) depend on the filter gains $G(0)$ and $\zeta$.
By increasing the bandwidth, the gain for SRC decreases slightly, whereas it increases for SDJ, where $\mu=0$ for both cases. The reason is that for these pulses by increasing $\alpha$, both $\overline{q}$ and the ratio $A_{\mathrm{ref}}/A$ decreases. We observe that for the regular Nyquist pulses in Sec. \[sec:Pulse-shaping\], the gain increases by increasing the bandwidth. The reason is that by increasing the roll-off factor, the required bias decreases much faster (see Fig. \[fig:RequiredDC\]) than the speed of increase in bandwidth. The BTN and the PL pulses have approximately similar gain, and the gain of the RC and Poly pulses are always smaller than the gain of the other two pulses. In case of the matched filter receiver, the noise variance does not depend on bandwidth. As a result, the ratio $A_{\mathrm{ref}}/A$ in (\[eq:Ratio\_MFRx\]) is not a function of the roll-off factor and the optical power gain only depends on the roll-off factor through its dependence on the required DC bias. In Fig. \[fig:OptPower\_BER\], the optical power gain of the RRC pulse increases for $0.5<BT_{\mathrm{b}}\le 0.71$, and a wide gap is maintained with respect to the Nyquist pulses. For $0.71<BT_{\mathrm{b}}\le 1$, since the required DC is slightly fluctuating, the same happens for the optical power gain of RRC, and the maximum optical power gain happens at $BT_{\mathrm{b}}=0.86$, where it is $\Upsilon = -0.22$ dB. The Xia pulse has a similar behavior, though it is not better than all Nyquist pulses.
For $\alpha\rightarrow1$, the optical power gain of the Xia, RC, and RRC pulses are approximately equal. In this case, although the output of matched filter will be equal to an RC pulse by either using RRC or Xia pulse, the performance will be different for other values of $\alpha$. By increasing the modulation level from binary to 4-PAM, for the same $\alpha$ and $\Delta a$, the optical power gain for all pulses decrease, since the required DC bias and symbol average increase while the ratio $A_{\mathrm{ref}}/A$ decreases. For $0.5<BT_{\mathrm{b}} < 1$, the optical power gain of the regular Nyquist pulses and root-Nyquist pulses with OOK modulation is significantly more than the gain for the all nonnegative Nyquist pulses with 4-PAM.
When the roll-off factor is equal to zero (i.e., the normalized bandwidth $BT_{\mathrm{b}}$ for the biased pulses with binary modulation is equal to 0.5 and for the biased pulses with 4-PAM is equal to 0.25), the regular Nyquist pulses discussed in Sec. \[sec:Pulse-shaping\] and the root-Nyquist pulses in Sec. \[sec:RootPulse-shaping\] will become equal to a sinc pulse with bandwidth $1/(2{T_\mathrm{s}})$. As discussed in Sec. \[sec:Required-DC-bias\], the required DC will be infinite for the sinc pulse. Hence, the gain $\Upsilon$ will asymptotically go to $-\infty$ when $\alpha\rightarrow0$.
Conclusions \[sec:CONCLUSIONS\]
===============================
In this work, a pulse shaping method for strictly bandlimited IM/DD systems is presented, in which the transmitted electrical signal must be nonnegative. The proposed approach adds a constant DC bias to the transmitted signal, which allows a wider selection of transmitted pulses without violating the nonnegativity constraint. This allows us to use Nyquist or root-Nyquist pulses for ISI-free transmission, with narrower bandwidth compared to previous works. It is possible to transmit with a bandwidth equal to that of ISI-free transmission in conventional coherent channels. To compare our proposed transmission schemes with previously designed schemes and to see the effect of increasing the modulation level, we evaluated analytically the average optical power versus bandwidth in two different scenarios. The optimization of modulation formats means a tradeoff between the two components of the optical power: the constellation power, which carries the data and is similar to the coherent case, and the bias power, which is constant. We prove the somewhat unexpected results that for narrowband transmission ($B{T_\mathrm{s}}\le 1 $), the two powers balance each other perfectly, so that OOK and BPSK have identical performance regardless of the pulse.
In the first scenario, the Nyquist pulses are compared when the noise-free eye opening is equal for all the pulses and modulation formats. Of the studied pulses, the SDJ pulse with OOK is the best known, as previously shown in [@4132995] over $B T_\mathrm{b}\ge 1$. At $0.5<B T_\mathrm{b}< 1$, the PL and BTN pulses with binary modulation have the best performance, being up to $2.39$ dB better than SDJ with 4-PAM modulation. Similarly, the 4-PAM BTN and PL pulses have highest gain over $0.25<B T_\mathrm{b}< 0.5$.
In the second scenario, all pulses have equal SER. Of the studied pulses, the SDJ with OOK modulation and sampling receiver has the highest gain for $B T_\mathrm{b}\ge 1$. At $0.869<B T_\mathrm{b}< 1$, the binary PL pulse has the best performance, whereas for $0.5<B T_\mathrm{b}\leq 0.869$, the RRC pulse with matched filter receiver achieves the highest gain. The gain of RRC in this scenario is up to $0.74$ dB over the best Nyquist pulse and $2.80$ dB over the best known results with unbiased PAM. It seems possible that further improvements can be achieved by utilizing the most recently proposed Nyquist pulses [@1369626; @beaulieu2002better; @chandan05; @alexandru11; @assimonis11], or their corresponding root-Nyquist pulses, and carefully optimizing their parameters.
Extensions to $M$-PAM systems with $M>4$ are straightforward, in order to gain even more spectral efficiency at the cost of reduced power efficiency. This might be important for designing power- and bandwidth-efficient short-haul optical fiber links (e.g., fiber to the home and optical interconnects) [@4528765; @5875697] and diffuse indoor wireless optical links [@gfeller2005wireless; @kahn2002wireless; @hranilovic2005design].
{width="1\linewidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Pulse Definition $q(t)$
------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
RC $\begin{cases}
\frac{\pi}{4}{\mathop{\mathrm{sinc}}\nolimits}\left(\frac{t}{{T_\mathrm{s}}}\right),& t = \pm \frac{{T_\mathrm{s}}}{2\alpha}, \\
{\mathop{\mathrm{sinc}}\nolimits}\left(\frac{ t}{{T_\mathrm{s}}}\right) \frac{\cos\left(\frac{\pi\alpha t}{{T_\mathrm{s}}}\right)}{1-(\frac{2\alpha t}{{T_\mathrm{s}}})^{2}},& \text{otherwise}
\end{cases}$
BTN ${\mathop{\mathrm{sinc}}\nolimits}\left(\frac{t}{{T_\mathrm{s}}}\right)\frac{\frac{2\pi\alpha t}{{T_\mathrm{s}}\ln2}\sin\left(\frac{\pi\alpha t}{{T_\mathrm{s}}}\right)+2\cos\left(\frac{\pi\alpha t}{{T_\mathrm{s}}}\right)-1}{\left(\frac{\pi\alpha t}{{T_\mathrm{s}}\ln2}\right)^2+1}$
PL ${\mathop{\mathrm{sinc}}\nolimits}\left(\frac{t}{{T_\mathrm{s}}}\right){\mathop{\mathrm{sinc}}\nolimits}\left(\frac{\alpha t}{{T_\mathrm{s}}}\right)$
Poly $\begin{cases}
1,& t = 0, \\
3{\mathop{\mathrm{sinc}}\nolimits}\left(\frac{ t}{{T_\mathrm{s}}}\right) \frac{{\mathop{\mathrm{sinc}}\nolimits}\left(\frac{\alpha t}{2{T_\mathrm{s}}}\right)^2-{\mathop{\mathrm{sinc}}\nolimits}\left(\frac{\alpha t}{{T_\mathrm{s}}}\right)}{\left(\frac{\pi\alpha t}{2{T_\mathrm{s}}}\right)^{2}},& \text{otherwise}
\end{cases}$
S2 ${\mathop{\mathrm{sinc}}\nolimits}^{2}\left(\frac{ t}{{T_\mathrm{s}}}\right)$
SRC $q_{\mathrm{RC}}^2(t)$, where $q_\mathrm{RC}$ is the RC pulse defined above
SDJ $\left[\left(\frac{1-\alpha}{2}\right){\mathop{\mathrm{sinc}}\nolimits}\left(\frac{(1-\alpha) t}{{T_\mathrm{s}}}\right)+\left(\frac{1+\alpha}{2}\right){\mathop{\mathrm{sinc}}\nolimits}\left(\frac{(1+\alpha) t}{{T_\mathrm{s}}}\right)\right]^{2}$
RRC $\begin{cases}
1-\alpha+\frac{4\alpha}{\pi}, & t=0,\\
\frac{\alpha}{\sqrt{2}}\left[(1+\frac{2}{\pi})\sin(\frac{\pi}{4\alpha})+(1-\frac{2}{\pi})\cos(\frac{\pi}{4\alpha})\right], & t=\pm\frac{{T_\mathrm{s}}}{4\alpha},\\
\frac{\sin\left(\frac{\pi(1-\alpha)t}{{T_\mathrm{s}}}\right)+\frac{4\alpha t}{{T_\mathrm{s}}}\cos\left(\frac{\pi(1+\alpha)t}{{T_\mathrm{s}}}\right)}{\frac{\pi t}{{T_\mathrm{s}}}\left(1-\left(\frac{4\alpha t}{{T_\mathrm{s}}}\right)^{2}\right)}, & \text{otherwise}\end{cases}$
Xia ${\mathop{\mathrm{sinc}}\nolimits}\left(\frac{t}{{T_\mathrm{s}}}\right)\frac{\cos\left(\frac{\pi \alpha t}{{T_\mathrm{s}}}\right)}{\frac{2\alpha t}{{T_\mathrm{s}}}+1}$
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Pulse Nyquist root-Nyquist $\overline{q}$ $q(0)$ $B {T_\mathrm{s}}$ ${E_q}/{T_\mathrm{s}}$
------- -------------- -------------- ---------------- ------------------------ -------------------- ------------------------
RC $\checkmark$ $1$ $1$ $(1+\alpha)/2$
BTN $\checkmark$ $1$ $1$ $(1+\alpha)/2$
PL $\checkmark$ $1$ $1$ $(1+\alpha)/2$
Poly $\checkmark$ $1$ $1$ $(1+\alpha)/2$
S2 $\checkmark$ $1$ $1$ $1$
SRC $\checkmark$ $1-\alpha/4$ $1$ $1+\alpha$
SDJ $\checkmark$ $1-\alpha/2$ $1$ $1+\alpha$
RRC $\checkmark$ $1$ $1-\alpha+4\alpha/\pi$ $(1+\alpha)/2$ $1$
Xia $\checkmark$ $\checkmark$ $1$ $1$ $(1+\alpha)/2$ $1$
![\[fig:TxSig\_RC\]The normalized transmitted signal $x(t)/A$ for $\mathcal{C}= \left\{0,1\right\}$ and using an RC pulse with $\alpha=0.6$ as $q(t)$. It can be seen that without using the bias $\mu=0.184$ (see Fig. \[fig:RequiredDC\]), the RC pulse would create a signal $x(t)$ that can be negative.](TransSignal_RC){width="100.00000%"}
![\[fig:TxSig\_SquareRC\]The normalized transmitted signal $x(t)/A$ for $\mathcal{C}= \left\{0,1\right\}$ and using an SRC pulse with $\alpha=0.6$ as $q(t)$. In this case, the required DC $\mu$ is zero.](TransSignal_SquareRC){width="100.00000%"}
![\[fig:RequiredDC\]The normalized minimum DC bias $\mu/\hat{a}$ vs. roll-off factor $\alpha$ for a variety of pulses and $M$-PAM. The dotted line represents the required bias for the RC pulse at $\alpha=0.6$, see Fig. \[fig:TxSig\_RC\].](RequiredDC){width="100.00000%"}
\[t\]\[t\][$t/{T_\mathrm{s}}$]{}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[![Noise-free eye diagrams for (a) RC, (b) PL, (c) BTN, and (d) Xia pulses with OOK modulation ($\mathcal{C} = \{0,1\}$) and sampling receiver. All pulses have $\alpha=0.60$ and are normalized to have the same optical power $\bar{q}=1$.[]{data-label="eye-diagrams"}](Eye_Raised_Cosine "fig:"){width="7cm"}]{} [![Noise-free eye diagrams for (a) RC, (b) PL, (c) BTN, and (d) Xia pulses with OOK modulation ($\mathcal{C} = \{0,1\}$) and sampling receiver. All pulses have $\alpha=0.60$ and are normalized to have the same optical power $\bar{q}=1$.[]{data-label="eye-diagrams"}](Eye_ParametricLinear "fig:"){width="7cm"}]{}
\[-2ex\] (a) (b)
[![Noise-free eye diagrams for (a) RC, (b) PL, (c) BTN, and (d) Xia pulses with OOK modulation ($\mathcal{C} = \{0,1\}$) and sampling receiver. All pulses have $\alpha=0.60$ and are normalized to have the same optical power $\bar{q}=1$.[]{data-label="eye-diagrams"}](Eye_BetterNyq "fig:"){width="7cm"}]{} [![Noise-free eye diagrams for (a) RC, (b) PL, (c) BTN, and (d) Xia pulses with OOK modulation ($\mathcal{C} = \{0,1\}$) and sampling receiver. All pulses have $\alpha=0.60$ and are normalized to have the same optical power $\bar{q}=1$.[]{data-label="eye-diagrams"}](Eye_Xia_Unmatched "fig:"){width="7cm"}]{}Ê
\[-2ex\] (c) (d)
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![\[fig:OptPower\_Eye\]The optical power gain $\Upsilon$ versus normalized bandwidth $BT_{\mathrm{b}}$ for various Nyquist pulses with a sampling receiver. The noiseless eye opening for all pulses is equal. The curves for $BT_{\mathrm{b}}\geq 1$ agree with [@4132995].](OptPower_Eye){width="100.00000%"}
![\[fig:OptPower\_BER\]The optical power gain versus normalized bandwidth $BT_{\mathrm{b}}$ for various pulses with a sampling receiver (S) or matched filter receiver (MF). The SER for all pulses is $10^{-6}$.](OptPower_BER){width="100.00000%"}
[^1]: Research supported in part by the Swedish Foundation for Strategic Research (SSF) under grant RE07-0026. The material in this paper was presented in part at the IEEE Global Communications Conference, Houston, TX, Dec. 2011.
M. Tavan is with the Wireless Information Network Laboratory (WINLAB), Department of Electrical and Computer Engineering, Rutgers University, Piscataway, NJ 08854 USA (e-mail: mt579@eden.rutgers.edu). E. Agrell and J. Karout are with the Department of Signals and Systems, Chalmers University of Technology, SE-41296 Gothenburg, Sweden (e-mail: agrell@chalmers.se, johnny.karout@chalmers.se).
|
---
abstract: 'We present new algorithm for growth of non-clustered planar graphs by aggregation of cells with given distribution of size and constraint of connectivity $k = 3$ per node. The emergent graph structures are controlled by two parameters—chemical potential of the cell aggregation and the width of the cell size distribution. We compute several statistical properties of these graphs—fractal dimension of the perimeter, distribution of shortest paths between pairs of nodes and topological betweenness of nodes and links. We show how these topological properties depend on the control parameters of the aggregation process and discuss their relevance for the conduction of current in self-assembled nanopatterns.'
author:
- Milovan Šuvakov and Bosiljka Tadić
title: 'Topology of Cell-Aggregated Planar Graphs'
---
Introduction
============
In recent years increased interests in various networks realizations [@SD_book; @nets_review] revealed that several new types of graphs termed [*structured graphs*]{} are more appropriate mathematical objects to describe complex network’s geometry than traditional [*random graphs*]{} [@BB_book]. The variety of structures was found to emerge through evolution processes in which nodes and links are added sequentially according to specified rules, in particular, the preferential attachment rules lead to strongly inhomogeneous [*scale-free graphs*]{} [@SD_book]. In contrast to the evolving networks, which comprise a class of [*causal*]{} graphs, the class of [*homogeneous*]{} graphs consists of graphs with fixed number of nodes and fluctuating or rewiring links according to given rules or certain optimization processes. Complex graph structures may emerge in this procedures, especially when certain global or local optimization constraints are imposed [@Stefan].
Planar graphs are special class of graphs that can be embedded in a Euclidean plane. A graph is planar [*iff it does not contain a subdivision of $K_5$ (5-clique) and $K_{3,3}$ (minimal non-planar graph with 6 nodes)*]{} [@BB_book]. Consequently, planar graphs fulfill Euler’s law: $N_p+N=E+1$, which is relating the number of nodes $N$, links $E$ and polygons $N_p$.
In this work we suggest a new method for growing a planar cellular graph by attachment of objects—cells (polygons) of length $n_p$, which are chosen from a given distribution $f(n_p)$. The polygons are added sequentially in time starting from an initial polygon. In addition, we strictly impose the constraint on number of links per node $k=3$, which is thus fulfilled everywhere in the interior of the graph and on some nodes on the graph boundary. The attachment of cells is controlled by two parameters—the width of the distribution of cell sizes $\mu_2$ and the parameter $\nu$ that plays the role of chemical potential of cell aggregation. In the limit of vanishing attachment potential $\nu \to 0$ the growth process resembles the one in diffusion-limited aggregation [@DLA]. However, aggregated are spatially extended cells of particles rather than single particles.
Emergent structures of cellular networks are resembling of soap froths [@Stabans_froth] or patterns of nano-particles self-assembled through nonlinear dynamic processes [@SA_book; @Philip_SA]. Typically, a pattern of cells appears when nano-particles are immersed in a liquid film, which is then allowed to evaporate until holes of different sizes open-up leaving particles in the walls between the holes [@Philip_SA; @Philip_REP]. Generally, the structure of the patterns effects the physical processes on them, such as current transport [@MSetal]. It is therefore important to understand the topology of the aggregated cellular networks in detail. Here we study the topological properties, such as shortest paths between nodes, topological centrality, and fractality of the graph’s perimeter in different cellular networks obtained by varying the control parameters of the aggregation processes.
Cell Aggregation
================
The basic idea is to make growing model of planar graph with given distribution of cell (polygon) sizes $f(n_p)$ and with degree of nodes $2 \le k \le 3$.
Topological constraint
----------------------
For this purpose we observe some topological constraints on the distribution of cell sizes: (i) $f(n_p)$ is defined for $n_p \geq 3$, for non-clustered graph we fix $f(3)=0$; (ii) planar graph obeys Euler’s law: $N_p+N=E+1$. Among these the homogeneous plane-filling structures are of special interest [@Stabans_froth]. For this class of graphs majority of nodes are in the interior of the graph, ie., nodes with degree $k=3$. Therefor $3N \approx 2E$ and Euler’s law becomes $$\label{Euler1}
6N_p=2E+6.$$ For large system with distribution of cell sizes $f(n_p)$ we have $$N=N_p\sum_{n_p} \frac{n_pf(n_p)}{3},\; E=N_p\sum_{n_p} \frac{n_pf(n_p)}{2}.$$ Substituting second relation into (\[Euler1\]): $6N_p=N_p\sum_{n_p} n_pf(n_p)+6$, then for large $N_p \gg 1$ one can find that the average cell size is equal to six $$\label{tc}
\langle n_p \rangle \equiv \sum_{n_p} n_pf(n_p)=6.$$
We use lognormal distribution of polygon size $$\label{lnormal}
f(n_p) = \frac{1}{s \sqrt{2\pi} x} e^{-\frac{\ln^2{x/x_0}}{2s^2}},$$ which is most often found in experiments [@Philip_SA]. Using the condition (\[tc\]) the number of independent parameters in (\[lnormal\]) is reduced $$\langle n_p \rangle = 6 \quad \Rightarrow \quad x_0=6e^{-s^2/2}, \; s^2 = \ln \left( 1+ \frac{\mu_2}{36} \right),$$ where second central moment $\mu_2$ remains as the control parameter in our case.
Model
-----
Starting from an initial cell, at each time step a new cell with size taken from the distribution $f(n_p)$ is attached to the graph boundary of the graph according to the following rules:
1. In the aggregation only nodes with degree 3 can be closed inside the graph;
2. New link can be assigned only to nodes with degree 2;
3. Potential nesting place is searched as an array of nodes on graph boundary with degree 3 limited with two nodes of degree 2. The nested part of the cell is identified with the nodes of the nesting string. Therefore, number of extra nodes to be added is $n=n_p-l$, where $l$ is length of the nesting string. We select the nesting place with probability $p \sim e^{-\nu n}$, where the parameter $\nu$ plays the role of the chemical potential for addition of new nodes.
$\begin{array}{c@{\hspace{0.1in}}c}
\includegraphics[width=0.48\textwidth]{suvakov_iccs06_fig1a.eps} &
\includegraphics[width=0.48\textwidth]{suvakov_iccs06_fig1b.eps}\\
\mbox{Open} & \mbox{Closed}
\end{array}$
We implemented this algorithm in *C++* according to following steps:
Initial graph: one polygon of size np taken from f(np)
For i=2 to Np
np = next random from distribution f(np)
If(there is no nodes on graph boundary with degree 2) exit(1)
For all j=(periphery node with degree 2)
d=distance to the next node on graph boundary with degree 2
Number of new nodes n = np-d-1
If(n>0) p(j)=exp(-nu*n)
End of loop j
If(there is no j with n>0) exit(2)
Normalize p(j)
j = next random from distribution p(j)
Add new polygon with size np linked with
node j and next node on graph boundary with degree 2
End of loop i
exit(0)
Depending on model parameters of the growth process and it stochasticity three possible exit cases are:
- exit(0) - Open structure (planar graph with $N_p$ polygons);
- exit(1) - Closed structure (after some number of step there are no more nodes of degree 2 and structure stops to grow, no nesting places of any size);
- exit(2) - No nesting place available for current cell. In this case one can take next cell, which in turn perturbs the actual distribution.
We never experienced the exit(2) situation for the range of parameters $\mu_2 \in [0.5, 2.0]$, $\nu \in [0,5]$ and $N_p=1000$ in huge number of samples. Two examples of the emergent open and closed structures are shown on Fig. \[ozslika\]. More examples of cellular networks are shown in Fig. \[velikaslika\] for varying parameters $\mu_2$ and $\nu$..
$\begin{array}{c@{\hspace{0.1in}}c@{\hspace{0.1in}}c}
\mbox{$\mu_2$} & \mbox{$\nu=0.0$} & \mbox{$\nu=1.0$} \\
\mbox{$0.5$} &
\includegraphics[width=0.45\textwidth]{suvakov_iccs06_fig2a.eps} &
\includegraphics[width=0.45\textwidth]{suvakov_iccs06_fig2b.eps} \\
\mbox{$2.0$} &
\includegraphics[width=0.45\textwidth]{suvakov_iccs06_fig2c.eps} &
\includegraphics[width=0.45\textwidth]{suvakov_iccs06_fig2d.eps} \\
\end{array}$
Fractal Dimension of Network Perimeter
======================================
![Scaling of the network perimeter for (a) $\mu_2=0$ hexagons only (b) $\mu_2=2.0$[]{data-label="FD"}](suvakov_iccs06_fig3.eps){width="100.00000%"}
During the nesting growing process in one step number of nodes $N$ increases by $n<n_p$. Number of added nodes $n$ at each step depends on the cell size and length of the nesting string. Therefore, $N \approx \kappa N_p$ where $\kappa=\langle n \rangle$ is average growth rate. For open structures (see Fig. \[ozslika\]-\[velikaslika\]) boundary of the graph becomes fractal, depending on the control parameters. In fact, length of the graph boundary $L$ grows as a power of the number of cells $N_p$ (or network size $N$) with fractal dimension $D$ defined by $L \sim N_p^D$.
In Fig. \[FD\] we show how number of nodes on the graph boundary increases with $N_p$. Each point is averaged over $10$ emergent growing networks. The dimension $D$ is in the range $\frac{1}{2} \le D \le 1$, when $D=1$ correspond to structures of high fractality, that is obtained for small values of the parameter $\nu$. $D=\frac{1}{2}$ correspond to planar “circle like” structures with reduced fractality. For $\mu_2>0$ we observe a continuous crossover between these two limits (see Fig. \[FD\]b for $\mu_2=2$). However, in structures with homogeneous cell distribution ($\mu_2=0$ - hexagons only) a sharp transition seems to occur at $\nu_c \approx 1.5$.
Shortest Paths and Centrality on Cellular Networks
==================================================
In this section we consider global topological properties of the cell-aggregated planar graphs and their dependence on the control parameters $\mu_2$ and $\nu$.
Shortest Paths
--------------
Shortest path between two nodes is defined as path along the smallest number of intermediate links [@BB_book]. We implemented an algorithm for counting shortest paths of Dijksta type [@NW_book]. In Fig. \[DSP\] we show distribution of lengths of shortest paths between all pairs of nodes on network. All networks are for fixed $\mu_2=1.0$ and approximately of the same size $N \approx 1000$ nodes. Each point in Fig. \[DSP\] is averaged over 100 sample networks. We found similar results for other $\mu_2$ values.
All these networks have similar topology at local level, because the number of links at all interior nodes is constant $k=3$. Therefore, distributions of shortest distances at small scale are similar for all values of parameter $\nu$. Differences in global topology appear on large scale for lengths larger then peak value $l_0 \sim 25$, which manifest in occurrence of additional peaks (see Fig. \[DSP\]). The probability of long paths increases for larger values of the parameter $\nu$. Whereas in the limit $\nu=0$ the distribution of length of shortest paths on large scale can be approximated with a normal distribution (inset on Fig. \[DSP\]).
Centrality Measures {#centrality}
-------------------
Betweenness centrality of a node in network is defined by [@Freeman; @BB_book] $$C_B(v)=\sum_{s\neq v \neq t} \frac{\sigma_{st}(v)}{\sigma_{st}}$$ where $\sigma_{st}$ is total number of shortest paths between nodes $s$ and $t$, and $\sigma_{st}(v)$ is number of these paths that node $v$ lies on. Betweenness of links is defined in analogous way. In our algorithm, we record number of shortest paths through each node and through each link on a network. In Fig. \[DBC\] we show distributions of betweenness of nodes and links averaged over 100 sample networks with fixed $\mu_2=1.0$ and size $N \approx 1000$ nodes.
For inhomogeneous networks, distributions of these two betweenness measures can be substantially different. In our case, however, they are similar because all interior nodes have fixed degree $k=3$. We find that distributions at small scale strongly depend on parameter $\nu$, which results from the graph boundary. Similar conclusions hold for other $\mu_2$ values.
In Fig. \[velikaslika\] thick lines represent links with large betweenness. For this type of networks the strongest lines, which make the skeleton of the graph, are connecting the nodes with largest centrality measure.
Conclusions
===========
We have introduced a new algorithm for growth of graphs by aggregation process of extended objects - polygons with size distribution. Depending on aggregation conditions, which are determined by two parameters controlling the distribution width and attachment potential, we can get a wide spectrum of emergent structures. In this paper we presented some results for the case of lognormal distribution of cells and additional constraints, leading to the emergent non-clustered planar graphs with a constant node connectivity. The algorithm works for variety of cell distributions and constraints, that may result in diverse opened (fractal) or closed structures. For instance, for a special set of parameters we can get closed structures of $C_{60}$ type.
We measured several topological properties of these networks in quantitative details—fractality of the graph boundary, shortest paths, and betweenness centrality. This properties are important for some dynamical process on networks such as electrical conductivity [@MSetal] via single electron tunneling [@Philip_SA; @Nano_book].
[**Acknowledgments:**]{}\
M.S. thanks financial support from the Marie Curie Research and Training Network MRTN-CT-2004-005728 project. B.T. is supported by the program P1-0044 of the Ministry of high education, science and technology (Slovenia).
[12]{}
Dorogovtsev, S. N., and Mendes, J. F. F.: [*Evolution of Networks.*]{} Oxford University Press 2003 Boccaleti S., Latora V. [*et al.*]{}: *Complex Networks: Structure and Dynamics*. Physics Reports (in press) Bollobás, B.: *Modern Graph Theory.* Springer (New York) 1998. Biely C., and Thurner S.: *Statistical Mechanics of Scale Free Networks at a Critical Point: Complexity Without Irreversibility.* cond-mat/0506140. Newman, M. E. J.: **The structure and function of complex networks.** SIAM Rev. [**46**]{} (2003) 167-225 Witten, T. A., and Sander L.M.: [*Diffusion-Limited Aggregation, a Critical Phenomenon.*]{} Phys. Rev. Lett. [**47**]{} (1981) 1400-1403
Stavans, J.: *Evolution of cellular structures.* Rep. Prog. Phys. [**56**]{} (1993) 733-789
B. Tadić, [*From Microscopic Rules to Emergent Cooperativity in Large-Scale Patterns*]{}, in [**Systems Self-Assembly: multidisciplinary snapshots**]{}, N. Krasnogor, S. Gustafson, D. Pelta, J. L. Verdegay (Eds.) Elsevier(2005). Moriarty, P., Taylor M. D. R., and Brust, M.: *Nanostructured Cellular Networks.* Phys. Rev. Lett. [**89**]{} (2002) 248303 Moriarty, P. : *Nanostructured materials.* Pep. Prog. Phys. [**64**]{} (2001) 297-381 M. Šuvakov [*et al.*]{} (in preparation)
Ahuja, R. K., Magnanti, T. L., and Orlin, J. B.: *Network Flows: Theory, Algorithms, and Applications.* Prentice Hall (New Jersey) 1993 Freeman, L.C.: *A set of measures of centrality based on beetweenness.* Sociometry [**40**]{} (1977) 35-41
Ferry, D. K. and Goodnick, S. M. *Transport in Nanostructures.* Cambridge University Press 1997
|
---
author:
- 'P. Delorme'
- 'L. Albert'
- 'T. Forveille'
- 'E. Artigau'
- 'X. Delfosse'
- 'C. Reylé'
- 'C. J. Willott'
- 'E. Bertin'
- 'S. M. Wilkins'
- 'F. Allard'
- 'D. Arzoumanian'
bibliography:
- 'bib.bib'
title: |
Extending the Canada-France brown Dwarfs Survey to the near-infrared:\
first ultracool brown dwarfs from CFBDSIR [^1]
---
Introduction
============
The significant improvement in detector technology, data storage, and analysis facilities in the past decade has made it possible to carry out wide-field surveys covering a large fraction of the sky instead of targeting specific sources. The wealth of data from these surveys necessitate a complex dedicated computer analysis to single out relevant scientific information. These surveys, such as DENIS [@Epchtein.1997], SDSS [@York.2000], 2MASS [@Skrutskie.2006], UKIDSS [@Lawrence.2007], and CFBDS [@Delorme.2008b] contain hundreds of millions of astrophysical sources and led to many advances in various fields, notably to identify extremely rare objects and build robust statistical studies.
The survey we are presenting here, the Canada-France Brown Dwarfs Survey-InfraRed aims at finding ultracool brown dwarfs (T$_{eff}<$ 650K) of which only 6 are currently published by @Warren.2007 [@Delorme.2008a; @Burningham.2008; @Burningham.2009; @Burningham.2010sub]. @Lucas.2010sub have very recently identified a probably even cooler object. These rare objects are in many ways the intermediate “missing link" between the cold atmospheres of the Solar System’s giant planets and cool stellar atmospheres. The physics and chemistry of their atmospheres, dominated by broad molecular absorption bands, are very planetary-like [see @Kulkarni.1997 for instance] and the cool brown dwarfs spectra are the key to constraining planetary and stellar atmosphere models. Nowadays, the Teff$<$700K atmosphere temperature regime is troublesome for modellers. A few ultracool late T brown dwarfs have now been discovered with effective temperatures below 650K. These discoveries step into unexplored territory and a new generation of models is emerging. It is facing several difficulties. 1) Out-of-equilibrium chemistry plays an important role, resulting for instance in NH$_3$ being less prevalent than expected [@Cushing.2006]. 2) Fine details of convection control both the L/T transition and the dredge up of hot chemical species in late T atmospheres. 3) Water cloud formation and dust nucleation play important roles. 4) Line opacities of several molecules, in particular NH$_3$ and, to a lesser extent, CH$_4$ are unknown and cause important spectroscopic feature mismatches. As a good example of the need for refined models, @Burningham.2009 have recently determined that a T8.5 dwarf companion to an M star was actually $\sim$15% cooler than model fitting would have predicted and @Dupuy.2009 and @Liu.2008 reached similar conclusions when studying brown dwarf binaries with dynamical masses.
Under those circumstances, observations are key to the development of the models. Only five brown dwarfs with temperatures below 650K (T8.5) are currently known from recent discoveries by UKIDSS and CFBDS, and this small number prevents discerning general trends from individual peculiarities. @Kirkpatrick.1999 and @Burgasser.2002 could rely on samples of 20-25 objects to respectively define the L and T spectral types. In this article we present the CFBDSIR, a near infrared (hereafter NIR) extension to the CFBDS that will provide a WIRCAM [@Puget.2004] $J$-band coverage down to $J_{vega}=20.0$ for 10$\sigma$ detections atop 335 square degrees of CFBDS MegaCam [@Boulade.2003proc] $z'$-band imaging with a 5$\sigma$ detection limit of $z'_{AB}$=23.25-24.05. All optical magnitudes presented in this article are AB magnitudes, while all NIR magnitudes are Vega magnitudes. When the CFBDSIR is complete, we hope to achieve a threefold increase in sample size to 15-20 characterised ultracool brown dwarfs and possibly find a few substellar objects significantly cooler than 500K of which none is known yet outside the solar system. This will define general trends and dispersions around them, permitting the study of ultracool dwarfs not only as individual interesting objects, but as a population. This will help define the T/Y spectral transition that is expected to occur in this temperature range. In section 1 we present the rationale of this new wide-field survey and the observations at its core. In section 2 we describe the data reduction and the data analysis methods we used to identify ultracool brown dwarfs candidates. Finally, we present the spectra and the photometry of the first ultracool brown dwarf identified with CFBDSIR in section 3.
Observations
============
Far-red and near infrared photometric properties of brown dwarfs
----------------------------------------------------------------
Field brown dwarfs are cool objects, with a temperature range that currently extends from $\sim$2500K (early L) to $\sim 525$K [latest objects identified, e.g. @Leggett.2009]. Even cooler, and still to be found, brown dwarfs should close the temperature gap between late type T-dwarfs and solar system Jovian planets($\sim$110K). These objects do exist [see @Burgasser.2009proc for instance] because low-mass brown dwarfs already observed in young clusters must cool down to this temperature range when they age. Brown dwarfs spectra differ significantly from a black body, and have considerable structure from deep absorption lines and bands. The spectral energy distribution of 500K brown dwarfs peaks in the in the $J$ band in the near infrared (even colder objects should emit more energy in the mid-infrared), and they are therefore most easily detected in that wavelength range. Their NIR $JHK$ colours are, however, not distinctive at a modest signal-to-noise ratio [see @Metchev.2008 for instance], and brown dwarfs are more easily recognised by including at least one photometric band blueward of 1 $\mu$m. At those wavelengths the steep slope of their spectra stands out, and they have very distinctively red $i'-z'$ and $z'-J$ colours. The CFBD Survey [@Delorme.2008b] took advantage of their distinctive $i'-z'$ colours to identify hundreds of L and T-dwarfs. However, the reddest and coolest brown dwarfs have extremely red $z'-J$ colours and are much easier to detect in the NIR than on original CFBDS far-red images. A WIRCam $J$-band coverage of CFBDS fields is a straightforward and efficient way to cumulate both the very effective selection criteria using far-red colour information and the NIR detection sensitivity to ultracool brown dwarfs. Since the overlap of CFBDS with existing NIR surveys with a relevant depth [such as UKIDSS, @Lawrence.2007] is marginal, it was impossible to rely on archive data.
The resulting survey, CFBDSIR, identifies astrophysical sources on $J$-band WIRCam [@Puget.2004] images and selects ultracool brown dwarfs candidates depending upon their $z'-J$ colours, using $z'$ magnitudes from CFBDS MegaCam [@Boulade.2003proc] images. The $z'-J$ colour has a wide range for brown dwarfs, varying from 2.6 for mid-L types to over 4.5 for late-T types (Fig. \[zJspT\]). It therefore provides (at least at a good signal-to-noise ratio) a good selection criteria to identify ultracool brown dwarfs. They usually have $z'-J>$3.8, as confirmed by atmosphere models such as BT-Settl [@Allard.2007], synthetic MegaCam/WIRCam colours derived from known objects [@Delorme.2008b] and direct observational data [@Delorme.2008a; @Burningham.2008; @Burningham.2009]. Models suggest that cooler objects, not yet discovered (T$<$500K), are even redder. Figure \[zJspT\] however shows that the colour-spectral type relation for late T-dwarfs has a wide spread, meaning that this $z'-J$ colour range also contains some warmer mid-T dwarfs. Follow-up photometry of candidates in $H$ and $K_{\rm s}$ band easily distinguishes the relatively blue very late T from the redder mid-T-dwarfs [see @Lodieu.2007 for instance]. Only very rare astrophysical objects share the same $z'JHK$ colour range: very high-redshift starburst-galaxies, extremely red Balmer break galaxies, and atypical broad absorption-lines quasars at moderately high redshift ($z\geq 2$), such as those described in @Hall.2002. However, those objects can be differenciated from brown dwarfs using only photometry: they are not as red in $i'-z'$ as ultracool brown dwarfs and will appear in the $i'$ CFBDS images whereas brown dwarfs will not. Quasars with redshift z$>$7 cannot be distinguished from cool brown dwarfs using the CFBDSIR $i',z',J$ and optical photometry, but we would be lucky to find just one of these extremely interesting objects. None has been discovered yet, and extrapolation of the z$=$6 luminosity function [@Willott.2010] suggests that wider or/and deeper surveys will be needed to find one. Other possible contaminants are detector artefacts and variable objects whose luminosity changed between $z'$ and $J$ exposures.
The survey
----------
[**WIRCam NIR data.**]{} This article focuses on the results obtained from a 66 square degrees pilot subset of CFBDSIR whose final coverage is expected to be 335 square degree. CFBDSIR WIRCam imaging goes to a depth of $J_{vega}=20.0$ for a point-source detection limit of 10$\sigma$, which ensures accurate photometry and rejects most spurious detections. Images were acquired at CFHT in QSO mode with seeing varying from 0.6to 1.0during semesters 2006B, 2007A, and 2009B. These WIRCam images overlap the existing $z'$-band images of CFBDS; i.e., each $\sim$ 1 square degree MegaCam field is covered by nine 21’ by 21’ WIRCam $J$-band 45-second exposures. For the sake of efficiency, no dithering is done. Instead, the 9 adjacent fields observed in sequence are used for sky construction. Each 9-field patch is actually observed twice, generally one night apart, in order to identify and remove any moving solar system object, for a total integration time of 90 seconds. The second sequence is observed with few pixels offset to average the fixed noise pattern seen on the WIRCam detectors. The 10$\sigma$ limiting depth of $J_{vega}=20.0$ ensures that all confirmed ultracool brown dwarf candidates can be observed with low-resolution spectrographs on 8-meter class telescopes.\
[**Optical data.**]{} The optical data consist of the $z'$-band images used for the CFBD Survey. This survey is made of public and P.I. data and associated 4 sub-surveys of different area shallowness, namely the Canada-France-Hawaii Telescope Legacy Survey-Very Wide,CFHTLS-Wide, CFHTLS-Deep and Red-sequence Cluster Survey-2 (RCS2), described in detail by @Delorme.2008b. The CFBDSIR only targets the 3 shallower components, as shown in Table \[maglim\]. This table compares the limiting magnitude, the maximum distances of detection for mid-L, mid T and the latest T-dwarfs, together with the areas covered by each survey. The similar table 1 in @Delorme.2008b incorrectly uses the maximal magnitude limit for the CFHTLS-WIDE, instead of the average one. The value of 23.3 stated in Table \[maglim\] here corrects this error. The 335 square degrees CFBDSIR probes about 2.8 times more volume for late T-dwarfs than the $\sim$900 square degree CFBDS, and even more for redder (and thus likely cooler) objects. Figure \[cfbdsir\_map\] shows the sky coverage of CFBDSIR as of March 2010. The optical data is public and available on the CFHT archive, while the NIR data becomes public after the standard one-year CFHT proprietary period.\
![Map of observed CFBDSIR fields as of March 2010. Dashed lines are +/-20 degrees of galactic latitude. \[cfbdsir\_map\]](14277fg1.ps){width="8cm"}
[|X|X|X|X|X|X|]{} Survey Name& $z'_{AB}$ $10\sigma$ ($5\sigma$) detection limit& mid-L maximal detection (pc) & mid-T maximal detection (pc) & T9 maximal detection (pc) & Final coverage (sq deg)\
CFBDS:RCS-2 & 22.5 (23.25) & 185 & 80 &16 & 600\
CFBDS:CFHTLS-Very Wide & 22.8 (23.55) & 215 & 90 & 18 & 150\
CFBDS:CFHTLS-Wide & 23.3 (24.05) & 270 & 115 & 22 & 170\
Survey Name& $J_{vega}$ $10\sigma$ detection limit & mid-L maximal detection (pc) & mid-T maximal detection (pc)& T9 maximal detection (pc) & Final coverage(sq deg)\
CFBDSIR & 20.0 & 235 & 145 & 35 & 335\
CFBDS initial detections are made in $z'$ and only better than $10\sigma$ detections are used. CFBDSIR makes the initial detection in $J$, using only better than $10\sigma$ detections in $J$, and is only limited in $z'$ by no detections. We define any $z'$-band object with a signal below the $5\sigma$ threshold as a no detection or ”z’-dropout".
[**Survey rationale.**]{} Ultracool brown dwarfs have $z'_{AB}-J_{vega}>3.8$ for the MegaCam and WIRCam photometric systems, with some already observed at $z'-J=4.5$. Given the relative depth of the optical and NIR exposures, most of the T-dwarfs earlier than T7 detected in the $J$-band image will have a $z'$-band counterpart. The objects detected at $J$ and not at $z'$ (hereafter $z$-dropouts) are thus good candidates for some of the coolest T-dwarfs known ($>$T7), and possibly even cooler Y dwarfs.
Since our most interesting candidates are likely to only be detected in $J$-band, we need to carefully eliminate most sources of contamination in this band. The main contaminants are detectors artefacts, variable objects and asteroids. As described in detail in section \[data\_analysis\], our point spread function (hereafter PSF) analysis removes most artefacts, while all asteroids are eliminated because they move between the two exposures taken at each pointing. Supernovae and other strongly variable objects have to be rejected during the follow-up, but their numbers are kept low thanks to the relative shallowness of the survey. Since we select all PSF-looking $z'$-band dropouts, CFBDSIR also very efficiently identifies high-proper motion objects: if the proper motion is high enough that a $J$-band source cannot be cross-matched with its $z'$-band counterpart, the automatic pipeline keeps the candidate in the selection as a $z'$-band dropout. Visual examination of all selected candidates subsequently identifies those high proper motion sources easily.\
[**NIR Follow-up.**]{} The NIR follow-up of CFBDSIR candidates removes the remaining candidates but also provides an efficient characterisation of the confirmed candidates. Ultracool brown dwarfs can be singled out by their blue $J-H$ and $H-K$ colours owing to stronger H$_2$O, CH$_4$ and probably NH$_3$ absorption than earlier T dwarfs. The CFBDSIR follow-up of the first 66 square degrees from CFBDSIR required about 3.5 nights on SOFI [@Moorwood.1998] at NTT. We were able to observe all 55 T-dwarfs candidates identified and to confirm 6 of them as T-dwarfs, including 3 ultracool brown dwarfs (later than T7 dwarfs and possible Y dwarfs) candidates. The rationale of the NIR follow-up is the following.
- A short (5 to 10 minute dithered image, with 30-s individual exposures), $J$-band exposure confirms the candidate and eliminates any remaining contaminants from variable sources, such as supernovae or detector artefacts. For confirmed candidates, this image also provides good signal-to-noise (from 20 to about 100, depending on the target magnitude) $J$-band photometry, better than the photometry on the detection image.
- All candidates confirmed with this SOFI $J$ photometry are imaged in $H$-band to characterise them as ultracool brown dwarfs or earlier T-dwarfs. A 10-20 minute dithered sequence, with 20-s individual exposures, usually achieve good signal-to-noise photometry, with an accuracy better than 5%.
- All confirmed candidates with measured $J-H<$0.1 are potential ultra-cool brown dwarfs that can be further characterised by acquiring $K_{\rm s}$-band photometry, if enough observing time is available. A 15-30 minute dithered sequence, with 15-s exposures typically achieves good signal-to-noise photometry in $K_{\rm s}$.
It should be noted that this very efficient confirmation and characterisation photometric follow-up is made possible by visitor-mode observation combined with the automatic image reduction by SOFI pipeline *Gasgano*. A simple photometric analysis can be carried out by the observer only minutes after the exposures allowing $J$ and possibly $H$ and $K_{\rm s}$ photometry to be derived on-the-fly. It is then possible to choose the follow-up strategy most suited to each individual candidate. The final astrometric and photometric calibration is done using 2MASS [@Skrutskie.2006] stars in the field as references.
Survey analysis
===============
Ultracool brown dwarfs are extremely red objects 25 to 60 times (3.5 to 4.5 magnitudes) brighter in $J$-band that in $z'$ band. This means that even if our $z'$ images are about 3 to 4 magnitudes deeper than the detection $J$-band images, many of our candidates with strong $J$ band detection will not appear at all in the $z'$ images. However, a non detection in $z'$ band is a very strong sign that the candidate is a very red astrophysical source with $z'-J>3.5$, exactly what we are looking for. The other alternative is that these sources are spurious, such as unrecognised artefacts (remaining detector cross-talk, unflagged hot pixel, optical ghost, etc.) on the $J$ image. Since ultracool brown dwarfs are extremely rare ( our preliminary estimate is $\sim$ 1 per 25 square degree, so fewer than 1 per $500~000$ sources down to $J$=20.0), false detection rates of even 1 per 10$^4$ real sources would greatly outnumber true brown dwarfs. We therefore need a very robust analysis of the discovery images to tie down the number of artefacts before we finally weed them out during the photometric follow-up of the candidates.
Image reduction
---------------
The WIRCam observations targeted 335 existing MegaCam fields from CFBDS, that have already been observed in $i'$ and $z'$. The reduction of MegaCam data is described by @Delorme.2008b.
Images were run through the ‘I‘iwi Interpretor of the WIRCam Images (‘I‘iwi - [www.cfht.hawaii.edu/Instruments/Imaging/WIRCam]{}) at CFHT, which does non-linearity correction, dark subtraction, flat-fielding, sky subtraction, bad pixel masking, photometric calibration, and rough astrometric calibration. Cross-talk subtraction was also performed on the 06B and 07A data, affected by amplifier to amplifier cross-talk.
Each set of 2 times 9 WIRCam images covering one $z'$ MegaCam image was then co-aligned, using the $z'$ image used as the astrometric reference. This astrometric calibration was performed with *Scamp* [@Bertin.2006], and the images were stacked using *Swarp*[@Bertin.2002]. This produced $\sim$1 square degree WIRCam images aligned with each targeted MegaCam field.
Image analysis {#data_analysis}
--------------
Since our stacks are only 2 exposures deep, the final frame is the average of both exposures. To easily reject moving solar system objects from our detection catalogues, we also produced a ”minimum“ image of both exposures. This combined image contains the lower of the pixel values of the two exposures, and is effectively devoid of all those objects that moved by more than one FWHM between the exposures. We carried out the analysis using the double-image mode of *Sextractor* [@Bertin.1996], with source detection on the ”minimum” image and photometric measurements on the average image, using PSF fitting. The different analysis steps are described here:
- Spatially variable PSF models of the images are built with PSFex [@Bertin.2010], using single stars within the image as prototypes..
- The sources on the minimal image are identified using *Sextractor*. Their astrometry and photometry are derived by PSF fitting. (The latter is discarded from thereon, but is included in the fit). Since moving objects are at different positions in the two exposures that were stacked to produce the minimal image, they do not appear in the resulting source list.
- A flux-only PSF fitting at the position of the sources identified on the minimal sum stacked image is then carried out on the average stacked image. This step produces the final WIRCam $J$-band catalogue.
- The MegaCam $z'$-band catalogue of the corresponding field is then produced using a similar position and flux PSF fitting with *Sextractor*.
- The $J$-band and $z'$-band catalogues are cross-matched so that each source in the $J$-band catalogue is associated with a $z'$ magnitude. In case there is no $z'$-band couterpart to a $J$-band detection, which is then a “$z$-dropout", the source is given the 5 sigma detection magnitude limit of the full $z'$ image as its $z'$ magnitude and the corresponding $z'-J$ colour is then considered as a lower limit of its actual colour.
Filtering and candidate selection
---------------------------------
The selection of ultracool brown dwarfs candidates in our catalogue aims at the greatest possible completeness (i.e. finding most of the ultracool dwarfs actually detectable on the images) while keeping the number of contaminants down to an acceptable level. Contaminants are particularly critical because the most promising candidates are the $z$-dropouts for which only a $J$ magnitude and a $z'-J$ lower limit on the colour is known. To keep the number of contaminants relatively low, we only selected sources with a signal-to-noise ratio above 10 in the $J$-band stack.
As visible in Fig. \[zJspT\], brown dwarfs later than T8 populate the $z'-J>$3.8 colour range. This figure as well as Fig. \[wilkins\] make use of spectra from http://www.jach.hawaii.edu/$\sim$skl/LTdata.html [@Martin.1999; @Kirkpatrick.2000; @Geballe.2001; @Leggett.2002; @Burgasser.2003; @Knapp.2004; @Golimowski.2004; @Chiu.2006]. As described in @Delorme.2008b these colours are synthesised from the spectra using the filters, atmosphere, telescope and detector transmission and sensitivity of the instruments used in CFBDSIR. Given the large colour spread of late T-dwarfs, we set a very conservative colour selection threshold of $z'-J>$3.5, to select as many ultracool dwarfs as possible. Since the 3.5$<z'-J<$3.8 colour range is mainly populated by warmer T-dwarfs and is still very far from any densely populated colour locus, as illustrated by Fig. \[colpop\], including it in our selection does not increase the number of contaminants significantly.
![$z'-J$ colour spectral type relation. The green symbols represent synthetic MegaCam/WIRCam colours for known objects with publicly available spectra. The black dashed line is the resulting averaged colour-spectral type relation. Red points show the observational colours and spectral types of CFBDS spectroscopically confirmed T-dwarfs. \[zJspT\] ](14277fg2.ps)
In addition to this colour selection, the main filtering criterion is based on the $\chi ^2$ from PSF fitting, to distinguish point sources from galaxies and artefacts, following the same method as described in @Delorme.2008b. Since brown dwarfs are point sources, this morphological rejection of all non-point-source-like objects effectively removes many non-stellar contaminants.
This automated selection produces an average of 6 candidates per square degree. Each of these is visually examined to remove remaining contaminants. In addition to inspecting both the $z'$-band and the $J$-band images, we examine the $i'$-band image (available from the CFBDS), to reject some variable contaminants, as well as possible contamination by atypical broad absorption line quasars, none of which is known to be very red in both $z'-J$ and $i'-z'$ [P.Hall, private communication and @Hall.2002]. Since the 5$\sigma$ detection limits of the CFBDS $i'$ images (between 24.5 and 25.5) are significantly deeper than the $z'$ images, broad absorption line quasars with $z'-J>3.5$ should be visible on the $i'$ image. We identified one such object, CFBDSIR232451-045852, with $z'-J=$4.7 and $i'-z'$=0.9.
Very high-redshift (z$>$6), star-forming galaxies and evolved galaxies with strong Balmer/4000Å breaks at z$\sim$2, shown in Fig. \[wilkins\], are other possible contaminants. While the surface density of J$\sim$20 high-redshift, star-forming galaxies is unknown (the brightest of these objects discovered so far have J$_{vega}\sim$25 [@Ouchi.2009], five magnitudes fainter than our detection limit), the expected exponential cut-off in the luminosity function suggests that the probability of finding any such galaxy in CFBDSIR is extremely low. Balmer/4000Å break galaxies at z$\sim$2.5 do enter the fringe of our selected colour range; however, their $i'-J$ colour is typically below 5, ensuring they are detected in the $i'$-band CFBDS images and are then eliminated from our candidate list.
After this final visual check, we end up with about 1 ultracool brown dwarf candidate per square degree, which has to be confirmed with pointed NIR observations.
![Colour-magnitude diagram of all sources detected with a signal-to-noise greater than 10 in a 1 square-degree stacked WIRCam image from CFBDSIR. Since no other filtering is applied, these sources can be stars, galaxies, or detector artefacts. In this specific instance, none of the red objects is a brown dwarf. \[colpop\]](14277fg3.ps)
![Colour-colour diagram of known stars and brown dwarfs from spectral type O (deep-blue dots) to T (dark-red dots). The blue line shows the colours of high-redshift-starburst galaxies, while the red line refers to red Balmer-break galaxies. These lines follow a redshift evolution and specific redshift steps are indicated by numbers near the lines. The colours are synthesised from spectra, and take into account the actual transmission and detector sensitivity of the MegaCam and WIRCam instruments used in CFBDSIR. \[wilkins\]](14277fg4.ps)
Results
=======
Photometric confirmation of 3 ultracool brown dwarfs
----------------------------------------------------
The 55 candidates found in a 66 square degree subset of the data were followed-up with photometry at the ESO New Technology Telescope, using the SOFI near infrared camera during visitor mode runs 083.C-0797(A) and 082.C-0506(A), in July 2009 and March 2009. These pointed NIR observations confirm six T-dwarfs, of which 3 are robust ultracool brown dwarf candidates. Those are likely to be either very late T-dwarfs ($>$T8) or cooler Y dwarfs. The brighter 2 of the 3 earlier T candidates are re-identifications of CFBDS-discovered brown dwarfs, spectroscopically confirmed as T3.5 and T5 (Albert et al., submitted.).
The candidates identified in the remaining 269 square degrees have not yet been followed-up, and they certainly include mostly contaminants. Extrapolating the results from the 66 square degree pilot subset of CFBDSIR to the full survey, we expect to find $\sim$15 ultracool brown dwarfs. Low-resolution NIR spectroscopy of the confirmed ultracool brown dwarfs will be needed to derive spectral types and characterise their cool atmosphere physics.\
[**Photometric properties of ultra-red brown dwarf candidates**]{}
![Colour-colour diagram of some of the latest brown dwarfs known, $>$T8, later than spectrum used for up to date spectral classification scheme together with the 3 CFBDSIR ultracool dwarfs candidates. The regular T-dwarf approximate colour range is highlighted in the upper left corner. Note that given the strong dispersion in colours of T dwarfs, some mid/late T do spread out of this indicative colour range. \[zJJH\]](14277fg5.ps){width="8cm"}
After the March 2009 $J, H$, and $K_s$ follow-up observations, one candidate stood out as particularly interesting. We then obtained additional WIRCam $J, H$, and $K_s$ photometry for this object, CFBDSIR J145829+101343 (hereafter CFBDSIR1458). The resulting magnitudes are presented in Table \[mag\]. The SOFI photometry in this table has additional uncertainty because its calibration is bootstrapped from a small number of good 2MASS reference stars in the narrow field of the SOFI NTT images, especially so in the $H$ and $Ks$ bands.
CFBDSIR1458 colours are detailed in Table \[colours\] and shown in Fig. \[f\_chart\], and can be summarised as follows:
- Very red far-red colours: $z'-J=$3.94, which Table \[colours\] compares to those of other ultracool brown dwarfs;
- Very blue NIR colours, with \[$J-H$;$J-K_{\rm s}$\]=\[-0.46;-0.94\], pointing to very strong molecular absorptions in the $H$ and $K$ bands. The $J-K_{\rm s}$ colour of -0.94 approximately translates into a $J-K$ colour of -1.02 [@Stephens.2004; @Leggett.2010].
 
------------------------ ------------------------
 
A second follow-up run at NTT on July 2009, confirmed 2 other ultra-red objects, CFBDSIR221903.07+002417.92 and CFBDSIR221505.06+003053.11 (hereafter CFBDSIR2219 and CFBDSIR2215), with even redder $z'-J$ colours, typical of ultracool brown dwarfs. These objects have redder $z'-J$ than any known T8+ brown dwarf. CFBDSIR2215, however, has only a moderately blue $J-H$ colour, which would be compatible with an earlier (i.e. T5-T7) spectral type, suggesting a peculiar spectrum. The photometric properties of these objects are described in Table \[colours\] and Fig. \[zJJH\].
Filter $z'$ $J$ $H$ $K$s
-------- ---------------- ---------------- ---------------- ----------------
Mag(1) $23.60\pm$0.24 19.66$\pm$0.02 20.12$\pm$0.13 20.60$\pm$0.37
Mag(2) - 19.72$\pm$0.04 19.96$\pm$0.16 20.13$\pm$0.32
: (1) $J,~H,~Ks$ CFHT WIRCam Vega photometry and $z'_{AB}$ MegaCam photometry of CFBDSIR1458. (2)$J,~H,~Ks$ NTT SOFI Vega photometry. \[mag\]
Object $z'-J$ $J-H$ $J-K$
-------------------- --------------- ----------------- ---------------- --
ULAS0034$~^{1,3}$ 3.90$\pm$0.06 -0.28$\pm$0.05 -0.12$\pm$0.06
CFBDS0059$~^{2}$ 3.86$\pm$0.06 -0.21$\pm$0.06 -0.67$\pm$0.06
ULAS1335$~^{1,5}$ 4.14$\pm$0.10 -0.35$\pm$0.01 -0.38$\pm$0.03
2MASS0939$~^{1,4}$ - 0.18$\pm$0.181 -0.58$\pm$0.10
WOLF940B$~^{6}$ 3.99$\pm$0.13 -0.61$\pm$0.04 -0.69$\pm$0.05
CFBDSIR1458 3.94$\pm$0.24 -0.46$\pm$0.13 -1.02$\pm$0.37
CFBDSIR2219 4.45$\pm$0.10 -0.20$\pm$0.085 -
CFBDSIR2215 4.30$\pm$0.12 -0.11$\pm$0.08 -
: Colours of ultracool brown dwarfs.\[colours\]
References:
$^1$@Leggett.2009, $^2$@Delorme.2008a $^3$@Warren.2007, $^4$@Tinney.2005, $^5$@Burningham.2008, $^6$@Burningham.2009
While CFBDSIR2219 and CFBDSIR2215 are queued for NIR spectroscopic characterisation, we already obtained a spectrum of CFBDSIR1458, which we discuss in section 4.2.
[**Proper motion of CFBDSIR1458**]{}\
Since the $z'$ MegaCam image and the $J$ WIRCam image are co-aligned by the CFBDSIR pipeline the proper motion is easily measured from these 2 images. CFBDSIR1458 moves by +0.58 in RA and -0.95 in DEC between the July 15, 2004 date of the $z'$ image and April 1, 2007 date of the $J$ image. The centroiding and image alignment uncertainties are low, at 0.035, but the error budget has to include chromatic refraction and the uncertain parallactic motion. The estimated spectroscopic distance is $\sim$23pc, with a likely range of 15–30 pc (Table \[pm\]), which translates to a 0.04maximum error from the parallax. Due to the very steep spectral energy distribution in the $z'$ band, chromatic refraction here is of the order of the measurement error. We also measured the proper motion between the $J$-band WIRCam discovery image and the $J$ band NTT follow-up image, finding +0.32 in RA and -0.80 in DEC between April 1, 2007 and March 5, 2009. This second measurement has smaller centroiding and image-alignment uncertainties, 0.03, minimal chromatic refraction uncertainties since the two $J$-band observations were obtained at similar airmasses, and a small unceretainty from the parallax because the two observations were coincidentally obtained at closely matched times of the year. It is therefore our preferred proper motion measurement, and is consistent with the previous one. Table \[pm\] presents the resulting yearly motion and kinematic parameters.
RA(.yr$^{-1}$) DEC(.yr$^{-1}$) RA(km.s$^{-1}$) DEC(km.s$^{-1}$)
----------------- ----------------- ----------------- ------------------
+0.17$\pm$0.016 -0.41$\pm$0.016 18 46
: Yearly proper motion for CFBDSIR1458. \[pm\]
The systematic errors induced by chromatic refraction and parallax effects are not corrected. The estimate of the absolute tangential velocity assumes a distance of 23pc.
Spectroscopic confirmation of a new ultracool brown dwarf
---------------------------------------------------------
[**Observations and reduction**]{}
After reduction and analysis of the March 7, data, an ESO Director Discretionary Time observation request was submitted on March 20, 2009 to obtain low-resolution NIR spectroscopy of CFBDSIR1458. This proposal for 4 hours of $H$-band observation, totalling 150 minutes of exposure on target, was accepted on April 22. The first observations were acquired on May 4, and the last on September 1, on average at high airmass, varying from 1.4 to more than 2.0.
The spectrum was extracted and calibrated using our own IDL procedures. The reduction proceeded as follows. The sequence of spectral images were flat-fielded using an internal flat taken immediately after the science frames. Since the trace was too faint for its position to be accurately determined, its curvature was derived from the reference star spectrum. The frames were then pair-subtracted, effectively removing most of the sky, dark current and hot pixels contributions. Each frame was collapsed along the spectral dimension to determine the positive and negative traces positions. We then extracted the spectra using positive and negative extraction boxes that have identical but opposite integrals; this minimised the contribution from residual sky line that would have remained from the pair subtraction. The same operation was performed on the A0 telluric calibration star. Spectra derived from individual image pairs were then median-combined into final target and calibration star spectra. A telluric absorption spectrum was derived using the calibration-star spectra. A black body spectrum with a temperature of 10 000K was assumed for the A0 star and hydrogen-lines were interpolated over. The target spectrum was then divided by the derived telluric transmission spectrum. A first-order wavelength calibration was obtained from an argon-lamp spectrum, and fine-tuned by registering bright OH lines obtained from a sum of the pair of images of interest.\
[**Spectroscopic properties**]{}
The resulting spectrum (Fig.\[spectra\]) has a low signal-to-noise owing to the faintness of the target ($H_{vega}=20.12$) and its relatively high airmass at the time of the observations. We plan to obtain better signal-to-noise observations, as well as $J$ and $K$-band low-resolution spectrum, but we were already able to derive from the present $H$-band spectrum the spectroscopic indices described by @Burgasser.2006 and @Delorme.2008a, which quantify the strength of key molecular absorption bands. As shown in Table \[indices\], this classifies CFBDSIR1458 as an ultracool brown dwarf with spectral type later than T8 and a temperature in the same range ($\sim $500-600K) as the coolest brown dwarfs known. Direct comparison of the CFBDSIR1458 $H$-band spectrum with other ultracool brown dwarfs (See Figs. \[spectra\] and \[spectra\_zoom\]) visually confirms that H$_2$O and CH$_4$ absorption in its atmosphere are significantly stronger than they are on 2MASS0415 [@Burgasser.2003] the T8 spectral template. This also strongly suggests that CFBDSIR1458 is indeed a later-than-T8 ultracool brown dwarf. The comparison with even cooler objects is less clear cut, mainly because of the low signal-to-noise of the spectrum, as emphasised by the strong variations in the spectral indices for 2 different -but both sufficiently sampled- binning of CFBDSIR1458 spectrum visible in Table \[indices\]. However, both the spectrum and the indices would tend to show that CFBDSIR1458 is not cooler than the coolest brown dwarfs already known.
This intermediate spectral feature would tentatively put CFBDSIR1458 in the same class as WOLF940B, that was classified as T8.5 by @Burningham.2009 who assigned a temperature of 550-600K to this object. Given the similarities in the $H$-band indices of both objects, a reasonable estimate would put CFBDSIR1458 in the same temperature range. However, this rough estimate will need to be confirmed by higher signal-to-noise multi-bands spectroscopic observations. In particular, its very blue $J-K$ colour could be due to sub-solar metallicity [@Leggett.2010]. Additional observations would also enable looking for the NH$_3$ absorption in the CFBDSIR1458 spectrum, such as the probable absorption band identified by @Delorme.2008a. IRAC imaging in the 4.5micron channel would also be very valuable since temperature can be reliably derived from the \[$H-4.5$\] colour [@Leggett.2010].
Object Sp. Type H$_2$O-H CH$_4$-H NH$_3$-H
----------------- ---------- ---------- ---------- ----------
Gl570D T7.5 0.208 0.137 0.672
2MASS0415 T8 0.183 0.104 0.625
WOLF940B T8+/Y? 0.141 0.091 0.537
CFBDSIR1458 (1) T8+/Y? 0.149 0.046 0.568
CFBDSIR1458 (2) T8+/Y? 0.146 0.087 0.582
ULAS0034 T8+/Y? 0.133 0.096 0.516
CFBDS0059 T8+/Y? 0.119 0.084 0.526
ULAS1335 T8+/Y? 0.114 0.077 0.564
The 2 values for CFBDSIR1458 were derived using a median-binning of the spectra over (1) 17 pixels (resolution $\sim$170) (2) 6 pixels(resolution 500). The other values are from @Burgasser.2006 [@Warren.2007; @Burningham.2008; @Delorme.2008a; @Burningham.2009]
![$H$-band spectra of CFBSIR1458 compared with CFBDS0059, one the 2 coolest brown dwarfs known. Both spectra have been median-binned over 6 pixels to match the full ISAAC spectral resolution of 500. \[spectra\]](14277fg7.ps){width="8cm"}
![Zoom on the $H$-band peak of the spectra of CFBSIR1458 compared with CFBDS0059 and 2MASS0415 [@Burgasser.2003], the T8 spectral template. The CFBSIR1458 ISAAC spectrum has been median-binned to a resolution of 170. \[spectra\_zoom\]](14277fg8.ps){width="8cm"}
Conclusion
==========
We have described CFBDSIR, a new NIR survey dedicated at finding ultracool brown dwarfs and using WIRCam camera on the the CFHT. Complementing existing deep far-red data by new $J$-band observations, we select brown dwarfs candidates on their very red $z'-J$ colour. A robust PSF analysis allows us to derive reliable colours and to distinguish point-source-like brown dwarfs from most contaminants. The candidates are then confirmed by follow-up pointed NIR observations in $J$-band and confirmed ultracool brown dwarfs are imaged in $H$ and $K_{\rm s}$ bands. We used these photometric measurements to identify several new brown dwarfs, including 3 objects likely as cool as and possibly even cooler than any published brown dwarfs.
We presented CFBSIR1458, the first CFBDSIR ultracool brown dwarf confirmed by spectroscopy. The analysis of its $H$-band spectra, though at relatively low-signal-to-noise, robustly confirms it as later than T8 spectral type and hints at a temperature in the 550-600K range, so among the coolest brown dwarfs discovered. When the 335 square degree survey is completed, we expect to discover a sample of 10 to 15 ultracool brown dwarfs, more than doubling the currently known population of later than T8 objects and enabling study of them as a population rather than extreme individual objects. This will put strong constraints on cool stellar and planetary atmosphere, and with additional mid-infrared follow-up, will help to define the selection criteria for the upcoming WISE survey.
Thanks go to the queue observers at CFHT who obtained data for this paper. This research has made use of the VizieR catalogue access tool, of SIMBAD database and of Aladin, operated at CDS, Strasbourg. This research has benefitted from the M, L, and T-dwarf compendium housed at DwarfArchives.org and maintained by Chris Gelino, Davy Kirkpatrick, and Adam Burgasser. Financial support from the ”Programme National de Physique Stellaire" (PNPS) of CNRS/INSU, France, is gratefully acknowledged.
[^1]: Based on observations obtained with WIRCam, a joint project of CFHT, Taiwan, Korea, Canada, France, and the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institute National des Sciences de l’Univers of the Centre National de la Recherche Scientifique of France, and the University of Hawaii. Based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Sciences de l’Univers of the Centre National de la Recherche Scientifique (CNRS) of France, and the University of Hawaii. This work is based in part on data products produced at TERAPIX and the Canadian Astronomy Data Centre as part of the Canada-France-Hawaii Telescope Legacy Survey, a collaborative project of NRC and CNRS. Based on observations made with the ESO New Technology Telescope at the La Silla Observatory under programs ID 082.C-0506(A) and 083.C-0797(A) with SOFI at NTT and ESO VLT Director Discretionary Time program 282.C-5075 with ISAAC.
|
---
abstract: '$\alpha$-cluster excited states in $^{32}$S are investigated with an extended $^{28}$Si+$\alpha$ cluster model, in which the $^{28}$Si core deformation and rotation, and the $\alpha$-cluster breaking are incorporated. In the GCM calculation with the extended $^{28}$Si+$\alpha$ cluster model, the $\alpha$-cluster excited states are obtained near the $^{28}$Si+$\alpha$ threshold energy. The $^{28}$Si core deformation and rotation effects, and also the $\alpha$-clusters breaking in the $^{28}$Si+$\alpha$ system are discussed. It is found that the rotation of the oblately deformed $^{28}$Si core gives a significant effect to the $\alpha$-cluster excited states whereas the $\alpha$-cluster breaking gives only a minor effect.'
address:
author:
- ', , and'
title: '$\alpha$-cluster excited states in $^{32}$S'
---
|
---
abstract: 'We provide an easily verifiable condition for local $k$-connectedness of an inverse limit of polyhedra.'
address:
- 'Department of Mathematics and Statistics, University of North Carolina at Greensboro, Greensboro, NC 27412, USA'
- 'Faculty of Mathematics, Informatics, and Mechanics, University of Warsaw, Banacha 2, 02-097 Warszawa, Poland'
author:
- 'G. C. Bell'
- 'A. Nagórko'
bibliography:
- 'references2.bib'
title: 'Local $k$-connectedness of an inverse limit of polyhedra'
---
[^1]
Introduction
============
The Nöbeling space characterization theorem [@ageev2007a; @ageev2007b; @ageev2007c; @levin2006; @nagorkophd] states that if a space is strongly universal in the class of $n$-dimensional separable complete metric spaces and is $k$-connected and locally $k$-connected for each $k < n$, then it is homeomorphic to the $n$-dimensional Nöbeling space $N^{2n+1}_n$.
In geometric group theory many spaces arise naturally as inverse limits of polyhedra (simplicial complexes endowed with the metric topology). In particular, boundaries at infinity of hyperbolic spaces can be expressed as such. Striking examples of applications of the Nöbeling space characterization theorem are proofs that the boundary at infinity of the curve complex of the $(n+5)$-punctured $2$-dimensional sphere is homeomorphic to the $n$-dimensional Nöbeling space $N^{2n+1}_n$ [@henselprzytycki2011; @gabai2014].
In the present paper we prove a condition for local $k$-connectedness of an inverse limit of polyhedra that is easy to verify. It is designed to aid detection of local $k$-connectedness in many examples of spaces arising in geometric group theory. We prove the following theorem.
Let $K$ and $L$ be simplicial complexes. We say that a map $p \colon K \to L$ is [[**$n$-regular**]{}]{} if it is quasi-simplicial (i.e. it is simplicial into the first barycentric subdivision $\beta L$ of $L$) and if for each simplex $\delta$ of $\beta L$ the inverse image $p^{-1}(\delta)$ has vanishing homotopy groups in dimensions less than $n$ (regardless of the choice of basepoint).
\[thm:main theorem\] Let $$X = \lim_{\longleftarrow}\big( K_1 \xleftarrow{p_1} K_2 \xleftarrow{p_2} \cdots \big).$$ Assume that for each $i$ the following conditions are satisfied.
1. $K_i$ is a locally finite-dimensional simplicial complex endowed with the metric topology.
2. $p_i$ is a quasi-simplicial map that is surjective and $n$-regular.
Then $X$ is locally $k$-connected for each $k < n$, and each short projection $\pi^m_i \colon K_m \to K_i$ and each long projection $\pi_i \colon X \to K_i$ is a weak $n$-homotopy equivalence (it induces homomorphisms on homotopy groups in dimensions less than $n$, regardless of the choice of basepoint).
Note that there is no assumption of local finiteness of complexes in the statement of Theorem \[thm:main theorem\].
It is known that (under suitable assumptions) if the bonding maps in the inverse limit are $n$-soft, then the inverse limit is $k$-connected for $k < n$, [@chigogidze1989]. The condition for local $k$-connectedness given in the present paper may be regarded as a combinatorial analog of this statement for inverse limits of polyhedra. Note that $n$-regular bonding maps need not be $n$-soft.
Preliminaries
=============
In this section we set the basic definitions and reference known results that will be used in the later sections.
Absolute extensors in dimension $n$
-----------------------------------
We say that a space $X$ is [[**$k$-connected**]{}]{} if each map $\varphi \colon S^k \to X$ from a $k$-dimensional sphere into $X$ is null-homotopic in $X$.
We say that a space $X$ is [[**locally $k$-connected**]{}]{} if for each point $x \in X$ and each open neighborhood $U \subset X$ of $x$ there exists an open neighborhood $V$ of $x$ such that each map $\varphi \colon S^k \to V$ from a $k$-dimensional sphere into $V$ is null-homotopic in $U$.
We say that a metric space $X$ is an [[**absolute neighborhood extensor in dimension $n$**]{}]{} if every map into $X$ from a closed subset $A$ of an $n$-dimensional metric space extends over an open neighborhood of $A$. The class of absolute neighborhood extensors in dimension $n$ is denoted by $ANE(n)$ and its elements are called $ANE(n)$-spaces.
We say that a metric space $X$ is an [[**absolute extensor in dimension $n$**]{}]{} if every map into $X$ from a closed subset of an $n$-dimensional metric space $Y$ extends over the entire space $Y$. The class of absolute extensors in dimension $n$ is denoted by $AE(n)$ and its elements are called $AE(n)$-spaces.
\[def:AE(C)\] Let $\mathcal{C}$ be a class of topological spaces. We let $AE(\mathcal{C})$ denote the class of [[**absolute extensors for all spaces from the class $\mathcal{C}$**]{}]{}. We write $AE(X)$ for $AE(\{X\})$.
Absolute extensors and absolute neighborhood extensors in dimension $n$ were characterized by Dugundji in the following theorem.
\[thm:dugundji\] Let $X$ be a metric space. Then,
1. $X\in ANE(n)\iff X$ is locally $k$-connected for all $k< n$; and
2. $X\in AE(n)\iff X\in ANE(n)$ and $X$ is $k$-connected for all $k< n$.
\[lem:union is ae\] Assume that $A_1 \subset A_2 \subset \ldots$ is a sequence of subsets of a metric space such that each $A_i$ is closed and for each $i$, $A_i \subset \operatorname{Int}A_{i+1}$. If for each $i$, $A_i$ is $AE(n)$, then $A = \bigcup_i A_i$ is $AE(n)$.
Polyhedra
---------
For a simplicial complex, the underlying polyhedron has two topologies, the Whitehead (weak) topology and the metric topology. The metric topology is the topology of point-wise convergence of barycentric coordinates [@hu1965 p. 100]. The weak topology is metrizable if and only if the complex is locally finite, and it coincides with the metric topology in this case. Since we work in the metric category with complexes that are not locally finite, [**we always assume the metric topology on simplicial complexes**]{} [@lundell1969; @hu1965].
Let $K$ be a simplicial complex. We let $\tau(K)$ denote the [[**triangulation of $K$**]{}]{} (the set of simplices of $K$). We let $V(K)$ denote the [[**vertex set of $K$**]{}]{}. We let $\beta K$ denote the [[**barycentric subdivision of $K$**]{}]{} (i.e., the same space but with a finer triangulation $\tau(\beta K)$).
\[def:polyhedron metric\] Let $K$ be a simplicial complex. Let $\kappa > 0$. Let $\ell_1(V(K))$ denote the Banach space $\{ x \in \mathbb{R}^{V(K)} \colon \sum_{v \in V(K)} | (x)_v | < \infty \}$, where $(x)_v$ denotes the $v$-th coordinate, equipped with the standard $\| \cdot \|_1$ norm. For $v \in v(K)$ define $e_v \colon V(K) \to \mathbb{R}$ by the formula $e_v(w) = 0$ for $w \neq v$ and $e_v(v) = \kappa$. We embed each vertex $v$ of $K$ as $e_v \in \ell_1(V(K))$ and extend this embedding to $K$ to be affine on each simplex of $K$. We consider $K$ to be a subspace of $\ell_1(V(K))$. We call the induced metric on $K$ the [[**metric of scale $\kappa$ on $K$**]{}]{}. For $\kappa=1$ it is the standard metric, as defined in [@hu1965 p. 100]. The topology induced by this metric is called [[**the metric topology**]{}]{}.
A [[**polyhedron**]{}]{} is a simplicial complex endowed with the metric topology.
Let $K$ and $L$ be polyhedra. We say that $K$ is a [[**full subpolyhedron**]{}]{} of $L$ if whenever the vertices $v_0,\ldots, v_n$ span a simplex in $K$ and each $v_i$ is a vertex in $L$, then the $v_i$ span a simplex in $L$.
\[lem:complex is ane\] A locally finite-dimensional polyhedron is a complete metric $ANE(\infty)$-space.
It is complete by [@hu1965 Lemma 11.5]. It is an $ANE(\infty)$-space by [@hu1965 Theorem 11.3].
Let $K$ and $L$ be simplicial complexes and let $p \colon K \to L$. We say that $p$ is [[**quasi-simplicial**]{}]{} if it is a simplicial map into $\beta L$.
\[lem:qs is lipschitz\] Assume that $K$ and $L$ are polyhedra endowed with metrics of scale $\kappa$ and $\lambda$ respectively. If $p \colon K \to L$ is quasi-simplicial, then it is $\frac{\lambda}{2\kappa}$-Lipschitz.
Weak $n$-homotopy
-----------------
We say that a map is a [[**weak $n$-homotopy equivalence**]{}]{} if it induces isomorphisms on homotopy groups of dimensions less than $n$, regardless of the choice of basepoint.
Let ${{\mathcal{F}_{}}}$ be a cover of a space $X$. We say that the two maps $f, g \colon Y \to X$ are [[**${{\mathcal{F}_{}}}$-close**]{}]{} if for each $y \in Y$ there exists $F \in {{\mathcal{F}_{}}}$ such that $f(y), g(y) \in F$.
Let ${{\mathcal{U}_{}}}$ be a cover of a space $X$. We say that two maps $f, g \colon Y \to X$ are [[**${{\mathcal{U}_{}}}$-homotopic**]{}]{} if there exists a homotopy $H \colon Y \times [0, 1] \to X$ whose paths refine ${{\mathcal{U}_{}}}$, i.e. for each $y \in Y$ there exists $U \in {{\mathcal{U}_{}}}$ such that $H(\{ y \} \times [0, 1])\subset U$.
Carrier Theorem {#sec:carrier theorem}
---------------
Let $\mathcal{C}$ be a class of topological spaces. We say that a cover ${{\mathcal{F}_{}}}$ of a topological space is a $\mathcal{C}$-cover, if for each $\mathcal{A} \subset \mathcal{F}$ the intersection $\bigcap \mathcal{A}$ is either empty or belongs to $\mathcal{C}$.
We say that a cover is [[**locally finite-dimensional**]{}]{} if its nerve is locally finite-dimensional.
A [[**carrier**]{}]{} is a function $C \colon {{\mathcal{F}_{}}} \to {{\mathcal{G}_{}}}$ from a cover ${{\mathcal{F}_{}}}$ of a space $X$ into a collection ${{\mathcal{G}_{}}}$ of subsets of a topological space such that for each $\mathcal{A} \subset {{\mathcal{F}_{}}}$ if $\bigcap_{A\in\mathcal{A}}A \neq \emptyset$, then $\bigcap_{A \in
\mathcal{A}} C(A) \neq \emptyset$. We say that a map $f$ is [[**carried by $C$**]{}]{} if it is defined on a closed subset of $X$ and $f(F) \subset C(F)$ for each $F \in {{\mathcal{F}_{}}}$. Here we write $f(F)$ to mean $f(F\cap \operatorname{dom}(f)).$
Assume that $C \colon {{\mathcal{F}_{}}} \to {{\mathcal{G}_{}}}$ is a carrier such that ${{\mathcal{F}_{}}}$ is a cover of a space $X$ and ${{\mathcal{G}_{}}}$ is an $AE(X)$-cover of another space. If ${{\mathcal{F}_{}}}$ is closed, locally finite, and locally finite-dimensional, then each map carried by $C$ extends to a map of the entire space $X$, also carried by $C$.
If ${{\mathcal{F}_{}}}$ is a closed locally finite locally finite-dimensional $AE(n)$-cover of a space $Y$, then any two ${{\mathcal{F}_{}}}$-close maps from a metric space of dimension less than $n$ into $Y$ are ${{\mathcal{F}_{}}}$-homotopic.
Covers
------
We regard covers as indexed collections of sets and use the usual notation, ${{\mathcal{F}_{}}} = \{ F_i \}_{i \in I}$, where $I$ denotes the indexing set.
\[def:stars\] Let $K$ be a polyhedron. Let $L \subset K$ be a subcomplex of $K$. The [[**open star $\operatorname{ost}_K L$ of $L$ in $K$**]{}]{} is the complement of the union of all simplices of $K$ that do not intersect $L$: $$\operatorname{ost}_K L = K \setminus \bigcup \{ \delta \in \tau(K) \colon \delta \cap L = \emptyset \}\text{.}$$ The [[**barycentric star $\operatorname{bst}_K L$ of $L$ in $K$**]{}]{} is the union of all simplices of $\beta K$ that intersect $L$: $$\operatorname{bst}_K L = \bigcup \{ \delta \in \tau(\beta K) \colon \delta \cap L \neq \emptyset \}\text{.}$$
We let $${{\mathcal{O}_{K}}} = \{ \operatorname{ost}_K \{ v \} \colon v \in V(K) \}$$ denote the [[**cover of $K$ by open stars of vertices**]{}]{}. We let $${{\mathcal{B}_{K}}} = \{ \operatorname{bst}_K \{ v \} \colon v \in V(K) \}$$ denote a [[**cover of $K$ by barycentric stars of vertices**]{}]{}.
\[lem:stars\] Let $K$ be a polyhedron. The cover ${{\mathcal{B}_{K}}}$ by barycentric stars of vertices is a closed locally finite $AE(\infty)$-cover of $K$. Moreover, if $K$ is locally finite-dimensional, then ${{\mathcal{B}_{K}}}$ is locally finite-dimensional. The cover ${{\mathcal{O}_{K}}}$ by open stars of vertices is an open $AE(\infty)$-cover of $K$.
Let $$X = \lim_{\longleftarrow} \left(
K_1\xleftarrow{p_1}K_2\xleftarrow{p_2}\cdots \right)$$ Let $v(K_i)$ denote the set of vertices of $K_i$. We let $${{\mathcal{O}_{K_i}}} = \{ O_v = \operatorname{ost}_{K_i} v \}_{v \in v(K_i)}$$ be the cover of $K_i$ by open stars of vertices of $K_i$ (see Definition \[def:stars\]) and $${{\mathcal{O}_{i}}} = \{ \pi^{-1}_i(\operatorname{ost}_{K_i} v) \}_{v \in v(K_i)}$$ be the cover of $X$ by sets of threads that pass through elements of ${{\mathcal{O}_{K_i}}}$.
Let $$X = \lim_{\longleftarrow} \left(
K_1\xleftarrow{p_1}K_2\xleftarrow{p_2}\cdots \right)$$ Let $v(K_i)$ denote the set of vertices of $K_i$. We let $${{\mathcal{B}_{K_i}}} = \{ \operatorname{bst}_{K_i} v \}_{v \in v(K_i)}$$ be the cover of $K_i$ by barycentric stars of vertices of $K_i$ (see Definition \[def:stars\]) and $${{\mathcal{B}_{i}}} = \{ \pi^{-1}_i(\operatorname{bst}_{K_i} v) \}_{v \in v(K_i)}$$ be the cover of $X$ by sets of threads that pass through elements of ${{\mathcal{B}_{K_i}}}$.
We say that a cover ${{\mathcal{F}_{}}} = \{ F_i \}_{i \in I}$ is [[**isomorphic**]{}]{} to a cover ${{\mathcal{G}_{}}} = \{ G_i \}_{i \in I}$ if for each $J \subset I$ we have $$\bigcap_{j \in J} F_j \neq \emptyset \iff
\bigcap_{j \in J} G_j \neq \emptyset.$$ Note the identical indexing set of ${{\mathcal{F}_{}}}$ and ${{\mathcal{G}_{}}}$.
Let $p \colon Y \to Z$ be a map. Let ${{\mathcal{F}_{}}} = \{ F_i \}_{i \in I}$ be a cover of $Z$. A [[**pull-back of ${{\mathcal{F}_{}}}$**]{}]{} is a cover $p^{-1}({{\mathcal{F}_{}}})$ of $Y$ defined by the formula $$p^{-1}({{\mathcal{F}_{}}}) = \{ p^{-1}(F_i) \}_{i \in I}.$$ The pull-back $p^{-1}({{\mathcal{F}_{}}})$ retains the indexing set $I$ of ${{\mathcal{F}_{}}}$.
\[lem:pullback properties\] Let $f \colon X \to Y$ be a map, let ${{\mathcal{G}_{}}}$ be a cover of $Y$ and let $f^{-1}({{\mathcal{G}_{}}})$ be a pull-back of ${{\mathcal{G}_{}}}$.
1. If $f$ is surjective, then the covers $f^{-1}({{\mathcal{G}_{}}})$ and ${{\mathcal{G}_{}}}$ are isomorphic.
2. If ${{\mathcal{G}_{}}}$ is open / closed / locally finite / locally finite-dimensional, then so is $f^{-1}({{\mathcal{G}_{}}})$.
Nerve Theorem
=============
A Nerve Theorem in its abstract form states that if two spaces admit isomorphic $AE(n)$-covers, then they are weak $n$-homotopy equivalent. For general spaces, some local finiteness and dimension restrictions are placed on the covers [@nagorko2007]. In our case we need such a theorem without these restrictions, which we are able to prove for polyhedra covered by subcomplexes.
The goal of this section is to prove the following theorem.
\[thm:nregular pullback\] If a quasi-simplicial map $p \colon K \to L$ of polyhedra is surjective and $n$-regular, then it is a weak $n$-homotopy equivalence and the pull-back $p^{-1}({{\mathcal{B}_{L}}})$ is an $AE(n)$-cover of $K$.
The main tool used in the proof is Theorem \[thm:nerve theorem\], which when applied to a canonical map into the nerve of a cover [@nagorko2007] implies the usual Nerve Theorem.
Nerve Theorem for non locally finite covers
-------------------------------------------
\[lem:deformation\] If $K$ is a full subpolyhedron of a polyhedron $L$, then
1. $K$ is a deformation retract of $\operatorname{bst}_L K$; and
2. $K$ is a deformation retract of $\operatorname{ost}_L K$.
We let $V_L$ denote the set of vertices of $L$ and $V_K \subset V_L$ denote the set of vertices of $K$. We regard $L$ as a subspace of $\ell_1(V_L)$ (see Definition \[def:polyhedron metric\]). Let $p \colon L \to \ell_1(V_L)$ be the map that sends all coordinates that do not belong to $K$ to $0$, defined by the formula $$\left( p(x) \right)_v =
\left\{
\begin{array}{ll}
0 & v \in V_L \setminus V_K \\
(x)_v & v \in V_K
\end{array}
\right..$$
We have $$\operatorname{ost}_L K = \{ x \in L \colon p(x) \neq 0 \}.$$ Let $q \colon \operatorname{ost}_L K \to L$ be defined by the formula $$q(x) = \frac{p(x)}{\| p(x) \|_1}.$$ We define a map $\varPhi \colon \operatorname{ost}_L K \times [0, 1] \to L$ by the formula $$\varPhi(x, t) = t \cdot q(x) + (1-t)\cdot x\text{.}$$ Observe that $\operatorname{supp}\varPhi(x, t) \subset \operatorname{supp}(p(x))$ and $V_K \cap \operatorname{supp}\varPhi(x, t) \neq \emptyset$, hence $\varPhi(x, t) \in \operatorname{ost}_L K$. Note that $\varPhi(x, 1) \in K$ because $K$ is a full subcomplex. Hence $\varPhi$ is a deformation retraction of $\operatorname{ost}_L K$ to $K$.
Observe that $$\operatorname{bst}_L K = \{ x \in L \colon \exists_{v \in V_K} (x)_v \geq \max \{ (x)_w \colon w \in V_L \} \}.$$ Since $\varPhi$ preserves the relation $\exists_{v \in V_K} (x)_v \geq \max \{ (x)_w \colon w \in V_L \}$, $\varPhi$ restricted to $\operatorname{bst}_L K \times [0,1]$ is into $\operatorname{bst}_L K$, hence it is a deformation retraction of $\operatorname{bst}_L K$ to $K$.
\[cor:open star swelling\] If $K$ is a polyhedron and $\mathcal{K} = \{ K_i \}_{i \in I}$ is an $AE(n)$-cover of $K$ by subcomplexes, then the collection $$\operatorname{ost}_{\beta K} \mathcal{K} = \{ \operatorname{ost}_{\beta K} K_i \}_{i \in I}$$ is an open $AE(n)$-cover of $K$. Moreover, $\operatorname{ost}_{\beta K} \mathcal{K}$ and $\mathcal{K}$ are isomorphic covers of $K$.
Observe that if $\{ A_j \}_{j \in J}$ is a collection of subcomplexes of a polyhedron $K$, then $$\bigcap_{j \in J} \operatorname{ost}_{\beta K} A_j =
\operatorname{ost}_{\beta K} \bigcap_{j \in J} A_j.$$ An application of Lemma \[lem:deformation\] finishes the proof.
\[lem:uclose uhomotopic\] If ${{\mathcal{U}_{}}}$ is an open $AE(n)$-cover of a space $Y$, then any two ${{\mathcal{U}_{}}}$-close maps from a metric space of dimension less than $n$ into $Y$ are ${{\mathcal{U}_{}}}$-homotopic.
Let ${{\mathcal{U}_{}}}$ be an open $AE(n)$-cover of a metric space $Y$. Let $X$ be a metric space of dimension less than $n$. Let $f, g \colon X \to Y$ be ${{\mathcal{U}_{}}}$-close. We have to show that $f$ and $g$ are homotopic by a homotopy whose paths refine ${{\mathcal{U}_{}}}$.
Let ${{\mathcal{V}_{}}} = \{ V_U \}_{U \in {{\mathcal{U}_{}}}}$ be the collection of subsets of $X$ defined by $V_U = f^{-1}(U) \cap g^{-1}(U)$. Since $f$ and $g$ are ${{\mathcal{U}_{}}}$-close, ${{\mathcal{V}_{}}}$ is an open cover of $X$.
Let ${{\mathcal{W}_{}}}$ be a closed, locally finite cover of $X$ with multiplicity at most $n$ that refines ${{\mathcal{V}_{}}}$. Let ${{\mathcal{F}_{}}} = \{ W \times [0,1] \colon W \in {{\mathcal{W}_{}}} \}$ be a cover of $X \times [0,1]$. Note that ${{\mathcal{F}_{}}}$ is a closed, locally finite, locally finite-dimensional cover of $X \times [0,1]$.
Let $H_0 \colon K \times \{ 0, 1 \} \to Y$ be the map defined by $H_0(x, 0) = f(x)$ and $H_0(x,1) = g(x)$.
Let $C \colon {{\mathcal{V}_{}}} \to {{\mathcal{U}_{}}}$ be the map defined by $C(V_U) = U$. It is a carrier and both $f$ and $g$ are carried by $C$, directly from the definition of ${{\mathcal{V}_{}}}$. Let $D \colon {{\mathcal{W}_{}}} \to {{\mathcal{V}_{}}}$ be a map such that for each $W \in {{\mathcal{W}_{}}}$ we have $W \subset D(W)$. The map $D$ exists because ${{\mathcal{W}_{}}}$ refines ${{\mathcal{V}_{}}}$. Let $E \colon {{\mathcal{F}_{}}} \to {{\mathcal{W}_{}}}$ be a map defined by $D(W \times [0,1]) = W$. The composition $C \circ D \circ E$ is a carrier and $H_0$ is carried by it. Observe that ${{\mathcal{U}_{}}}$ is an $AE(X \times [0,1])$-cover of $Y$ (see Definition \[def:AE(C)\]). By the Carrier Theorem, $H_0$ can be extended to a map $H \colon X \times [0,1]$ that is carried by $C \circ D \circ E$. Clearly, $H$ is a homotopy between $f$ and $g$. For each path $\{ x \} \times [0, 1] \subset X \times [0,1]$ there exists element $W \times [0, 1] \in {{\mathcal{F}_{}}}$ such that $\{ x \} \times [0, 1] \subset W \times [0, 1]$, as ${{\mathcal{W}_{}}}$ is a cover of $X$. Since $H$ is carried into ${{\mathcal{U}_{}}}$, the whole path lies in an element of ${{\mathcal{U}_{}}}$. Therefore $H$ is a ${{\mathcal{U}_{}}}$-homotopy.
\[thm:nerve theorem\] Let $\mathcal{K} = \{ K_i \}_{i \in I}$ be a cover of a polyhedron $K$ by subcomplexes. Let $\mathcal{L} = \{ L_i \}_{i \in I}$ be a cover of a polyhedron $L$ by subcomplexes. Let $p \colon K \to L$ be a surjective simplicial map that maps elements of $\mathcal{K}$ into the corresponding elements of $\mathcal{L}$ (i.e. $p(K_i) \subset L_i$ for each $i \in I$). If $\mathcal{K}$ and $\mathcal{L}$ are isomorphic $AE(n)$-covers, then $p$ is a weak $n$-homotopy equivalence.
To begin, we will show that $p$ induces monomorphisms on homotopy groups of dimensions less than $n$. Let $m < n$. Let $\varphi \colon S^m \to K$. Assume that $p \circ \varphi$ is null-homotopic in $L$. Let $\varPhi \colon B^{m+1} \to L$ denote such a null-homotopy ($\varPhi_{\mid S^m} = p \circ \varphi$). We have to show that $\varphi$ is null-homotopic in $K$.
Let ${{\mathcal{F}_{}}}$ be a finite closed cover of $B^{m+1}$ with mesh small enough so that for each $F \in {{\mathcal{F}_{}}}$ we can pick $i_F \in I$ such that $$(\ast)\ \varphi(F \cap S^m) \subset \operatorname{ost}_{\beta K} K_{i_F} \text{ and }
(\ast\ast)\ \varPhi(F) \subset \operatorname{ost}_{\beta L} L_{i_F}.$$ By $(\ast\ast)$, a map $C \colon \mathcal{F} \to \operatorname{ost}_{\beta L} \mathcal{L}$ defined by $C(F) = \operatorname{ost}_{\beta L} L_{i_F}$ is a carrier. Since $\mathcal{K}$ and $\mathcal{L}$ are isomorphic, the map $C' \colon \mathcal{F} \to \operatorname{ost}_{\beta K} \mathcal{K}$ defined by $C'(F) = \operatorname{ost}_{\beta K} K_{i_F}$ is a carrier as well. By $(\ast)$, $\varphi$ is carried by $C'$. By Corollary \[cor:open star swelling\], $\operatorname{ost}_{\beta K} \mathcal{K}$ is an $AE(n)$-cover. By the Carrier Theorem, $\varphi$ extends to a map $\tilde \varphi \colon B^{m+1} \to K$ that is carried by $C'$. This is a null-homotopy of $\varphi$ in $K$ and we are done with the proof that $p$ induces monomorphisms on homotopy groups of dimensions less than $n$.
Next we will show that $p$ induces epimorphisms on homotopy groups of dimensions less than $n$. Let $m < n$. Let $\psi \colon S^m \to L$. Let ${{\mathcal{G}_{}}}$ be a closed finite cover of $S^m$ that refines $\psi^{-1}(\operatorname{ost}_{\beta L} \mathcal{L})$. For each $G \in {{\mathcal{G}_{}}}$ pick $i_G \in I$ such that $\psi(G) \subset \operatorname{ost}_{\beta L} L_{i_G}$. Then the map $D \colon {{\mathcal{G}_{}}} \to \operatorname{ost}_{\beta L} \mathcal{L}$ defined by $D(G) = \operatorname{ost}_{\beta L} L_{i_G}$ is a carrier and $\psi$ is carried by $D$. Since $\mathcal{K}$ and $\mathcal{L}$ are isomorphic, the map $D'(G) = \operatorname{ost}_{\beta K} K_{i_G}$ is a carrier as well. By Corollary \[cor:open star swelling\], $\operatorname{ost}_{\beta K} \mathcal{K}$ is an $AE(n)$-cover. By the Carrier Theorem, there exists a map $\tilde \psi \colon S^m \to K$ that is carried by $D'$. Observe that if $x \in G \in {{\mathcal{G}_{}}}$, then $\tilde \psi(x) \in \operatorname{ost}_{\beta_K} K_{i_G}$ and $\psi(x) \in \operatorname{ost}_{\beta L} L_{i_G}$. Since $p(K_{i_G}) \subset L_{i_G}$, the maps $\psi$ and $p \circ \tilde \psi$ are $\operatorname{ost}_{\beta L} \mathcal{L}$-close. By Corollary \[cor:open star swelling\], $\operatorname{ost}_{\beta L} \mathcal{L}$ is an $AE(n)$-cover and by Lemma \[lem:uclose uhomotopic\], $\psi$ and $p \circ \tilde \psi$ are homotopic. Hence $p$ induces epimorphisms on homotopy groups of dimensions less than $n$. This concludes the second part of the proof.
Proof of Theorem \[thm:nregular pullback\]
------------------------------------------
\[lem:simplicial regular\] Let $K$ and $L$ be polyhedra. If $p \colon K \to L$ is a simplicial, surjective map such that for each $\delta \in \tau(L)$ the inverse image $p^{-1}(\delta)$ is an $AE(n)$, then $p$ is a weak $n$-homotopy equivalence.
Apply Theorem \[thm:nerve theorem\] with the cover $\mathcal{L} = \tau(L)$ of $L$ and the cover $\mathcal{K} = p^{-1}(\mathcal{L})$ of $K$.
(Proof of Theorem \[thm:nregular pullback\]) The map $p$ is simplicial onto $\beta K$ and satisfies the conditions of Lemma \[lem:simplicial regular\], hence it is a weak $n$-homotopy equivalence.
Let $\mathcal{A} \subset {{\mathcal{B}_{L}}}$ such that $\bigcap_{A \in \mathcal{A}} A \neq \emptyset$. By Lemma \[lem:stars\], $\bigcap_{A \in \mathcal{A}} A$ is an absolute extensor in dimension $n$. We have $B = \bigcap_{A \in \mathcal{A}} p^{-1}(A) = p^{-1}(\bigcap_{A \in \mathcal{A}} A)$. The map $p_{|B} \colon B \to \bigcap_{A \in \mathcal{A}} A$ satisfies the conditions of Lemma \[lem:simplicial regular\], hence it is a weak $n$-homotopy equivalence. Therefore $B$ has vanishing homotopy groups in dimensions less than $n$ and by Theorem \[thm:dugundji\], it is an absolute extensor in dimension $n$. Hence $p^{-1}({{\mathcal{B}_{L}}})$ is an $AE(n)$-cover.
A lifting condition
===================
\[lem:single lift\] Let $p \colon Y \to Z$ be a surjective map of metric spaces. Let ${{\mathcal{F}_{}}}$ be a cover of $Z$. Assume that either ${{\mathcal{F}_{}}}$ is an open cover or ${{\mathcal{F}_{}}}$ is a closed, locally finite, locally finite-dimensional cover. Assume that $p^{-1}({{\mathcal{F}_{}}})$ is an $AE(n)$-cover.
Let $X$ be an at most $n$-dimensional metric space. Let $f \colon X \to Z$. Let $A$ be a closed subset of $X$. Let $g_0 \colon A \to Y$ be a map such that $p \circ g_0 = f_{|A}$ (a lift of $f$ on $A$).
Then there exists a map $g \colon X \to Y$ that satisfies the following conditions:
1. $g_{|A} = g_0$ ($g$ extends $g_0$); and
2. $p\circ g$ is ${{\mathcal{F}_{}}}$-close to $f$.
We have the following commutative diagram. $$\xymatrix@M=8pt@C=35pt{
Y \ar[r]^{p} & Z \\
A \ar@{^{(}->}[r] \ar[u]_{g_0} & X \ar[u]^{f} \ar@{.>}[ul]_{g} \\
}$$ Let $\mathcal{P} = p^{-1}({{\mathcal{F}_{}}})$ and ${{\mathcal{G}_{}}} = f^{-1}({{\mathcal{F}_{}}})$. Let $C \colon {{\mathcal{G}_{}}} \to \mathcal{P}$ be defined by the formula $C(f^{-1}(F)) = p^{-1}(F)$. Since $p$ is surjective, $C$ is a carrier. For each $F \in {{\mathcal{F}_{}}}$ and each $x \in f^{-1}(F) \cap A$ we have $g_0(x) \in p^{-1}(f(x)) \in p^{-1}(F)$; hence $g_0$ is carried by $C$.
Let $\mathcal{H}$ be a closed locally finite locally finite-dimensional cover of $X$ such that $\mathcal{H}$ refines ${{\mathcal{G}_{}}}$. If ${{\mathcal{F}_{}}}$ is an open cover, then $\mathcal{H}$ exists as a refinement of an open cover ${{\mathcal{G}_{}}}$. If ${{\mathcal{F}_{}}}$ is closed locally finite locally finite-dimensional, then we can take $\mathcal{H} = {{\mathcal{G}_{}}}$, which has the required properties by Lemma \[lem:pullback properties\]. Let $D \colon \mathcal{H} \to {{\mathcal{G}_{}}}$ be any map such that $H \subset D(H)$ for each $H \in \mathcal{H}$.
Observe that $C \circ D \colon \mathcal{H} \to \mathcal{P}$ is a carrier, $g_0$ is carried by $C \circ D$, $\mathcal{P}$ is an $AE(n)$-cover, $X$ is at most $n$-dimensional metric space and $\mathcal{H}$ is closed, locally finite and locally finite-dimensional. By the Carrier Theorem, there exists a map $g \colon X \to Y$ that extends $g_0$ and that is carried by $C \circ D$. This implies that for each $x \in X$ there exists $F \in {{\mathcal{F}_{}}}$ such that $x \in f^{-1}(F)$ and $g(x) \subset p^{-1}(F)$. Therefore $p(g(x)) \in F$ so $p \circ g$ is ${{\mathcal{F}_{}}}$-close to $f$.
A lifting condition for inverse limits
--------------------------------------
\[def:limit\] Let $$Z=\lim_{\longleftarrow}\left( Z_1\xleftarrow{p_1}Z_2\xleftarrow{p_2}\cdots \right).$$ Let ${{\mathcal{F}_{i}}}$ be a cover of $Z_i$. We define the following conditions.
1. For each $i$, $Z_i$ is a complete metric space.
2. For each $i$, the bonding map $p_i \colon Z_{i+1} \to Z_i$ is surjective and $1$-Lipschitz.
3. For each $i$, ${{\mathcal{F}_{i}}}$ is either an open cover or a closed, locally finite, locally finite-dimensional cover.
4. For each $i$, the pull-back $p_i^{-1}({{\mathcal{F}_{i}}})$ is an $AE(n)$-cover.
5. $\sum_i \operatorname{mesh}{{\mathcal{F}_{i}}} < \infty$.
We denote short projections in the inverse limit by $\pi^k_i \colon Z_k \to Z_i (k > i)$ and long projections in the inverse limit by $\pi_k \colon Z \to Z_k$.
\[lem:limit lift\] Let $$Z=\lim_{\longleftarrow}\left( Z_1\xleftarrow{p_1}Z_2\xleftarrow{p_2}\cdots \right).$$ Let ${{\mathcal{F}_{i}}}$ be a cover of $Z_i$. Assume that conditions of Definition \[def:limit\] are satisfied.
Let $X$ be an at most $n$-dimensional metric space. Let $f \colon X \to Z_1$. Let $A$ be a closed subset of $X$. Let $g_0 \colon A \to Z$ be a map such that $\pi_1 \circ g_0 = f_{|A}$ (a lift of $f$ on $A$).
Then there exists a map $g \colon X \to Z$ such that $g_{|A} = g_0$ ($g$ extends $g_0$).
Let $f_1 = f$. Applying Lemma \[lem:single lift\] for each $i>1$ we construct a map $f_i \colon X \to Z_i$ that satisfies the following conditions.
1. $p_{i-1} \circ f_{i}$ is ${{\mathcal{F}_{i-1}}}$-close to $f_{i-1}$.
2. $\pi_i \circ g_0 = f_i\mid_A$.
For each $k$ and each $m > k$ we let $$a_m^k=\pi^m_k\circ f_m: X \to Z_k.$$ By (i), we have ${{d_{\sup\nolimits}}}(f_m, p_m \circ f_{m+1}) \leq \operatorname{mesh}{{\mathcal{F}_{m}}}$. By (B) the short projection $\pi^m_k$ is $1$-Lipschitz, hence ${{d_{\sup\nolimits}}}(a^k_m, a^k_{m+1})
= {{d_{\sup\nolimits}}}(\pi^m_k \circ f_m, \pi^m_k \circ p_m \circ f_{m+1}) \leq \operatorname{mesh}{{\mathcal{F}_{m}}}$. Therefore for $l > m$ we have $${{d_{\sup\nolimits}}}(a^k_l, a^k_m) \leq {{d_{\sup\nolimits}}}(a^k_l, a^k_{l-1}) + {{d_{\sup\nolimits}}}(a^k_{l-1}, a^k_{l-2})
+ \cdots + {{d_{\sup\nolimits}}}(a^k_{m+1}, a^k_m) \leq$$ $$\leq \operatorname{mesh}{{\mathcal{F}_{l-1}}} + \operatorname{mesh}{{\mathcal{F}_{l-2}}} + \cdots + \operatorname{mesh}{{\mathcal{F}_{m+1}}} + \operatorname{mesh}{{\mathcal{F}_{m}}}.$$ By (E), the sequence $a_m^k$ is uniformly convergent. Let $$a^k = \lim_{m \to \infty} a_m^k.$$ By (A), $Z_k$ is complete, hence $a^k \colon X \to Z_k$ is well defined.
It follows from the definition that $p_k \circ a^{k+1}_m = a^k_m$. Passing to the limit we have $$p_k \circ a^{k+1} = \lim_{m \to \infty} p_k \circ a^{k+1}_m = \lim_{m \to\infty} a^k_m = a^k.$$ Therefore, for each $x$ the sequence $(a^k(x))_k$ is a thread in $Z$ and we can define a map $g \colon X \to Z$ by the formula $$(g(x))_k = a^k(x).$$
Observe that by (ii) and by the definition of $a^k_m$ we have $$a^k_m |_A = \pi^m_k \circ f_m |_A = \pi^m_k \circ \pi_m \circ g_0 |_A = \pi_k \circ g_0 |_A.$$ Therefore $a^k\mid_A = \pi_k \circ g_0$ for each $k$, hence $g |_A = g_0 |_A$.
\[thm:ane condition\] Let $$Z=\lim_{\longleftarrow}\left( Z_1\xleftarrow{p_1}Z_2\xleftarrow{p_2}\cdots \right).$$ Let ${{\mathcal{F}_{i}}}$ be a cover of $Z_i$. If the conditions of Definition \[def:limit\] are satisfied and $Z_1$ is an $ANE(n)$, then $Z$ is an $ANE(n)$.
We will show that $Z$ is an $ANE(n)$ directly from the definition. Let $X$ be an at most $n$-dimensional metric space and let $A\subset X$ be a closed subset. Take $g_0\colon A\to Z$. By assumption, $Z_1$ is an $ANE(n)$ so we can extend $\pi_1\circ g_0:A\to Z_1$ to a map $f_1:U\to Z_1$, where $U$ is an open neighborhood of $A$ in $X$. By Lemma \[lem:limit lift\], $f_1$ can be lifted to a map $g \colon U \to Z$ such that $g_{|A} = g_0$. This is the extension of $f$ we sought.
Local $k$-connectedness of inverse limits of polyhedra
======================================================
Let $$\label{eq:sequence}\tag{L}
X = \lim_{\longleftarrow}\left( K_1 \xleftarrow{p_1} K_2 \xleftarrow{p_2} \cdots.\right)$$ Fix $m$ and let $A \subset K_m$ be a subcomplex of $K_m$. A [[**restriction of to $A$**]{}]{} is the inverse limit $$\label{eq:restriction}\tag{R}
X' = \lim_{\longleftarrow} \left(K'_m \xleftarrow{p'_m} K'_{m+1} \xleftarrow{p'_{m+1}} \cdots\right),$$ where $K'_m =\ A$ and for each $j \geq m$, $K'_{j+1} = p^{-1}_j(K'_j)$ and $p'_j = p_j \mid_{K'_{j+1}}$.
\[lem:restriction\] Let be the restriction of to a subcomplex $A \subset K_m$.
If, for each $i$, satisfies the conditions:
1. $K_i$ is a polyhedron and
2. $p_i$ is a quasi-simplicial map that is surjective and $n$-regular,
then so does . The inverse limit $X'$ is homeomorphic to $\pi_m^{-1}(A)$, where $\pi_m \colon X \to X_m$ denotes the long projection.
We have $$\pi^{-1}_m(A) = \lim_{\longleftarrow} \left(K'_1 \xleftarrow{p'_1} K'_{2} \xleftarrow{p'_{2}} \cdots\right),$$ where $K'_j = p_j(K'_{j+1})$ and $p'_j = p_j \mid_{K'_{j+1}}$ for $j < m$. The restriction is the same sequence with first $m-1$ elements removed. This changes the metric on the limit, but not the topology, hence $X'$ is homeomorphic to $\pi_m^{-1}(A)$.
The other conditions are trivial to verify.
\[thm:ane limit of polyhedra\] Let $$X = \lim_{\longleftarrow} \left(K_1 \xleftarrow{p_1} K_2 \xleftarrow{p_2} \cdots\right).$$ Assume that for each $i$ the following conditions are satisfied:
1. $K_i$ is a locally finite-dimensional polyhedron; and
2. $p_i$ is a quasi-simplicial map that is surjective and $n$-regular.
Then
1. $X$ is an $ANE(n)$;
2. each short projection $\pi^k_i \colon K_k \to K_i$ and each long projection $\pi_i \colon X \to X_i$ is a weak $n$-homotopy equivalence;
3. for each $i$, the covers ${{\mathcal{O}_{i}}}$ and ${{\mathcal{B}_{i}}}$ are $AE(n)$-covers of $X$.
Fix a metric of scale $2^{-i}$ on $K_i$ (see Definition \[def:polyhedron metric\]). Let $(*)$ denote the inverse limit $$X = \lim_{\longleftarrow} \left( K_1 \xleftarrow{p_1} K_2 \xleftarrow{p_2} \cdots\right)$$ along with a sequence ${{\mathcal{B}_{K_i}}}$ of covers, where ${{\mathcal{B}_{K_i}}}$ is a cover of $K_i$ by barycentric stars of vertices.
We will verify that $(*)$ satisfies conditions (A)–(E) of Definition \[def:limit\].
Condition (A): By Lemma \[lem:complex is ane\], every $K_i$ is a complete metric space.
Condition (B): By condition (II), $p_i$ is surjective. By the choice of scale on $K_i$ and Lemma \[lem:qs is lipschitz\], each bonding map is $1$-Lipschitz.
Condition (C): By Lemma \[lem:stars\], each ${{\mathcal{B}_{K_i}}}$ is a closed, locally finite, locally finite-dimensional cover.
Condition (D): By assumption $p_i$ is quasi-simplicial and $n$-regular. Hence by Theorem \[thm:nregular pullback\], the pull-back $p_i^{-1}({{\mathcal{B}_{K_i}}})$ is an $AE(n)$-cover.
Condition (E): By the choice of the metric on $K_i$, we have $\sum_i \operatorname{mesh}{{\mathcal{B}_{K_i}}} \leq \sum_i \operatorname{diam}K_i < \infty$.
To prove (1) it is enough to verify the conditions of Theorem \[thm:ane condition\] for $(*)$. We just verified conditions of Definition \[def:limit\]. By Lemma \[lem:complex is ane\], $K_1$ is an $ANE(n)$. We are done.
By Theorem \[thm:nregular pullback\] each $p_i$ is a weak $n$-homotopy equivalence and therefore all the short projections $\pi^k_i$ are weak $n$-homotopy equivalences. To finish the proof of (2), we must show that
- for fixed $i>0$ and $m<n$, the long projection $\pi_i:X\to K_i$ induces a monomorphism on the homotopy group of dimension $m$, regardless of the choice of base point; and
- for fixed $i>0$ and $m<n$, the long projection $\pi_i:X\to K_i$ induces an epimorphism on the homotopy group of dimension $m$, regardless of the choice of base point.
We prove (mono). Fix $i > 0$ and $m < n$. Let $\varphi \colon S^m \to X$. Assume that $\pi_i \circ \varphi$ is null-homotopic in $K_i$. Let $\varPhi \colon B^{m+1} \to K_i$ denote the null-homotopy ($B^{m+1}$ denotes the $m+1$-dimensional unit ball in $\mathbb{R}^{m+1}$). We have the following commutative diagram: $$\xymatrix@M=8pt@C=35pt{
X \ar[r]^{\pi_i} & K_i \\
S^m \ar@{^{(}->}[r] \ar[u]_{\varphi} & B^{m+1} \ar[u]^{\varPhi} \ar@{.>}[ul]_{\tilde \varPhi} \\
}$$ The inverse limit $$X = \lim_{\longleftarrow} \left(K_i \xleftarrow{p_i} K_{i+1} \xleftarrow{p_{i+1}} \cdots \right)$$ along with the sequence of covers by barycentric stars of vertices (which we obtain by truncating (\*)) satisfies the conditions of Definition \[def:limit\]. Since $B^{m+1}$ is at most $n$-dimensional, by Lemma \[lem:limit lift\], there exists a lift $\tilde \varPhi$ such that the diagram is commutative. This lift is a null-homotopy of $\varphi$ in $X$, hence $\pi_i$ induces a monomorphism on the homotopy group of dimension $m$. We will show (epi) at the end of the proof.
Next we prove (3). To begin, we will show that for each $i$, ${{\mathcal{B}_{i}}}$ is an $AE(n)$-cover. Let $\mathcal{A}$ be a collection of elements of ${{\mathcal{B}_{i}}}$ such that the intersection $\bigcap \mathcal{A}$ is non-empty. By the definition of ${{\mathcal{B}_{i}}}$, we have $\mathcal{A} = \{ \pi_i^{-1}(\operatorname{bst}_{K_i} v) \}_{v \in V}$ for some set of vertices $V$ of $K_i$. Let $A = \bigcap_{v \in V} p_i^{-1}(\operatorname{bst}_{K_i} v)$. Since $p_i$ is quasi-simplicial, $A$ is a subcomplex of $K_{i+1}$. By Theorem \[thm:nregular pullback\], $A$ has vanishing homotopy groups in dimensions less than $n$. Let $$\tag{**}
X' = \lim_{\longleftarrow} \left( K'_{i+1} \xleftarrow{p'_{i+1}} K'_{i+2} \xleftarrow{p'_{i+2}} \cdots\right),$$ where $K'_{i+1} =\ A$ and for each $j \geq i+1$, $K'_{j+1} = p^{-1}_j(K'_j)$ and $p'_j = p_j \mid_{K'_{j+1}}$, be the restriction of (\*) to $A$. By Lemma \[lem:restriction\] it satisfies assumptions (I) and (II) and is homeomorphic to $\bigcap \mathcal{A}$. Hence from what we have already proven, $X' = \bigcap \mathcal{A}$ is an $ANE(n)$ and the long projection $\pi_{i+1} \colon \bigcap \mathcal{A} \to A$ induces monomorphisms on homotopy groups of dimensions less than $n$. Since $A$ has vanishing homotopy groups in these dimensions, so does $\bigcap \mathcal{A}$. By Theorem \[thm:dugundji\], $\bigcap \mathcal{A}$ is an $AE(n)$ hence ${{\mathcal{B}_{i}}}$ is an $AE(n)$-cover.
It follows from Lemma \[lem:union is ae\] that ${{\mathcal{O}_{i}}}$ is an $AE(n)$-cover, as open stars are (infinite) unions of iterated barycentric stars.
Finally, we prove (epi). Fix $i > 0$ and $m < n$. Let $\varphi \colon S^m \to K_i$. Let $\mathcal{O}$ be a cover of $K_i$ by open stars of vertices. We have just shown that $\pi_i^{-1}(\mathcal{O})$ is an $AE(n)$-cover. Hence by Lemma \[lem:single lift\], there exists a map $\psi \colon S^m \to X$ such that $\pi_i \circ \psi$ and $\varphi$ are $\mathcal{O}$-close. By Lemma \[lem:uclose uhomotopic\], these maps are homotopic. This shows that $\pi_i$ induces epimorphisms on homotopy groups of dimensions less than $n$. Hence $\pi_i$ is a weak $n$-homotopy equivalence.
Let $$X = \lim_{\longleftarrow} \left(K_1 \xleftarrow{p_1} K_2 \xleftarrow{p_2} \cdots\right).$$ Assume that for each $i$, (I) $K_i$ is a locally finite-dimensional polyhedron; (II) $p_i$ is a quasi-simplicial map that is surjective and $n$-regular. Let $Y$ be a metric space of dimension less than $n$ and let $f, g \colon Y \to X$. Then if $f$ is ${{\mathcal{O}_{i}}}$-close to $g$, then $f$ is ${{\mathcal{O}_{i}}}$-homotopic to $g$.
By Theorem \[thm:ane limit of polyhedra\](3), ${{\mathcal{O}_{i}}}$ is an $AE(n)$-cover of $X$. By Lemma \[lem:uclose uhomotopic\], $f$ and $g$ are ${{\mathcal{O}_{i}}}$-homotopic.
[^1]: This research was supported by the NCN (Narodowe Centrum Nauki) grant no. 2011/01/D/ST1/04144.
|
---
abstract: 'We analyze backward step control globalization for finding zeros of Gâteaux-differentiable functions that map from a Banach space to a Hilbert space. The results include global convergence to a distinctive solution characterized by propagating the initial guess by a generalized Newton flow with guaranteed bounds on the discrete nonlinear residual norm decrease and an (also numerically) easily controllable asymptotic linear residual convergence rate. The convergence theory can be exploited to construct efficient numerical methods, which we demonstrate for the case of a Krylov–Newton method and an approximation-by-discretization multilevel framework. Both approaches optimize the asymptotic linear residual convergence rate, either over the Krylov subspace or through adaptive discretization, which in turn yields practical and efficient stopping criteria and refinement strategies that balance the nonlinear residuals with the relative residuals of the linear systems. We apply these methods to the class of nonlinear elliptic boundary value problems and present numerical results for the Carrier equation and the minimum surface equation.'
author:
- Andreas Potschka
title: '[Backward step control for Hilbert space problems]{}'
---
Introduction {#sec:introduction}
============
Let $U$ be a Banach space with norm ${\left\lVert.\right\rVert}_U$ and $V$ be a Hilbert space (we discuss generalizations to Banach spaces in section \[sec:Banach\]) with inner product ${\left(.,.\right)}_V$ and norm $\smash[t]{{\left\lVertv\right\rVert}_V = \smash[b]{\sqrt{{\left(v,v\right)}_V}}}$ for $v \in V$. For some open subset $D \subseteq U$, let $F: D \to V$ be continuously Gâteaux differentiable [@Hamilton1982 3.1.1. Def.] with derivative $F': D
\times U \to V$, i.e., $$F'(u, \delta u) = \lim_{h \to 0} \frac{1}{h} \left( F(u+h \delta u) - F(u)
\right)$$ exists for all $u \in D$, $\delta u \in U$ and $F'$ is continuous as a function from the product space $D \times U$ to $V$ (which is a weaker requirement than continuity of $F'$ considered as a map from $D$ to $\mathcal{L}(U, V)$, the Banach space of all bounded linear operators from $U$ to $V$). Note that continuous Gâteaux differentiability implies linearity of $F'$ in the second argument [@Hamilton1982 3.2.5 Thm.]. In the following, we write $A(u) \delta u$ shorthand for $A(u, \delta u)$ whenever an operator $A$ is linear in the second argument. We consider the problem of finding an unknown $u \in D$ such that $$\label{eqn:FOfxIsZero}
F(u) = 0_V$$ with a Newton-type iteration: Given $u_0 \in D$, find a suitable approximation $M: D \times V \to U$ (linear in the second argument) of the inverse of $F'(u)$ and a step size sequence $(t_k)_{k \in \mathbb{N}}$ satisfying $t_k \in [0, 1]$ such that the iteration $$\label{eqn:Newton}
u_{k+1} = u_k + t_k \delta u_k \quad
\text{with } \delta u_k = -M(u_k) F(u_k)$$ converges to a solution $u^\ast \in D$ of . The first and main part of this article is devoted to finding a suitable step size sequence $(t_k)_{k \in \mathbb{N}}$ in section \[sec:convergence\]. Out of the many ways to construct $M(u)$, we elaborate on two choices in section \[sec:designM\]. For convenience, we define the negative generalized Newton flow $f: D \to U$ as $$f(u) = M(u)F(u).$$ As in [@Potschka2016], our convergence analysis will be based on generalized Newton paths $u^k: [0, \infty) \to U$, which are defined as the solutions of the initial value problems $$\label{eqn:genDavidenko}
\frac{{{\mathrm d}}u^k}{{{\mathrm d}}t}(t) = -f(u^k(t)) \quad
\text{for } t \in [0, \infty) \quad
\text{with } u^k(0) = u_k.$$ We shall prove existence and uniqueness of solutions to in our setting in Theorem \[thm:arclength\]. Note that is an explicit Euler discretization of with step sizes $(t_k)_{k \in \mathbb{N}}$. If $M$ is chosen as the inverse of $F'$ then is a damped Newton method and is the Davidenko differential equation [@Davidenko1953]. The equivalence of the Newton method with explicit Euler on the Davidenko differential equation has been exploited by various authors (see, e.g., [@Deuflhard1974; @Brezinski1975; @Ascher1987; @Bock2000b; @Deuflhard2006]).
In the theory and the numerics below, the operator $M$ does not appear explicitly anymore and only the function $f$ will be required, which implicitly defines $M(u)$ in the direction $F(u)$. As it turns out, all other directions of $M(u)$ are not important. This observation leads to crucial improvements in the assumptions stated in [@Potschka2016].
The convergence theory below lends itself immediately to the construction of numerical algorithms for the approximate solution of via . In particular, it can be used to construct $M(u)$ from $F'(u)$ by finite-dimensional approximation, which can then be exploited to construct adaptive discretization schemes that optimize the contraction rate of the algorithm in $V$. In the case of Finite Element analysis, our approach delivers a multi-level Newton adaptive mesh refinement algorithm [@Hohmann1994] that can be used straight-forward as another tool complementing refinement strategies based on a posteriori error estimation (see, e.g., [@Graetsch2005; @Ainsworth2000; @Becker2001]).
Historically, Newton-type methods come in two flavors: Either, we do not solve the linearized systems with operator $F'(u_k)$ and right-hand side $-F(u_k)$ exactly, which is usually known under the name *inexact Newton* method [@Dembo1982; @Ypma1984]. Alternatively, we apply an approximation $M(u)$ of the inverse of $F'(u)$ directly, an approach which is sometimes called *approximate Newton* method and which is the classical form of Quasi-Newton methods [@Dennis1977]. As pointed out by Bank and Rose [@Bank1981], the two flavors are in fact different formulations of the same class of methods. As mentioned above, our analysis is based on the approximate Newton formulation in order to define the generalized Newton path , but the control of the linearized residual in the sense of inexact Newton methods emerges as the $\kappa$-condition A\[ass:kappa\], which is identical to a choice of $\eta_k = \kappa < 1$ for the residual forcing sequence $(\eta_k)$ first proposed in [@Dembo1982]. In turn, the classical analysis of the local rates of convergence in [@Dembo1982] carries over to our setting if $\kappa$ is allowed to be reduced from iteration to iteration in the sense of a forcing sequence once we are close to a solution. This shall, however, not be the focus of this paper, where we focus on the preasymptotic global convergence and are content with locally linear convergence rates.
Similar to [@Deuflhard1991], we base our globalization approach on a continuous curve, but we substitute here the exact Newton path (obtained with the choice $M(u) = \smash[t]{F'(u)}^{-1}$ from ) by a generalized Newton path, which is allowed to have $M(u) \neq
\smash[t]{F'(u)}^{-1}.$ Following ideas of [@Hohmann1994], we can use the (contravariant) $\kappa$-condition A\[ass:kappa\] to design a multilevel Newton-type method for the construction of $M$ based on adaptive discretization. This leads to adaptive discretizations solely based on balancing the discretization residual with the nonlinear residual of the Newton-type method. We remark that this convenient black-box approach might be inferior to more involved schemes that balance the discretization errors and nonlinear errors (instead of residuals) [@Rannacher2013] or exploit underlying (energy) minimization properties [@Deuflhard1998] for particular problem classes. Nonetheless, our method is the first for which convergence to the closest solution in the sense of the generalized Newton flow can be proven.
#### Contributions
In this article, we extend the convergence analysis of backward step control for from the finite-dimensional to the Hilbert space setting. We provide reasonable assumptions and convergence results for with backward step control. The main result is convergence to a distinct solution characterized by the propagation of the initial guess by the generalized Newton flow provided that no singularity of the problem interferes. In addition, we prove an a priori bound on the nonlinear reduction of the residual norm. The convergence theory can be exploited to construct efficient numerical algorithms, which we discuss for the case of a Krylov–Newton method and a Finite Element approximation. Both are based on the optimization of the residual contraction constant, which yields in the latter case an efficient adaptive mesh refinement strategy. The results are demonstrated for the numerical solution of the Carrier equation and the minimal surface equation.
#### Overview
In section \[sec:convergence\], we discuss the general assumptions, derive and motivate the method of backward step control, provide reasons why implicit and higher-order time stepping methods for are not advisable, provide step size bounds, and establish the notion of generalized Newton paths, which are the central ingredient for the analysis of local and global convergence of backward step control. We then discuss extensions to Banach spaces and present an algorithmic realization with a minimal working Matlab example code for the solution of $\arctan(u) = 0$. In section \[sec:designM\], we exploit the convergence analysis to construct two Newton-type methods, a Krylov–Newton method and a method based on approximation-by-discretization. We apply these methods in section \[sec:ellipticPDE\] to the class of nonlinear elliptic boundary value problems and provide numerical results for the Carrier equation and the minimum surface equation.
#### Notation
We denote the open ball of radius $r>0$ around $u\in U$ by $B(u,r)$ and the Laplace operator by ${\Delta}= \nabla \cdot \nabla$. As usual, we write $C^0$ for the space of continuous functions, $H^1_0(\Omega)$ for the Sobolev space of square integrable functions on a bounded domain $\Omega \subset \mathbb{R}^n$ that vanish at the boundary and admit square integrable derivatives, and $H^{-1}(\Omega)$ for its dual space. The Euler number is denoted by $e =
\sum_{k=0}^{\infty} \frac{1}{k!}$.
Convergence analysis {#sec:convergence}
====================
The overall structure of the backward step control convergence analysis in Hilbert spaces is similar to the finite-dimensional case [@Potschka2016]. The intricate interplay of the changes in the details, however, advises us to present the convergence analysis in a self-contained fashion.
Discussion of assumptions
-------------------------
We start with the following definitions.
The *level function* $T: D \to \mathbb{R}$ is $T(u) = \frac{1}{2} {\left\lVertF(u)\right\rVert}_V^2.$
The *level set of $u \in D$* is $\smash[t]{\widetilde{\mathcal{T}}}(u) = \left\{ \bar{u} \in D \mid
T(\bar{u}) \le T(u) \right\}.$
The *path connected level set of $u \in D$* is $$\mathcal{T}(u) = \left\{ \bar{u} \in \widetilde{\mathcal{T}}(u)
\mid \exists
c \in C^0\left([0, 1], \widetilde{\mathcal{T}}(u)\right)
\text{with } c(0) = u, c(1) = \bar{u} \right\}.$$
For $r \in (1, \infty)$ the *set of $r$-regular points* is $$\mathcal{R}_r = \left\{ u \in D \mid r^{-1} {\left\lVertF(u)\right\rVert}_V <
{\left\lVertf(u)\right\rVert}_U < r {\left\lVertF(u)\right\rVert}_V \right\}.$$
The *set of $\infty$-regular points* is $\mathcal{R}_\infty = \bigcup_{r \in (1, \infty)} \mathcal{R}_r$.
We remark that if $u \in D \setminus \mathcal{R}_\infty$, which means that $u
\not\in \mathcal{R}_r$ for all $r
\in (1, \infty)$, then $M(u)$ is either not bounded or does not admit a bounded inverse [@Yosida1978 §I.6, Cor. 2, 3]. The contrary is, however, not true: $M(u)$ may be unbounded or not admit a bounded inverse although $u \in
\mathcal{R}_r$ for some $r \in (1, \infty)$, because in the definition of $\mathcal{R}_r$ only the action of $M(u)$ in direction $F(u)$ is of interest.
We require the following assumptions to hold true:
\[ass:validIni\] There exists an $r \in (1, \infty)$ such that $u_0 \in \mathcal{R}_r$, and ${\left\lVertF(u_0)\right\rVert}_V > 0$.
\[ass:kappa\] There exists a $\kappa < 1$ such that for all $u \in \mathcal{R}_r \cap
\mathcal{T}(u_0)$ $${\left\lVertF(u) - F'(u) f(u)\right\rVert}_V \le \kappa {\left\lVertF(u)\right\rVert}_V.$$
\[ass:omega\] There exists an $\omega < \infty$ such that for all $u \in \mathcal{T}(u_0), t
\in [0, 1]$ $${\left\lVert\left[ F'(u) - F'(u-tf(u)) \right] f(u)\right\rVert}_V \le \omega t {\left\lVertf(u)\right\rVert}_U
{\left\lVertF(u)\right\rVert}_V.$$
\[ass:fLipschitz\] There exists an $L < \infty$ such that for all $u, \bar{u} \in
\mathcal{T}(u_0)$ $${\left\lVertf(u) - f(\bar{u})\right\rVert}_U \le L {\left\lVertu - \bar{u}\right\rVert}_U.$$
\[ass:gamma\] For all $\eta > 0$ there exist constants $\gamma$, $t_\gamma > 0$ such that for all $t \in [0, t_\gamma]$, $u \in \mathcal{R}_r \cap \mathcal{T}(u_0)$ with ${\left\lVertf(u)\right\rVert}_U > \eta$ $${\left\lVertf(u-tf(u)) - f(u)\right\rVert}_U \ge \gamma t.$$
The main difference in the assumptions here compared to the finite-dimensional setting in [@Potschka2016] is the weakening of A\[ass:kappa\] and A\[ass:omega\] from a formulation with matrices to a formulation which requires the properties to hold only in the direction of the residual $F(u)$. Thus, all requirements can be postulated without using norms for operators that map between $U$ and $V$. Apart from the avoidance of operator norms, we had to replace all arguments based on compactness of bounded sets by other means for the proofs in the Hilbert space case. The discussion of the assumptions in [@Potschka2016] still applies to a large extent here: We require in A\[ass:validIni\] that $u_0$ is an $r$-regular point but not a solution. The central $\kappa$-condition A\[ass:kappa\] is a contravariant version of Bock’s covariant $\kappa$-condition [@Bock1987] in the sense that it quantifies on the one hand the deviation of the inexact increment $\delta u = -M(u) F(u)$ from the Newton increment $\delta
u^\mathrm{Newton} = -F(u)^{-1} F(u)$ in the $V$-norm $${\left\lVertF'(u) \left[ \delta u^\mathrm{Newton} - \delta u \right]\right\rVert}_V
= {\left\lVertF(u) - F'(u)f(u)\right\rVert}_V
\le \kappa {\left\lVertF(u)\right\rVert}_V$$ in comparison to Bock’s covariant $\kappa^\mathrm{cov} < 1$ in the $U$-norm (see also [@Potschka2013 section 5.2]) $$\label{eqn:kappa_Bock}
{\left\lVertM(u - f(u)) \left[ F(u) - F'(u)f(u) \right]\right\rVert}_U
\le \kappa^\mathrm{cov} {\left\lVertf(u)\right\rVert}_U.$$ On the other hand, $\kappa$ in A\[ass:kappa\] characterizes the asymptotic Q-linear convergence rate of the residual norms ${\left\lVertF(u_k)\right\rVert}_V$, whereas $\kappa^\mathrm{cov}$ in characterizes the asymptotic R-linear convergence rate of the error ${\left\lVertu_k - u_\ast\right\rVert}_U$ if $u_\ast =
\lim_{k \to \infty} u_k$ (for a discussion of different affine invariances see [@Deuflhard2006]). The $\omega$-condition A\[ass:omega\] measures a combination of the nonlinearity and the well-posedness of because if $F'$ is Lipschitz continuous with Lipschitz constant $L'$, then we obtain $${\left\lVert\left[ F'(u) - F'(u - tf(u)) \right] f(u)\right\rVert}_V \le L' t {\left\lVertf(u)\right\rVert}_U^2$$ and boundedness of $M(u)$ in direction $F(u)$ with constant $C$ implies A\[ass:omega\] with $\omega = CL'$. The Lipschitz condition A\[ass:fLipschitz\] is classical. The nonstandard assumption A\[ass:gamma\] follows, for instance, if $f$ is bi-Lipschitz with constant $\ell$ $${\left\lVertf(u - tf(u)) - f(u)\right\rVert}_U \ge \ell t {\left\lVertf(u)\right\rVert}_U$$ with $\gamma = \eta \ell$ and $t_\gamma$ arbitrary.
Backward step control
---------------------
Newton-type methods are explicit Euler discretizations with step sizes $t_k$ of the generalized Newton flow . Thus, the convergence of Newton-type methods is strongly connected to the stability problem of the explicit Euler method. Implict Euler, in contrast, has ideal stability properties: It is an L-stable method (see, e.g., [@Hairer1996]). Hence, in order to determine $t_k$, we consider the backward iterate $$\bar{u}_k(t_k) := u_{k+1} + t_k f(u_{k+1}) = u_k + t_k g(u_k, t_k) \quad
\text{with } g(u, t) := f(u - tf(u)) - f(u).$$ The point $\bar{u}_k(t_k)$ is the starting point of a (stable) implicit Euler step for that arrives exactly at $u_{k+1}$, the result of a possibly unstable explicit Euler step starting from $u_k$. The idea of backward step control is based on a backward error argument: If a small perturbation of the starting point $u_k$ can be found from which a stable implicit Euler step arrives exactly at $u_{k+1}$, we can accept the step size. We thus require that the distance between $u_k$ and $\bar{u}_k(t_k)$ is bounded by some fixed constant $H>0$ through the choice $$\label{eqn:BSC} \tag{BSC}
t_k = \min \mathcal{B}_H(u_k) \quad \text{where }
\mathcal{B}_H(u) = \left\{ t \in [0, 1] \mid H = t {\left\lVertg(u,t)\right\rVert}_U
\right\} \cup \{1\},$$ which implies ${\left\lVert\bar{u}_k(t_k) - u_k\right\rVert}_U \le H$ (with equality for $t_k <
1$) by continuity of $g$.
Implicit and higher-order time stepping methods
-----------------------------------------------
The question whether explicit Euler is really the best method to solve arises naturally. We can answer this question affirmatively for two reasons: First, all implicit methods have the drawback that an approximated inverse of an operator involving derivatives of the approximated inverse $M(u)$ would be required, e.g., in the case of the implicit Euler method $$\begin{aligned}
0 &= u_{k+1} + t_k f(u_{k+1}) - u_k,
\intertext{with a local Newton corrector}
u_{k+1}^{i+1} &= u_{k+1}^i - \left[ {{\mathrm I}}_U + t_k f'(u_{k+1}^i) \right]^{-1}
\left( u_{k+1}^i + t_k f(u_{k+1}^i) - u_k \right),\end{aligned}$$ which is not readily available and would require higher regularity of $M$ than guaranteed by the assumptions above. Second, higher order methods would destroy the well-known locally quadratic convergence of the Newton method, where $M(u) = (F'(u))^{-1}$. This can be seen from the homotopy formulation $$\label{eqn:uHomotopy}
F(u(t)) - e^{-t} F(u_0) = 0,$$ which we can differentiate with respect to $t$ to arrive exactly at provided that $F'(u(t))$ stays invertible. Thus, the second order truncation error of explicit Euler is required to obtain locally quadratic convergence, because higher consistency orders would result in the locally linear convergence dictated by .
We have not explored multi-step methods of order one further. The possible outcome of this line of research is unfortunately unclear at present and exceeds the scope of this paper.
Step size bounds
----------------
\[lem:localFullSteps\] If A\[ass:validIni\] and A\[ass:fLipschitz\] hold, then delivers full steps $t_k = 1$ in the vicinity of a solution $u^\ast \in
\mathcal{R}_r \cap \mathcal{T}(u_0)$.
Let $u_k \in B(u^\ast, L^{-2} H)$. Hence, it holds for all $t \in [0, 1]$ that $$t {\left\lVertg(u_k, t)\right\rVert}_U
\overset{\text{A\ref{ass:fLipschitz}}}{\le} L t^2 {\left\lVertf(u_k)\right\rVert}_U
= L t^2 {\left\lVertf(u_k) - f(u^\ast)\right\rVert}_U
\overset{\text{A\ref{ass:fLipschitz}}}{\le} L^2 t^2 {\left\lVertu_k - u^\ast\right\rVert}_U < H.$$ Thus, $\mathcal{B}_H(u_k) = \{1\}$ and $t_k = 1$ by virtue of .
\[lem:lowerStepsizeBound\] If A\[ass:validIni\] and A\[ass:fLipschitz\] hold, then generates for all $u_k \in \mathcal{R}_r \cap \mathcal{T}(u_0)$ step sizes that are either $t_k = 1$ or have the lower bounds $$t_k \ge \frac{\sqrt{H}}{\sqrt{L {\left\lVertf(u_k)\right\rVert}_U}}
> \frac{\sqrt{H}}{\sqrt{rL{\left\lVertF(u_k)\right\rVert}_V}}
\ge \frac{\sqrt{H}}{\sqrt{rL{\left\lVertF(u_0)\right\rVert}_V}}.$$
*Proof * If $t_k < 1$, then $$t_k^2 \overset{\eqref{eqn:BSC}}{=} \frac{t_k H}{{\left\lVertg(u_k, t_k)\right\rVert}_U}
\overset{\text{A\ref{ass:fLipschitz}}}{\ge}
\frac{H}{L {\left\lVertf(u_k)\right\rVert}_U}
> \frac{H}{rL{\left\lVertF(u_k)\right\rVert}_V}
\ge \frac{H}{rL{\left\lVertF(u_0)\right\rVert}_V}. \qedhere$$
\[lem:upperStepsizeBound\] Let A\[ass:validIni\] and A\[ass:gamma\] hold and let $\bar{t}
\in (0, 1)$ and $\eta > 0$. Then there exists an ${\overline{H}} > 0$ such that for all $H \in (0, {\overline{H}}]$ and $u \in \mathcal{R}_r \cap
\mathcal{T}(u_0)$ with ${\left\lVertf(u)\right\rVert}_U \ge \eta$ it holds that $\min
\mathcal{B}_H(u) \le \bar{t}$.
*Proof by contradiction * We assume to the contrary that for all ${\overline{H}} > 0$ there exists an $H
\in (0, {\overline{H}}]$ and a $u \in \mathcal{R}_r \cap \mathcal{T}(u_0)$ satisfying ${\left\lVertf(u)\right\rVert}_U \ge \eta$ and $\min \mathcal{B}_H(u) > \bar{t}$. Then, A\[ass:gamma\] guarantees the existence of $\gamma$, $t_\gamma > 0$ such that for $t := \min \{ t_\gamma, \bar{t} \} < \min \mathcal{B}_H(u)$ we obtain from that $${\overline{H}} \ge H \ge t {\left\lVertg(u, t)\right\rVert}_U \ge \gamma t^2 > 0.$$ Because $\eta$ and thus $\gamma$ and $t$ are independent of ${\overline{H}}$, we obtain a contradiction for ${\overline{H}} \to 0$.
\[lem:dampedStepBound\] Let A\[ass:validIni\], A\[ass:fLipschitz\], and A\[ass:gamma\] hold and let $\eta > 0$. Then there exists an ${\overline{H}} > 0$ such that for all $H
\in (0, {\overline{H}}]$ and $u \in \mathcal{R}_r \cap \mathcal{T}(u_0)$ it holds that $${\left\lVertf(u)\right\rVert}_U \min \mathcal{B}_H(u) \le \eta.$$
We choose $\bar{t} \in (0, 1)$ sufficiently small so that it satisfies $r
{\left\lVertF(u_0)\right\rVert}_V \bar{t} \le \eta.$ Then, Lemma \[lem:upperStepsizeBound\] yields the existence of an ${\overline{H}} > 0$ such that for all $H \in (0, {\overline{H}}]$ and all $u \in
\mathcal{R}_r \cap \mathcal{T}(u_0)$ with ${\left\lVertf(u)\right\rVert}_U \ge \eta$ it holds that $\min \mathcal{B}_H(u) \le \bar{t}$. Hence, $${\left\lVertf(u)\right\rVert}_U \min \mathcal{B}_H(u)
\overset{\text{A\ref{ass:validIni}}}{<}
r {\left\lVertF(u)\right\rVert}_V \bar{t} \le r {\left\lVertF(u_0)\right\rVert} \bar{t} \le \eta.$$ For the remaining $u \in \mathcal{R}_r \cap \mathcal{T}(u_0)$ the assertion holds by virtue of ${\left\lVertf(u)\right\rVert}_U < \eta$.
Finite arclength of generalized Newton paths
--------------------------------------------
In the next step, we study the generalized Newton paths given by .
\[lem:Fdescent\] If A\[ass:validIni\], A\[ass:kappa\], A\[ass:fLipschitz\] and $u_k \in
\mathcal{R}_r \cap \mathcal{T}(u_0)$ hold, then there exists $\bar{t} > 0$ such that has a unique local solution $u^k(t) \in
\mathcal{R}_r \cap \mathcal{T}(u_k)$ for $t \in [0, \bar{t})$ which satisfies $${\left\lVertF(u^k(t))\right\rVert}_V \le e^{-(1-\kappa)t} {\left\lVertF(u_k)\right\rVert}_V \quad
\text{for all } t \in [0, \bar{t}).$$
The Picard–Lindelöf theorem [@Amann1990 II.7, exercise 3] yields with A\[ass:fLipschitz\] the existence of a unique local solution $u^k(t)$ to in some neighborhood $(-\bar{t}, \bar{t})$ of $t =
0$. Without loss of generality, $\bar{t} > 0$ is small enough to ensure $u^k(t) \in \mathcal{R}_r$ for $t \in [0, \bar{t})$ because $\mathcal{R}_r$ is open. For ease of notation, we abbreviate $u^k(t)$ by $u$. The Cauchy–Schwarz inequality and A\[ass:kappa\] show that the level function is nonincreasing along this solution because $$\begin{aligned}
\frac{{{\mathrm d}}}{{{\mathrm d}}t} T(u) &= {\left(F(u), F'(u) \frac{{{\mathrm d}}u}{{{\mathrm d}}t}\right)}_V
= -{\left(F(u), F'(u) f(u)\right)}_V\\
&= -{\left\lVertF(u)\right\rVert}_V^2 + {\left(F(u), F(u) - F'(u) f(u)\right)}_V\\
&\le -{\left\lVertF(u)\right\rVert}_V^2 + \kappa {\left\lVertF(u)\right\rVert}_V^2
= -2(1-\kappa) T(u) \le 0.
\end{aligned}$$ Gronwall’s inequality (see, e.g., [@Amann1990]) yields $$T(u^k(t)) \le e^{-2 (1-\kappa) t} T(u_k)$$ and thus $u^k(t) \in \mathcal{T}(u_k)$ for $t \in [0, \bar{t})$. The assertion follows after multiplication by two and taking square roots.
We show in Theorem \[thm:arclength\] below that the quantities in the following definition are well-defined under suitable assumptions.
\[def:regularpart\] For $r \in (1,\infty)$, we define the *$r$-regular part $u_r^k$* of the generalized Newton path $u^k$ as the solution to the initial value problem $$\frac{{{\mathrm d}}u_r^k}{{{\mathrm d}}t}(t) =
\begin{cases}
-f(u_r^k(t)) & \text{for } u_r^k(t) \in \mathcal{R}_r,\\
0 & \text{otherwise},
\end{cases}
\quad \text{for } t \in [0, \infty),
\quad \text{with } u_r^k(0) = u_k.$$ We denote its limit by $u_k^\ast = \lim_{t \to \infty} u^k_r(t)$ and define $t_k^\ast = \inf \{t \in [0, \infty) \mid u_r^k(t) \not\in \mathcal{R}_r \}$ with the usual convention that $\inf \varnothing = \infty$.
\[thm:arclength\] Let A\[ass:validIni\], A\[ass:kappa\], and A\[ass:fLipschitz\] hold. If $u_k \in \mathcal{R}_r \cap \mathcal{T}(u_0)$, then the $r$-regular part of the generalized Newton path exists uniquely and has a finite arclength $\ell(u_r^k)$ satisfying $${\left\lVertu_k - u_k^\ast\right\rVert}_U \le
\ell(u_r^k) < \frac{r}{1 - \kappa} {\left\lVertF(u_k)\right\rVert}_V
< \frac{r^2}{1 - \kappa} {\left\lVertf(u_k)\right\rVert}_U.$$ If $u_k^\ast \in \mathcal{R}_r$, then $F(u_k^\ast) = 0$.
*Proof * The unique local solution of Lemma \[lem:Fdescent\] can be extended uniquely in $\mathcal{T}(u_k)$ by repeated application of the Picard–Lindelöf theorem either until $u^k(t) \not \in \mathcal{R}_r$ for some $t = t_k^\ast$ or to the whole interval $t \in [0, \infty)$. In the first case, the $r$-regular part is uniquely determined by $u_r^k(t) = u_r^k(t_k^\ast)$ for all $t \ge t_k^\ast$. We can now use the definition of $\mathcal{R}_r$ and Lemma \[lem:Fdescent\] in order to show $$\begin{aligned}
\ell(u_r^k) &= \int_{0}^{\infty} {\left\lVert\frac{{{\mathrm d}}u_r^k}{{{\mathrm d}}t}(t)\right\rVert}_U {{\mathrm d}}t
= \int_{0}^{t_k^\ast} {\left\lVertf(u^k(t))\right\rVert}_U {{\mathrm d}}t
< r \int_{0}^{t_k^\ast} {\left\lVertF(u^k(t))\right\rVert}_V {{\mathrm d}}t \\
&\le r \int_{0}^{\infty} e^{-(1-\kappa)t} {{\mathrm d}}t {\left\lVertF(u_k)\right\rVert}_V
= \frac{r}{1-\kappa} {\left\lVertF(u_k)\right\rVert}_V < \frac{r^2}{1-\kappa}
{\left\lVertf(u_k)\right\rVert}_U.
\end{aligned}$$ We obtain the lower arclength bound by noting that the shortest path between $u_k$ and $u_k^\ast$ has arclength ${\left\lVertu_k - u_k^\ast\right\rVert}_U$. If $u_k^\ast \in
\mathcal{R}_r$, then $t_k^\ast = \infty$ and Lemma \[lem:Fdescent\] reveals $${\left\lVertF(u_k^\ast)\right\rVert}_V \le \lim_{t \to \infty} e^{-(1-\kappa)t}{\left\lVertF(u_k)\right\rVert}_V =
0.
\qedhere$$
Local convergence
-----------------
We use the next lemma to prove discrete descent of the residual norm.
\[lem:kappaScaled\] A\[ass:kappa\] holds if and only if for all $u \in \mathcal{R}_r \cap
\mathcal{T}(u_0)$ and $t \in [0, 1]$ $${\left\lVertF(u) - t F'(u) f(u)\right\rVert}_V
\le \left[ 1 - (1-\kappa) t \right] {\left\lVertF(u)\right\rVert}_V.$$
*Proof * As in [@Potschka2016], the nontrivial direction of the proof follows from the convexity of the functional $\varphi(t) = {\left\lVertF(u) - t F'(u) f(u)\right\rVert}_V$ and A\[ass:kappa\] according to $$\varphi(t) \le (1-t)\varphi(0) + t \varphi(1)
\le \left[ (1-t) + \kappa t \right] {\left\lVertF(u)\right\rVert}_V.
\qedhere$$
\[lem:Fdecrease\] Let A\[ass:validIni\], A\[ass:kappa\] and A\[ass:omega\] hold. If $u_k
\in \mathcal{R}_r \cap \mathcal{T}(u_0)$, then $${\left\lVertF(u_{k+1})\right\rVert}_V \le \left[ 1 - (1-\kappa) t_k + \frac{\omega}{2}
{\left\lVertf(u_k)\right\rVert}_U t_k^2 \right] {\left\lVertF(u_k)\right\rVert}_V.$$ Furthermore, if there exists a $\theta < 1$ such that the step size sequence satisfies $$\omega t_k {\left\lVertf(u_k)\right\rVert}_U \le 2 \theta (1 - \kappa),$$ then $${\left\lVertF(u_{k+1})\right\rVert}_V \le \left[ 1 - (1-\theta) (1-\kappa) t_k \right]
{\left\lVertF(u_k)\right\rVert}_V.$$
Because $F'$ is continuously Gâteaux differentiable, we can apply [@Hamilton1982 3.2.2. Thm. and 2.1.4. Thm.] to show that for $u \in D$ and $\delta u \in U$ $$F(u + t_k \delta u) - F(u) =
\int_{0}^{1} F'(u + \tau t_k \delta u) t_k \delta u {{\mathrm d}}\tau
\overset{\tau = \frac{t}{t_k}}{=} \int_{0}^{t_k} F'(u + t \delta u) \delta
u {{\mathrm d}}t.$$ Using Lemma \[lem:kappaScaled\] and A\[ass:omega\] we obtain the first assertion from $$\begin{aligned}
& \quad\, {\left\lVertF(u_{k+1})\right\rVert}
= {\left\lVertF(u_k) - \int_{0}^{t_k} F'(u_k - \tau f(u_k)) f(u_k) {{\mathrm d}}\tau\right\rVert}_V\\
&= {\left\lVertF(u_k) - t_k F'(u_k) f(u_k) + \int_{0}^{t_k} \left[ F'(u_k) -
F'(u_k - \tau f(u_k)) \right] f(u_k) {{\mathrm d}}\tau\right\rVert}_V\\
&\le {\left\lVertF(u_k) - t_k F'(u_k) f(u_k)\right\rVert}_V
+ \int_{0}^{t_k} {\left\lVert\left[ F'(u_k) - F'(u_k - \tau f(u_k)) \right]
f(u_k)\right\rVert}_V {{\mathrm d}}\tau\\
&\le \left[ 1 - (1 - \kappa) t_k + \frac{\omega}{2} {\left\lVertf(u_k)\right\rVert}_U t_k^2
\right] {\left\lVertF(u_k)\right\rVert}_V.
\end{aligned}$$ The second assertion follows immediately.
We can now state a local convergence theorem.
\[thm:localConvergence\] Let A\[ass:validIni\], A\[ass:kappa\], and A\[ass:omega\] hold. If there exists a $\bar{t} \in (0,1)$ such that $t_k \ge \bar{t}$ for all $k \in \mathbb{N}$ and if there exists a $\theta \in (0, 1)$ such that for some $k \in
\mathbb{N}$ the iterate $u_k$ satisfies $$\mathcal{T}(u_k) \subseteq \mathcal{R}_r
\quad \text{and} \quad
\omega r {\left\lVertF(u_k)\right\rVert}_V \le 2 \theta (1 - \kappa),$$ then $(u_k)_{k \in \mathbb{N}}$ converges to some point $u^\ast \in
\mathcal{T}(u_k)$ with $F(u^\ast) = 0$.
Because $t_k \in [0, 1]$, we have $$\omega t_k {\left\lVertf(u_k)\right\rVert}_U \le \omega {\left\lVertf(u_k)\right\rVert}_U
\overset{\text{A\ref{ass:validIni}}}{\le} \omega r {\left\lVertF(u_k)\right\rVert}_V
\le 2 \theta (1 - \kappa).$$ Hence, repeated application of Lemma \[lem:Fdecrease\] yields for all $j \in
\mathbb{N}$ $$\label{eqn:geomFConv}
{\left\lVertF(u_{k+j})\right\rVert}_V
\le q^j {\left\lVertF(u_k)\right\rVert}_V \quad \text{with }
q := 1 - (1-\theta)(1-\kappa) \bar{t}.$$ Because $q < 1$, ${\left\lVertF(u_k)\right\rVert}_V$ converges geometrically. In addition, we obtain that $(u_k)_{k \in
\mathbb{N}}$ is a Cauchy sequence by virtue of $$\begin{aligned}
{\left\lVertu_k - u_{k+j}\right\rVert}_U &\le \sum_{i=0}^{j-1} {\left\lVertu_{k+i} - u_{k+i-1}\right\rVert}_U
= \sum_{i=0}^{j-1} t_{k+i} {\left\lVertf(u_{k+i})\right\rVert}_U\\
&\le r \sum_{i=0}^{j-1} {\left\lVertF(u_{k+i})\right\rVert}_V
\le r {\left\lVertF(u_{k})\right\rVert}_V \sum_{i=0}^{\infty} q^i
= \frac{r}{1-q} {\left\lVertF(u_k)\right\rVert} \overset{k \to \infty}{\longrightarrow} 0.
\end{aligned}$$ Thus, $(u_{k})_{k \in \mathbb{N}}$ converges to some $u^\ast \in \mathcal{T}(u_k)
\subseteq \mathcal{R}_r$ and implies $F(u^\ast) = 0$.
For the rate of convergence, we obtain the following result:
\[lem:linearconvergence\] Under the assumptions of Theorem \[thm:localConvergence\], ${\left\lVertF(u_k)\right\rVert}_V$ converges linearly with asymptotic linear convergence rate $\kappa < 1$.
Because $u_k \in \mathcal{R}_r$, it follows that ${\left\lVertf(u_k)\right\rVert}_U \le r
{\left\lVertF(u_k)\right\rVert}_V \to 0$. Hence, there is a sequence $(\theta_K)_{K \in
\mathbb{N}}$ with $\theta_K \in [0, 1)$ and $\theta_K \to 0$ such that $$\omega t_k {\left\lVertf(u_k)\right\rVert}_U \le 2 \theta_K (1 - \kappa) \quad \text{for all } k
\ge K.$$ Lemma \[lem:localFullSteps\] and repeated application of Lemma \[lem:Fdecrease\] then deliver $${\left\lVertF(u_{K+1})\right\rVert}_V \le \left[ 1 - (1-\theta_K)(1-\kappa) \right]
{\left\lVertF(u_K)\right\rVert}_V,$$ where $1 - (1-\theta_K)(1-\kappa) \to \kappa$ as $K \to \infty$.
In order to obtain methods with guaranteed superlinear or quadratic local convergence, it is necessary to appropriately drive $\kappa$ to zero as $F(u_k)
\to 0$ as exhaustively described by the means of forcing sequences $\eta_k =
\kappa$ in [@Dembo1982].
Global convergence
------------------
The following lemmas are required for the main theorem, which assures convergence of the iterates to $u_0^\ast$. As a prerequisite, we prove that every neighborhood of an isolated zero $u^\ast$ of $F$ contains a path connected level set that contains a neighborhood of $u^\ast$.
\[lem:boundedLevelSets\] Let A\[ass:validIni\], A\[ass:kappa\], and A\[ass:fLipschitz\] hold. If there exist $\varepsilon > 0$ and $u^\ast \in D$ such that $u^\ast$ is the only zero of $F$ on $B(u^\ast, \varepsilon) \subseteq \mathcal{R}_r \cap
\mathcal{T}(u_0),$ then there exists an $\tilde{\varepsilon} > 0$ with $$\bigcup_{u \in B(u^\ast, \tilde{\varepsilon})} \mathcal{T}(u)
\subseteq B(u^\ast, \varepsilon).$$
*Proof by contradiction * We assume to the contrary that there exists a sequence $(u_n)_{n \in
\mathbb{N}}$ with ${\left\lVertu_n - u^*\right\rVert}_U < \frac{\varepsilon}{2n}$ and $\mathcal{T}(u_n) \not\subseteq B(u^*, \varepsilon)$. Hence, there exists a sequence $(\tilde{v}_n)_{n \in \mathbb{N}}$ with $\tilde{v}_n \in \mathcal{T}(u_n)$ and ${\left\lVert\tilde{v}_n - u^\ast\right\rVert}_U \ge
\varepsilon$. Because $\mathcal{T}(u_n)$ is path connected, there exist continuous functions $c_n: [0, 1] \to \mathcal{T}(u_n)$ with $c_n(0) = u_n$ and $c_n(1) = \tilde{v}_n$. Because ${\left\lVertc_n(0) - u^*\right\rVert}_U <
\frac{\varepsilon}{2}$ and ${\left\lVertc_n(1) - u^*\right\rVert}_U \ge \varepsilon$, the intermediate value theorem yields the existence of $v_n = c_n(\tau_n) \in
\mathcal{T}(u_n)$ for some $\tau_n \in [0, 1]$ satisfying $$\label{eqn:annulus}
{\left\lVertv_n - u^*\right\rVert}_U = \frac{\varepsilon}{2}.$$ By Theorem \[thm:arclength\], we obtain for the distance to the limit $v_n^\ast$ of the $r$-regular part of the generalized Newton path emanating from $v_n$ that $$\begin{aligned}
{\left\lVertv_n - v_n^\ast\right\rVert}_U
&\le \ell(v_r^n) < \frac{r}{1-\kappa} {\left\lVertF(v_n)\right\rVert}_V
\le \frac{r}{1-\kappa} {\left\lVertF(u_n)\right\rVert}_V
< \frac{r^2}{1-\kappa} {\left\lVertf(u_n)\right\rVert}_U\\
&= \frac{r^2}{1-\kappa} {\left\lVertf(u_n) - f(u^*)\right\rVert}_U
\le \frac{r^2L}{1-\kappa} {\left\lVertu_n - u^*\right\rVert}_U <
\frac{r^2L\varepsilon}{2(1-\kappa)n} \to 0,
\end{aligned}$$ which implies for some sufficiently large $n$ that $v^*_n \in B(u^*,
\varepsilon) \subseteq \mathcal{R}_r$ and thus $F(v^*_n) = 0$. By we get $v_n^* \neq u^*$ in contradiction to the uniqueness of $u^*$. We also need a bound on the deviation of two neighboring generalized Newton paths emanating from $u_k$ and $u_{k+1}$.
\[lem:oneStepError\] Let A\[ass:fLipschitz\] hold. If $u^k(t_k + \tau)$, $u^{k+1}(\tau) \in
\mathcal{T}(u_k)$ for all $\tau \in [0, t]$, then $${\left\lVertu^k(t_k + t) - u^{k+1}(t)\right\rVert}_U
\le \frac{1}{2} {\left\lVertf(u_k)\right\rVert}_U L e^{L(t_k+t)} t_k^2.$$
We use the integral form of the Gronwall inequality as in [@Potschka2016 Lem. 8.5].
In order to prove an a priori bound on the decrease of the nonlinear residual for , we need the following Lemma.
\[lem:sqrtBound\] Let $h > 0$. If a sequence $(a_k)_{k \in \mathbb{N}}$ of nonnegative numbers satisfies $a_{k+1}^2 \le a_k^2 - 2 h a_k$ for all $k \in \mathbb{N},$ then $a_k \le \max \{a_0 - k h, 0\}$ for all $k \in \mathbb{N}.$
If $a_k \le 2 h$, it follows that $a_{k+1} = 0$ by virtue of $0 \le a_{k+1}^2 \le a_k (a_k - 2h) \le 0.$ If $a_k > 2 h$, we immediately obtain $a_{k+1}^2 \le a_k^2 - 2 h a_k + h^2 = (a_k - h)^2.$ Taking the square root completes the proof.
We can now prove the main theorem of this article.
(0,0) circle (1); (-0.9,0) circle (0.7); (0,-2) arc (-90:90:2); (0,2) arc (90:270:2); (0,2) – (-1,2); (-1,2) arc (90:150:5); (0,-2) – (-1,-2); (-1,-2) arc (90:150:1); (0,0) – (-1,0); (-1,0) arc (90:150:3); (-1,-0.1) – (-1,0.1); (-0.8,-0.1) – (-0.8,0.1); (-0.8,-0.7) arc (-90:90:0.7); (-0.8,-0.7) – (-1,-0.7); (-0.8,0.7) – (-1,0.7); (-1,0.7) arc(90:150:3.7); (-1,-0.7) arc(90:150:2.3); (0,0) circle (2pt) node\[right\] [$u_0^\ast$]{}; (-0.9,-1.2) – (-0.9,0) node [(0,0) circle (1pt);]{}; at (0,-1.4) [$u^0\left( \sum_{i=0}^{k-1} t_i \right)$]{}; at (0,1) [$B(u_0^\ast, \frac{\varepsilon}{2})$]{}; at (-3.8,1.5) [$\mathcal{N}_\varepsilon$]{}; at (-3,0.5) [$\mathcal{N}_{\tilde{\varepsilon}}$]{}; at (-2.5,-0.1) [$u^0(t)$]{};
\[thm:convergenceBSC\] Let A\[ass:validIni\], A\[ass:kappa\], A\[ass:omega\], A\[ass:fLipschitz\], and A\[ass:gamma\] hold. If there exists an $\varepsilon > 0$ such that $$\mathcal{N}_\varepsilon := \bigcup_{t \in [0, \infty)} B(u^0(t), \varepsilon)
\subseteq \mathcal{R}_r$$ and $u_0^\ast$ is an isolated zero of $F$, then there exist constants ${\overline{H}} >0$, $K < \infty$, and $c > 0$ such that the following statements hold true:
1. For all $H \in (0, {\overline{H}}]$ iteration with step size selection converges to $u_0^\ast$.
2. After $\smash[t]{\bar{k}} = \lceil K / \smash[t]{\sqrt{H}} \rceil$ steps, $u_{\bar{k}}$ lies in the region of local full step convergence.
3. The sequence of residual norms ${\left\lVertF(u_k)\right\rVert}_V$ decreases geometrically and satisfies the additional a priori bound $$\sqrt{{\left\lVertF(u_k)\right\rVert}_V}
\le \sqrt{{\left\lVertF(u_0)\right\rVert}_V} - k c \sqrt{H}$$ for all $k \in \{ i \in \mathbb{N} \mid t_j < 1, j = 0, \dotsc, i \}$, which includes $k = \smash[t]{\bar{k}}.$
Because $u_0^\ast$ is an isolated zero of $F$, we can assume without loss of generality that $\varepsilon > 0$ was chosen small enough such that $u_0^\ast$ is the only zero of $F$ in $\mathcal{N}_\varepsilon$. Let now $\theta \in (0, 1)$. We further assume without loss of generality $\varepsilon$ to be sufficiently small to satisfy $\omega r^2 L \varepsilon \le 2 \theta (1-\kappa)$. It follows with A\[ass:validIni\] and A\[ass:fLipschitz\] that for all $u
\in B(u_0^\ast, \varepsilon)$ $$\begin{aligned}
\omega r {\left\lVertF(u)\right\rVert}_V
&< \omega r^2 {\left\lVertf(u)\right\rVert}_U
= \omega r^2 {\left\lVertf(u) - f(u_0^\ast)\right\rVert}_U\\
&\le \omega r^2 L {\left\lVertu - u_0^\ast\right\rVert}_U
< \omega r^2 L \varepsilon
\le 2 \theta (1 - \kappa).
\end{aligned}$$ Thus, Theorem \[thm:localConvergence\] establishes that $B(u_0^\ast,
\varepsilon)$ is contained in the region of local full step convergence. By possibly reducing $\varepsilon$ further according to Lemma \[lem:boundedLevelSets\], we can guarantee that if $u_{k} \in
B(u_0^\ast, \varepsilon)$ for some $k \in \mathbb{N}$, then $(u_k)_k$ converges to the unique zero $u_0^\ast$ in $\mathcal{N}_\varepsilon.$ For the first statement, it now remains to show that $u_k \in B(u_0^\ast,
\varepsilon)$ for some $k \in \mathbb{N}.$
To this end, we choose $T_\ast < \infty$ such that $${\left\lVertu^0(t) - u_0^\ast\right\rVert}_U \le \frac{\varepsilon}{2} \quad \text{for all } t
\ge T_\ast - 1.$$ Because $u_0^\ast$ is the only zero of $F$ on $\mathcal{N}_\varepsilon$, we can choose an $\tilde{\varepsilon} \in (0, \frac{\varepsilon}{2})$ such that there is an $\eta > 0$ satisfying $${\left\lVertf(u)\right\rVert}_U > \eta \quad \text{for all }
u \in \mathcal{T}(u_0) \cap \mathcal{N}_{\tilde{\varepsilon}}, \text{ where }
\mathcal{N}_{\tilde{\varepsilon}} := \bigcup_{t \in [0, T_\ast]} B(u^0(t),
\tilde{\varepsilon}).$$ From A\[ass:gamma\] we obtain the existence of constants $\gamma > 0$ and $t_\gamma \in (0, 1)$ such that $$\label{eqn:lbg}
{\left\lVertg(u, t)\right\rVert}_U \ge \gamma t \quad \text{for all } t \in [0, t_\gamma]
\text{ and } u \in \mathcal{N}_{\tilde{\varepsilon}}.$$ We can then use $\bar{t} = t_\gamma$ in Lemma \[lem:upperStepsizeBound\] to obtain a constant ${\overline{H}} > 0$ such that $$\label{eqn:ubtk}
\min \mathcal{B}_H(u) \le t_\gamma < 1 \quad \text{for all } H \in (0,
{\overline{H}}] \text{ and } u \in \mathcal{N}_{\tilde{\varepsilon}}.$$ In anticipation of a later argument in the proof, we can assume without loss of generality that ${\overline{H}}$ is sufficiently small to satisfy $$\label{eqn:smallHbar}
T_\ast e^{L T_\ast}\left( rL {\left\lVertF(u_0)\right\rVert}_V \right)^{\frac{3}{2}}
\smash[t]{{\overline{H}}}^{\frac{1}{2}} \le 2 \gamma \tilde{\varepsilon}.$$ Combining the inequalities and with , we see that $$\label{eqn:lbtk2}
\gamma t_k^2 \le t_k {\left\lVertg(u_k, t_k)\right\rVert}_U \le H \quad \text{for all } H \in
(0, {\overline{H}}] \text{ and } u_k \in \mathcal{N}_{\tilde{\varepsilon}}.$$ We now choose an $H$-dependent $\bar{k} \in \mathbb{N}$ that satisfies $$T_\ast - 1 \le \sum_{i=0}^{\bar{k}-1} t_i \le T_\ast$$ and show by induction that $u_k \in \mathcal{T}(u_0) \cap
\mathcal{N}_{\tilde{\varepsilon}}$ for all $k \le \bar{k}$, which clearly holds true for $k=0$ because $u_0 = u^0(0) \in B(u_0, \tilde{\varepsilon}).$ We can now assume inductively that $u_i \in \mathcal{T}(u_0) \cap
\mathcal{N}_{\tilde{\varepsilon}}$ for all $i \le k-1$ with $k \le \bar{k}$ in order to show $u_k \in \mathcal{T}(u_0) \cap
\mathcal{N}_{\tilde{\varepsilon}}.$ Because $\mathcal{N}_{\tilde{\varepsilon}} \subseteq \mathcal{R}_r$, it follows from that $t_i < 1$. Hence, Lemma \[lem:lowerStepsizeBound\] yields $$T_\ast \ge
\sum_{i=0}^{k-1} t_i \ge k \frac{\sqrt{H}}{\sqrt{r L {\left\lVertF(u_0)\right\rVert}_V}},$$ which implies the bound $$\label{eqn:ubkbar}
k \le \frac{T_\ast \sqrt{rL{\left\lVertF(u_0)\right\rVert}_V}}{\sqrt{H}}.$$ We then obtain by a telescope argument, Lemma \[lem:oneStepError\], A\[ass:validIni\], , , and that $$\begin{aligned}
{\left\lVertu^0 \left( \sum_{i=0}^{k-1} t_i \right) - u_k\right\rVert}_U
&\le \sum_{j=0}^{k-1} {\left\lVertu^j \left( \sum_{i=j}^{k-1} t_i \right) -
u^{j+1} \left( \sum_{i=j+1}^{k-1} t_i \right)\right\rVert}_U\\
&\le
\sum_{j=0}^{k-1} \frac{1}{2} {\left\lVertf(u_{j})\right\rVert}_U L e^{L T_\ast} t_j^2
< \frac{rL e^{LT_\ast} {\left\lVertF(u_0)\right\rVert}_V}{2 \gamma} H k\\
&\le \frac{T_\ast e^{LT_\ast} \left( rL {\left\lVertF(u_0)\right\rVert}_V
\right)^{\frac{3}{2}} }{2 \gamma} \sqrt{H} \le \tilde{\varepsilon}.
\end{aligned}$$ Because $\sum_{i=0}^{k-1} t_i \le T_\ast$, we have established by induction that $u_k \in \mathcal{T}(u_0) \cap \mathcal{N}_{\tilde{\varepsilon}}$ for all $k \le \bar{k}$. Finally, $u_{\bar{k}} \in B(u_0^\ast, \varepsilon)$ by virtue of $${\left\lVertu_0^\ast - u_{\bar{k}}\right\rVert}_U
\le {\left\lVert u_0^\ast - u^0 \left( \sum_{i=0}^{\bar{k}-1} t_i \right)\right\rVert}_U
+ {\left\lVert u^0 \left( \sum_{i=0}^{\bar{k}-1} t_i \right) - u_{\bar{k}}\right\rVert}_U
< \frac{\varepsilon}{2} + \tilde{\varepsilon} < \varepsilon.$$ This proves the first statement. The second assertion follows from by choosing $K = T_\ast
\sqrt{r L {\left\lVertF(u_0)\right\rVert}_V}$ for $k = \bar{k}$. For the third statement, we obtain from the lower step size bound of Lemma \[lem:lowerStepsizeBound\] and Lemma \[lem:Fdecrease\] that for all $k$ with $t_k < 1$ it holds that $$\begin{gathered}
{\left\lVertF(u_{k+1})\right\rVert}_V \le \left[ 1 - (1-\theta)(1-\kappa)t_k \right]
{\left\lVertF(u_k)\right\rVert}_V
\le {\left\lVertF(u_k)\right\rVert}_V - 2 c \sqrt{H {\left\lVertF(u_k)\right\rVert}_V},\\
\text{with } c = \frac{(1-\theta)(1-\kappa)}{2 \sqrt{rL}} > 0.
\end{gathered}$$ Hence, the residual norm sequence converges geometrically and Lemma \[lem:sqrtBound\] yields the a priori bound with $a_k =
\sqrt{{\left\lVertF(u_k)\right\rVert}_V}$ and $h = c \sqrt{H}$. Because we established $u_k \in
\mathcal{N}_{\tilde{\varepsilon}}$ for all $k \le \smash[t]{\bar{k}}$ in the proof of the first statement, we obtain $t_k < 1$ for all $k \le
\smash[t]{\bar{k}}$ from , which completes the proof.
A note on generalizations to Banach spaces {#sec:Banach}
------------------------------------------
The only step in the convergence proof that exploits the Hilbert space structure of $V$ is Lemma \[lem:Fdescent\]. All remaining steps can be carried out even if $V$ is only a Banach space. The general approach here is to modify the used level function and to require an additional assumption akin to condition A\[ass:kappa\].
In this section, we shortly present the necessary modifications for the special case of the Lebesgue space $V =
L^p(\Omega)$ of real-valued $p$-integrable functions with $ 2 < p <
\infty$ and $\Omega \subset \mathbb{R}^n$. First, we need to consider a different level function $$T_p(u) = \frac{1}{p} {\left\lVertF(u)\right\rVert}_V^p
= \frac{1}{p} \int_{\Omega} {\left\lvertF(u)(s)\right\rvert}^p {{\mathrm d}}s,$$ with analogously defined level sets $\mathcal{T}_p(u)$. In addition to A\[ass:kappa\], we require pointwise that $$\left[F(u) - F'(u)f(u)\right](s) \le \kappa {\left\lvertF(u)(s)\right\rvert} \quad \text{for almost all } s
\in \Omega \text{ and all } u \in \mathcal{R}_r \cap \mathcal{T}_p(u_0).$$ Standard arguments on the differentiability of $p$-norms (see, e.g., [@Lieb2001 Thm. 2.6]) then deliver with the abbreviation $u = u^k(t)$ that $$\begin{aligned}
& \quad\, \frac{{{\mathrm d}}}{{{\mathrm d}}t} T_p(u^k(t))
= -\int_{\Omega} {\left\lvertF(u)(s)\right\rvert}^{p-2} \left[ F(u)(s) \cdot ( F'(u) f(u)) (s)
\right] {{\mathrm d}}s\\
&= -\int_{\Omega} {\left\lvertF(u)(s)\right\rvert}^p {{\mathrm d}}s
+ \int_{\Omega} {\left\lvertF(u)(s)\right\rvert}^{p-2} \left[ F(u)(s) \cdot (F(u) - F'(u) f(u))
(s) \right] {{\mathrm d}}s\\
&\le -p T_p(u)
+ \int_{\Omega} {\left\lvertF(u)(s)\right\rvert}^{p-1} (F(u) - F'(u) f(u)) (s) {{\mathrm d}}s\\
&\le -p(1-\kappa) T_p(u) \le 0.\end{aligned}$$ After application of Gronwall’s inequality, multiplication by $p$ and taking the $p$-th root, we establish the result of Lemma \[lem:Fdescent\] $${\left\lVertF(u^k(t))\right\rVert}_V \le e^{-(1-\kappa)t} {\left\lVertF(u_k)\right\rVert}_V.$$
Algorithmic realization {#sec:realization}
-----------------------
The algorithmic realization of can be carried over verbatim from the finite-dimensional setting laid out in [@Potschka2016 section 10] with the use of ${\left\lVert.\right\rVert}_U$ for all occuring norms. As in [@Potschka2016], we do not use monotone iterations [@Galantai2014] for numerical computations here but use the bisection procedure with exponentially smoothed step size prediction. For convenience, we sketch it again: In order to determine $t_k$ from , we approximately compute a zero of the Lipschitz continuous scalar function $t \mapsto t {\left\lVertg(u_k, t)\right\rVert}_U - H$ by a bracketing procedure. Numerically, we are content with a $t_k$ that satisfies $$\label{eqn:HlHu}
H'_k := t_k {\left\lVertg(u_k, t_k)\right\rVert}_U \in [H^\mathrm{l}, H^\mathrm{u}]
\quad \text{or} \quad
t_k \approx 1 \text{ and } H'_k < H^\mathrm{l},$$ where $H^\mathrm{l} < H$ and $H^\mathrm{u} > H$ are close to $H$.
For illustration purposes, we provide in Table \[alg:example\_atan\] example code in Matlab, which computes for the real-valued function $F(u) =
\arctan(u)$ the iterates $u_1, \dotsc, u_5$ of iteration with $M(u) = F'(u)^{-1} = u^2 + 1$ and step sizes $t_k$ satisfying , starting from $u_0 = 2$ for $H = 0.8$. Full step Newton diverges for this choice of $u_0$. For the sake of brevity, the remaining algorithmic parameters suggested in [@Potschka2016 section 10.2] are used as explict values in the code and termination and error checks are omitted.
k t u du dup Hprime
0 1.0000 2.0e+00 -5.5e+00 1.7e+01 2.3e+01 decrease t
0 0.5000 2.0e+00 -5.5e+00 1.0e+00 3.3e+00 decrease t
0 0.2500 2.0e+00 -5.5e+00 -7.6e-01 1.2e+00 accept t
1 0.2335 6.2e-01 -7.6e-01 -4.9e-01 6.3e-02 increase t
1 0.6168 6.2e-01 -7.6e-01 -1.5e-01 3.8e-01 accept t
2 0.7543 1.5e-01 -1.5e-01 -3.4e-02 8.6e-02 accept t
3 1.0000 3.4e-02 -3.4e-02 2.7e-05 3.4e-02 accept t
4 1.0000 -2.7e-05 2.7e-05 -1.3e-14 2.7e-05 accept t
5 1.0000 1.3e-14 -1.3e-14 -0.0e+00 1.3e-14 accept t
We display the output in Table \[tab:output\]. Each line in the output corresponds to one evaluation of $f(u) = M(u) F(u)$, not counting the initial evaluation of $f(u_0)$. We observe that for $k = 0$, two bisection steps are required to reduce $t_0$ to $\frac{1}{4}$. For $k = 1$, one bisection step is required to increase $t_1$ to above $0.6$. For $k \ge 2$, the predicted step sizes are accepted without further bisection steps. The iterate $u_5$, which is known already at the end of iteration 4, is an acceptable solution candidate, because ${\left\lVert\delta u_5\right\rVert}_2 \approx 1.3 \cdot 10^{-14}$.
The smoothed step size prediction in line 5 is vital for reducing the number of bracketing steps. The predicted step size often already satisfies in all but a few iterations and thus almost no extra computational effort in terms of increment evaluations $f(u_k)$, which typically comprise setting up and solving one linear system, is required for the globalization procedure in most iterations.
Discussion of drawbacks
-----------------------
Backward step control appears to suffer from two substantial drawbacks: First, it depends on a problem specific parameter $H > 0$ that needs to be chosen sufficiently small for convergence, but not too small to cause unecessarily many iterations and thus inefficiency of the method. Second, it may not converge in one step on affine linear problems. We argue here that these two seemingly detrimental properties are actually *necessary* for the class of methods that converge to the closest solution (in the sense of the Newton flow): Let us assume we want to solve a one-dimensional nonlinear equation $F(u) = 0$, $F:
\mathbb{R} \to \mathbb{R}$, with a certain *parameter-free* nonlinear method. Assume it produces a sequence of iterates $(u^k)_{k \in \mathbb{N}}
\subset \mathbb{R}$ that converges to some $u^\ast \neq u^0$. Because $u^0$ cannot be an accumulation point of $(u^k)_{k \in \mathbb{N}}$, we can find the closest iterate $u^j$ to $u^0$ for some $j \in \mathbb{N}$. Any function that differs from $F$ only on the open interval $I$ between $u^0$ and $u^j$ will inevitably lead to the same iterates $(u^k)_{k \in \mathbb{N}}$ if we apply the parameter-free nonlinear method to it. Thus, we can modify $F$ smoothly on $I$ to introduce zeros of $F$ that are closer to $u^0$ than $u^\ast$. We illustrate this construction for an affine linear $F$ in Fig. \[fig:drawback\]. We conclude that any method that provably converges to the closest zero must depend on a parameter such as $H$ to account for problem specific quantities that are virtually impossible to estimate numerically.
Moreover, existing globalization methods [@Deuflhard1991; @Hohmann1994; @Deuflhard1998; @Deuflhard2006] are not parameter free either, because they require an initial step size guess $t_0$ that needs to be sufficiently small.
Design of Newton-type methods {#sec:designM}
=============================
The convergence analysis of backward step control lends itself immediately to the design of globally convergent Newton-type methods. The two required steps are:
1. Define $M$ (or directly $f$) respecting the $\kappa$-condition A\[ass:kappa\].
2. Use to determine the step size sequence $(t_k)$.
The second step is generic. For the first step, we give two important examples in the following two sections. Both methods find approximations $\delta u_k$ of the Newton increment $\delta
u_k^\mathrm{Newton}$ determined by the (infinite-dimensional) linear system $$\label{eqn:HilbertSpaceNewton}
F'(u_k) \delta u_k^\mathrm{Newton} = -F(u_k).$$ We emphasize that the operator $M(u_k)$ is implicitly determined by requiring $\delta u_k = -M(u_k) F(u_k)$. It is not required to actually compute $M(u_k)$, as long as we have $\delta u_k$. However, we need to make sure that $f$ is Lipschitz continuous with respect to $u$.
Krylov–Newton methods {#sec:KrylovNewton}
---------------------
Krylov subspace methods like CG, MINRES, or GMRES [@Hestenes1952; @Paige1975; @Saad1986] for the iterative solution of linear systems, originally developed for large but finite-dimensional sparse systems, can also be stated for infinite-dimensional linear operators. The convergence theory is more complicated in the infinite-dimensional case (see, e.g., [@Nevanlinna1993; @Gasparo2008]). The structure of the linear mapping $F'(u_k): U \to V$ does usually not immediately admit the application of Krylov subspace methods, unless a left preconditioner $P_k: V \to U$ is available, such that $A_k := P_k F'(u_k)$ is an endomorphism on $U$. Then, the $m$-th iterate $\delta u_k^{m}$ of a Krylov subspace method applied to is an approximate solution restricted to the (at most $m$-dimensional) $m$-th Krylov subspace $$\mathcal{K}^m(A_k, P_k F(u_k)) = \left\{ q(A_k) P_k F(u_k) \mid
q \text{ is a polynomial of degree less than } m \right\}.$$ We now focus on residual minimizing Krylov subspace like GMRES and MINRES, which are constructed on the basis of the additional optimality requirement $$\label{eqn:residual_minimizing}
\delta u_k^m = \operatorname*{arg~min}_{\delta u \in \mathcal{K}^m(A_k, P_k F(u_k))}
{\left\lVertP_k F(u_k) + A_k \delta u_k)\right\rVert}_U.$$ As in [@Guennel2014], we investigate the important special case where $V = U^\ast$. In this setting, we can choose $P_k$ as the Riesz isomorphism and immediately obtain $$\label{eqn:Krylov}
\begin{aligned}
{\left\lVertP_k F(u_k) + A_k \delta u_k)\right\rVert}_U &=
{\left\lVertP_k \left[ F(u_k) + F'(u_k) \delta u_k \right]\right\rVert}_U\\
&= {\left\lVertF(u_k) + F'(u_k) \delta u_k)\right\rVert}_V.
\end{aligned}$$ In the light of and , the $\kappa$-condition A\[ass:kappa\] is nothing but the classical relative termination condition used by Krylov subspace methods with given relative tolerance $\kappa$. In other words, GMRES and MINRES choose from the Krylov subspace the increment that achieves the smallest left-hand side in the condition of A\[ass:kappa\] and thus yield (with Lemma \[lem:linearconvergence\]) an asymptotic linear convergence rate bounded above by $\kappa$ for the nonlinear Krylov–Newton method. Also for other Krylov space methods like preconditioned CG, the most commonly used termination criterion is that of the relative residual, even though the relative residual is not guaranteed to decrease monotonically. Thus, the backward step convergence theory of section \[sec:convergence\] delivers suitable termination criteria for the inner linear iterations. In particular, a rather loose relative stopping criterion of, say, $\kappa =
\frac{1}{10}$ delivers asymptotically already one decimal digit of accuracy per nonlinear iteration.
With this approach, there is one theoretic gap we need to close: The increment $-f(u_k) = -M(u_k) F(u_k) = \delta u_k^m$ depends on the number of Krylov subspace iterations $m$, which is determined adaptively. As a concatenation of a finite number of Lipschitz continuous operations, the $m$-th Krylov iterate depends Lipschitz continuously on $u_k$, but changes in $m$ from one nonlinear $k$-iteration to another can lead to discontinuities in the operator $M$. However, a small modification of the above approach can make sure that $M(u)$ and thus $f(u)$ are Lipschitz continuous with respect to $u$, as required for A\[ass:fLipschitz\]: Instead of using the final iterate $\delta u_k^m$, we could use a linear combination $f(u_k) = (1-\alpha_k) \delta
u_k^{m-1} + \alpha_k \delta u_k^m$ of the two last iterates such that instead of the inequality A\[ass:kappa\] the equality $$\nu(\alpha_k) := {\left\lVertF(u_k) + F'(u_k) \left[ (1-\alpha_k) \delta u_k^{m-1} +
\alpha_k \delta u_k^m \right]\right\rVert}_V = \kappa {\left\lVertF(u_k)\right\rVert}_V$$ holds for some $\alpha_k \in [0, 1]$. This is always possible by virtue of the intermediate value theorem applied to the continuous function $\nu(\alpha)$, which satisfies $\nu(0) > \kappa {\left\lVertF(u_k)\right\rVert}$ and $\nu(1) \le \kappa
{\left\lVertF(u_k)\right\rVert}$ because the Krylov subspace method has terminated in step $m$ but not yet in step $m-1$. If we now assume that there is an upper bound on the number of Krylov iterations required to satisfy the relative termination criterion A\[ass:kappa\] on the level set of $u^0$, we can establish Lipschitz continuity of $M$.
Based on our experience in practical computations, however, it is more efficient to always use $\alpha_k = 1$ and robustify the bisection procedure for the approximate solution to against discontinuities of $g$ by relaxing the lower bound $H^\mathrm{l}$ in closer to zero. This might give rise to smaller than necessary step sizes $t_k$, which has not been observed to be problematic in practical computations, but usually delivers faster local residual contraction once $t_k = 1$.
We report numerical results of a Krylov–Newton method for the Carrier equation in section \[sec:carrier\].
Approximation by discretization {#sec:M_by_discretization}
-------------------------------
Following the multilevel Newton approach of [@Hohmann1994], we can also compute an approximate solution $\delta u_k$ by first discretizing and then (approximately) solving the discretized system. The $\kappa$-condition A\[ass:kappa\] yields a computable criterion for checking if the approximation is accurate enough to ensure convergence in $U$. If not, we need to improve the discretization (and possibly the accuracy of the approximate solution to the resulting finite-dimensional linear system).
In this conceptually simple approach, challenges can arise in the evaluation of the $V$-norms in A\[ass:kappa\]. We address these issues for the case of an elliptic partial differential equation in section \[sec:minsurf\]. Moreover, the evaluation of the $V$-norm in A\[ass:kappa\] can provide a means to adaptively discretize and in turn also the original nonlinear problem.
As in section \[sec:KrylovNewton\], the procedure for approximating a solution to usually involves some discrete decisions, for instance the marking and refinement of certain discretization cells as $k$ increases. Thus, the so constructed $f(u)$ is not Lipschitz continuous. A smoothed formulation akin to the interpolation construction in section \[sec:KrylovNewton\] exceeds the scope of this article and shall be investigated in future work.
Numerical examples from nonlinear elliptic boundary value problems {#sec:ellipticPDE}
==================================================================
In this section, we illustrate the general paradigms presented in section \[sec:M\_by\_discretization\] for the class of elliptic boundary value problems on a bounded domain $\Omega \subset {{\mathbb R}}^n$ with continuously differentiable boundary $\partial \Omega$ that can be cast as nonlinear root-finding problems with the Sobolev spaces $U = H^1_0(\Omega)$ and $V =
H^{-1}(\Omega)$.
Based on the Poincaré inequality (see, e.g., [@Evans2010]), we can use the inner product $${\left(u,v\right)}_U = \int_{\Omega} \nabla u \cdot \nabla v$$ for the Hilbert space $U$. For the Hilbert space $V$, we can then compute norms via the Riesz representation theorem [@Yosida1978 §III.3]: For every $v
\in V$, there exists a uniquely determined $r_v \in U$ such that $$\label{eqn:RieszRepresentation}
{\left(u, r_v\right)}_U = \int_{\Omega} v u \text{ for all } u \in U
\quad \text{and} \quad
{\left\lVertv\right\rVert}_V = {\left\lVertr_v\right\rVert}_U.$$
We investigate the numerical performance of backward step control on two nonlinear elliptic boundary value problems on bounded domains $\Omega$ with continuously differentiable boundary.
All algorithmic parameters of backward step control are chosen as in [@Potschka2016] unless otherwise stated.
Preconditioned Krylov subspace methods {#sec:PGMRES}
--------------------------------------
We first illustrate the general Krylov subspace method approach presented in section \[sec:KrylovNewton\]. As our focus in the case of Krylov–Newton methods here lies on a concise algorithmic statement rather than ultimate computational speed, we use Chebfun [@Battles2004; @Driscoll2014] as an algorithmic tool for the numerical results in section \[sec:carrier\], because it allows to compute numerically with functions (represented as adaptively truncated Chebyshev expansions) instead of numbers [@Trefethen2007] and supports the automatic computation of Fréchet derivatives by the use of automatic differentiation in function space [@Birkisson2012].
Because the linear operators in Chebfun are implemented in strong form, we also use the strong form of the inner product in $U$ $${\left(u,v\right)}_U = -\int_{-1}^{1} u {\Delta}v = -\int_{-1}^{1} v {\Delta}u,$$ (which requires $u$ or $v$ to have square integrable second derivatives). It follows from that $$-\int_{-1}^{1} u {\Delta}r_v = {\left(u, r_v\right)}_U = \int_{-1}^{1} v u
\quad \text{for all } u \in U,$$ and, thus, $r_v = -{\Delta}^{-1} v$ and $${\left\lVertv\right\rVert}_V^2 = {\left\lVertr_v\right\rVert}_U^2 = {\left\lVert-{\Delta}^{-1} v\right\rVert}_U^2.
$$ In Chebfun, ${\Delta}^{-1}$ can be evaluated efficiently using ultraspherical spectral collocation [@Olver2013]. Based on these prerequisites, we modified Chebfun’s builtin GMRES to use the inner product and norm of $U$ (instead of $L^2(\Omega)$) in combination with ${\Delta}^{-1}$ as a preconditioner, which yields the correct residual norm ${\left\lVert{\Delta}^{-1}v\right\rVert}_U = {\left\lVertv\right\rVert}_V$. We note that we could have used MINRES or even CG because the resulting left-preconditioned linear system is self-adjoint and positive definite in the case at hand [@Guennel2014]. This does not affect our results dramatically, because GMRES produces the same iterates as MINRES for self-adjoint systems. However, the current version of Chebfun does not ship a MINRES implementation. The numerical results for the Carrier equation in section \[sec:carrier\] indicate that using the Riesz isomorphism as a preconditioner works satisfactorily in our example.
### Application to the Carrier equation {#sec:carrier}
For $\varepsilon > 0$, we want to determine a real-valued function $u(x)$ on $x \in [-1, 1]$ that solves the nonlinear second order boundary value problem $$\label{eqn:carrier}
\varepsilon {\Delta}u + 2 (1-x^2) u + u^2 = 1, \quad u(\pm 1) = 0,$$ which—according to [@Bender1999 §9.7]—is due to Carrier. For small $\varepsilon$, it becomes challenging to solve because of the existence of many local solutions (compare Fig. \[fig:carrier\_solution\]).
We apply the GMRES–Newton method described in section \[sec:PGMRES\] to and illustrate the convergence of the method in Fig. \[fig:carrier\_convergence\] for $\varepsilon
= 10^{-3}$, $\kappa = 10^{-2}$, and varying values of $H = H_\mathrm{rel}
{\left\lVert\delta u_0\right\rVert}_U$, where we choose $H_\mathrm{rel} \in \{ 0.5, 0.1, 0.05,
0.01 \}$. The initial guess is $u_0 = 0$ and the termination criterion is ${\left\lVertF(x_k)\right\rVert}_V \le 10^{-11}$. The relative GMRES termination tolerance $\kappa$ was chosen rather large but at the same time small enough to ensure sufficiently fast local convergence with two decimal digits per iteration. We first observe that no convergence can be obtained for $H_\mathrm{rel} = 0.5$, even though ${\left\lVertF(u_{41})\right\rVert}_V$ drops below $2.6 \cdot 10^{-5}$ and ${\left\lVert\delta u_7\right\rVert}_U \approx 0.28$. From these numbers we can estimate the nonlinearity of the problem in terms of $\omega$ and its well-posedness in terms of $r$ based on Lemma \[lem:Fdecrease\], which yields $$\omega \ge 2 (1 - \kappa) / (t_7 {\left\lVertf(u_7)\right\rVert}_U) \approx 7.1
\quad \text{and} \quad
\omega r \ge 2 (1 - \kappa) / (t_{41} {\left\lVertF(u_{41})\right\rVert}_V) \approx 7.6 \cdot
10^{4}$$ and shows that the problem is highly nonlinear.
For the remaining choices of $H_\mathrm{rel}$ we obtain convergence, albeit a different local solution is found for $H_\mathrm{rel} = 0.1$ than for the others (compare Fig. \[fig:carrier\_solution\]), which nicely confirms the theory of Theorem \[thm:convergenceBSC\]. As guaranteed by Lemma \[lem:localFullSteps\], full steps $t_k = 1$ are taken in the vicinity of a solution and we can clearly observe the asymptotic linear convergence rate of $\kappa = 10^{-2}$ for the residual norm predicted by Lemma \[lem:linearconvergence\]. The final increment norms seem rather large because we do not compute $\delta u_k$ if ${\left\lVertF(u_k)\right\rVert}_V$ is already below $10^{-11}$. Thus, the last increment norm lags behind by one iteration and would be much smaller if we computed it again for the final iterate.
In Fig. \[fig:carrier\_gmres\], we see that the number of GMRES iterations needed in each nonlinear iteration stays moderately small. In each GMRES iteration, one operator-vector-multiplication must be carried out, which we compute via Chebfun as a directional derivative of $F$. In total, 1255 ($H_\mathrm{rel} = 0.1$), 1455 ($H_\mathrm{rel} = 0.05$), and 2471 ($H_\mathrm{rel} = 0.01$) directional derivatives of $F$ are required to solve the problem.
In addition, we can observe from Fig. \[fig:carrier\_gmres\] that the predicted step size $t_k$ needs to be corrected only in few iterations $k$ by the backward step control bisection procedure outlined in section \[sec:realization\]. In the first iteration, four ($H_\mathrm{rel} = 0.1, 0.05$) and five ($H_\mathrm{rel} = 0.01$) bisection steps are required to reduce the initial step size guess $t_0 = 1$. In all other iterations marked by $\bullet$ in Fig. \[fig:carrier\_gmres\], only one additional bisection step is necessary, except in iteration $k=8$ for $H_\mathrm{rel} = 0.1$ and $k=21$ for $H_\mathrm{rel} = 0.01$, where two bisection steps need to be taken. This backs up our claim that the computational overhead of backward step control is small.
Approximation by Finite Elements {#sec:FE}
--------------------------------
In contrast to section \[sec:PGMRES\], we now explicitly discretize the increment $\delta u_k$ and the step determination equation by Finite Elements: To this end, let $\mathcal{C}$ be a partition of $\Omega$ into cells $C \in
\mathcal{C}$. We can then construct the finite-dimensional Finite Element subspace $$U_{\mathcal{C}}^p = \left\{ u \in C^0(\Omega) \mid u \text{ is a polynomial
of degree $p$ on each } C \in \mathcal{C} \right\} \subset U.$$ The increment is then determined by finding $\delta u_k \in U_{\mathcal{C}}^p$ such that $$\label{eqn:discretizedincrement}
(F'(u_k) \delta u_k) \varphi = -F(u_k) \varphi \quad
\text{for all } \varphi \in U_{\mathcal{C}}^p.$$ By fixing a nodal basis of $U_{\mathcal{C}}^p$, we obtain a linear system with a large but sparse (typically symmetric positive definite) ${\left\lvertU_{\mathcal{C}}^p\right\rvert}$-by-${\left\lvertU_{\mathcal{C}}^p\right\rvert}$ matrix.
### Computation of norms in $V$
The discretization in the previous paragraph is completely standard. We now describe the non-standard part, which comprises the computation of $$\label{eqn:kappak}
{\left\lVertF(u_k)\right\rVert}_V \quad \text{and} \quad
\kappa_k := \frac{{\left\lVertF(u_k) + F'(u_k) \delta u_k\right\rVert}_V}{{\left\lVertF(u_k)\right\rVert}_V}$$ in the space $V = H^{-1}(\Omega)$. To this end, we use again the Riesz representation . However, using the same Finite Element subspace $U_{\mathcal{C}}^p$ for the discretization of would yield the wrong value $\kappa_k = 0$ because the numerator vanishes. Moreover, using $U_{\mathcal{C}}^p$ would also give wrong results for ${\left\lVertF(u_k)\right\rVert}_V$ because the residual projected on $U_{\mathcal{C}}^p$ converges locally quadratic (as a Newton method on a finite-dimensional space) if we solve exactly, but does not see the discretization error. Thus, we need to solve on richer Finite Element spaces. The numerical results in section \[sec:minsurf\] indicate that choosing $U_{\mathcal{C}}^{p+1}$ seems to be sufficient for good estimates of the required $V$-norms.
### Adaptive mesh refinement to minimize the contraction rate $\kappa$
Using $p$-refinement for the solution of instead of refinement of $\mathcal{C}$ has two advantages: First, the increase in the degrees of freedom ${\left\lvert\smash[b]{\smash[t]{U_{\mathcal{C}}^{p+1}}}\right\rvert}$ with respect to ${\left\lvertU_{\mathcal{C}}^p\right\rvert}$ is only moderate if $p$ is moderately large, e.g., $p = 3$. Second, the squared norm $\smash[t]{{\left\lVertr_v\right\rVert}}_U^2$ can then be written as a sum of contributions from each cell $C \in \mathcal{C}$, which indicate which cells should ideally be refined if $\kappa_k$ is larger than a desired residual contraction rate $\kappa < 1$ prescribed by the user. The cellwise contributions $\kappa_k$ can be treated in the same fashion as existing cellwise error indicators.
### Application to the minimum surface equation {#sec:minsurf}
In this section, we consider the classical minimum surface problem in the following special form: Let ${\left\lvert.\right\rvert}$ denote the Euclidean norm in $\mathbb{R}^2$ and let $\Omega = \left\{ x \in \mathbb{R}^2 \mid
{\left\lvertx\right\rvert}_2 < 1 \right\}$ and $u^\partial(x) = \sin(2 \pi (x_1 + x_2))$. We seek a function $u$ on $\Omega$ that equals $u^\partial$ on $\partial
\Omega$ and minimizes the area of its graph $$\min I(u) = \int_{\Omega} \sqrt{1 + \smash[t]{{\left\lvert\nabla u\right\rvert}}^2} \quad
\text{s.t.} \quad u\big|_{\partial\Omega} = u^\partial\big|_{\partial\Omega}.$$ With the spaces $U = H^1_0(\Omega)$ and $V = H^{-1}(\Omega)$ as before, the minimum is described as the solution $u \in u^\partial + U$ to the variational problem $$F(u) \varphi := \int_{\Omega} \nabla \varphi \cdot {\mathbf{g}}(\nabla u)
= 0 \quad \text{for all } \varphi \in U,
\quad \text{where } {\mathbf{g}}({\mathbf{v}}) := \left(1 +
\smash[t]{{\left\lvert{\mathbf{v}}\right\rvert}}^2\right)^{-\frac{1}{2}} {\mathbf{v}}.$$ Thus, $F$ maps $u^\partial + U$ to $V$. Its Gâteaux derivative $F':
(u^\partial + U) \times U \to V$ can then be expressed as $$\left( F'(u) \delta u \right) \varphi
= \int_{\Omega} \nabla \varphi \cdot {\mathbf{g}}'(\nabla u) \nabla \delta u,$$ where the Jacobian of ${\mathbf{g}}$ is given by $${\mathbf{g}}'({\mathbf{v}}) = \frac{1}{\sqrt{1 + \smash[t]{{\left\lvert{\mathbf{v}}\right\rvert}}^2}} \left( {{\mathrm I}}_2 -
\frac{1}{1 + \smash[t]{{\left\lvert{\mathbf{v}}\right\rvert}}^2} {\mathbf{v}} {\mathbf{v}}^T \right).$$ We recall that $F$ is continuously Gâteaux differentiable from $U = H^1_0(\Omega)$ (with norm $\smash[t]{{\left\lVertu\right\rVert}}_U^2 = \int_{\Omega}
\smash[t]{{\left\lvert\nabla u\right\rvert}}^2$) to $V = H^{-1}(\Omega)$. To see this, we use the chain rule [@Hamilton1982 3.3.4. Thm.] on the continuously Gâteaux differentiable $L^2(\Omega, \mathbb{R}^2)$ inner product and the Nemytskii operator defined by ${\mathbf{g}}$. The Nemytskii operator defined by ${\mathbf{g}}$ mapping from $L^2(\Omega,
\mathbb{R}^2)$ to itself is continuously Gâteaux differentiable by virtue of [@Goldberg1992 Thm. 8, Rem. 6], because ${\mathbf{g}}$ and its Jacobian are uniformly bounded ${\left\lvert{\mathbf{g}}({\mathbf{v}})\right\rvert} \le 1$ and ${\left\lvert{\mathbf{g}}'({\mathbf{v}})\right\rvert}_{2 \times 2}
\le 1$ (where ${\left\lvert.\right\rvert}_{2 \times 2}$ denotes the spectral norm of 2-by-2 matrices), implied by the eigenvalues $$\frac{1}{ \sqrt{1 + \smash[t]{{\left\lvert{\mathbf{v}}\right\rvert}}^2}}
\left( 1 - \frac{\smash[t]{{\left\lvert{\mathbf{v}}\right\rvert}}^2}{1 + \smash[t]{{\left\lvert{\mathbf{v}}\right\rvert}}^2} \right)
\quad \text{and} \quad
\frac{1}{\sqrt{1 + \smash[t]{{\left\lvert{\mathbf{v}}\right\rvert}}^2}},$$ of ${\mathbf{g}}'({\mathbf{v}})$ corresponding to the eigenspace spanned by ${\mathbf{v}}$ and its complement. We remark here that $F$ does not satisfy the stronger property of being continuously Fréchet differentiable as a mapping from $H^1_0(\Omega)$ to $H^{-1}(\Omega)$ as noted in [@Wachsmuth2014].
We solve the resulting system only approximately with a preconditioned Conjugate Gradient (PCG) method [@Hestenes1952]. The resulting inexactness also contributes to the computations of $\kappa_k$ in .
All computations were obtained with the software package deal.II [@Bangerth2016; @Bangerth2007].
### Efficient computation of Riesz representations
As described in section \[sec:FE\], the Riesz representations $r_v$ need to be computed from in order to compute the $V$-norms entering $\kappa_k$. The following algorithmical and computational approaches are important to prevent the computation times for the solution of on the richer Finite Element subspace $U_{\mathcal{C}}^{p+1}$ from dominating the overall computational effort:
1. We found that PCG with a symmetric Gauss–Seidel smoother as preconditioner delivers good results. In the computations reported below, we employed a Chebyshev smoother of degree four because it yields a slightly better performance than Gauss–Seidel when running the code on several processors in parallel. The use of multigrid does not pay off because it usually involves a rather expensive setup machinery on unstructured meshes (see, e.g., [@Janssen2011]) and the right-hand sides of consist mainly of high-frequency residuals in $U_{\mathcal{C}}^{p+1} \setminus U_{\mathcal{C}}^p$, the low frequency residuals having been mostly eliminated on $U_{\mathcal{C}}^p$ already.
2. In order to avoid a possible memory bottleneck caused by storing the stiffness matrix discretized on the high-dimensional space $U_{\mathcal{C}}^{p+1}$, we use a matrix-free realization of the Laplacian [@Kronbichler2012].
3. Because the resulting computation times are then dominated by the bandwidth of the access to main memory, we perform all computations involved in the (matrix-free) matrix-vector products for the solution of only with single instead of double precision floating point arithmetic. The numerical results below indicate that this approach is still sufficiently accurate, while being considerably faster.
4. Because $\kappa_k$ only steers the algorithm but does not affect the quality of the iterates $x_k$ directly, it can be computed with rather low accuracy requirements in the PCG method. We use relative stopping criteria of $0.1$ and $0.05$ for the numerator and the denominator of $\kappa_k$ in .
### Details about the numerical setup
![Left: The initial mesh for the solution of the minimum surface equation has 425 cells. With $p=3$, the resulting Finite Element space has 2929 degrees of freedom. Right: The accepted and discarded values of $\kappa_k$ in the numerical solution of the minimum surface equation. Discarded values are marked with $\circ$ at fractional iteration numbers, the values of $\kappa_k$ on the finally successful mesh are marked with $\times$ at integer iteration numbers. The discarded values of $\kappa_k$ converge to 1, while the accepted ones approach approximately 0.08.[]{data-label="fig:initialmesh"}](initialmesh){height="30mm"}
We discretize and by nodal Finite Elements of order $p = 3$ and $p
+ 1 = 4$, respectively, on quadrilaterals with tensor product polynomials using Gauss–Lobatto nodes. For the elements on the curved boundary, we employ polynomial tranformations of degree seven from the reference cell to the physical cells. The relative tolerance of PCG for the solution of is $0.001$ (in the Euclidean norm on the discretized vectors).
As a starting guess, we let $u_0$ be the Finite Element interpolation of $u^\partial$ on the coarsest mesh depicted in Fig. \[fig:initialmesh\]. In a first phase, we iterate until the increment norm in $U$ is below 0.01 without computing any $V$-norms. This first phase is a finite-dimensional Newton method ($\kappa = 0$) globalized with backward step control ($H_{\mathrm{rel}} = 0.05$, $t_0 = 1$) on the Finite Element subspace $U_{\mathcal{C}}^p$ belonging to the initial mesh.
For the successive phase of nonlinear adaptive mesh refinement described in section \[sec:FE\], we choose $\kappa = 0.5$. If $\kappa_k > \kappa$, we mark all cells for refinement that have a contribution of more than $2^{-p}$ times the maximum cell contribution to $\kappa_k$, up to a given maximum number of 200,000 cells. The resulting number of cells might be slightly higher in the final mesh due to mesh smoothing in deal.II. In the first eight iterations, the step size is gradually increased from $t_0 =
0.0625$ to $t_7 = 0.9738$. All other iterations are performed with full steps $t_k = 1$. The mesh is refined for the first time in iteration 11, kept for iteration 12, and then successively refined in each further step until the maximum number of cells is reached in iteration 21. We can furthermore observe from Fig. \[fig:initialmesh\] that from iteration 16 on, the first trial value of $\kappa_k$ before the refinement is always one (up to three decimal digits), which shows that no further improvement can be achieved by performing more nonlinear iterations on the current discretization, which is automatically detected correctly by the algorithm.
We compare in Fig. \[fig:minsurfConvergence\] the convergence of the residual norm ${\left\lVertF(u_k)\right\rVert}_V$ (computed afterwards with high accuracy on a refined mesh, which is generated by one additional global refinement step of the triangulation of the final mesh) of backward step control $\kappa$-optimizing adaptive mesh refinement with the convergence when using mesh refinement with the deal.II builtin Kelly error indicator [@Kelly1983; @Gago1983]. In contrast to the theory of backward step control, there is no theoretical guideline for the Kelly indicator on how many nonlinear iterations to run before another round of refinement is triggered. We choose to refine as soon as the increment norm ${\left\lVert\delta
u_k\right\rVert}_U$ becomes less than or equal to a factor $\rho > 0$ of the error estimate returned on the last mesh by the Kelly indicator. Fig. \[fig:minsurfConvergence\] shows that $\kappa$-optimizing adaptive mesh refinement delivers the best ratio of residual norm versus CPU time and versus the number of degrees of freedom compared to mesh refinement based on the Kelly indicator for varying values of $\rho = 0.5, 0.1, 0.01$.
The computations for the solution of the minimum surface equation with a final number of 1.5 million degrees of freedom using $\kappa$-optimizing mesh refinement took 112s wall clock time on the four cores of a mid 2012 MacBook Pro, 2.3 GHz Intel Core i7, 8 GB. Out of this grand total, the computations necessary for estimating $\kappa_k$ took only 18s, even though they need to be performed on the high-dimensional Finite Element space $U_{\mathcal{C}}^{p+1}$.
Conclusions
===========
We presented a comprehensive convergence analysis for , a method that globalizes the convergence of Newton-type methods for the solution of in a Hilbert space setting. We proved that under the reasonable assumptions A\[ass:validIni\]–A\[ass:gamma\] the iterates $u_k$ either leave the region of $r$-regular points $\mathcal{R}_r$ (in which case we need to adjust $M$ or embed $F$ in a suitable homotopy in order to prevent attraction to singularities) or converge to the distinctive solution $u_0^\ast$ (the initial guess $u_0$ propagated by the generalized Newton flow ) provided that $H > 0$ is chosen sufficiently small. Moreover, we provided an $H$-dependent a priori bound on the decrease of ${\left\lVertF(u_k)\right\rVert}_V$ and characterized the asymptotic linear residual convergence rate by $\kappa$. We provided efficient numerical methods based on the blueprint of bounding and optimizing $\kappa$ in each iteration, either over a finite-dimensional subspace in a Krylov–Newton method or through an adaptive Finite Element discretization, in order to balance the nonlinear residual norm with the residual norm of the linear systems. We applied these methods to the class of nonlinear elliptic boundary value problems and presented numerical results for the Carrier equation in a Chebfun implementation and for the minimum surface equation in deal.II. The challenge to efficiently compute norms in $V = H^{-1}(\Omega)$ via the Riesz preconditioner can be addressed by suitable numerical methods and techniques.
The author is grateful to Felix Lenders and to Gerd Wachsmuth for comments on an earlier draft of this manuscript and to the anonymous reviewers for their fruitful comments. This work was funded by the European Research Council through S. Engell’s and H.G. Bock’s ERC Advanced Investigator Grant MOBOCON (291 458) and by the German Federal Ministry of Education and Research under grants 05M2013-GOSSIP and 05M2016-MOPhaPro.
[10]{} \[1\][[\#1]{}]{} urlstyle \[1\][DOI \#1]{}
Ainsworth, M., Oden, J.: A posteriori error estimation in finite element analysis. Pure and Applied Mathematics (New York). Wiley-Interscience \[John Wiley & Sons\], New York (2000)
Amann, H.: Ordinary differential equations, *de Gruyter Studies in Mathematics*, vol. 13. Walter de Gruyter & Co., Berlin (1990)
Ascher, U., Osborne, M.: [A]{} note on solving nonlinear equations and the natural criterion function. ournal of [O]{}ptimization [T]{}heory and [A]{}pplications **55**(1), 147–152 (1987)
Bangerth, W., Davydov, D., Heister, T., Heltai, L., Kanschat, G., Kronbichler, M., Maier, M., Turcksin, B., Wells, D.: The `deal.II` library, version 8.4. Journal of Numerical Mathematics **24** (2016)
Bangerth, W., Hartmann, R., Kanschat, G.: [deal.II]{} – a general purpose object oriented finite element library. ACM Trans. Math. Softw. **33**(4), 24/1–24/27 (2007)
Bank, R., Rose, D.: Global approximate [N]{}ewton methods. Numer. Math. **37**(2), 279–295 (1981)
Battles, Z., Trefethen, L.: An extension of [MATLAB]{} to continuous functions and operators. SIAM J. Sci. Comput. **25**(5), 1743–1770 (2004)
Becker, R., Rannacher, R.: An optimal control approach to a posteriori error estimation in finite element methods. Acta Numer. **10**, 1–102 (2001)
Bender, C., Orszag, S.: Advanced mathematical methods for scientists and engineers. [I]{}. Springer-Verlag, New York (1999). Asymptotic methods and perturbation theory, Reprint of the 1978 original
Birkisson, A., Driscoll, T.: Automatic [F]{}réchet differentiation for the numerical solution of boundary-value problems. ACM Trans. Math. Software **38**(4) (2012)
Bock, H.: [R]{}andwertproblemmethoden zur [P]{}arameteridentifizierung in [S]{}ystemen nichtlinearer [D]{}ifferentialgleichungen, *Bonner Mathematische Schriften*, vol. 183. Universität Bonn, Bonn (1987)
Bock, H., Kostina, E., Schlöder, J.: [O]{}n the [R]{}ole of [N]{}atural [L]{}evel [F]{}unctions to [A]{}chieve [G]{}lobal [C]{}onvergence for [D]{}amped [N]{}ewton [M]{}ethods. In: M. Powell, S. Scholtes (eds.) [S]{}ystem [M]{}odelling and [O]{}ptimization. [M]{}ethods, [T]{}heory and [A]{}pplications, pp. 51–74. Kluwer (2000)
Brezinski, C.: Numerical stability of a quadratic method for solving systems of nonlinear equations. Computing **14**(3), 205–211 (1975)
Davidenko, D.: [O]{}n a new method of numerical solution of systems of nonlinear equations. oklady [A]{}kad. [N]{}auk [SSSR]{} ([N]{}.[S]{}.) **88**, 601–602 (1953)
Dembo, R., Eisenstat, S., Steihaug, T.: Inexact [N]{}ewton methods. SIAM J. Numer. Anal. **19**(2), 400–408 (1982)
Dennis Jr., J., Moré, J.: Quasi-[N]{}ewton methods, motivation and theory. SIAM Rev. **19**(1), 46–89 (1977)
Deuflhard, P.: [A]{} [M]{}odified [N]{}ewton [M]{}ethod for the [S]{}olution of [I]{}ll-conditioned [S]{}ystems of [N]{}onlinear [E]{}quations with [A]{}pplications to [M]{}ultiple [S]{}hooting. umer. [M]{}ath. **22**, 289–311 (1974)
Deuflhard, P.: Global inexact [N]{}ewton methods for very large scale nonlinear problems. Impact Comput. Sci. Engrg. **3**(4), 366–393 (1991)
Deuflhard, P.: [N]{}ewton methods for nonlinear problems, *Springer Series in Computational Mathematics*, vol. 35. Springer, Heidelberg (2011). Affine invariance and adaptive algorithms
Deuflhard, P., Weiser, M.: Global inexact [N]{}ewton multilevel [FEM]{} for nonlinear elliptic problems. In: Multigrid methods [V]{} ([S]{}tuttgart, 1996), *Lect. Notes Comput. Sci. Eng.*, vol. 3, pp. 71–89. Springer, Berlin (1998)
Driscoll, T., Hale, N., Trefethen, L. (eds.): Chebfun guide. Pafnuty Publications, Oxford (2014)
Evans, L.: Partial differential equations, *Graduate Studies in Mathematics*, vol. 19, second edn. American Mathematical Society, Providence, RI (2010)
Gago, J.d.S., Kelly, D., Zienkiewicz, O., Babu[š]{}ka, I.: A posteriori error analysis and adaptive processes in the finite element method. [II]{}. [A]{}daptive mesh refinement. Internat. J. Numer. Methods Engrg. **19**(11), 1621–1656 (1983)
Galántai, A., Abaffy, J.: Always convergent iteration methods for nonlinear equations of [L]{}ipschitz functions. Numer. Algor. pp. 1–11 (2014)
Gasparo, M., Papini, A., Pasquali, A.: Some properties of [GMRES]{} in [H]{}ilbert spaces. Numer. Func. Anal. Opt. **29**(11–12), 1276–1285 (2008)
Goldberg, H., Kampowsky, W., Tröltzsch, F.: On [N]{}emytskij operators in [$L_p$]{}-spaces of abstract functions. Math. Nachr. **155**, 127–140 (1992)
Gr[ä]{}tsch, T., Bathe, K.J.: A posteriori error estimation techniques in practical finite element analysis. Comput. & Structures **83**(4-5), 235–265 (2005)
Günnel, A., Herzog, R., Sachs, E.: A note on preconditioners and scalar products in [K]{}rylov subspace methods for self-adjoint problems in [H]{}ilbert space. Electron. Trans. Numer. Anal. **41**, 13–20 (2014)
Hairer, E., Wanner, G.: Solving Ordinary Differential Equations [I]{}[I]{}, *Springer Series in Computational Mathematics*, vol. 14, second edn. Springer, Berlin (1996)
Hamilton, R.: The inverse function theorem of [N]{}ash and [M]{}oser. Bull. Amer. Math. Soc. (N.S.) **7**(1), 65–222 (1982)
Hestenes, M., Stiefel, E.: Methods of conjugate gradients for solving linear systems. J. Res. Nat. Bur. Standards **49**, 409–436 (1952)
Hohmann, A.: Inexact [G]{}auss [N]{}ewton methods for parameter dependent nonlinear problems. Ph.D. thesis, Freie Universität Berlin (1994)
Janssen, B., Kanschat, G.: Adaptive multilevel methods with local smoothing for [$H^1$]{}- and [$H^{\rm curl}$]{}-conforming high order finite element methods. SIAM J. Sci. Comput. **33**(4), 2095–2114 (2011)
Kelly, D., Gago, J.d.S., Zienkiewicz, O., Babu[š]{}ka, I.: A posteriori error analysis and adaptive processes in the finite element method. [I]{}. [E]{}rror analysis. Internat. J. Numer. Methods Engrg. **19**(11), 1593–1619 (1983)
Kronbichler, M., Kormann, K.: A generic interface for parallel cell-based finite element operator application. Comput. & Fluids **63**, 135–147 (2012)
Lieb, E., Loss, M.: Analysis, *Graduate Studies in Mathematics*, vol. 14, second edn. American Mathematical Society, Providence, RI (2001)
Nevanlinna, O.: Convergence of iterations for linear equations. Lectures in mathematics. Birkh[ä]{}user, Basel, Boston, Berlin (1993)
Olver, S., Townsend, A.: A fast and well-conditioned spectral method. SIAM Rev. **55**(3), 462–489 (2013)
Paige, C., Saunders, M.: [S]{}olutions of sparse indefinite systems of linear equations. . [N]{}umer. [A]{}nal. **12**(4), 617–629 (1975)
Potschka, A.: A direct method for parabolic [PDE]{} constrained optimization problems. Advances in [N]{}umerical [M]{}athematics. Springer (2013)
Potschka, A.: Backward step control for global [N]{}ewton-type methods. SIAM J. Numer. Anal. **54**(1), 361–387 (2016)
Rannacher, R., Vihharev, J.: Adaptive finite element analysis of nonlinear problems: balancing of discretization and iteration errors. J. Numer. Math. **21**(1), 23–61 (2013)
Saad, Y., Schultz, M.: [GMRES]{}: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM J. Sci. Stat. Comput. **7**(3), 856–869 (1986)
Trefethen, L.: Computing numerically with functions instead of numbers. Math. Comput. Sci. **1**(1), 9–19 (2007)
Wachsmuth, G.: Differentiability of implicit functions: [B]{}eyond the implicit function theorem. J. Math. Anal. Appl. **414**(1), 259–272 (2014)
Yosida, K.: Functional analysis, fifth edn. Springer-[V]{}erlag (1978)
Ypma, T.: Local convergence of inexact [N]{}ewton methods. SIAM J. Numer. Anal. **21**(3), 583–590 (1984)
|
---
abstract: 'Let $K/F$ be a quadratic tamely ramified extension of a non-Archimedean local field $F$ of characteristic zero. In this paper, we give an explicit formula for Langlands’ lambda function $\lambda_{K/F}$.'
address: |
Chennai Mathematical Institute\
H1, Sipcot It Park, Siruseri\
Kelambakkam, 603103\
India
author:
- '**Sazzad Ali Biswas**'
title: 'Langlands’ Lambda function for quadratic tamely ramified extensions'
---
**Introduction**
================
Let $K/F$ be a finite subextension (but need not be Galois) in $\overline{F}/F$, where $\overline{F}$ is an algebraic closure of a non-Archimedean local field $F$ of characteristic zero. Let $\psi$ be a nontrivial additive character of $F$. Then Langlands’s lambda function (or simply $\lambda$-function) (cf. [@RL]) of the extension $K/F$ is: $$\lambda_{K/F}(\psi):=W(\text{Ind}_{G_K}^{G_F}(1_K),\psi),$$ where $1_K$ is the trivial representation of $G_K:=\text{Gal}(\overline{F}/K)$. Here $W$ denotes for local constant (or epsilon factor) (cf. [@JT1]). We also can define the $\lambda$-function via Deligne’s constant $c(\rho):=\frac{W(\rho)}{W(\det(\rho))}$, where $\rho$ is a finite dimensional representation of $G_F$ and $\det(\rho)$ is the determinant of $\rho$.
Langlands has shown (cf. Theorem 1 on p. 105 of [@JT1]) that the local constants are [**weakly**]{} extendible functions. Therefore, to compute the local constant of any induced local Galois representation, we have to compute the $\lambda$-function explicitly. Since the local Langlands correspondence preserves local constants, the explicit computation of local constants is an important part of the Langlands program. In the automorphic side of the local Langlands correspondence, we have local converse theorem (cf. [@JWC], [@GH]), but so far we do not have any such converse theorem in the Galois side [**because**]{} explicit computation of $\lambda$-functions (hence epsilon factors of local Galois representations) are not known. In the previous paper [@SABNT] the author gave an explicit computation of the lambda function for any tamely ramified Galois extension assuming the computation of the lambda function for a tamely ramified quadratic extension. In this paper we provide a formula for the tamely ramified quadratic case, thus completing the work in [@SABNT].
We should mention that in [@TS] Saito has computed the $\lambda$-function for an arbitrary extension assuming the residual characteristic of the base field is not equal to 2 (cf. Theorem on p. 508 of [@TS]). In Theorem II 2B on p. 508 of [@TS], when ramification index is [**even**]{}, Saito has computed the lambda functions for even degree extensions via the [**Legendre symbol**]{} and [**Hilbert symbol**]{}.
In this paper, we also compute this $\lambda$-functions for quadratic tamely ramified extensions. In our computation, we use the classical quadratic [**Gauss sums**]{} and these computations are different from the Saito’s result and explicit. The main idea for tamely ramified quadratic extension case is to reduce the $\lambda$-functions computation to the classical quadratic Gauss sums computations.
In this paper, as mentioned above, we compute lambda functions for quadratic tamely ramified extensions explicitly. This, together with the work in [@SABNT], yields an explicit computation of lambda functions for all tamely ramified extensions.
We now state the main theorem of this paper.
\[Theorem 3.21\] Let $K$ be a tamely ramified quadratic extension of $F/{\mathbb{Q}}_p$ with $q_F=p^s$. Let $\psi_F$ be the canonical additive character of $F$. Let $c\in F^\times$ with $-1=\nu_F(c)+d_{F/{\mathbb{Q}}_p}$, and $c'=\frac{c}{\text{Tr}_{F/F_0}(pc)}$, where $F_0/{\mathbb{Q}}_p$ is the maximal unramified extension in $F/{\mathbb{Q}}_p$. Let $\psi_{-1}=c'\cdot\psi_F$, then $$\lambda_{K/F}(\psi_F)=\Delta_{K/F}(c')\cdot\lambda_{K/F}(\psi_{-1}),$$ where $$\lambda_{K/F}(\psi_{-1})=\begin{cases}
(-1)^{s-1} & \text{if $p\equiv 1\pmod{4}$}\\
(-1)^{s-1}i^{s} & \text{if $p\equiv 3\pmod{4}$}.
\end{cases}$$ If we take $c=\pi_{F}^{-1-d_{F/{\mathbb{Q}}_p}}$, where $\pi_F$ is a norm for $K/F$, then $$\Delta_{K/F}(c')=\begin{cases}
1 & \text{if $\overline{\text{Tr}_{F/F_0}(pc)}\in k_{F_0}^{\times}=k_{F}^{\times}$ is a square},\\
-1 & \text{if $\overline{\text{Tr}_{F/F_0}(pc)}\in k_{F_0}^{\times}=k_{F}^{\times}$ is not a square}.
\end{cases}$$ Here “overline” stands for modulo $P_{F_0}$ and $c'\cdot \psi_F(x):=\psi_F(c'x)$ for all $x\in F$.
\[Remark 1.2\]
But in general, computation of $\lambda_{K/F}$, where $K/F$ is a wildly ramified quadratic extension, seems subtle. When $F={\mathbb{Q}}_2,$ in Example 3.4.14, pp. 60-63 of [@SABT], we have explicit computation for $\lambda_{K/{\mathbb{Q}}_2}$. In [@MW], one also can find some particular cases (cf. on pp. 252-255 of [@MW]). But if $F/{\mathbb{Q}}_2$ is an arbitrary finite extension and $K/F$ is [**quadratic**]{} extension, then computation of $\lambda_{K/F}$ is still [**open.**]{} And if $K$ is an abelian extension with $N_{K/F}(K^\times)={F^\times}^2$, we have the following theorem.
\[Theorem 3.26\] Let $F$ be an extension of ${\mathbb{Q}}_2$. Let $K$ be the abelian extension for which $N_{K/F}(K^\times)={F^\times}^2$. Then $\lambda_{K/F}=1$.
On pp. 7-8, Theorem \[Theorem 3.21\] and Theorem \[Theorem 3.26\] are proven.
**Notations and Preliminaries**
===============================
Let $F$ be a non-Archimedean local field of characteristic zero, i.e., a finite extension of the field $\mathbb{Q}_p$ (field of $p$-adic numbers), where $p$ is a prime. Let $O_F$ be the ring of integers in the local field $F$ and $P_F=\pi_F O_F$ is the unique prime ideal in $O_F$ and $\pi_F$ is a uniformizer, i.e., an element in $P_F$ whose valuation is one, i.e., $\nu_F(\pi_F)=1$. Let $q_F$ be the cardinality of the residue field $k_F$ of $F$. Let $U_F=O_F-P_F$ be the group of units in $O_F$. Let $P_{F}^{i}=\{x\in F:\nu_F(x){\geqslant}i\}$ and for $i{\geqslant}0$ define $U_{F}^i=1+P_{F}^{i}$ (with proviso $U_{F}^{0}=U_F=O_{F}^{\times}$). We also consider that $a(\chi)$ is the conductor of nontrivial character $\chi: F^\times\to \mathbb{C}^\times$, i.e., $a(\chi)$ is the smallest integer $m{\geqslant}0$ such that $\chi$ is trivial on $U_{F}^{m}$. We say $\chi$ is unramified if the conductor of $\chi$ is zero and otherwise ramified. Throughout the paper, when $K/F$ is unramified we choose uniformizers $\pi_K=\pi_F$. And when $K/F$ is ramified (both tame and wild) we choose uniformizers $\pi_F=N_{K/F}(\pi_K)$, where $N_{K/F}$ is the norm map from $K^\times$ to $F^\times$. In this paper $\Delta_{K/F}:=\det(\text{Ind}_{K/F}(1))$.
The conductor of any nontrivial additive character $\psi$ of the field $F$ is an integer $n(\psi)$ if $\psi$ is trivial on $P_{F}^{-n(\psi)}$, but nontrivial on $P_{F}^{-n(\psi)-1}$.
Local constant formula for character
------------------------------------
For a nontrivial multiplicative character $\chi$ of $F^\times$ and nontrivial additive character $\psi$ of $F$, we have (cf. [@JT1], p. 94): $$\label{eqn 2.5}
W(\chi,\psi)=\chi(c)q_F^{-a(\chi)/2}\sum_{x\in\frac{U_F}{U_{F}^{a(\chi)}}}\chi^{-1}(x)\psi(x/c),$$ where $c=\pi_{F}^{a(\chi)+n(\psi)}$.
\[Definition of canonical additive character\]
We define the non trivial additive character of $F$, $\psi_F:F\to\mathbb{C}^\times$ as the composition of the following four maps:
$F\xrightarrow{\mathrm{Tr}_{F/\mathbb{Q}_p}}\mathbb{Q}_p\xrightarrow{\alpha}\mathbb{Q}_p/\mathbb{Z}_p
\xrightarrow{\beta}\mathbb{Q}/\mathbb{Z}\xrightarrow{\gamma}\mathbb{C}^\times$,
where
1. $\mathrm{Tr}_{F/\mathbb{Q}_p}$ is the trace from $F$ to $\mathbb{Q}_p$,
2. $\alpha$ is the canonical surjection map,
3. $\beta$ is the canonical injection which maps $\mathbb{Q}_p/\mathbb{Z}_p$ onto the $p$-component of the divisible group $\mathbb{Q}/\mathbb{Z}$ and
4. $\gamma$ is the exponential map $x\mapsto e^{2\pi i x}$, where $i=\sqrt{-1}$.
For every $x\in\mathbb{Q}_p$, there is a rational $r$, uniquely determined modulo $1$, such that $x-r\in\mathbb{Z}_p$. Then $\psi_{{\mathbb{Q}}_p}(x)=\psi_{{\mathbb{Q}}_p}(r)=e^{2\pi i r}$. The nontrivial additive character $\psi_F=\psi_{{\mathbb{Q}}_p}\circ \rm{Tr}_{F/{\mathbb{Q}}_p}$ of $F$ is called the **canonical additive character** (cf. [@JT1], p. 92).
Classical Gauss sums
--------------------
Let $k_q$ be a finite field. Let $p$ be the characteristic of $k_q$; then the prime field contained in $k_q$ is $k_p$. The structure of the [**canonical**]{} additive character $\psi_q$ of $k_q$ is the same as the structure of the canonical (see the definition \[Definition of canonical additive character\]) character $\psi_F$, namely [**it comes by trace**]{} from the canonical character of the base field, i.e.,
$\psi_q=\psi_p\circ \text{Tr}_{k_q/k_p}$,
where
$\psi_p(x):=e^{\frac{2\pi i x}{p}}$ for all $x\in k_p$.
**Gauss sums:** Let $\chi, \psi$ be a multiplicative and an additive character respectively of $k_q$. Then the Gauss sum $G(\chi,\psi)$ is defined by $$G(\chi,\psi)=\sum_{x\in k_{q}^{\times}}\chi(x)\psi(x).$$ For computation of $\lambda_{K/F}$, where $K/F$ is a tamely ramified quadratic extension, we will use the following theorem.
\[Theorem 2.7\] Let $k_q$ be a finite field with $q=p^s$, where $p$ is an odd prime and $s\in\mathbb{N}$. Let $\chi$ be the quadratic character of $k_q$ and let $\psi$ be the canonical additive character of $k_q$. Then $$G(\chi,\psi)=\begin{cases}
(-1)^{s-1}q^{\frac{1}{2}} & \text{if $p\equiv 1\pmod{4}$},\\
(-1)^{s-1}i^sq^{\frac{1}{2}} & \text{if $p\equiv 3\pmod{4}$}.
\end{cases}$$
**Explicit computation of $\lambda_{K/F}$, where $K/F$ is a quadratic extension**
=================================================================================
Let $K/F$ be a quadratic extension of the field $F/{\mathbb{Q}}_p$. Let $G=\mathrm{Gal}(K/F)$ be the Galois group of the extension $K/F$. Let $t$ be the **ramification break or jump** (cf. [@JPS]) of the Galois group $G$ (or of the extension $K/F$). Then it can be proved that the conductor of $\omega_{K/F}$ (the quadratic character of $F^\times$ associated to $K$ by class field theory) is $t+1$. When $K/F$ is unramified we have $t=-1$, therefore the conductor of a quadratic character $\omega_{K/F}$ of $F^\times$ is zero, i.e., $\omega_{K/F}$ is unramified. And when $K/F$ is tamely ramified we have $t=0$, then $a(\omega_{K/F})=1$. In the wildly ramified case (which occurs if $p=2$) it can be proved that $a(\omega_{K/F})=t+1$ is, [**up to the exceptional case $t=2\cdot e_{F/{\mathbb{Q}}_2}$**]{}, always an **even number** which can be seen by the filtration of $F^\times$ (cf. p. 50 of [@SABT]).
**Computation of $\lambda_{K/F}$, where $K/F$ is a tamely ramified quadratic extension**
----------------------------------------------------------------------------------------
The existence of a tamely ramified quadratic character $\chi$ (which is not unramified) of a local field $F$ implies $p\ne2$ for the residue characteristic. Then
$F^\times/{F^\times}^2\cong V$
is isomorphic to Klein’s $4$-group. So we have only $3$ nontrivial quadratic characters in that case, corresponding to $3$ quadratic extensions $K/F$. One is unramified and other two are ramified. The unramified case is well settled. The two ramified quadratic characters determine two different quadratic ramified extensions of $F$.\
In the ramified case we have $a(\chi)=1$ because it is tame, and we take $\psi$ of conductor $-1$. Then we have $a(\chi)+n(\psi)=0$ and therefore in the formula of $W(\chi,\psi)$ (cf. equation (\[eqn 2.5\])) we can take $c=1$. So we obtain: $$\label{eqn 3.33}
W(\chi,\psi)=q_{F}^{-\frac{1}{2}}\sum_{x\in U_F/U_{F}^{1}}\chi^{-1}(x)\psi(x)=
q_{F}^{-\frac{1}{2}}\sum_{x\in k_{F}^{\times}}\chi'^{-1}(x)\psi'(x),$$ where $\chi'$ is the quadratic character of the residue field $k_{F}^{\times}$, and $\psi'$ is an additive character of $k_F$. When $n(\psi)=-1$, we observe that both [**the ramified characters $\chi$ give the same $\chi'$, hence the same $W(\chi,\psi)$**]{}, because one is different from other by a quadratic unramified character twist. To compute an explicit formula for $\lambda_{K/F}(\psi_{-1})$, where $K/F$ is a tamely ramified quadratic extension and $\psi_{-1}$ is an additive character of $F$ with conductor $-1$, we need to use [**classical quadratic Gauss sums**]{}.
Let $\psi_{-1}$ be an additive character of $F/{\mathbb{Q}}_p$ of conductor $-1$, i.e., $\psi_{-1}:F/P_F\to{\mathbb{C}}^\times$. Now restrict $\psi_{-1}$ to $O_F$, it will be one of the characters $a\cdot\psi_{q_F}$, for some $a\in k_{q_F}^{\times}$ and usually it will not be $\psi_{q_F}$ itself. Therefore, choosing $\psi_{-1}$ is very important and we have to choose $\psi_{-1}$ such a way that its restriction to $O_F$ is exactly $\psi_{q_F}$. Then we will be able to use the quadratic classical Gauss sum in the $\lambda$-function computation. We also know that there exists an element $c\in F^\times$ such that $$\label{eqn 3.34}
\psi_{-1}=c\cdot\psi_F$$ induces the canonical character $\psi_{q_F}$ on the residue field $k_F$.
Now question is: [**Finding proper $c\in F^\times$ for which $\psi_{-1}|_{O_F}=c\cdot\psi_F|_{O_F}=\psi_{q_F}$**]{}, i.e., the canonical character of the residue field $k_F$.
From the definition of conductor of the additive character $\psi_{-1}$ of $F$, we obtain from the construction (\[eqn 3.34\]) $$\label{eqn 3.35}
-1=\nu_F(c)+n(\psi_F)=\nu_F(c)+d_{F/{\mathbb{Q}}_p},$$ where $d_{F/{\mathbb{Q}}_p}$ is the exponent of the different $\mathcal{D}_{F/{\mathbb{Q}}_p}$. In the next two lemmas we choose the proper $c$ for our requirement.
\[Lemma 3.20\] Let $F/{\mathbb{Q}}_p$ be a local field and let $\psi_{-1}$ be an additive character of $F$ of conductor $-1$. Let $\psi_F$ be the canonical character of $F$. Let $c\in F^\times$ be any element such that $-1=\nu_F(c)+d_{F/{\mathbb{Q}}_p}$, and $$\label{eqn 3.36}
\text{Tr}_{F/F_0}(c)=\frac{1}{p},$$ where $F_0/{\mathbb{Q}}_p$ is the maximal unramified subextension in $F/{\mathbb{Q}}_p$. Then the restriction of $\psi_{-1}=c\cdot\psi_F$ to $O_F$ is the canonical character $\psi_{q_F}$ of the residue field $k_F$ of $F$.
Since $F_0/{\mathbb{Q}}_p$ is the maximal unramified subextension in $F/{\mathbb{Q}}_p$, we have $\pi_{F_0}=p$, and the residue fields of $F$ and $F_0$ are isomorphic, i.e., $k_{F_0}\cong k_{F}$, because $F/F_0$ is totally ramified extension. Then every element of $O_F/P_F$ can be considered as an element of $O_{F_0}/P_{F_0}$. Moreover, since $F_0/{\mathbb{Q}}_p$ is the maximal unramified extension, then from Proposition 2 of [@AW] on p. 140, for $x\in O_{F_0}$ we have $$\rho_{p}(\text{Tr}_{F_0/{\mathbb{Q}}_p}(x))=\text{Tr}_{k_{F_0}/k_{{\mathbb{Q}}_p}}(\rho_0(x)),$$ where $\rho_0,\,\rho_p$ are the canonical homomorphisms of $O_{F_0}$ onto $k_{F_0}$, and of $O_{{\mathbb{Q}}_p}$ onto $k_{{\mathbb{Q}}_p}$, respectively. Then for $x\in k_{F_0}$ we can write $$\label{eqn 3.50}
\text{Tr}_{F_0/{\mathbb{Q}}_p}(x)=\text{Tr}_{k_{F_0}/k_{{\mathbb{Q}}_p}}(x).$$ Furthermore, since $F/F_0$ is totally ramified, we have $k_{F}=k_{F_0}$, then the trace map for the tower of the residue fields $k_{F}/k_{F_0}/k_{{\mathbb{Q}}_p}$ is: $$\label{eqn 3.51}
\text{Tr}_{k_F/k_{{\mathbb{Q}}_p}}(x)=\text{Tr}_{k_{F_0}/k_{{\mathbb{Q}}_p}}\circ\text{Tr}_{k_F/k_{F_0}}(x)=\text{Tr}_{k_{F_0}/k_{{\mathbb{Q}}_p}}(x),$$ for all $x\in k_F$. Then from the equations (\[eqn 3.50\]) and (\[eqn 3.51\]) we obtain $$\label{eqn 3.52}
\text{Tr}_{F_0/{\mathbb{Q}}_p}(x)=\text{Tr}_{k_F/k_{{\mathbb{Q}}_p}}(x)$$ for all $x\in k_F$.
Since the conductor of $\psi_{-1}$ is $-1$, for $x\in O_F/P_F(=O_{F_0}/P_{F_0}$ because $F/F_0$ is totally ramified) we have $$\begin{aligned}
\psi_{-1}(x)
&=c\cdot\psi_F(x)
=\psi_{F}(cx)
=\psi_{{\mathbb{Q}}_p}(\text{Tr}_{F/{\mathbb{Q}}_p}(cx))
=\psi_{{\mathbb{Q}}_p}(\text{Tr}_{F_0/{\mathbb{Q}}_p}\circ\text{Tr}_{F/F_0}(cx))\\
&=\psi_{{\mathbb{Q}}_p}(\text{Tr}_{F_0/{\mathbb{Q}}_p}(x\cdot\text{Tr}_{F/F_0}(c)))\\
&=\psi_{{\mathbb{Q}}_p}(\text{Tr}_{F_0/{\mathbb{Q}}_p}(\frac{1}{p}x)), \quad\text{since $x\in O_F/P_F=O_{F_0}/P_{F_0}$ and $\text{Tr}_{F/F_0}(c)=\frac{1}{p}$}\\
&=\psi_{{\mathbb{Q}}_p}(\frac{1}{p}\text{Tr}_{F_0/{\mathbb{Q}}_p}(x)), \quad\text{because $\frac{1}{p}\in{\mathbb{Q}}_p$}\\
&=e^{\frac{2\pi i \text{Tr}_{F_0/{\mathbb{Q}}_p}(x)}{p}},\quad\text{because $\psi_{{\mathbb{Q}}_p}(x)=e^{2\pi i x}$}\\
&=e^{\frac{2\pi i\text{Tr}_{k_F/k_{{\mathbb{Q}}_p}}(x)}{p}},\quad \text{using equation $(\ref{eqn 3.52})$}\\
&=\psi_{q_F}(x).
\end{aligned}$$ This competes the lemma.
The next step is to produce good elements $c$ more explicitly. By using Lemma \[Lemma 3.20\], in the next lemma we see more general choices of $c$.
\[Lemma 3.21\] Let $F/{\mathbb{Q}}_p$ be a tamely ramified local field and let $\psi_{-1}$ be an additive character of $F$ of conductor $-1$. Let $\psi_F$ be the canonical character of $F$. Let $F_0/{\mathbb{Q}}_p$ be the maximal unramified subextension in $F/{\mathbb{Q}}_p$. Let $c\in F^\times$ be any element such that $-1=\nu_F(c)+d_{F/{\mathbb{Q}}_p}$, then
$c'=\frac{c}{\text{Tr}_{F/F_0}(pc)}$,
fulfills conditions (\[eqn 3.35\]), (\[eqn 3.36\]), and hence $\psi_{-1}|_{O_F}=c'\cdot\psi_F|_{O_F}=\psi_{q_F}$.
By the given condition we have $\nu_{F}(c)=-1-d_{F/{\mathbb{Q}}_p}=-1-(e_{F/{\mathbb{Q}}_p}-1)=-e_{F/{\mathbb{Q}}_p}$. Then we can write $c=\pi_{F}^{-e_{F/{\mathbb{Q}}_p}}u(c)=p^{-1}u(c)$ for some $u(c)\in U_F$ because $F/{\mathbb{Q}}_p$ is tamely ramified, hence $p=\pi_{F}^{e_{F/{\mathbb{Q}}_p}}$. Then we can write $$\text{Tr}_{F/F_0}(pc)=p\cdot\text{Tr}_{F/F_0}(c)=p\cdot p^{-1}u_0(c)=u_0(c)\in U_{F_0}\subset U_{F},$$ where $u_0(c)=\text{Tr}_{F/F_0}(u(c))$, hence $\nu_{F}(\text{Tr}_{F/F_0}(pc))=0$. Then the valuation of $c'$ is:
$\nu_F(c')=\nu_F(\frac{c}{\text{Tr}_{F/F_0}(pc)})=\nu_F(c)-\nu_F(\text{Tr}_{F/F_{0}}(pc))$\
$=\nu_F(c)-0=\nu_{F}(c)=-1-d_{F/{\mathbb{Q}}_p}$.
Since $\text{Tr}_{F/F_0}(pc)=u_0(c)\in U_{F_0}$, we have
$\text{Tr}_{F/F_0}(c')=\text{Tr}_{F/F_0}(\frac{c}{\text{Tr}_{F/F_0}(pc)})
=\frac{1}{\text{Tr}_{F/F_0}(pc)}\cdot\text{Tr}_{F/F_0}(c)=\frac{1}{p\cdot\text{Tr}_{F/F_0}(c)}\cdot\text{Tr}_{F/F_0}(c)=\frac{1}{p}$.
Thus we observe that here $c'\in F^\times$ satisfies equations (\[eqn 3.35\]) and (\[eqn 3.36\]). Therefore, from Lemma \[Lemma 3.20\] we can see that $\psi_{-1}|_{O_F}=c'\cdot\psi_{F}|_{O_F}$ is the canonical additive character of $k_F$.
By Lemmas \[Lemma 3.20\] and \[Lemma 3.21\] we get many good (in the sense that $\psi_{-1}|_{O_F}=c\cdot\psi_{F}|_{O_F}=\psi_{q_F}$) elements $c$ which we will use in our next theorem to calculate $\lambda_{K/F}$, where $K/F$ is a tamely ramified quadratic extension.
From [@BH], p. 190, part (2) of the Proposition, we have $$\lambda_{K/F}(\psi_{-1})=\lambda_{K/F}(c'\psi_{F})=\Delta_{K/F}(c')\cdot \lambda_{K/F}(\psi_F).$$ Since $\Delta_{K/F}$ is quadratic, we can write $\Delta_{K/F}=\Delta_{K/F}^{-1}$. So we obtain $$\lambda_{K/F}(\psi_F)=\Delta_{K/F}(c')\cdot\lambda_{K/F}(\psi_{-1}).$$ Now we have to compute $\lambda_{K/F}(\psi_{-1})$, and which we do in the following:\
Since $[K:F]=2$, we have $\text{Ind}_{K/F}(1)=1_F\oplus\omega_{K/F}$. The conductor of $\omega_{K/F}$ is $1$ because $K/F$ is a tamely ramified quadratic extension, and hence $t=0$, so $a(\omega_{K/F})=t+1=1$. Therefore, we can consider $\omega_{K/F}$ as a character of $F^\times/U_{F}^{1}$. So the restriction of $\omega_{K/F}$ to $U_{F}$, $\text{res}(\omega_{K/F}):=\omega_{K/F}|_{U_F}$, we may consider as the uniquely determined character of $k_{F}^{\times}$ of order $2$. Since $c'$ satisfies equations (\[eqn 3.35\]), (\[eqn 3.36\]), then from Lemma \[Lemma 3.21\] we have $\psi_{-1}|_{O_F}=c'\cdot\psi_F|_{O_F}=\psi_{q_F}$, and this is the canonical character of $k_F$. Then from equation (\[eqn 3.33\]) we can write $$\begin{aligned}
\lambda_{K/F}(\psi_{-1})
&=q_{F}^{-\frac{1}{2}}\sum_{x\in k_{F}^{\times}}\text{res}(\omega_{K/F})(x)\psi_{q_F}(x)\\
&=q_{F}^{-\frac{1}{2}}\cdot G(\text{res}(\omega_{K/F}),\psi_{q_F}).\end{aligned}$$ Moreover, by Theorem \[Theorem 2.7\] we have $$G(\text{res}(\omega_{K/F}),\psi_{q_F})=\begin{cases}
(-1)^{s-1}q_{F}^{\frac{1}{2}} & \text{if $p\equiv 1\pmod{4}$}\\
(-1)^{s-1}i^{s}q_{F}^{\frac{1}{2}} & \text{if $p\equiv 3\pmod{4}$}.
\end{cases}$$ By using the classical quadratic Gauss sum we obtain $$\lambda_{K/F}(\psi_{-1})=\begin{cases}
(-1)^{s-1} & \text{if $p\equiv 1\pmod{4}$}\\
(-1)^{s-1}i^{s} & \text{if $p\equiv 3\pmod{4}$}.
\end{cases}$$
We also can write $\Delta_{K/F}=\det(\text{Ind}_{K/F}(1))=\det(1_F\oplus \omega_{K/F})=\omega_{K/F}.$ So we have $$\Delta_{K/F}(\pi_F)=\omega_{K/F}(\pi_F)=1,$$ because $\pi_F\in N_{K/F}(K^\times)$.
Under the assumption of the Theorem \[Theorem 3.21\] we have $\pi_F\in N_{K/F}(K^\times)$, $\Delta_{K/F}=\omega_{K/F}$ and $c'=\frac{c}{\text{Tr}_{F/F_0}(pc)}$, where $c\in F^\times$ with $\nu_F(c)=-1-d_{F/{\mathbb{Q}}_p}$. Then we can write $$\begin{aligned}
\Delta_{K/F}(c')
=\omega_{K/F}(c')
&=\omega_{K/F}\left(\frac{c}{\text{Tr}_{F/F_0}(pc)}\right)\\
&=\omega_{K/F}\left(\frac{\pi_{F}^{-e_{F/{\mathbb{Q}}_p}}u(c)}{u_0(c)}\right),
\quad\text{where $c=\pi_{F}^{-e_{F/{\mathbb{Q}}_p}}u(c)$, $\text{Tr}_{F/F_0}(pc)=u_0(c)\in U_{F_0}$}\\
&=\omega_{K/F}(\pi_{F}^{-e_{F/{\mathbb{Q}}_p}})\omega_{K/F}(v),\quad\text{where $v=\frac{u(c)}{u_0(c)}\in U_F$}\\
&=\omega_{K/F}(x)\\
&=\begin{cases}
1 & \text{when $x$ is a square element in $k_{F}^{\times}$}\\
-1 & \text{when $x$ is not a square element in $k_{F}^{\times}$},
\end{cases}\end{aligned}$$ where $v=xy$, with $x=x(\omega_{K/F},c)\in U_{F}/U_{F}^{1}$, and $y\in U_{F}^{1}$.
In particular, if we choose $c$ such a way that $u(c)=1$, i.e., $c=\pi_{F}^{-1-d_{F/{\mathbb{Q}}_p}}$, then we have $\Delta_{K/F}(c')=\Delta_{K/F}(\text{Tr}_{F/F_0}(pc)).$ Since $\text{Tr}_{F/F_0}(pc)\in O_{F_0}$ is a unit and $\Delta_{K/F}=\omega_{K/F}$ induces the quadratic character of $k_{F}^{\times}=k_{F_0}^{\times}$, then for this particular choice of $c$ we obtain $$\Delta_{K/F}(c')=\begin{cases}
1 & \text{if $\overline{\text{Tr}_{F/F_0}(pc)}$ is a square in $k_{F_0}^{\times}$}\\
-1 & \text{if $\overline{\text{Tr}_{F/F_0}(pc)}$ is not a square in $k_{F_0}^{\times}$}.\\
\end{cases}$$
**Computation of $\lambda_{K/F}$, where $K/F$ is a wildly ramified extension**
------------------------------------------------------------------------------
In the case $p=2$, the square class group of $F$, i.e., $F^\times/{F^\times}^2$ can be very large (cf. Theorem 2.29 on p. 165 of [@TYM]), so we can have many quadratic characters but they are wildly ramified, not tame. In Remark \[Remark 1.2\], we mention the current status of the this wildly ramified quadratic case.
Let $G=\mathrm{Gal}(K/F)$. From Theorem 2.29 on p. 165 of [@TYM], if $F/{\mathbb{Q}}_2$, we have $|F^\times/{F^\times}^2|=2^m,\,(m{\geqslant}3)$ and the $2$-rank of $G$ (i.e., the dimension of $G/G^2$ as a vector space over ${\mathbb{F}}_2$) $\text{rk}_2(G)\ne 1$ and $G$ is not metacyclic. Then from Bruno Kahn’s result, the second Stiefel-Whitney class $s_2(\text{Ind}_{K/F}(1))=0$ (cf. Theorem 1 of [@BK]). Since $s_2(\text{Ind}_{K/F}(1))=0$, the Deligne constant $c(\text{Ind}_{K/F}(1))=1$ (cf. Theorem 3 on p. 129 of [@JT1]). Again since here $\text{rk}_2(G)\ne 1$ and $G$ is not metacyclic, we have $\Delta_{K/F}\equiv 1$. Therefore, we can conclude that $$\lambda_{K/F}(\psi)=c(\text{Ind}_{K/F}(1))\cdot W(\Delta_{K/F},\psi)=1,$$ where $\psi$ is a nontrivial additive character of $F$.
\[Example wild\]
In this case, we have (cf. pp. 60-63 of [@SABT]): $$\lambda_{{\mathbb{Q}}_2(\sqrt{5})/{\mathbb{Q}}_2}=1, \lambda_{{\mathbb{Q}}_2(\sqrt{-1})/{\mathbb{Q}}_2}=i, \lambda_{{\mathbb{Q}}_2(\sqrt{-5})/{\mathbb{Q}}_2}=i,
\lambda_{{\mathbb{Q}}_2(\sqrt{2})/{\mathbb{Q}}_2}=1,$$ $$\lambda_{{\mathbb{Q}}_2(\sqrt{10})/{\mathbb{Q}}_2}=-1,
\lambda_{{\mathbb{Q}}_2(\sqrt{-2})/{\mathbb{Q}}_2}=i, \lambda_{{\mathbb{Q}}_2(\sqrt{-10})/{\mathbb{Q}}_2}=-i.$$
**Acknowledgements.** I would like to thank Prof. E.-W. Zink, Humboldt University, Berlin for suggesting this problem and his constant valuable advice and comments. I express my gratitude to the referee for his/her valuable comments and suggestions for the improvement of the paper.
[99]{}
S.A. Biswas, Local Constants for Galois Representations - Some Explicit Results, Ph.D. Thesis, <http://arxiv.org/pdf/1603.06089.pdf>.
S.A. Biswas, Computation of the Lambda function for a finite Galois extension, Journal of Number Theory, Volume 186, May 2018, Pages 180-201, <https://doi.org/10.1016/j.jnt.2017.09.025.>
C.J. Bushnell, G. Henniart, The local Langlands conjecture for $GL(2)$, Springer-Verlag, 2006.
J. W. Cogdell, L-functions and converse theorems for $GL_n$, IAS/Park City Mathematics Series, Volume [**12**]{}, 2002, pp. 97-177.
G. Henniart, Charactérisation de la correspondence de Langlands locale par les facteurs $\epsilon$ de paires, Invent. Math [**113**]{}(1993), no. 2, 339-350.
B. Kahn, La deuxième classe de Stiefel-Whitney d’une représentation régulière. I, C. R. Acad. Sci. Paris Sér. I Math. [**297**]{} (1983), no. 6, 313-316.
T.Y. Lam, Introduction to quadratic forms over fields, Graduate Studies in Mathematics, Volume [**67**]{}, AMS, Providence, Rhode Island, 2004.
R.P. Langlands, On the functional equation of the Artin $L$-functions, unpublished article, <https://publications.ias.edu/sites/default/files/a-ps.pdf>.
R. Lidl, H. Niederreiter, Finite fields, Encyclopedia of Mathematics and its applications, Cambridge University press 2000.
C.J. Moreno, A. Wan, Unusual applications of quadratic Gaussian sums, DIMACA Series in Discrete Mathematics and Theoretical Computer Science, Volume [**64**]{}, 2004, pp. 227-264.
T. Saito, Local constant of $\text{Ind}_{K}^{L}1$, Comment. Math. Helvetici [**70**]{} (1995), 507-515.
J-P. Serre, Local field, Springer-Verlag, 1979.
J. Tate, Local Constants, Algebraic Number Fields (L-functions and Galois properties), in: A. Fröhlich (Ed.), Proceedings of Symposium, 1977, pp. 89-131.
A. Weil, Basic number theory, Third edition, Springer-Verlag, 1974.
|
---
abstract: 'We present a concept for control of the ion polarization, called a transparent spin method. The spin transparency is achieved by designing such a synchrotron structure that the net spin rotation angle in one particle turn is zero. The polarization direction of any ions including deuterons can be efficiently controlled using weak quasi-static fields. These fields allow for dynamic adjustment of the polarization direction during an experiment. The main features of the Transparent Spin method are illustrated in a figure-8 collider. The results are relevant to the Electron-Ion Collider considered in the US, the ion-ion collider NICA constructed in Russia, and a polarized Electron-ion collider planned in China.'
author:
- 'Yu.N. Filatov'
- 'A.M. Kondratenko'
- 'M.A. Kondratenko'
- 'Ya.S. Derbenev'
- 'V.S. Morozov'
title: Transparent Spin Method for Spin Control of Hadron Beams in Colliders
---
*Introduction.* — Polarized beam experiments have been and remain a crucial tool in understanding particle and nuclear structure and reactions from the first principles [@b:principles]. In particular, polarized light ion ($p$, $d$, ${}^3He$) and electron beams are necessary for the successful operation of a proposed high-luminosity polarized *Electron-Ion Collider* (EIC) that is currently under active design [@b:EIC_MEIC; @b:EIC_eRHIC; @b:EIC_China]. Technologies for production of polarized light ions already exist or will be operational in the near future at several colliders worldwide [@b:IonSource].
After a polarized beam is generated by a source, the first task is to preserve the beam’s polarization in the process of its acceleration to the final energy. The next task is to maintain the polarization during a long-term experimental run. There are presently complete and efficient solutions to these tasks for future colliders under design and consideration. In general, they are based on using the *Siberian Snake* [@b:AccSnake1; @b:AccSnake2; @b:SYLee] and *figure-8 orbit* techniques [@b:AccFig8]. Both techniques eliminate depolarizing *spin resonances* [@b:SYLee] by making the spin tune energy independent. Presently, polarization preservation using helical snakes [@b:RHICSnakes] has been successfully demonstrated up to high energies in a proton-proton collider, RHIC. Other techniques for suppressing depolarization during acceleration have also been proposed, for example, by employing high super-periodicity in the design of a synchrotron [@b:HighPeriodicity].
A typical task of a collider setup is adjustment of the required polarization direction at the interaction point (IP). In RHIC, *fixed* longitudinal polarization at a specific IP is set using a pair of *spin rotators* turning the spins by $\pm 90^\circ$ [@b:RHICSnakes]. Another task is change of the polarization direction during an experiment and, in particular, frequent flips of the polarization to minimize systematic errors. Experimentally verified spin-flipping schemes are based on adiabatically sweeping an RF magnet’s frequency through an induced spin resonance [@b:RFflip]. This technique is used in RHIC for spin flipping with an efficiency of 97% in the energy range of 24 to 255 GeV [@b:RFflipRHIC]. However, every single crossing of an RF resonance causes some polarization loss that limits the admissible number of spin flips during an experiment.
This paper presents a new method for ion polarization control called a *Transparent Spin* (TS) technique. This technique allows one to first preserve the polarization during acceleration and maintain it in the collider mode. It then enables flexible and efficient manipulation of the polarization direction of any particle species. The polarization can be adjusted to any direction at any orbital location during the whole time of an experiment. Such an adjustment causes practically no polarization loss.
The ideas of the TS method were first formulated in the design process of the figure-8 booster and collider synchrotrons of the JLEIC project [@b:baseJLEIC]. They were further developed and applied in the design of the racetrack rings of the NICA hadron collider project [@b:NICA1].
*Transparent spin concept.* — According to the basic theorem about the spin motion of a particle moving along an arbitrary periodic closed orbit [@b:axisN], there always exists a periodic spin axis $\vec{n}(z)=\vec{n}(z+C)$ such that the spin motion can be in general represented as precession about $\vec{n}(z)$ with a spin tune $\nu$. Here $z$ is the distance along the closed orbit, $C$ is the orbit circumference, and $2\pi\nu$ is the phase advance of the spin precession per particle turn. The $\vec{n}(z)$ axis is unique if the spin precession frequency is not a harmonic of the particle circulation frequency. Due to the spin tune spread associated with the beam emittances, the resulting beam polarization is established along $\vec{n}$ [@b:AlongAxisN; @b:Derbenev1973polkin].
The TS method is based on designing such a magnetic structure of a synchrotron that the net effect of the synchrotron elements on the spin for motion along the design closed orbit is compensated over a single particle turn. The magnetic lattice of the synchrotron becomes “transparent” to the spin. A natural example of such a structure is a flat figure-8 ring. In “transparent” structures, the spin motion becomes degenerate: any spin direction at any orbital location repeats every particle turn. This means that the particles are in a *TS resonance* with $\nu=0$. The spin motion in such a situation is highly sensitive to small perturbations of the magnetic fields along the orbit. At the same time, this sensitivity allows one to implement a simple and efficient spin control system using a *spin navigator* (SN). An SN is a flexible device consisting of elements with weak *constant* or *quasi-stationary* magnetic fields rotating the spins about a desirable direction $\vec{n}_N$ by a small angle $2\pi\nu_N$. Such a navigator has practically no effect on the orbital beam dynamics [@b:SmallSol; @b:smallField].
In an ideal synchrotron lattice, the stable polarization axis $\vec{n}$ in the straight housing the SN coincides with the SN axis $\vec{n}_N$ and the resulting spin tune $\nu$ equals $\nu_N$. Clearly, to control the spin, one must use a sufficiently strong SN to dominate over the effect of the small perturbative fields around a particle’s trajectory. There are two sources of these fields: lattice imperfections and focusing magnetic and bunching RF fields experienced by particles during their free transverse and longitudinal (betatron and synchrotron) oscillations.
Let us illustrate the TS method using a figure-8 orbit as an example.
*Spin motion in a figure-8 synchrotron.* — In an ideal figure-8 synchrotron, rotation of the spins in one arc is compensated by an opposite rotation in the other arc. When a particle is moving on a flat design orbit, the spin tune is zero for any particle energy, so the spin motion is degenerate.
When a particle’s trajectory deviates from the design orbit, the particle experiences perturbing magnetic fields. The effect of these fields on the spin in a large number of particle turns can be expressed in terms of the *TS resonance field* $\vec{\omega}$ (TS-field). In the absence of SNs, the direction of $\vec{\omega}$ defines the direction of the stable spin axis $\vec{n}=(\vec{\omega}/\omega)$ while its absolute value, or the *TS resonance strength*, $\omega=|\vec{\omega}|$ defines the particle’s spin tune: $\nu=\omega$, i.e. the spin completes a full rotation about the $\vec{n}$ axis in $1/\omega$ particle turns around the orbit [@b:EIC_MEIC]. The TS resonance spin field is a 3D extension of the conventional 2D resonance strength concept for cases when the spin motion is degenerate. Conventional codes, such as DEPOL [@b:DEPOL], are valid for synchrotrons with a distinct polarization direction and calculate the spin field component transverse to this direction.
To set the required polarization direction, an SN is inserted into the synchrotron lattice. For example, a weak solenoid inserted into an experimental straight stabilizes the longitudinal polarization direction in that straight.
*Strength of the TS resonance.* — The TS-field $\vec{\omega}$ consists of two parts: the coherent field associated with the closed orbit distortion caused by lattice imperfections and the incoherent field associated with the beam emittances.
With *fixed* element alignment errors, the field of a synchrotron is constant and does not change during an experiment. Since alignment errors are random, our analysis uses an rms strength obtained statistically assuming independent distributions of element misalignments [@b:EIC_MEIC; @b:statModel]. The magnitude of is non-zero only in the second order of the particle’s oscillation amplitudes and is proportional to the beam emittances. In practice, significantly exceeds [@b:EIC_MEIC].
Figure \[f:W\_JLEIC\] shows an analytic calculation of and for deuterons and protons as functions of the beam momentum for JLEIC’s figure-8 collider ring [@b:AccJLEIC]. is calculated assuming random uncorrelated transverse shifts of all quadrupoles giving an rms closed orbit excursion of 100 $\mu$m. The calculation of assumes normalized transverse emittances of 1 mm$\cdot$mrad.
![\[f:W\_JLEIC\] TS resonance strengths for deuterons and protons in JLEIC versus the beam momentum.](W_JLEIC_deut "fig:"){width="\columnwidth"}\
![\[f:W\_JLEIC\] TS resonance strengths for deuterons and protons in JLEIC versus the beam momentum.](W_JLEIC_prot "fig:"){width="\columnwidth"}
The deuteron grows with momentum but does not exceed $3\cdot 10^{-5}$. The proton is $ 10^{-4}$ – $10^{-3}$ in the whole momentum range with exception of narrow interference peaks where the spin effects of the arc magnets add up coherently. In both cases, is about two orders of magnitude lower than in the whole momentum range.
*Polarization control condition.* — In the presence of a navigator, the particle spins are precessing about an effective spin field $\vec{h}$ consisting of the navigator and TS fields [@b:axisN]: $$\vec{h}(z)=\nu_N \,\vec{n}_N(z)+\vec{\omega}(z)\,.$$ At the SN location, $\vec{n}_N(z)$ coincides with the SN’s spin rotation axis. In the rest of the ring, $\vec{n}_N(z)$ evolves according to the Thomas-BMT equation.
To stabilize and control the polarization, the navigator strength must significantly exceed the TS resonance strength (the control condition) $$\nu_N \gg \omega \,.$$
Polarization then points along the navigator field direction $\vec{n}_N$. The angle of uncontrolled deviation of the polarization from $\vec{n}_N$ is of the order of $\omega/\nu_N$.
In the above example, depending on the beam momentum, the SN strength must be large compared to for deuterons and $10^{-4}$–$10^{-2}$ for protons to control their polarization.
*Acceleration in TS mode.* — Note that $\vec{h}$ is a function of energy. Typically, during acceleration, $\vec{h}$ changes adiabatically slowly, i.e. in a characteristic time of the field change, the spin makes a large number of turns about the field. In that case, the spin initially oriented along $\vec{h}$ follows it during the whole acceleration process and polarization is preserved. Since, in figure-8 accelerators as well as in accelerators with two snakes, the spin tune does not explicitly depend on energy, synchrotron oscillations have practically no effect on the spin dynamics. We showed analytically and numerically that, in the presence of an appropriate spin navigator, effects of imperfections, coupling, betatron and synchrotron oscillations can be neglected [@b:EIC_MEIC].
Figure \[f:Sz\_deut\] shows the longitudinal spin component of a deuteron during acceleration in JLEIC [@b:AccJLEIC]. The simulation is done using a spin tracking code Zgoubi [@b:Zgoubi]. The simulation parameters are chosen based on a prediction of the same model that is used to analytically calculate the TS-resonance strength in Fig. \[f:W\_JLEIC\]. The field ramp rate is 3 T/min. The longitudinal polarization direction is stabilized by a solenoid with a maximum field integral of about 7 T$\cdot$m. Such a solenoid provides a navigator strength of $\nu_N=3\cdot 10^{-3}$ and the control condition is satisfied with a large margin. The change in the longitudinal spin component does not exceed a value of $2\cdot 10^{-5}$ as shown in Fig. \[f:Sz\_deut\].
![\[f:Sz\_deut\] Longitudinal spin component of a deuteron as a function of the beam momentum during acceleration in JLEIC.](Sz_deut){width="\columnwidth"}
When accelerating protons in the same lattice, the same solenoid provides an SN strength of . Figure \[f:Sz\_prot\] shows the longitudinal spin components of three protons with $\Delta p/p=0$ (green line), $\Delta p/p=10^{-3}$ (red line) and $\Delta p/p=-10^{-3}$ (blue line). The graphs of the longitudinal spin components practically do not differ from each other (the red line covers up the blue and green lines), i.e. the synchrotron energy modulation has no noticeable effect on the ion spin motion.
![\[f:Sz\_prot\] Longitudinal spin component during acceleration of three protons in JLEIC.](Sz_prot){width="\columnwidth"}
The control condition in Fig. \[f:Sz\_prot\] is met almost everywhere except for narrow energy regions of the interference peaks (see Fig. \[f:W\_JLEIC\]) where the navigator strength is comparable to . However, crossing of these peaks during acceleration does not depolarize the beam. As illustrated in Fig. \[f:Sz\_prot\], they only cause coherent deviation of the spins from the longitudinal direction by angles determined by both the navigator and the coherent part of the TS-resonance strengths. The beam polarization restores its longitudinal direction in places where the control condition is again satisfied. This indicates that, in this example, the change in spin occurs adiabatically during acceleration. These tracking results agree well with the analytic calculation of the TS resonance strength in Fig. \[f:W\_JLEIC\].
This example suggests that if the goal is to simply preserve the polarization degree, the spin control condition can be relaxed to a weaker one of $\nu_N \gg {\mbox{$\omega _\text{emitt}$}}$. Direction control can be restored by accounting for in the SN setting. can be measured experimentally.
*Spin navigators.* — Spin navigators can be technically realized in different ways using longitudinal and transverse fields [@b:baseJLEIC; @b:smallField]. As an example, let us consider a spin navigator design based on weak solenoids, which have no effect on the closed orbit.
A single solenoid stabilizes the longitudinal polarization direction at its location. One can set any polarization direction in the horizontal plane ($xz$) at the IP by introducing a second solenoid into the collider’s lattice. Figure \[f:SN\_2sol\] shows a schematic of such a spin navigator where the two solenoids are separated by an arc dipole [@b:SmallSol].
![\[f:SN\_2sol\] Schematic of a spin navigator for control of the ion polarization in the collider’s plane using small solenoids.](SN_2sol_2){width="0.9\columnwidth"}
One can similarly design a 3D spin navigator for control of all three polarization direction components [@b:baseJLEIC].
*TS collider features.* — Let us briefly formulate some of the new capabilities that become available when operating polarized beams in the TS mode.
[**1.**]{} Acceleration of the polarized beams. — Energy independence of the spin tune in figure-8 TS synchrotrons and racetracks with two snakes allows for acceleration of beams without polarization loss. Besides, the figure-8 design solves a serious problem of accelerating polarized $p$ and ${}^3 He$ beams in the booster energy range and polarized deuteron beams to the EIC energies, where full Siberian Snakes are technically challenging [@b:AccJLEIC].
[**2.**]{} Long-term polarization maintenance. — One of the critical issues for any collider is a potential depolarization caused by higher order stationary spin resonances associated, in particular, with the beam-beam space charge interaction in the presence of a large spin tune spread. A long-term polarization maintenance was experimentally demonstrated in RHIC with two Siberian Snakes, where the polarization is sustained for many hours [@b:pol_RHIC]. This success is achieved due to compensation of the spin tune spread by the snakes and a fine selection of the betatron tunes. Accordingly, with a proper tuning of the orbital oscillations, one should observe a similar stabilizing effect of the figure-8 configuration for all polarized ion species of an EIC.
[**3.**]{} Polarization control. — An SN allows for flexible manipulation of the polarization direction not only at the IP but at any orbital location. It is also an efficient instrument for matching of the required polarization direction at injection of the beam into any of the synchrotrons of the collider complex. In addition, an SN simplifies the polarimetry by allowing adjustment of an optimal polarization orientation at a polarimeter.
[**4.**]{} On-line monitoring of the polarization. — The TS mode gives a unique opportunity of quickly determining the polarization during an experiment, or on-line monitoring of the polarization. When adiabatically manipulating the polarization, its value is preserved at a high precision. The polarization vector is then a function of the SN magnetic fields and can be monitored using their set values. The relative accuracy $\Delta$ of the polarization direction obtained this way is determined by the ratio of the TS resonance strength to the SN strength: $\Delta\sim\omega/\nu_N \ll 1$.
[**5.**]{} Spin flipping. — Spin flipping can be realized using a couple of SN solenoids that allow one to simultaneously control the polarization direction as well as the spin tune. The spin tune can be maintained constant during a spin flip thus avoiding any possibility of crossing the TS resonance or any of the higher-order spin resonances, which prevents polarization loss. To preserve the polarization degree during spin manipulation, the spin direction must change adiabatically. This condition can be specified as: $\tau \gg 1/\nu$, where $\tau$ is the number of particle turns necessary to flip the spin. For JLEIC, the flip time requirements are ms for protons and $t_d\gg 0.1$ s for deuterons [@b:JLEICflip].
[**6.**]{} Ultra-high precision experiments. — It may be of interest to consider compensation of the coherent part of the TS resonance strength at a selected energy using an SN. Furthermore, an appropriate design of the magnetic lattice may significantly suppress the emittance-dependent part as well. This may create new opportunities for ultra-high precision experiments with polarized beams in synchrotrons such as search for a permanent electric dipole moment of charged spinning particles.
*TS racetrack with two identical Siberian snakes.* — The TS mode of polarized colliding beams can also be realized in a synchrotron with a racetrack orbit geometry in its entire energy range by installing two identical Siberian Snakes in the opposite straights of the collider. For example, it is planned to use two solenoidal snakes in the NICA collider to operate it in the TS mode with polarized protons and deuterons in the momentum range of up to 13.5 GeV/c [@b:NICA2].
*TS mode in high-energy hadron colliders.* — The TS mode can also be efficiently used in high energy polarized $pp$ colliders such as RHIC, LHC and future ultra-high energy projects. Earlier studies have shown that, using a system of many snakes around a collider ring [@Derbenev:1983tt; @Derbenev:1988tt; @b:SSC] or, alternatively, using spin-compensated quadrupoles, one could stabilize the coherent spin to hundreds of TeV and higher energies [@b:spin_comp_quad]. Such systems can be adjusted to accommodate the TS spin dynamics with its advantages for spin control and manipulation discussed above.
*Proof-of-principle TS experiment in RHIC.* — RHIC is currently the only operating collider allowing experiments with polarized protons. In normal operation, the axes of its two helical snakes are orthogonal to each other and the spin tune is a half-integer [@b:RHICSnakes]. The snake design allows one to change the snake axis orientations by adjusting the currents of helical magnet pairs. To experimentally test the TS mode, we plan to configure the snake axes to have the same orientations [@b:testRHIC]. Operation of RHIC in the TS mode may expand its capabilities for polarized proton experiments.
*TS mode at an integer spin resonance.* — One may consider the possibility of arranging the TS mode at an integer spin resonance $\nu=\gamma G=k$ (an imperfection resonance) in a conventional racetrack synchrotron even without snakes. However, in contrast to the above examples, the energy dependence of the spin tune limits the efficacy of the TS mode at integer resonances. Nevertheless, integer resonances may be used to test and tune polarization manipulation devices (i.e. spin navigators) in the TS mode at low energies available at conventional synchrotrons such as COSY (Julich), AGS (Brookhaven), and Nuclotron (Dubna). This mode of operation is also of interest for the EIC at BNL for manipulation of the deuteron polarization direction using spin navigators near the points of integer spin resonances, which occur about every 13 GeV. However, this question requires further study.
*Conclusions.* — We presented a concept of a Transparent Spin method for control of hadron polarization in colliders and identified its main features. The Transparent Spin mode in figure-8 colliders and in racetracks with two identical snakes allows one to:
- Control of the polarization by weak magnetic fields not affecting the orbital dynamics,
- Accelerate the beam without polarization loss,
- Maintain stable polarization during an experiment,
- Set any required polarization direction at any orbital location in a collider,
- Change the polarization direction using a spin navigator during an experiment,
- Monitor the polarization on-line during an experiment,
- Do frequent coherent spin flips of the beam to reduce experiment’s systematic errors,
- Carry out ultra-high precision experiments.
The TS mode allows one to significantly expand the capabilities of polarized beam experiments at the EIC in the US, NICA in Russia, EicC in China, and other future facilities. It makes it possible for the RHIC-based EIC to manipulate the polarization during experiments in the whole energy ranges for protons and ${}^3He$ as well as for deuterons near energies corresponding to integer spin resonances. Solenoidal snakes of the NICA collider allow for operation with polarized protons and deuterons in the whole momentum range of up to 13.5 GeV/c.
A proof-of-principle experimental test of the Transparent Spin mode is currently in preparation. Its goal is to validate the TS concept as a new effective tool in addition to existing polarized beam accelerator techniques.
This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics under contract DE-AC05-06OR23177.
[99]{} A. Accardi *et al*., Eur. Phys. J. **A52**(9), 268 (2016). A.M. Kondratenko *et al*., arXiv:1604.05632 (2016).
E.C. Aschenauer *et al*., arXiv:1409.1633 (2014).
X. Chen, arXiv:1809.00448 (2018).
A. Sy, ICFA-BD Newsletter **74**, 99 (2018).
Ya.S. Derbenev and A.M. Kondratenko, Sov. Phys. Doklady **20**, 562 (1976).
Ya.S. Derbenev and A.M. Kondratenko, AIP Conf. Proc. **51** (1979), p. 292.
S.Y. Lee, *Spin Dynamics and Snakes in Synchrotrons* (World Scientific, Singapore, 1997).
Ya.S. Derbenev, University of Michigan report, UM HE 96-05 (1996).
V.I. Ptitsin and Yu.M. Shatunov, Nucl. Instr. Meth. A **398**, 126 (1997).
V.H. Ranjbar *et al*., Phys. Rev. Accel. Beams **21**, 111003 (2018).
V.S. Morozov *et al*., Phys. Rev. Lett. **91**, 214801 (2003).
H. Huang *et al*., Phys. Rev. Lett. **120**, 264804 (2018).
V.S. Morozov *et al*., in Proc. IPAC’15 (2015), p. 2301.
A.D. Kovalenko *et al.*, PEPAN **45**, 321 (2014).
Yu.N. Filatov *et al*., EPJ Web Conf. **204**, 10014 (2019).
Ya.S. Derbenev, A.M. Kondratenko, and A.N. Skrinskii, Sov. Phys. JETP **33**, 658 (1971).
Ya.S. Derbenev and A.M. Kondratenko, Sov. Phys. JETP **35**, 1230 (1972).
Ya.S. Derbenev and A.M. Kondratenko, Sov. Phys. JETP **37**, 968 (1973); Zh. Eksp. Teor. Fiz. **64**, 1918 (1973).
V.S. Morozov *et al*., PoS (PSTP 2013) 026 (2013).
A.M. Kondratenko *et al*., in Proc. Inl. Workshop EIC’14, JLab, Newport News, VA, MOCAUD3 (2014).
E.D. Courant, AIP Conf. Proc. **42**, 94 (1978);\
E.D. Courant and R.D. Ruth, BNL Rep. BNL-51270 (1980).
V.S. Morozov *et al*., in Proc. of IPAC’19 (2019), p. 2791. V.S. Morozov *et al*., in Proc. of IPAC’17 (2017), p. 3014.
F. M$\acute{\text{e}}$ot, Nucl. Instr. Meth A **427**, 353 (1999).
M. Harrison, S. Peggs, and T. Roser, Ann. Rev. Nucl. Part. Sci. **52**, 425 (2002).
V.S. Morozov *et al.*, in Proc. of NA-PAC’16 (2016), p. 558. Ya.S. Derbenev and A.M. Kondratenko, Conf. Proc. C **830811** (1983), p. 413.
Ya.S. Derbenev and A.M. Kondratenko, in Proc. 8th Int’l Symposium on High-energy Spin Physics, Minneapolis, MN, September 12-17, 1988 (1988), p.1474.
S.Y. Lee and E.D. Courant, Phys. Rev. D **41**, 292 (1990).
A.W. Chao and Ya.S. Derbenev, Part. Acc. **36**, 25 (1991).
V.S. Morozov *et al*., in Proc. of IPAC’19 (2019), p. 2783.
|
---
abstract: 'Within the composite operator formalism we derive a formula for the pion decay constant $f_{\pi}$, as defined directly from the residue at the pion pole of the meson propagator, rather than from the matrix element of the axial current. The calculation is performed under some simplifying assumptions, and we verify the complete consistency with soft-pion results, in particular with the Adler-Dashen relation. The formula one obtains for (the pole-defined) $f_{\pi}^2$ differs from the previous Pagels-Stokar expression by an additive term, and it still provides $f_{\pi}^2$ in terms of the quark self-energy. We make some numerical estimates leading to $(30 \div 40)\%$ deviation for $f_{\pi}^2$ with respect to the Pagels-Stokar formula.'
author:
- |
A.Barducci, R.Casalbuoni, M.Modugno, G.Pettini\
[Dipartimento di Fisica, Univ. di Firenze and I.N.F.N., Sezione di Firenze ]{}\
R.Gatto\
[Département de Physique Théorique, Univ. de Genève]{}
date: |
UGVA-DPT 1997/03-972\
Firenze Preprint - DFF-273/03/1997
title: |
UNIVERSITÉ DE GENÈVE\
SCHOLA GENEVENSIS MDLIX
[ **CORRECTIONS TO THE PAGELS-STOKAR FORMULA FOR $f_{\pi}$ [^1]**]{}
---
458.5pt -21pt -1pt
Introduction
============
In quantum chromodynamics the pion, as well-known, is a pseudo Nambu-Goldstone boson associated with the spontaneous breaking of chiral symmetry. Its decay constant $f_{\pi}$ plays a key dynamical role in the chiral symmetry breaking mechanism of QCD, and analogous quantities appear in other theories which use similar paradigms, such as electroweak symmetry breaking through a new strong sector. In 1979 Pagels and Stokar proposed an approximate expression for the calculation of $f_{\pi}$ [@pagels]. Their derivation used a sum rule due to Jackiw and Johnson [@jackiw], plus additional assumptions within the so-called dynamical perturbation theory, and allowed for an approximate expression for $f_{\pi}$, as defined from the matrix element of the axial current, in terms of the quark self-energy. The formula is of great utility in QCD and also in theories derived from the old technicolour concept.
There has been a vast literature on the Pagels-Stokar (PS) formula with the general conclusion that it leads to a sensible result for $f_{\pi}$, within uncertainties not easily controllable in view of the theoretical approximations present in the derivation [@miranski]. The Pagels-Stokar expression for the pion decay constant is
$$\begin{aligned}
f_{PS}^2&=&{d(\underline{r})\over(2\pi)^2}\int_0^\infty d k^2~ k^2~
{\Sigma_{0}^2(k^2)-{1\over 2} k^2 \Sigma_{0}(k^2)\displaystyle{
d \Sigma_{0}(k^2)\over d k^2}\over
\left[\displaystyle k^2+\Sigma_{0}^2(k^2)\right]^2}\\
&&\nonumber \end{aligned}$$
where $d(\underline{r})$ is the dimension of the quark colour representation ($d(\underline{r})=3$ in QCD) and $\Sigma_{0}$ is the dynamical quark self-energy in the chiral limit.
We shall present below a new formula for $f_{\pi}$, which shares with the PS formula the advantage of only depending on the self-energy $\Sigma_{0}$, and that we derive within the composite operator formalism developed in ref. [@cjt] and as modified in ref. [@bcd]. Within such schemes one could think of two different calculations of the pion decay constant. One can evaluate the coupling of the pseudo-Goldstone to the axial-vector current. Alternatively, one can directly evaluate the residue at the pion pole of the meson propagator. These procedures correspond to different definitions, but only the second one agrees with the current-algebra and soft-pion results [@genrev] (in particular the Adler-Dashen relation). We shall follow the second procedure, that is calculating the residue at the pion pole. Within the composite operator formalism, for a first approximate understanding, we discuss here the so-called “rigid case”, in which the presumably small logarithmic corrections coming from the renormalization-group analysis are neglected. For the asymptotic behaviour of the self-energy function we choose the one dictated by the operator product expansion (OPE) [@miranski] up to a logarithm which we neglect. The motivation for such approximate study is mainly that in this way it is possible to derive analitically the complete expression for the effective action at two fermion loops. Furthermore a phenomenological analysis we have previously performed of the pseudoscalar masses [@bcd] in this contest did not show any major inadequacy of such a treatment. Our expression for $f_{\pi}^2$ is
$$\begin{aligned}
f_{\pi}^2 &=& {d(\underline{r})\over(2\pi)^2}
\int_0^\infty d k^2~\left[
k^2~{\Sigma_{0}^2(k^2)-{1\over 2} k^2 \Sigma_{0}(k^2)
{\displaystyle {d \Sigma_{0}(k^2)\over{d k^2}}}\over
\big[k^2+\Sigma_{0}^2(k^2)\big]^2}
\right.\nonumber\\
&&\qquad\qquad+\left.{\displaystyle k^6
\left({d\Sigma_{0}(k^2)\over d k^2}\right)^2
- k^4\Sigma_{0}^2(k^2)\left({d \Sigma_{0}(k^2)\over d k^2 }\right)^2 - k^4
\Sigma_{0}(k^2){d \Sigma_{0}(k^2) \over d k^2 }\over 2
{\big[k^2+\Sigma_{0}^2(k^2)\big]^2}
}\right]\end{aligned}$$
The first term in Eq. (2) is $f_{PS}^2$ of Eq. (1). By writing $f_{\pi}^2$ = $f_{PS}^2$ $(1+{\delta}^2)$, to get an evaluation of $\delta$ we can go back to two different Ansatz for $\Sigma_{0}$ we had used in the past to study low energy QCD [@bcd; @giulio]. We find for $\delta^2$ values such as 0.35 and 0.37, suggesting that the correction to the Pagels-Stokar expression is presumably large $(30 \div 40)\%$ and presumably rather insensitive to the modification of the self-energy shape, provided the ultra-violet behaviour in $k^2$ is roughly maintained.
The effective action
====================
We start from the effective Euclidean action for the composite operator formalism $$\Gamma( {\bf\Sigma} ) = - {\rm Tr}\ln\left[{\bf S}_0^{-1} + {\delta
\Gamma_2 \over\delta{\bf S}}\right] + {\rm Tr}\left[ {\delta \Gamma_2
\over\delta{\bf S} }{\bf S}\right] -{\bf \Gamma}_2({\bf S})
+ counterterms$$ where ${\bf S}^{-1}_0= (i\hat{p}- {\bf m})$, [**m**]{} is the bare quark mass matrix, ${\bf \Gamma}_2({\bf S})$ is the sum of all two-particle irreducible vacuum diagrams with fermionic propagator ${\bf S}$ and ${\bf \Sigma} =
-\delta{\bf \Gamma}_2/\delta{\bf S}$. Eq. (3) is the modification of the effective action of Cornwall, Jackiw and Tomboulis [@cjt] which was introduced in ref. [@bcd] to account for the correct stability properties of the theory. At two-loop level ${\bf \Gamma}_2=
{1\over2}{\rm Tr}({\bf S}\Delta{\bf S})$, where $\Delta$ is the gauge boson propagator, so that ${\bf \Sigma} = - \Delta{\bf S}$, ${\rm Tr}\left[{\delta{\bf \Gamma}_2/\delta{\bf S}}\right] =
2{\bf \Gamma}_2$, and one can rewrite Eq. (3) in terms of ${\bf \Sigma}$ $$\begin{aligned}
\Gamma( {\bf\Sigma} ) &=& - {\rm Tr}\ln\left[{\bf S}_0^{-1} -{\bf \Sigma}
\right] + {\bf \Gamma}_2({\bf \Sigma}) + counterterms\nonumber\\
&=&- {\rm Tr}\ln\left[{\bf S}_0^{-1} -{\bf \Sigma}
\right] + {1\over 2}{\rm Tr}\left({\bf \Sigma}\Delta^{-1}{\bf \Sigma}
\right) + counterterms\end{aligned}$$
Here the variable ${\bf \Sigma}$ plays the role of a dynamical variable. At the minimum of the functional action, that is when the Schwinger-Dyson equation is satisfied, ${\bf \Sigma}$ is nothing but the fermion self-energy . A parametrization for ${\bf \Sigma}$, employed in [@bcd], was $${\bf\Sigma}=({\bf s} + i \gamma_5{\bf p})f(k)\equiv {\bf\Sigma}_s
+i\gamma_5 {\bf\Sigma}_p$$ with a suitable Ansatz for $f(k)$, and with [**s**]{} and [**p**]{} scalar and pseudoscalar constant fields respectively.
The effective potential
=======================
The effective potential one obtains from Eq. (4) (see ref. [@bcd]) is $$\begin{aligned}
V={\Gamma\over\Omega}&=& -{8\pi^2 d(\underline{r})\over 3 C_2(\underline{r})
g^2}
\int{d^4 k\over(2\pi)^4}{\rm tr}\left[{\bf \Sigma}_s\Box_k{\bf \Sigma}_s
+{\bf \Sigma}_p\Box_k{\bf \Sigma}_p\right]-\nonumber\\
&&- d(\underline{r})~{\rm Tr}\ln\left[i\hat{k} - \left({\bf m}+{\bf \Sigma}_s
\right)-i\gamma_5{\bf \Sigma}_p\right] + \delta Z~{\rm tr}({\bf m}{\bf s})\end{aligned}$$ where $C_2(\underline{r})$ is the quadratic Casimir of the fermion colour representation (for $SU(3)_c$ $C_2 = 4/3$) and ${\bf \Sigma}_s = \lambda_{\alpha} s_{\alpha} f(k)/ \sqrt{2}$, ${\bf \Sigma}_p = \lambda_{\alpha} p_{\alpha} f(k)/ \sqrt{2}$, ${\bf m} = \lambda_{\alpha}m_{\alpha}/ \sqrt{2}$ ($\alpha=0,\cdots,8$, $\lambda_0 = \sqrt{2/3}$, $\lambda_i =$ Gell-Mann matrices, $i=1,\cdots,8$). Furthermore $\delta Z$ has a divergent piece to compensate the leading divergence proportional to ${\rm tr}({\bf m}{\bf s})$ in the logarithmic term. For a quark of mass $m$ the effective potential is $$\begin{aligned}
V(s, p, m) &=& -{ d(\underline{r}) c}
\int{d^4 k\over(2\pi)^4}\left[{\Sigma}_s\Box_k{ \Sigma}_s
+{ \Sigma}_p\Box_k{ \Sigma}_p\right]-\nonumber\\
&&- 2d(\underline{r})\int{d^4 k\over(2\pi)^4}
\ln\left[k^2 + \left( m+{\Sigma}_s\right)^2 + {\Sigma}_p^2\right]
+ \delta Z m s\end{aligned}$$ where we have defined $c=2\pi^2/g^2$. In ref. [@bcd] after fixing $\delta Z$ so as to cancel the leading divergence proportional to $m s$ in the logarithm, we had imposed the normalization condition $$\lim_{ m\rightarrow0}{1\over m}\left.{\partial V\over
\partial\langle\bar{\psi}\psi\rangle}\right|_{extr}=1$$ or, with $\langle\bar{\psi}\psi\rangle =
\left(d(\underline{r})M^3/2\pi^2\right) c \bar{s}$ $$\lim_{ m\rightarrow0}{1\over m}\left.{\partial V\over
\partial s}\right|_{extr}={d(\underline{r})M^3\over2\pi^2}c$$ where $M$ is a momentum scale for the self energy and $\bar{s}$ is the extremum of the effective potential. The extrema of the effective potential in the massless case $m=0$ depend only on $c$. Therefore Eq. (9) becomes, in this case, an equation for $c$ and M is left undetermined. This is nothing but the usual dimensional transmutation. The numerical values for $c$ and $s_0$, the minimum of the effective potential in the massless case, are obtained once one has fixed the Ansatz for $f(k)$.
With $\Sigma_0(k) = s_0 f(k)$ we shall write in general $$\delta Z = d(\underline{r})\left[{M^3\over2\pi^2}c +
{4\over s_0}\int{d^4 k\over(2\pi)^4}{
{\Sigma}_{0s}\over k^2+ {\Sigma}_{0s}^2}\right]$$ The gap equation, from $\displaystyle{d V\over d s}=0$, is $${d(\underline{r})\over \bar{s}}\left[-2c \int{d^4 k\over(2\pi)^4}
\bar{\Sigma}_s\Box_k\bar{\Sigma}_s - 4\int{d^4 k\over(2\pi)^4}
{(m + \bar{\Sigma}_s)\bar{\Sigma}_s\over k^2+ (m+\bar{\Sigma}_s)^2}\right]
+m\delta Z = 0$$ where $\bar{\Sigma}_s=\bar{s}f(k)$ and $\bar{s}$ is the value at the minimum in the presence of the bare mass.
Let us now turn to the effective action. The fields [**s**]{} and [**p**]{} depend in this case on the space coordinates and we shall use the Weyl symmetrization prescription $${\bf\Sigma}=({\bf s} + i \gamma_5{\bf p})f(k)\rightarrow
{1\over2}\left[
{\bf s}({\bf x}) + i\gamma_5 {\bf p}({\bf x}),f({\bf k})
\right]_{+}$$ We are interested in oscillations around the minimum of the effective potential, so we introduce $$\begin{aligned}
&&{{\mbox{\boldmath $\chi$}}}({\bf x})={\bf s}({\bf x})-\bar{\bf s},\qquad
{{\mbox{\boldmath $\pi$}}}({\bf x})={\bf p}({\bf x})-\bar{\bf p}\equiv{\bf p}({\bf x})
\nonumber\\
&&{\bf v}({\bf x})={{\mbox{\boldmath $\chi$}}}({\bf x})+i\gamma_5{{\mbox{\boldmath $\pi$}}}({\bf x})
\nonumber\\
&&\bar{\bf S }(k)=i\hat{k}- \left( ({\bf m} +\bar{\bf \Sigma}_s(k)\right)\end{aligned}$$ The [*Tr ln*]{} term in Eq. (4), which we denote as $\Gamma_{log}$, becomes $$\Gamma_{log}=-{\rm Tr}\ln \left[i\hat{k} - {\bf m} -{\bf \Sigma}\right]
=-{\rm Tr}\ln \left[\bar{\bf S }^{-1} -{1\over2}\left[
{\bf v}({\bf x}),f({\bf k})\right]_{+}\right]$$ As we are interested in the 2-points function, we expand to second order in ${\bf v}({\bf x})$ and after some calculation we obtain for the Fourier transform of $\Gamma_{log}$ in Eq. (14) $$\begin{aligned}
\Gamma_{log}&=&d(\underline{r})\Big\{
-2\Omega\int{d^4k\over(2\pi)^4}
{\rm tr}\ln\left[
k^2 + \left({\bf m} +\bar{\bf \Sigma}_s(k)\right)^2
\right] +\int{d^4k\over(2\pi)^4}f(k){\rm tr}
\left[\bar{\bf S}(k){{\mbox{\boldmath $\chi$}}}(0)\right]
\nonumber\\
&&+
{1\over2}\int{d^4k\over(2\pi)^4}\int{d^4q\over(2\pi)^4}
{\rm tr}\left[
\bar{\bf S}(k)i\gamma_5 V(k, k+q){{\mbox{\boldmath $\pi$}}}(-q)\bar{\bf S}(k+q)
i\gamma_5 V(k+q, k){{\mbox{\boldmath $\pi$}}}(q)\right. \nonumber\\
&&+\Big. \Big. (pseudoscalar\leftrightarrow scalar,
i\gamma_5\leftrightarrow {\bf 1})\Big]+ \cdots\Big\}\end{aligned}$$ where $$V(k_1,k_2)\equiv {1\over2}\left[f(k_1) + f(k_2)\right]$$
For $\Gamma_2$ we obtain (working in Landau’s gauge) $$\begin{aligned}
\Gamma_2&=& -{d(\underline{r})\; c}\;\Omega
\int{d^4 k\over(2\pi)^4}{\rm tr}\left[\bar{\bf \Sigma}_s
\Box_k\bar{\bf \Sigma}_s\right]\nonumber\\
&& -{2d(\underline{r}) c}
\left(\int{d^4 k\over(2\pi)^4}f(k)\Box_k f(k)\right){\rm tr}
(\bar{\bf s}{{\mbox{\boldmath $\chi$}}}(0))\nonumber\\
&& +\int{d^4 q\over(2\pi)^4}{\rm tr}
\Big\{
-{d(\underline{r}) c}{{\mbox{\boldmath $\pi$}}}(-q)
\left( \int{d^4 k\over(2\pi)^4}f(k)\Box_k f(k) \right)
{{\mbox{\boldmath $\pi$}}}(q)
\nonumber\\
&&+ (pseudoscalar \leftrightarrow scalar)\Big\}\end{aligned}$$
For the counterterm one has $$\Gamma_{ct}={\rm tr}({\bf m}\bar{\bf s})\Omega\delta Z +
{\rm tr}({\bf m}{{\mbox{\boldmath $\chi$}}}(0))\delta Z$$
The improved expression for $f_{\pi}$
=====================================
We note that each term in the effective action consists of a constant, a linear and a quadratic term in the fields. The constant term gives back the original potential Eq. (6) at the minimum. Such a term controls the normalization. The linear term vanishes by virtue of the gap equation, Eq. (11). The quadratic term stands up for the effective action up to the second order in the fields. In space-time coordinates $$\begin{aligned}
\Gamma&=&\int{d^4 x}{d^4 y}\int{d^4 q\over(2\pi)^4}
{\rm e}^{\displaystyle -iq(x - y)}\pi_{\alpha}(x)\cdot\nonumber\\
&&\cdot{\rm tr}\left\{
-{d(\underline{r}) c}
{\lambda_{\alpha}\over\sqrt{2}}
\int{d^4 k\over(2\pi)^4}f(k)
\Box_k f(k) {\lambda_{\beta}\over\sqrt{2}}+\right.\nonumber\\
&&+\left.
{1\over2}d(\underline{r})\int{d^4k\over(2\pi)^4}
\left[
\bar{\bf S}(k)i\gamma_5{\lambda_{\alpha}\over\sqrt{2}}
V(k, k+q)\bar{\bf S}(k+q)
i\gamma_5{\lambda_{\beta}\over\sqrt{2}}
V(k+q, k)\right]\right\} \nonumber\\
&&\cdot\pi_{\beta}(y) +\left. (pseudoscalar\leftrightarrow scalar,
i\gamma_5\leftrightarrow {\bf 1}\right)+ \cdots\end{aligned}$$
From $$G^{-1}_{\alpha\beta}(x-y)={\delta^2\Gamma\over\delta\pi_{\alpha}(x)
\delta\pi_{\beta}(y)}$$
one finds for the Fourier transform of $G^{-1}_{\alpha\beta}(x-y)$ $$\begin{aligned}
G^{-1}_{\alpha\beta}(q^2)&=&{\rm tr}\left\{
-{2 d(\underline{r}) c}{\lambda_{\alpha}\over\sqrt{2}}
\int{d^4 k\over(2\pi)^4} f(k)
\Box_k f(k) {\lambda_{\beta}\over\sqrt{2}}+\right.\nonumber\\
&&+\left.
d(\underline{r})\int{d^4k\over(2\pi)^4}
\left[
\bar{\bf S}(k)i\gamma_5{\lambda_{\alpha}\over\sqrt{2}}
V(k, k+q)\bar{\bf S}(k+q)
i\gamma_5{\lambda_{\beta}\over\sqrt{2}}
V(k+q, k)\right]\right\}\end{aligned}$$
By using the gap equation to eliminate $c$ we get, for a quark of mass $m$ $$\begin{aligned}
G^{-1}_{\alpha\beta}(q^2)&=&
d(\underline{r})\int{d^4k\over(2\pi)^4}
{\rm tr}\left[
\bar{S}(k)i\gamma_5 V(k, k+q)\bar{S}(k+q)
i\gamma_5 V(k+q, k)\right]\nonumber\\
&&+{4~d(\underline{r})\over\bar{s}^2}\int{d^4 k\over(2\pi)^4}
{(m + \bar{\Sigma}_s)\bar{\Sigma}_s\over
k^2+ (m+\bar{\Sigma}_s)^2}
-{m\over\bar{s}}\delta Z\end{aligned}$$
We can eliminate $\delta Z$ by using the normalization condition, see Eq. (10), and the relation $\langle\bar{\psi}\psi\rangle =
\left(d(\underline{r})M^3/2\pi^2\right) c \bar{s}$ to obtain $$\begin{aligned}
G^{-1}(q^2)&=&-{m\over\bar{s}^2}\langle\bar{\psi}\psi\rangle
+ d(\underline{r})\left\{-{4 m\over\bar{s}s_0}
\int{d^4 k\over(2\pi)^4}
{{\Sigma}_{0}(k)\over k^2+{\Sigma}_{0}^2(k)}\right.\nonumber\\
&&+{4\over\bar{s}^2}\int{d^4 k\over(2\pi)^4}
{(m + \bar{\Sigma}_s)\bar{\Sigma}_s\over k^2+ (m+\bar{\Sigma}_s)^2}
\nonumber\\
&&+\left.\int{d^4k\over(2\pi)^4}
{\rm tr}\left[
\bar{S}(k)i\gamma_5V(k, k+q)\bar{S}(k+q)
i\gamma_5V(k+q, k)\right]\right\}\end{aligned}$$
Note that the second and third terms in Eq. (23) regularize each other in the ultraviolet.
In the limit of small quark masses, expanding in $q^{\mu}$, we find $$G^{-1}(q^2)\equiv \left({F\over\sqrt{2}s_0}\right)^2\cdot
\left(q^2 - {2m\over F^2}\langle\bar{\psi}\psi\rangle_0\right)$$
with $$\begin{aligned}
F^2 &=& {d(\underline{r})\over(2\pi)^2}
\int_0^\infty d k^2~\left[
k^2~{\Sigma_{0}^2(k^2)-{1\over 2} k^2 \Sigma_{0}(k^2)
{\displaystyle {d \Sigma_{0}(k^2)\over{d k^2}}}\over
\big[k^2+\Sigma_{0}^2(k^2)\big]^2}
\right.\nonumber\\
&&\qquad\qquad+\left.{\displaystyle k^6
\left( {d\Sigma_{0}(k^2)\over d k^2}\right)^2
- k^4\Sigma_{0}^2(k^2)\left({d \Sigma_{0}(k^2)\over d k^2 }\right)^2 - k^4
\Sigma_{0}(k^2){d \Sigma_{0}(k^2) \over d k^2 }\over 2
{\big[k^2+\Sigma_{0}^2(k^2)\big]^2}
}\right]\end{aligned}$$
In Minkowski metrics the propagator (24) has a pole at $q^2 = m^2_{\pi} = -2 m \langle\bar{\psi}\psi\rangle_0/F^2$, with residue $(\sqrt{2}s_0/F)^2$, where $\langle\bar{\psi}\psi\rangle_0=(d(\underline{r})M^3/2\pi^2) c s_0$. The Adler-Dashen relation (which follows from the symmetries and current algebra) requires the identification $f^2_{\pi}= F^2$, so that $q^2 = m^2_{\pi} = -2 m \langle\bar{\psi}\psi\rangle_0/f^2_{\pi}$. The rescaling factor $b$ relating the canonical field $\varphi_{\pi}$ (with unit residue at the pole) to the field $\pi$, $\varphi_{\pi}=b\pi$, is then $b=-f_{\pi}/\sqrt{2}s_0$, as indeed follows from current algebra and soft pions theorem (see ref. [@bcd]). Comparison of (25) with the Pagels-Stokar formula (1) leads to our new formula (2) of the introduction.
To get a numerical insight into the problem we use the dynamical calculations in ref. [@bcd; @giulio], where overall fits to low energy QCD were made on the basis of two alternative Ansatz for $\Sigma(k)= s f(k)$: a smooth Ansatz $f(k) =
M^3 /(M^2 + k^2)$ for which the relevant parameters took values $c=0.554$, $s_0=-4.06$, and a step-function Ansatz $f(k) =M [\theta(M^2 -k^2)
+(M^2/k^2) \theta(k^2 - M^2)]$ for which one found $c=0.32$, $s_0=-2.69$. Our new expression for $f_{\pi}$ has $f^2_{\pi} = f^2_{PS}(1 +\delta^2)$, where $\delta$ follows from Eq. (25). We find $\delta^2 = 0.347$ in the case of the smooth Ansatz, and $\delta^2 = 0.376$ for the step-function Ansatz. We want to remark that in the massless limit we consider, the correction $\delta$ depends only on $s_0$ (or $c$) and the shape of $f(k)$, but not on $M$. In particular, because of the fact that the relevant contribution to the chiral symmetry breaking phenomenon comes from relatively short-distance effects, the corrections will depend mainly on the ultraviolet behaviour of the self-energy. That is, the correction does not depend on the fit we have made to low energy QCD.
It therefore seems that: $(i)$ the corrections are relevant with respect to the old Pagels-Stokar formula; $(ii)$ the corrections do not seem to vary in a sensible way when varying the Ansatz for the self energy, at least within the Ansatz we have used.
It is obvious that the next step one should take is to see whether the neglected corrections, which we know must be there, can modify the results. However, due to previous experience from the study of the pseudoscalar masses [@bcd], we would not expect substantial changes in the overall picture of dynamical symmetry breaking.
Finally we may stress that the new formula (2) we have obtained for $f_{\pi}$, within the approximations made, does not require additional inputs beyond those already present in the Pagels-Stokar formula.
[*This work has been carried out within the program Human Capital and Mobility (BBW/OFES 95.0200; CHRXCT 94-0579)*]{}
[99]{}
H. Pagels and S. Stokar, Phys.Rev. [**D20**]{} (1979) 2947. R. Jackiw and K. Johnson, Phys. Rev. [**D8**]{} (1973) 2386. For a review and for additional references: V. A. Miransky, “Dynamical Symmetry Breaking in Quantum Field Theories", World Scientific, Singapore (1993); K. Aoki, M. Bando, T. Kugo, M. G. Mitchard and H. Nakatani, Prog. Theor. Phys. [**84**]{} (1990) 683; P. Jain and H. J. Munczek, Phys. Rev [**D44**]{} (1991) 1873, Phys. Lett. [**B282**]{} (1992) 157 ; T. Kugo and M. G. Mitchard, Phys Lett. [**B282**]{} (1992) 162, Phys. Lett. [**B286**]{} (1992) 355. J. M. Cornwall, R. Jackiw and E. Tomboulis, Phys. Rev. [**D10**]{} (1974) 2428. A. Barducci, R. Casalbuoni, S. De Curtis, D. Dominici and R. Gatto, Phys. Lett. [**147B**]{} (1984) 460, Phys. Rev. [**D38**]{} (1988) 238; R. Casalbuoni, S. De Curtis, D. Dominici and R. Gatto, Phys. Lett. [**150B**]{} (1985) 295; R. Casalbuoni, in [*Proceedings of the International Symposium on Composite Models of Quarks and Leptons*]{}, Tokyo, Japan, 1985, edited by H. Terazawa and M. Yasué (Institute for Nuclear Study, University of Tokyo, Tokyo, 1985). For general reviews of the field and references see: S. L. Adler and R. F. Dashen, “Current Algebra and Applications”, Benjamin, New York (1968); S. B. Treiman, R. Jackiw and D. J. Gross, “Lectures on Current Algebra and Applications”, Princeton University Press, Princeton, New Jersey (1972); V. De Alfaro, S. Fubini, G. Furlan and C. Rossetti, “Current in Hadron Physics”, North Holland, Amsterdam (1973); E. Commins and P. H. Bucksbaum, “Weak Interactions of Leptons and Quarks”, Cambridge University Press, Cambridge, England (1983); H. Georgi, “Weak Interactions and Modern Particle Theory”, Benjamin-Cummings, Menlo Park, California (1984). A. Barducci, R. Casalbuoni, S. De Curtis, R. Gatto and G. Pettini, Phys. Rev. [**D46**]{} (1992) 2203.
[^1]: Partially supported by the Swiss National Foundation
|
---
abstract: |
An alternative approximation scheme has been used in solving the Schrödinger equation for the exponential-cosine-screened Coulomb potential. The bound state energies for various eigenstates and the corresponding wave functions are obtained analytically up to the second perturbation term.
Keywords: Exponential-cosine-screened Coulomb potential, Perturbation theory
PACS NO: 03.65.Ge
address: |
$^{\ast }$Department of Electrical and Electronic Engineering, Near East\
University, Nicosia, North Cyprus, Mersin-10, Turkey
author:
- 'Sameer M. Ikhdair[^1]'
title: 'A perturbative treatment for the exponential-cosine-screened Coulomb potential '
---
Introduction
=============
The generalized exponential-cosine-screened Coulomb (GECSC) potential or the generalized cosine Yukawa (GCY) potential: $$V\left( r\right) =-\left( \frac{A}{r}\right) \exp (-\delta r)\cos (g\delta
r),$$ where $A$ is the strength coupling constant and $\delta $ is the screening parameter, is known to describe adequately the effective interaction in many-body enviroment of a variety of fields such as atomic, nuclear, solid-state, plasma physics and quantum field theory \[1,2\]. It is also used in describing the potential between an ionized impurity and an electron in a metal \[3,4\] or a semiconductor \[5\] and the electron-positron interaction in a positronium atom in a solid \[6\]. The potential in (1) with $g=1$ is known as a cosine-screened Coulomb potential. The static screened Coulomb (SSC) potential ($g=0$ case) is well represented by Yukawa form: $V(r)=-(\alpha Ze^{2})\exp (-\delta r)/r$ which emerges as a special case of the ECSC potential in (1) with $A=\alpha Ze^{2},$ where $\alpha
=(137.037)^{-1}$ is the fine-structure constant and $Z$ is the atomic number, is often used for the description of the energy levels of light to heavy neutral atoms \[7\]. It is known that SSC potential yields reasonable results only for the innermost states when $Z$ is large. However, for the outermost and middle atomic states, it gives rather poor results. Although the bound state energies for the SSC potential with $Z=1$ have been studied \[7\].
The Schrödinger equation for such a potential does not admit exact solutions. So various approximate methods \[8\] both numerical and analytical have been developed Hence, the bound-state energies of the ECSC potential were first calculated for the $1s$ state using numerical \[3,8,9\] and analytical \[10,11\] methods and for the $s$ states by a variational method \[12\]. Additionally, the energy eigenvalues of the ECSC potential \[13\] have been recalculated for the $1s$ state with the use of the ground-state logarithmic perturbation theory \[14,15\] and the Padé approximant method. The problem of determining the critical screening parameter $\delta _{c}$ for the $s$ states was also studied \[16\].
It has also been shown that the problem of screened Coulomb potentials can be solved to a very high accuracy \[17\] by using the hypervirial relations \[18,19,20\] and the Padé approximant method. The bound-state energies of the ECSC potential for all eigenstates were accurately determined within the framework of the hypervirial Padé scheme \[21\]. Further, the large-N expansion method of Mlodinow and Shatz \[22\] was also applied to obtain the energies of the ground and first excited $s-$ states and the corresponding wave functions. Recently, we studied the bound-states of the ECSC potential for all states using the shifted large $N-$ expansion technique \[23\].
In this paper, we investigate the bound-state properties of ECSC potential using a new perturbative formalism \[24\] which has been claimed to be very powerful for solving the Schrödinger equation to obtain the bound-state energies as well as the wave functions in Yukawa or SSC potential problem \[24,25\] in both bound and continuum regions. This novel treatment is based on the decomposition of the radial Schrödinger equation into two pieces having an exactly solvable part with an addiitional piece leading to either a closed analytical solution or approximate treatment depending on the nature of the perturbed potential.
The contents of this paper is as follows. In section \[TM\] we breifly outline the method with all necessary formulae to perform the current calculations. In section \[A\] we apply the approach to the Schrödinger equation with the ECSC potential and present the results obtained analytically and numerically for the bound-state energy values. Finally, in section \[CR\] we give our concluding remarks.
The Method {#TM}
==========
For the consideration of spherically symmetric potentials, the corresponding Schrödinger equation, in the bound state domain, for the radial wave function reads
$$\frac{\hbar ^{2}}{2m}\frac{\psi _{n}^{\prime \prime }(r)}{\psi _{n}\left(
r\right) }=V(r)-E_{n},$$
with
$$V\left( r\right) =\left[ V_{0}(r)+\frac{\hbar ^{2}}{2m}\frac{\ell (\ell +1)}{r^{2}}\right] +\Delta V(r),$$
where $\Delta V(r)$ is a perturbing potential and $\psi _{n}(r)=\chi
_{n}(r)u_{n}(r)$ is the full radial wave function, in which $\chi _{n}(r)$ is the known normalized eigenfunction of the unperturbed Schrödinger equation whereas $u_{n}(r)$ is a moderating function corresponding to the perturbing potential. Following the prescription of Ref. 24, we may rewrite (2) as
$$\frac{\hbar ^{2}}{2m}\left( \frac{\chi _{n}^{\prime \prime }(r)}{\chi _{n}(r)}+\frac{u_{n}^{\prime \prime }(r)}{u_{n}(r)}+2\frac{\chi _{n}^{\prime
}(r)u_{n}^{\prime }(r)}{\chi _{n}(r)u_{n}(r)}\right) =V(r)-E_{n}.$$
The logarithmic derivatives of the unperturbed $\chi _{n}(r)$ and perturbed $u_{n}(r)$ wave functions are given by
$$W_{n}(r)=-\frac{\hbar }{\sqrt{2m}}\frac{\chi _{n}^{\prime }(r)}{\chi _{n}(r)}\text{ \ \ and \ \ }\Delta W_{n}=-\frac{\hbar }{\sqrt{2m}}\frac{u_{n}^{\prime }(r)}{u_{n}(r)},$$
which leads to
$$\frac{\hbar ^{2}}{2m}\frac{\chi _{n}^{\prime \prime }(r)}{\chi _{n}(r)}=W_{n}^{2}(r)-\frac{\hbar }{\sqrt{2m}}W_{n}^{^{\prime }}(r)=\left[ V_{0}(r)+\frac{\hbar ^{2}}{2m}\frac{\ell (\ell +1)}{r^{2}}\right] -\varepsilon _{n},$$
where $\varepsilon _{n}$ is the eigenvalue for the exactly solvable potential of interest, and
$$\frac{\hbar ^{2}}{2m}\left( \frac{u_{n}^{\prime \prime }(r)}{u_{n}(r)}+2\frac{\chi _{n}^{\prime }(r)u_{n}^{\prime }(r)}{\chi _{n}(r)u_{n}(r)}\right)
=\Delta W_{n}^{2}(r)-\frac{\hbar }{\sqrt{2m}}\Delta W_{n}^{\prime
}(r)+2W_{n}(r)\Delta W_{n}(r)=\Delta V(r)-\Delta \varepsilon _{n},$$
in which $\Delta \varepsilon _{n}=E_{n}^{(1)}+E_{n}^{(2)}+\cdots $ is the correction term to the energy due to $\Delta V(r)$ and $E_{n}=\varepsilon
_{n}+\Delta \varepsilon _{n}.$ If Eq. (7), which is the most significant piece of the present formalism, can be solved analytically as in (6), then the whole problem, in Eq. (2) reduces to the following form
$$W_{n}(r)+\Delta W_{n}(r))^{2}-\frac{\hbar }{\sqrt{2m}}(W_{n}(r)+\Delta
W_{n}(r))^{\prime }=V(r)-E_{n},$$
which is a well known treatment within the frame of supersymmetric quantum theory (SSQT) \[26\]. Thus, if the whole spectrum and corresponding eigenfunctions of the unperturbed interaction potential are known, then one can easily calculate the required superpotential $W_{n}(r)$ for any state of interest leading to direct computation of related corrections to the unperturbed energy and wave function.
For the perturbation technique, we can split the given potential in Eq.(2) into two parts. The main part corresponds to a shape invariant potential, Eq. (6), for which the superpotential is known analytically and the remaining part is treated as a perturbation, Eq. (7). Therefore, it is obvious that ECSC potential can be treated using this prescription. In this case, the zeroth-order term corresponds to the Coulomb potential while higher-order terms consitute the perturbation. However, the perturbation term in its present form cannot be solved exactly through Eq. (7). Thus, one should expand the functions related to the perturbation in terms of the perturbation parameter $\lambda $,
$$\Delta V(r;\lambda )=\sum_{i=1}^{\infty }\lambda _{i}V_{i}(r),\text{ \ \ }\Delta W_{n}(r;\lambda )=\sum_{i=1}^{\infty }\lambda _{i}W_{n}^{(i)}(r),\text{ \ }E_{n}^{(i)}(\lambda )=\sum_{i=1}^{\infty }\lambda _{i}E_{n}^{(i)},$$
where $i$ denotes the perturbation order. Substitution of the above expansions into Eq. (7) and equating terms with the same power of $\lambda $ on both sides up to $O(\lambda ^{3})$ gives
$$2W_{n}(r)W_{n}^{(1)}(r)-\frac{\hbar }{\sqrt{2m}}\frac{dW_{n}^{(1)}(r)}{dr}=V_{1}(r)-E_{n}^{(1)},$$
$$W_{n}^{(1)2}(r)+2W_{n}(r)W_{n}^{(2)}(r)-\frac{\hbar }{\sqrt{2m}}\frac{dW_{n}^{(2)}(r)}{dr}=V_{2}(r)-E_{n}^{(2)},$$
$$2(W_{n}(r)W_{n}^{(3)}(r)+W_{n}^{(1)}(r)W_{n}^{(2)}(r))-\frac{\hbar }{\sqrt{2m}}\frac{dW_{n}^{(3)}(r)}{dr}=V_{3}(r)-E_{n}^{(3)}.$$
Hence, unlike the other perturbation theories, Eq. (7) and its expansion, Eqs. (10-12), give a flexibility for the easy calculations of the perturbative corrections to energy and wave functions for the $nth$ state of interest through an appropriately chosen perturbed superpotential.
Application to the ECSC Potential {#A}
=================================
Considering the recent interest in various power-law potentials in the literature, we work through the article within the frame of low screening parameter. In this case, the ECSC potential can be expanded in power series of the screening parameter $\delta $ as \[10\]
$$V(r)=-\left( \frac{A}{r}\right) \exp (-\delta r)\cos (\delta r)=-\left(
\frac{A}{r}\right) \sum_{i=0}^{\infty }V_{i}(\delta r)^{i},$$
where the perturbation coefficients $V_{i}$ are given by
$$V_{1}=-1,\text{ }V_{2}=0,\text{ }V_{3}=1/3,\text{ }V_{4}=-1/6,\text{ }V_{5}=1/30,\cdots .$$
We now apply this approximation method to the ECSC potential with the angular momentum barrier
$$V(r)=-\left( \frac{A}{r}\right) e^{-\delta r}\cos (\delta r)+\frac{\ell
(\ell +1)\hbar ^{2}}{2mr^{2}}=\left[ V_{0}(r)+\frac{\ell (\ell +1)\hbar ^{2}}{2mr^{2}}\right] +\Delta V(r),$$
where the first piece is the shape invariant zeroth-order which is an exactly solvable piece corresponding to the unperturbed Coulomb potential with $V_{0}(r)=-A/r$ while $\Delta V(r)=A\delta -(A\delta
^{3}/3)r^{2}+(A\delta ^{4}/6)r^{3}-(A\delta ^{5}/30)r^{4}+\cdots $ is the perturbation term. The literature is rich with examples of particular solutions for such power-law potentials employed in different fields of physics, for recent applications see Refs. \[27,28\]. At this stage one may wonder why the series expansion is truncated at a lower order. This can be understood as follows. It is widely appreciated that convergence is not an important or even desirable property for series approximations in physical problems. Specifically, a slowly convergent approximation which requires many terms to achieve reasonable accuracy is much less valuable than the divergent series which gives accurate answers in a few terms. This is clearly the case for the ECSC problem \[29\]. However, it is worthwhile to note that the main contributions come from the first three terms. Thereby, the present calculations are performed up to the second-order involving only these additional potential terms, which suprisingly provide highly accurate results for small screening parameter $\delta .$
Ground State Calculations $\left( n=0\right) $
----------------------------------------------
In the light of Eq. (6), the zeroth-order calculations leading to exact solutions can be carried out readily by setting the ground-state superpotential and the unperturbed exact energy as
$$W_{n=0}\left( r\right) =-\frac{\hbar }{\sqrt{2m}}\ \frac{\ell +1}{r}+\sqrt{\frac{m}{2}}\frac{A}{(\ell +1)\hbar },\text{ \ \ }E_{n}^{(0)}=-\frac{mA^{2}}{2\hbar ^{2}(n+\ell +1)^{2}},\text{ \ \ \ }n=0,1,2,....$$
and from the literature, the corresponding normalized Coulomb bound-state wave function \[30\]
$$\chi _{n}(r)=N_{n,l}^{(C)}r^{\ell +1}\exp \left[ -\beta r\right] \times
L_{n}^{2\ell +1}\left[ 2\beta r\right] ,$$
in which $N_{n,l}^{(C)}=\left[ \frac{2mA}{\left( n+\ell +1\right) \hbar ^{2}}\right] ^{\ell +1}\frac{1}{(n+\ell +1)}\frac{1}{\sqrt{\frac{\hbar ^{2}}{mAn!}(n+2\ell +1)!}}$ is a normalized constant, $\beta =\frac{mA}{\left(
n+\ell +1\right) \hbar ^{2}}$ and $L_{n}^{k}\left( x\right)
=\sum_{m=0}^{n}(-1)^{m}\frac{(n+k)!}{\left( n-m\right) !(m+k)!m!}x^{m}$ is an associate Laguarre polynomial function \[31\].
For the calculation of corrections to the zeroth-order energy and wavefunction, one needs to consider the expressions leading to the first- and second-order perturbation given by Eqs. (10–11). Multiplication of each term in these equations by $\chi _{n}^{2}(r)$, and bearing in mind the superpotentials given in Eq. (5), one can obtain the straightforward expressions for the first-order correction to the energy and its superpotential: $$E_{n}^{(1)}=\int_{-\infty }^{\infty }\chi _{n}^{2}(r)\left( -\frac{A\delta
^{3}}{3}r^{2}\right) dr,\text{ }W_{n}^{(1)}\left( r\right) =\frac{\sqrt{2m}}{\hbar }\frac{1}{^{X_{n}^{2}(r)}}\int^{r}\chi _{n}^{2}(x)\left[ E_{n}^{(1)}+\frac{A\delta ^{3}}{3}x^{2}\right] dx,$$ and also for the second-order correction and its superpotential:
$$E_{n}^{(2)}=\int_{-\infty }^{\infty }\chi _{n}^{2}(r)\left[ \frac{A\delta
^{4}}{6}r^{3}-W_{n}^{(1)2}\left( r\right) \right] dr,\text{ }$$ $$W_{n}^{(2)}\left( r\right) =\frac{\sqrt{2m}}{\hbar }\frac{1}{^{X_{n}^{2}(r)}}\int^{r}\chi _{n}^{2}(x)\left[ E_{n}^{(2)}+W_{n}^{(1)2}(x)-\frac{A\delta ^{4}}{6}x^{3}\right] dx\ ,$$ for any state of interest. The above expressions calculate $W_{n}^{(1)}(r)$ and $W_{n}^{(2)}(r)$ explicitly from the energy corrections $E_{n}^{(1)}$ and $E_{n}^{(2)}$ respectively, which are in turn used to calculate the moderating wave function $u_{n}(r).$
Thus, through the use of Eqs. (18) and (19), after some lengthy and tedious integrals, we find the zeeroth order energy shift and their moderating superpotentials as
$$E_{0}^{(1)}\ =-\frac{\hbar ^{4}\left( \ell +1\right) ^{2}\left( \ell
+2\right) \left( 2\ell +3\right) }{6Am^{2}}\delta ^{3},$$
$$\begin{aligned}
E_{0}^{(2)} &=&\frac{\hbar ^{6}\left( \ell +1\right) ^{3}\left( \ell
+2\right) \left( 2\ell +3\right) \left( 2\ell +5\right) }{24A^{2}m^{3}}\delta ^{4} \\
&&-\frac{\hbar ^{10}\left( \ell +1\right) ^{6}\left( \ell +2\right) \left(
2\ell +3\right) \left( 8\ell ^{2}+37\ell +43\right) }{72A^{4}m^{5}}\delta
^{6},\end{aligned}$$
$$W_{0}^{(1)}(r)=-\frac{\hbar \left( \ell +1\right) \delta ^{3}r}{3\sqrt{2m}}\left\{ r-\frac{\hbar ^{2}\left( \ell +1\right) \left( \ell +2\right) }{Am}\right\} ,$$ $$W_{0}^{(2)}(r)=-\frac{\hbar \delta ^{4}cr}{2\sqrt{2m}}\left\{ \delta
^{2}r^{3}+ar^{2}+b\left[ r+\frac{\hbar ^{2}(\ell +1)(\ell +2)}{Am}\right]
\right\} -\frac{\hbar \left( \ell +1\right) }{\sqrt{2m}A}E_{0}^{(2)},$$ in which
$$\begin{aligned}
a &=&\frac{\hbar ^{2}(\ell +1)(3\ell +7)\delta ^{2}}{Am}-\frac{3Am}{\hbar
^{2}(\ell +1)^{2}},\text{ \ }b=\left[ \frac{\hbar ^{4}(\ell +1)^{2}(8\ell
^{2}+37\ell +43)\delta ^{2}}{2A^{2}m^{2}}-\frac{3}{2}\frac{(2\ell +5)}{(\ell
+1)}\right] ,\text{ } \nonumber \\
\text{c} &=&\frac{\hbar ^{2}(\ell +1)^{3}}{9Am}\end{aligned}$$
Therefore, the analytical expressions for the lowest energy and full radial wave function of an ECSC potential are then given by
$$E_{n=0,\ell }=E_{n=0,\ell }^{(0)}+A\delta +E_{0}^{(1)}+E_{0}^{(2)}+\cdots ,\text{ }\psi _{n=0,\ell }(r)\approx \chi _{n=0,\ell }(r)u_{n=0,\ell }(r),$$
in which
$$u_{n=0,\ell }(r)\approx \exp \left( -\frac{\sqrt{2m}}{\hbar }\int^{r}\left(
W_{0}^{(1)}\left( x\right) +W_{0}^{(2)}\left( x\right) \right) dx\right) .$$
Hence, the explicit form of the full wave function in (22) for the ground state is
$$\psi _{n=0,\ell }(r)=\left[ \frac{2mA}{(\ell +1)\hbar ^{2}}\right] ^{\ell +1}\frac{1}{(\ell +1)^{2}}\sqrt{\frac{Am}{\hbar ^{2}(2\ell +1)!}}r^{\ell
+1}\exp (P(r)),$$
with $P(r)=\sum_{i=1}^{5}p_{i}r^{i}$ is a polynomial of fifth order having the following coefficients: $$p_{1}=\frac{(\ell +1)}{A}E_{0}^{(2)}-\frac{Am}{(\ell +1)\hbar ^{2}},\text{ \
}p_{2}=\frac{9}{4}\frac{(\ell +2)}{(\ell +1)^{2}}c^{2}d\delta ^{4},\text{ \ }p_{3}=\frac{1}{6}cd\delta ^{4},\text{ }p_{4}=\frac{1}{8}ac\delta ^{4},\text{
}p_{5}=\frac{1}{10}c\delta ^{6},\text{\ }$$ in which $d=b+\frac{6Am}{\hbar ^{2}(\ell +1)^{2}\delta }$ and other parameters are given in (21).
Excited state calculations $(n\geq 1)$
--------------------------------------
The calculations procedures lead to a handy recursion relations in the case of ground states, but becomes extremely cumbersome in the description of radial excitations when nodes of wavefunctions are taken into account, in particular during the higher order calculations. Although several attempts have been made to bypass this difficulty and improve calculations in dealing with excited states, (cf. e.g. \[32\], and the references therein) within the frame of supersymmetric quantum mechanics.
Using Eqs. (5) and (17), the superpotential $W_{n}(r)$ which is related to the excited states can be readily calculated through Eqs. (18) and (19). So the first-order corrections in the first excited state $(n=1)$ are
$$E_{1}^{(1)}=-\frac{\hbar ^{4}\left( \ell +2\right) ^{2}\left( \ell +7\right)
\left( 2\ell +3\right) }{6Am^{2}}\delta ^{3},$$
$$W_{1}^{(1)}(r)\approx -\frac{\hbar \left( \ell +2\right) \delta ^{3}}{3\sqrt{2m}}\left\{ r^{2}+\frac{\hbar ^{2}(\ell +2)(\ell +3)}{Am}r-\frac{2\hbar
^{4}(\ell +1)(\ell +2)^{2}}{A^{2}m^{2}}\right\} .$$
Consequently, the use of the approximated first two terms of $W_{1}^{(1)}(r)$ in the preceeding equation in (19) gives the energy correction in the second-order as
$$\begin{aligned}
\ E_{1}^{(2)} &\approx &\frac{\hbar ^{6}\left( \ell +2\right) ^{3}\left(
\ell +11\right) \left( 2\ell +3\right) \left( 2\ell +5\right) }{24A^{2}m^{3}}\delta ^{4} \nonumber \\
&&-\frac{\hbar ^{10}\left( \ell +2\right) ^{6}\left( \ell +3\right) \left(
2\ell +3\right) \left( 7\ell ^{2}+101\ell +211\right) }{72A^{4}m^{5}}\delta
^{6},\end{aligned}$$
or
$$\begin{aligned}
\ E_{1}^{(2)} &\approx &\frac{\hbar ^{6}\left( \ell +2\right) ^{3}\left(
\ell +11\right) \left( 2\ell +3\right) \left( 2\ell +5\right) }{24A^{2}m^{3}}\delta ^{4} \nonumber \\
&&-\frac{\hbar ^{10}\left( \ell +2\right) ^{5}\left( 16\ell ^{5}+294\ell
^{4}+1795\ell ^{3}+5085\ell ^{2}+6878\ell +3568\right) }{72A^{4}m^{5}}\delta
^{6},\end{aligned}$$
if all the terms in (26) are used. Therefore, the approximated energy value of the ECSC potential corresponding to the first excited state is
$$E_{n=1,\ell }=E_{1}^{(0)}+A\delta +E_{1}^{(1)}+E_{1}^{(2)}+\cdots .$$
The related radial wavefunction can be expressed in an analytical form in the light of Eqs (18), (19) and (22), if required. The appromation used in this work would not affect considerably the sensitivity of the calculations. On the other hand, it is found analytically that our investigations put forward an interesting hierarchy between $W_{n}^{(1)}(r)$ terms of different quantum states in the first order after circumventing the nodal difficulties elegantly,
$$W_{n}^{(1)}(r)\approx -\frac{\hbar \left( n+\ell +1\right) \delta ^{3}}{3\sqrt{2m}}\left\{ r^{2}+\frac{\hbar ^{2}(n+\ell +1)(n+\ell +2)}{Am}r-\frac{2\hbar ^{4}(n+\ell )(n+\ell +1)^{2}}{A^{2}m^{2}}\right\} ,$$
which, for instance, for the second excited state $\left( n=2\right) $ leads to the first-order correction
$$\ E_{2}^{(1)}=-\frac{\hbar ^{4}\left( \ell +3\right) ^{2}\left( \ell
+2\right) \left( 2\ell +23\right) }{6Am^{2}}\delta ^{3},$$
$$W_{2}^{(1)}(r)\approx -\frac{\hbar \left( \ell +3\right) \delta ^{3}}{3\sqrt{2m}}\left\{ r^{2}+\frac{\hbar ^{2}(\ell +3)(\ell +4)}{Am}r-\frac{2\hbar
^{4}(\ell +2)(\ell +3)^{2}}{A^{2}m^{2}}\right\} .$$
Thus, the use of the approximated first two terms of $W_{2}^{(1)}(r)$ in the preceeding equation(19) gives the energy correction in the second-order as
$$\begin{aligned}
\ E_{2}^{(2)} &=&\frac{\hbar ^{6}\left( \ell +2\right) \left( \ell +3\right)
^{2}\left( 2\ell +5\right) \left( 2\ell ^{2}+45\ell +153\right) }{24A^{2}m^{3}}\delta ^{4} \nonumber \\
&&-\frac{\hbar ^{10}\left( \ell +2\right) \left( \ell +3\right) ^{5}(16\ell
^{4}+474\ell ^{3}+3879\ell ^{2}+12118\ell +12873)}{72A^{4}m^{5}}\delta ^{6}.\end{aligned}$$
Therefore, the approximated energy eigenvalue of the ECSC potential corresponding to the second excited state is
$$E_{n=2,\ell }=E_{2}^{(0)}+A\delta +E_{2}^{(1)}+E_{2}^{(2)}+\cdots .$$
For the numerical work, some numerical values of the perturbed energies of the $1s$ and $2s$ states, in the atomic units we take $\hbar =m=A=1,$ for different values of the screening parameter $\delta $ in the range $0\leq
\delta \leq 0.10$ are displayed in Tables 1 and 2, respectively. The results are consistent to order $\delta ^{6\text{ }}$with earlier results obtained by applying different methods in Refs. \[9,22,23\]. Further, we display the results for the energy eigenvalues of $2s,$ $2p,$ $3s,$ $3p,$ and $3d$ states in Tables 3 and 4. Our results are then compared with accurate energy eigenvalues obtained by other authors. Thus, through the comparison of our results with those of Refs. \[9,10,22,23\] for large $n$ and $\ell -$ values and small screening parameter values yields indeed excellent results.
On the other hand, we take $A=\sqrt{2}$ and $\delta =\sqrt{2}G.$ Cosequently, we compute the binding energies $(-E_{n,\ell })$ of the lowest-lying states, $1s$ to $3d,$ for various values of $\delta .$ Hence, the detailed analysis of the results in terms of various domains of parameters $A$ and $\delta $ of ECSC potential are displayed in Table 5. For further study of the bound-state energies and normalizations with analytical perturbation calculation in Table 6. We consider $A=Z=4,$ $8,$ $16,$ 24 in order to cover the range of low to high atomic numbers. For low strength of $A=Z,$ the energy eigenvalues nobtained are in good agreement with the other methods for low values of the screening parameter $\delta .$ Obviously, when $\delta $ is small the Coulomb field character prevails and the method has been adjusted to that. However, the results become gradually worse as $A$ and/or $\delta $ are large.
Concluding Remarks {#CR}
==================
We have shown that the bound-state energies of the exponential cosine screened Coulomb (ECSC) potential for all eigenstates can be accurately determined within the framework of a new approximation formalism. Avoiding the disadvantages of the standard non-relativistic perturbation theories, the present formulae have the same simple form both for ground and excited states and provide, in principle, the calculation of the perturbation corrections up to any arbitrary order in analytical or numerical form.
Additionally, the application of the present technique to ECSC potential is really of great interest leading to analytical expressions for both energy eigenvalues and wave functions. Comparing various energy levels with different works in the literature we find that this treatment is quite reliable and further analytical calculations with this non-perturbative scheme would be useful. In particular, the method becomes more reliable as the potential strength increases.
The author wishes to dedicate this work to his son Musbah for his love and assistance.
P. Anderson, Phys. Rev. 86, 694 (1952); R. Kubo, Phys. Rev. 87, 568 (1952).
R. A. Ferrell and D. J. Scalapino, Phys. Rev. A 9, 846 (1974); A. J. Bray, J. Phys. A 7, 2144 (1974); E. Brezin, J. Phys. A 12, 759 (1979).
V. L. Bonch-Bruevich and V. B. Glasko, Sov. Phys. Dokl. 4, 147 (1959).
N. Takimoto, J. Phys. Soc. Jpn. 14, 1142 (1959).
V. L. Bonch-Bruevich and Sh. M. Kogan, Sov. Phys. Solid State 1,1118 (1960) C. Weisbuch and B. Vinter, Quantum Semiconductor Heterostructures (Academic Press, New York, 1993); P. Harrison, Quantum Wells, Wires and Dots (John Wiley and Sons, England, 2000).
E. P. Prokopev, Sov. Phys. Solid State 9, 993 (1967).
R. Dutt and Y. P. Varshni, Z. Phys. D 2, 207 (1986).
R. Dutt, A. Ray and P. P. Ray, Phys. Lett. A 83, 65 (1981); C. S. Lam and Y. P. Varshni, Phys. Rev. A 4, 1875 (1971); D. Singh and Y. P. Varshni, Phys. Rev. A 28, 2606 (1983).
H. de Meyer[* et al*]{}., J. Phys. A 18, L 849 (1985).
C. S. Lai, Phys. Rev. A 26, 2245 (1982).
R. Dutt [*et al*]{}., J. Phys. A 18, 1379 (1985).
C. S. Lam and Y. P. Varshni, Phys. Rev. A 6, 1391 (1972).
V. L. Eletsky, V. S. Popov, and V. M. Weinberg, Phys. Lett. A 84, 235 (1981).
A. D. Dolgov and V. S. Popov, Phys. Lett. B 79, 403 (1978).
Y. Aharanov and C. K. Au, Phys. Rev. Lett. 42, 1582 (1979)
P. P. Ray and A. Ray, Phys. Lett. B 78, 443 (1981).
C. S. Lai, Phys. Rev. A 23, 455 (1981).
J. D. Hirschfelder, J. Chem. Phys. 33, 1462 (1960).
J. Killingbeck, Phys. Lett. A 65, 87 (1978).
M. Grant and C. S. Lai, Phys. Rev. A 20, 718 (1979).
C. S. Lai, Phys. Rev. A 26, 2245 (1982).
R. Sever and C. Tezcan, Phys. Rev. A 35, 2725 (1987).
S. M. Ikhdair and R. Sever, Z. Phys. D 28,1 (1993).
B. Gönül, K. Köksal and E. Bakir, \[arXiv:quant-ph/0507098\]; B. Gönül, Chin. Phys. Lett. 21, 1685 (2004).
B. Gönül, Chin. Phys. Lett. 21, 2330 (2004); B. Gönül and M. Koçak, Mod. Phys. Lett. A 20, 355 (2005); B. Gönül, N. Çelik and E. Olğar, Mod. Phys. Lett. A 20, 1683 (2005); B. Gönül and M. Koçak, \[arXiv: quant-ph/0412161\]; ibid. \[arXiv: quant-ph/0503055\].
F. Cooper, A. Khare and U. P. Sukhatme, Phys. Rep. 251, 267 (1995).
M. Zonjil, J. Math. Chem. 26, 157 (1999).
M. Alberg and L. Wilets, Phys. Lett. A 286, 7 (2001).
D. J. Doren and D. R. Herschbach, Phys. Rev. A 34, 2665 (1986).
H. A. Bethe and E. E. Salpeter, Quantum Mechanics of One- and Two-Electron Atoms (springer, Berlin, 1957).
L. S. Gradshteyn and I. M. Ryzhik, Tables of Integrals, Series and Products (Academic, New York, 1965).
C. Lee, Phys. Lett. A 267, 101 (2000).
----------- ---------------- -------------------- ---------------------- --------------------
$\delta $ $1/N$ \[22\] Dynamical \[9\] Shifted $1/N$ \[23\] $E_{n,\ell }$
0.01 $-0.490$ $001$ $-0.490$ $001$ $0$ $-0.490$ $000$ $9$
0.02 $-0.480$ $008$ $-0.480$ $007$ $8$ $-0.480$ $007$ $83$ $-0.480$ $007$ $8$
0.03 $-0.470$ $026$ $-0.470$ $026$ $0$ $-0.470$ $025$ $9$
0.04 $-0.460$ $061$ $-0.460$ $060$ $9$ $-0.460$ $061$ $01$ $-0.460$ $060$ $8$
0.05 $-0.450$ $117$ $-0.450$ $117$ $4$ $-0.450$ $117$ $2$
0.06 $-0.440$ $200$ $-0.440$ $200$ $4$ $-0.440$ $200$ $57$ $-0.440$ $200$ $0$
0.07 $-0.430$ $313$ $-0.430$ $313$ $4$
0.08 $-0.420$ $461$ $-0.420$ $463$ $6$ $-0.420$ $463$ $86$ $-0.420$ $461$ $7$
0.09 $-0.410$ $647$ $-0.410$ $648$ $8$
0.1 $-0.400$ $875$ $-0.400$ $883$ $9$ $-0.400$ $884$ $21$ $-0.400$ $878$ $5$
----------- ---------------- -------------------- ---------------------- --------------------
: Comparison of bound energy eigennvalues for $0\leq \protect\delta
\leq 0.1$ for the $1s$ state in atomic units.
----------- ---------------- -------------------- ---------------------- --------------------
$\delta $ $1/N$ \[22\] Dynamical \[9\] Shifted $1/N$ \[23\] $E_{n,\ell }$
0.01 $-0.115$ $013$ $-0.115$ $013$ $5$ $-0.115$ $013$ $4$
0.02 $-0.105$ $103$ $-0.105$ $103$ $6$ $-0.105$ $103$ $61$ $-0.105$ $103$ $3$
0.03 $-0.095$ $334$ $-0.095$ $336$ $6$ $-0.095$ $334$ 6
0.04 $-0.085$ $755$ $-0.085$ $769$ $0$ $-0.085$ $769$ $59$ $-0.085$ $762$ $1$
0.05 $-0.076$ $406$ $-0.076$ $449$ $7$ $-0.076$ $432$ 6
0.06 $-0.067$ $311$ $-0.067$ $421$ $7$ $-0.067$ $426$ $08$ $-0.067$ 390 0
0.07 $-0.058$ $482$ $-0.058$ 680 0
0.08 $-0.049$ $915$ $-0.050$ $392$ $2$ $-0.050$ $408$ $25$ $-0.050$ 357 6
0.09 $-0.041$ $598$ $-0.042$ 494 5
0.1 $-0.033$ $500$ $-0.034$ $967$ $7$ $-0.035$ $004$ $67$ $-0.035$ 188 0
----------- ---------------- -------------------- ---------------------- --------------------
: Comparison of bound energy eigennvalues for $0\leq \protect\delta
\leq 0.1$ for the $2s$ state in atomic units.
------- ----------- ------------------- ------------------- ---------------- -------------------- ---------------------- --------------------
State $\delta $ $E[10,10]$ \[10\] $E[10,11]$ \[10\] Pertur.\[10\] Variational \[10\] Shifted \[23\] $E_{n,\ell }$
$2s$ $0.10$ $-0.034$ $941$ $-0.034$ $941$ $-0.034$ $425$ $-0.034$ $935$ $-0.035$ $004$ $67$ $-0.035$ 188 0
$2p$ $-0.032$ $469$ $-0.032$ $469$ $-0.032$ $042$ $-0.032$ $470 $ $15$ $-0.032$ 673 3
$2s$ $0.08$ $-0.050$ $387$ $-0.050$ $387$ $-0.050$ $222$ $-0.050$ $384$ $-0.050$ $408$ $25$ $-0.050$ 357 6
$2p$ $-0.048$ $997$ $-0.048$ $997$ $-0.048$ $996$ $93$ $-0.048$ $993$ $9$
$2s$ $0.06$ $-0.067$ $421$ $-0.067$ $421$ $-0.067$ $385$ $-0.067$ $421$ $-0.067$ $426$ $08$ $-0.067$ 390 0
$2p$ $-0.066$ $778$ $-0.066$ $778$ $-0.066$ $777$ $29$ $-0.066$ $761$ $1$
$2s$ $0.04$ $-0.085$ $769$ $-0.085$ $769$ $-0.085$ $767$ $-0.085$ $769$ $-0.085$ $769$ $59$ $-0.085$ $762$ 1
$2p$ $-0.085$ $591$ $-0.085$ $591$ $-0.085$ $559$ $13$ $-0.085$ $552$ $0$
$2s$ $0.02$ $-0.105$ $104$ $-0.105$ $104$ $-0.105$ $104$ $-0.105$ $104$ $-0.105$ $103$ $61$ $-0.105$ $103$ $3$
$2p$ $-0.105$ $075$ $-0.105$ $075$ $-0.105$ $075$ $-0.105$ $074 $ $64$ $-0.105$ $074$ 4
------- ----------- ------------------- ------------------- ---------------- -------------------- ---------------------- --------------------
: Energy eigenvalues as a function of screening parameter $\protect\delta $ for the $2s$ and $2p$ states in atomic units.
------- ----------- ------------------- ------------------- ---------------- -------------------- ---------------------- --------------------
State $\delta $ $E[10,10]$ \[10\] $E[10,11]$ \[10\] Pertur. \[10\] Variational \[10\] Shifted \[23\] $E_{n,\ell }$
$3s$ $0.06$ $-0.005$ $461$ $-0.005$ $462$ $-0.004$ $538$ $-0.005$ $454$ $-0.005$ $666$ $38$ $-0.007$ $077$ $8$
$3p$ $-0.004$ $471$ $-0.004$ $472$ $-0.004$ $492$ $33$ $-0.005$ 405 8
$3d$ $-0.002$ $308$ $-0.002$ $309$ $-0.002$ $313$ $56$ $-0.002$ $924$ $0$
$3s$ $0.05$ $-0.011$ $576$ $-0.011$ $576$ $-0.011$ $685$ $44$ $-0.011$ $952$ $3$
$3p$ $-0.010$ $929$ $-0.010$ $929$ $-0.010$ $538$ $-0.010$ $939 $ $85$ $-0.011$ $111$ $7$
$3d$ $-0.009$ $555$ $-0.009$ $555$ $-0.009$ $292$ $-0.009$ $555 $ $42$ $-0.009$ $694$ $0$
$3s$ $0.04$ $-0.018$ $823$ $-0.018$ $823$ $-0.018$ $707$ $-0.018$ $822$ $-0.018$ $867$ $16$ $-0.018$ $858$ $6$
$3p$ $-0.018$ $453$ $-0.018$ $453$ $-0.018$ $457$ $05$ $-0.018$ $450$ $5$
$3d$ $-0.017$ $682$ $-0.017$ $682$ $-0.017$ $682$ $08$ $-0.017$ $691$ $0$
$3s$ $0.02$ $-0.036$ $025$ $-0.036$ $025$ $-0.036$ $022$ $-0.036$ $025$ $-0.036$ $027$ $38$ $-0.036$ $021$ $3$
$3p$ $-0.035$ $968$ $-0.035$ $968$ $-0.035$ $965$ $-0.035$ $967 $ $71$ $-0.035$ $964$ $0$
$3d$ $-0.035$ $851$ $-0.035$ $851$ $-0.035$ $849$ $-0.035$ $850 $ $66$ $-0.035$ $849$ $0$
------- ----------- ------------------- ------------------- ---------------- -------------------- ---------------------- --------------------
: Energy eigenvalues as a function of screening parameter $\protect\delta $ for the $3s,$ $3p$ and $3d$ states in atomic units.
--------- ------- ------------------- ------- ------------------- ------- ------------------- ------- ------------------- ------- --------------------
G State $-E_{0,0}$ State $-E_{1,0}$ State $-E_{0,1}$ State $-E_{1,1}$ State $-E_{0,2}$
$0.002$ $1s$ $0.996$ $000$ $0$ $2s$ $0.246$ $000$ $2$ $2p $ $0.246$ $000$ $1$ $3p$ $0.107$ $112$ $0$ $3d$ $0.107$ $111$ $4 $
$0.005$ $0.990$ $000$ $2$ $0.240$ $003$ $4$ $0.240$ $002$ $4$ $0.101$ $125$ $5$ $0.101$ $116$ $0$
$0.010$ $0.980$ $001$ $9$ $0.230$ $026$ $9$ $0.230$ $019$ $3$ $0.091$ $221$ $7$ $0.091$ $147$ $5$
$0.020$ $0.960$ $015$ $6$ $0.210$ $206$ $6$ $0.210$ $148$ $9$ $0.071$ $928$ $1$ $0.071$ $361$ $7$
$0.025$ $0.950$ $030$ $2$ $0.200$ $395$ $3$ $0.200$ $285$ $7$ $0.062$ $648$ $5$ $0.061$ $566$ $5$
$0.050$ $0.900$ $234$ $4$ $0.152$ $865$ $2$ $0.152$ $099$ $1$ $0.022$ $223$ $5$ $0.014$ $137$ $4$
--------- ------- ------------------- ------- ------------------- ------- ------------------- ------- ------------------- ------- --------------------
: Energy eigenvalues of the ECSC potential in units of $\hbar =m=1,$ $A=2^{1/2}$ and $\protect\delta =GA.$
------ --------- ----- ---------------- ------ --------- ----- ----------------
$A$ $\ell $ $n$ $-E_{n,\ell }$ $A$ $\ell $ $n$ $-E_{n,\ell }$
$4$ $0$ $0$ $3.207$ $029$ $16$ $0$ $1$ $12.825$ $303$
$8$ $0$ $0$ $14.403$ $752$ $0$ $2$ $4.023$ $139$
$1$ $0$ $2.433$ $587$ $1$ $1$ $4.009$ $505$
$16$ $0$ $0$ $60.801$ $938$ $24$ $0$ $1$ $31.217$ $455$
$1$ $0$ $12.818$ $287$ $0$ $2$ $11.279$ $786$
$24$ $0$ $0$ $139.201$ $31$ $1$ $1$ $11.269$ $899$
$1$ $0$ $31.212$ $563$ $1$ $2$ $4.412$ $177$
$2$ $0$ $11.249$ $961$ $2$ $1$ $4.380$ $887$
$2$ $2$ $1.411$ $568$
------ --------- ----- ---------------- ------ --------- ----- ----------------
: Energy eigenvalues of the ECSC potential for all states in units of $\hbar =2m=1,$ and $\protect\delta =0.2$ $fm^{-1}.$
[^1]: sameer@neu.edu.tr
|
---
abstract: |
In this paper we consider bubbling solutions to the general Liouville system: $$\label{abeq1}
\Delta_g u_i^k+\sum_{j=1}^n a_{ij}\rho_j^k\left(\frac{h_j e^{u_j^k}}{\int h_j e^{u_j^k}}-1\right)=0\quad\text{in \,}M,\,\, i=1,...,n \,\, (n\ge 2)$$ where $(M,g)$ is a Riemann surface, and $A=(a_{ij})_{n\times n}$ is a constant non-negative matrix and $\rho_j^k\to \rho_j$ as $k\to \infty$. Among other things we prove the following sharp estimates.
1. The location of the blowup point.
2. The convergence rate of $\rho_j^k-\rho_j$, $j=1,..,n$.
These results are of fundamental importance for constructing bubbling solutions. It is interesting to compare the difference between the general Liouville system and the $SU(3)$ Toda system on estimates (1) and (2).
address:
- |
Department of Mathematics\
Taida Institute of Mathematical Sciences\
National Taiwan University\
Taipei 106, Taiwan
- |
Department of Mathematics\
University of Florida\
358 Little Hall P.O.Box 118105\
Gainesville FL 32611-8105
author:
- 'Chang-shou Lin'
- Lei Zhang
title: 'On Liouville systems at critical parameters, Part 1: one bubble'
---
[^1]
Introduction
============
Let $(M,g)$ be a compact Riemann surface whose volume is normalized to be $1$, $h_1,...,h_n$ be positive $C^3$ functions on $M$, $\rho_1,..,\rho_n$ be nonnegative constants. In this article we continue our study of the following Liouville system defined on $(M,g)$: $$\label{mainsys}
\Delta_g u_i+\sum_{j=1}^n\rho_j a_{ij} (\frac{h_je^{u_j}}{\int_Mh_je^{u_j}dV_g}-1)=0,\quad i\in I:=\{1,..,n\}$$ where $dV_g$ is the volume form, $A=(a_{ij})$ is a non-negative constant matrix, $\Delta_g$ is the Laplace-Beltrami operator ($-\Delta_g\ge 0$). When $n=1$ and $a_{11}=1$, equation is the mean field equation of the Liouville type: $$\label{equfromsys}
\Delta_g u+\rho\left(\frac{h e^u}{\int_M h e^udV_g}-1\right)=0\quad\text{in \,}M.$$ Therefore, the Liouville system is a natural extension of the classical Liouville equation, which has been extensively studied for the past three decades. Both the Liouville equation and the Liouville system are related to various fields of geometry, Physics, Chemistry and Ecology. For example in conformal geometry, when $\rho=8\pi$ and $M$ is the sphere $\mathbb S^2$, equation is equivalent to the famous Nirenberg problem. For a bounded domain in $\mathbb R^2$ and $n=1$, a variant of can be derived from the mean field limit of Euler flows or spherical Onsager vortex theory, as studied by Caglioti, Lions, Marchioro and Pulvirenti[@lion1; @lion2], Kiessling[@kiess], Chanillo and Kiessling[@chanillo1] and Lin [@linarch]. In classical gauge field theory, equation is closely related to the Chern-Simons-Higgs equation for the abelian case, see [@caffarelli; @hong; @jackiw; @yang]. Various Liouville systems are also used to describe models in the theory of self-gravitating systems [@aly], Chemotaxis[@childress; @keller], in the physics of charged particle beams[@bennet; @debye; @kiess2; @kiessling2], in the non-abelian Chern-Simons-Higgs theory[@dunne; @jostlinwang; @yang] and other gauge field models [@dziar; @phys; @kimleelee]. For recent developments of these subjects or related Liouville systems in more general settings, we refer the readers to [@barto1; @barto3; @ChenLin1; @chenlin2; @chenlinnew; @clwang; @CSW; @CSW1; @licmp; @linduke; @linarch; @lwy; @lwz; @linzhang1; @nolasco2; @rubinstein; @spruck; @wolansky1; @wolansky2; @zhangcmp; @zhangccm] and the references therein.
For any solution $u$ of , clearly adding any constant to $u$ gives another solution. So it is nature to assume $u\in \,{^\text{\r{}}\hspace{-.33cm}H}^1(M)$, where $${^\text{\r{}}\hspace{-.33cm}H}^1(M)=\left\{
u\in L^2(M)\,\Big|\, |\nabla_g u|\in L^2(M)\text{ and }\int_M u\,dV_g=0
\right\}.$$ Corresponding to (\[mainsys\]) we set $${^\text{\r{}}\hspace{-.33cm}H}^{1,n}=\,{^\text{\r{}}\hspace{-.33cm}H}^1(M)\times\cdots\times\,{^\text{\r{}}\hspace{-.33cm}H}^1(M)$$ to be the space for solutions. For any $\rho=(\rho_1,\cdots, \rho_n)$, $\rho_i>0 (i\in I=\{1,...,n\})$, let $\varPhi_\rho$ be a nonlinear functional defined in ${^\text{\r{}}\hspace{-.33cm}H}^{1,n}$ by $$\varPhi_\rho(u)=\frac{1}{2}\sum_{i,j\in I} a^{ij}\int_M \nabla_g u_i\cdot \nabla_g u_j dV_g-\sum_{j\in I}\rho_j\log\int_M h_j e^{u_j}dV_g$$ where $(a^{ij})_{n\times n}$ is the inverse of $A=(a_{ij})_{n\times n}$. It is easy to see that equation is the Euler-Lagrangian equation of $\varPhi_\rho$.
If the matrix $A$ satisfies the following two assumptions: $$\begin{aligned}
&& (H1):\quad \mbox{ $A$ is symmetric, nonnegative, irreducible and invertible}. \\
&&(H2): \quad a^{ii}\leq 0,\,\, \forall i\in I, \quad a^{ij}\geq 0, \,\, \forall i\ne j,
\quad \sum_{j\in I}a^{ij}\geq 0, \,\, \forall i\in I,\end{aligned}$$ the authors prove in [@linzhang2] that for $\rho$ satisfying $$\label{gammaN}
8\pi N\sum_{i\in I}\rho_i<\sum_{i,j\in I}a_{ij}\rho_i\rho_j<8\pi (N+1)\sum_{i\in I}\rho_i,$$ there is a priori estimate for all solutions $u$ to (\[mainsys\]), and the Leray-Schauder degree $d_{\rho}$ for equation (\[mainsys\]) is $$d_{\rho}=
\frac{1}{N!}\bigg ((-\chi_M+1)...(-\chi_M+N) \bigg )\quad \text{if } \mbox{ (\ref{gammaN}) holds }$$ where $\chi_M$ is the Euler characteristic of $M$. Moreover, if $\rho^k$ tends to the hyper-surface $\{\rho;\quad 8\pi N\sum_{i\in I}\rho_i=\sum_{i,j\in I}a_{ij}\rho_i\rho_j\}$, there exist exactly $N$ disjoint blowup points (see [@linzhang2]).
The proof of the a priori bound in [@linzhang2] relies on the sharp estimate for a sequence of bubbling solutions to (\[mainsys\]). Let $u^k$ be the blowup solutions corresponding to $\rho^k$ and $B(p_t,\delta_0)$ ($t=1,..,N$) be disjoint balls around distinct blowup points in $M$. Then under assumptions $(H1)$ and $(H2)$, the behavior of $u^k$ around any $p_t$ is fully bubbling, that is, the maximum values of any components of $u^k$ in any of the balls are of the same magnitude: $$\max_{B(p_t,\delta_0)} u_i^k = \max_{B(p_t,\delta_0)} u_j^k +O(1), \quad \forall i, j\in I.$$ Moreover, after a suitable scaling around each blowup point $p_t$, $u^k$ converges to an entire solution $\tilde U=(\tilde U_1,..,\tilde U_n)$ of the following Liouville system: $$\label{12apr23e2}
\left\{\begin{array}{ll}
\Delta \tilde U_i+\sum_{j=1}^n a_{ij}e^{\tilde U_j}=0,\quad \mbox{in }\,\, \mathbb R^2, \\
\\
\int_{\mathbb R^2}e^{\tilde U_i}<\infty, \quad \tilde U_i \mbox{ is radial},\,\, \forall i\in I.
\end{array}
\right.$$
One may expect the limiting entire solution to be different around each blowup point, however the authors proved that $\tilde U$ is independent of blowup points, and only depends on the ratio of $\rho_1^k-\rho_1:\rho_2^k-\rho_2:..\rho_n^k-\rho_n$ (see [@linzhang1]). Naturally it leads to the question: how to construct bubbling solutions with the help of this information?
In this paper and subsequent ones, we are devoted to study the bubbling phenomenon of Liouville systems: how to accurately estimate the bubbling solutions of (\[mainsys\]) and how to construct them. These are quite challenging analytic problems. In general, blowup analysis for a system of equations is much harder than that for the single equation. One reason is that the Pohozaev identity, a balancing condition, is no longer so powerful as in the scalar case. Another reason is that there are too many entire solutions: the parameter $\sigma=(\sigma_1,...,\sigma_n)$ ($\displaystyle{\sigma_i=\frac 1{2\pi}\int_{\mathbb R^2}e^{u_i}}$), which represents the energy of the entire solution, forms a submanifold of $n-1$ dimension. However, for the Liouville equation, the energy is just one number: $\int_{\mathbb R^2}e^u=8\pi$.
In this article we consider the case of one blowup point, and always assume $(H1)$ only. Let $\rho=(\rho_1,...,\rho_n)$ and $$\label{12jun20e1}
\Lambda_J(\rho)=8\pi \sum_{i\in J} \rho_i-\sum_{i,j\in J}a_{ij}\rho_i\rho_j$$ for any $J\subset I:=\{1,...,n\}$. Define $$\Gamma_1=\{\rho; \,\, \Lambda_I(\rho)=0\quad \mbox{ and }\Lambda_J(\rho)>0 \mbox{ for all } \emptyset \subsetneq J \subsetneq I \,\, \}.$$ Note that if $(H2)$ also holds also, then $\Lambda_I(\rho)=0$ implies $\Lambda_J(\rho)>0$ for all $J\subsetneq I$ ( see [@linzhang2]). For any $\rho$ we define $(m_1,...,m_n)$ by $$\label{12apr23e1}
m_i=\frac 1{2\pi} \sum_{j\in I} a_{ij}\rho_j.$$ The quantity $m_i$ can be interpreted by the entire solution $\tilde U$ of (\[12apr23e2\]). In fact $$\rho_i=\int_{\mathbb R^2} e^{\tilde U_i}dx, \quad i\in I,$$ and $$\label{12apr23e3}
\tilde U_i(x)=-m_i \log |x| +O(1),\quad \mbox{ for $|x|$ near infinity. }$$ The integrability of $e^{\tilde U_i}$ implies $m_i>2$ for all $i$. On the other hand $\Lambda_I(\rho)=0$ can be written as $\sum_{i\in I}(m_i-4)\rho_i=0$. Thus either $\min\{m_1,...,m_n\}<4$ or $m_i=4$ for all $i\in I$. We also note that (\[12apr23e3\]) implies that $\Gamma_1$ is a smooth submanifold because the normal vector at $\rho\in \Gamma_1$: $$(\sum_{j\in I} a_{ij} \rho_j-4\pi,...,\sum_{j\in I} a_{nj}\rho_j -4\pi),$$ has all its components positive.
The asymptotic behavior of $\tilde U_i(x)$ shows that the decay rate of $e^{\tilde U_i(x)}$ is $O(|x|^{-m})$, where $$m=\min\{m_1,...,m_n\}.$$ In this article we define $Q\in \Gamma_1$ with $m=4$, i.e. $m_i=4$ for all $i$. Thus the decay rate of $e^{\tilde U_i}$ for $\rho=Q$ is $O(|x|^{-4})$. The difference on the decay rate for $Q$ and $\rho\neq Q$ will have great effects on bubbling analysis later.
Let $u^k=(u_1^k,...,u_n^k)$ be a sequence of blow up solutions to (\[mainsys\]) with $\rho=\rho^k$ such that $\rho^k\to \rho\in \Gamma_1$. The point $Q$ defined above is of particular importance, the readers will see that in our main theorems below, the asymptotic behavior of blowup solutions, the nature of $\Lambda_I(\rho^k)$ and the location of blowup point are all significantly different depending on $\rho=Q$ or not.
Let $p$ be the blowup point of $u^k$ and $$\label{12apr23e4}
M_k=\max_{B(p,\delta)} \bigg ( u_1^k(x)-\log \int_M h_1e^{u_1^k}dV_g \bigg ),$$ $$\label{12apr23e5}
\epsilon_k=e^{-\frac 12 M_k}.$$ Since $\rho^k\to \Gamma_1$, there is only one blowup point $p$. It is easy to see that $u^k$ fully blows up at $p$ (see Lemma \[ful\]): $$\label{12apr23e6}
\max_{B(p, \delta_0)}(u_i^k(x)-\log \int_M h_i e^{u_i^k}dV_g)=M_k+O(1), \quad i\in I.$$
Our first result is on the location of the blowup point $p$. Let $p_k\to p$ be where the maximum of $\{u_1^k,..,u_n^k\}$ is attained, then we have
\[locations\] Let $\rho^k\to \rho\in \Gamma_1$ and all $\rho_i^k-\rho_i$ have the same sign.
1. If $\rho\neq Q$, then $$\label{11july13e7}
\sum_{i\in I} \bigg (\nabla (\log h_i)(p_k)+2\pi m\nabla_1 \gamma(p_k,p_k)\bigg )
\rho_{i} =O(\epsilon_k^{m-2}).$$
2. If $\rho=Q$, then $$\label{11july13e8}
\sum_i\bigg (\nabla (\log h_i)(p_k)+8\pi \nabla_1 \gamma(p_k,p_k) \bigg )\rho_{i}=O(\epsilon_k^2\log \epsilon_k^{-1}).$$ where $\nabla_1$ denotes the derivative with respect to the first variable, and $\gamma(x,y)$ stands for the regular part of the Green’s function.
Our second result is about the decay rate of $\Lambda_I(\rho^k)$. To state the result, we need to define the following quantity $D_i$ ($i\in I=\{1,...,n\}$): $$\label{leadingD}
D_i=\lim_{\delta_0\to 0}\big (\delta_0^{2-m}-\frac{m-2}{2\pi}
\int_{M\setminus B(p,\delta_0)}\frac{h_i(x)}{h_i(p)}e^{2\pi m (G(x,p)-\gamma(p,p))}dV_g \big )$$ provided that $m<4$. The limit is well defined if $m<4$, see section 6.
\[rhoto1mle4\] Suppose $\rho^k\to \rho\in \Gamma_1$ and $\rho\neq Q$, if all $\rho_i^k-\rho_i$ have the same sign, then $$\Lambda_I(\rho^k)
=8\pi^2\sum_{i\in I_1}(e^{c_i}D_i+o(1))\epsilon_k^{m-2},$$ where $I_1$ is a subset of $I$ where $m_i=m$ for all $i\in I_1$, $c_i$ is a constant determined in (\[12jun29e1\]). $o(1)\to 0$ as $k\to \infty$.
If $M$ is a flat torus with fundamental cell domain $\Omega\subset \mathbb R^2$, then $D_i$ can be written as $$D_i=\frac{m-2}{2\pi }\bigg (
\int_{\mathbb R^2\setminus \Omega} \frac{1}{|x-p|^m}dx-\lim_{\delta_0\to 0}\int_{\Omega\setminus B_{\delta_0}}
\frac{H_i(x,p)}{|x-p|^m}dx\bigg )$$ where $$H_i(x,p)=\frac{h_i(x)}{h_i(p)}e^{2\pi m(\gamma(x,p)-\gamma(p,p))}-1, \quad i\in I.$$ See [@ccl] and [@clwang] for related discussions.
The assumption that all $\rho_i^k-\rho_i$ have the same sign seems due to some technical difficulties. When $n=2$ and $\rho\neq Q$, this assumption is not needed for both Theorem \[locations\] and Theorem \[rhoto1mle4\].
\[rhotome4\] Suppose $\rho^k\to \rho\in \Gamma_1$ and $\rho=Q$. If all $\rho_i^k-Q_i$ have the same sign, then $$\label{12apr23e10}
\Lambda_I(\rho^k)
=-16\pi^2(\sum_{i\in I} b_i e^{c_i}+o(1))\epsilon_k^2\log \epsilon_k^{-1}.$$ where $$b_i=\frac 14\bigg (\Delta \, \log h_i(p)-2K(p)+8\pi+|\nabla \, \log h_i(p)+8\pi \nabla_1\gamma(p,p)|^2\bigg ),$$ $c_i$ is determined in (\[12jun29e1\]).
Important information on bubbling solutions can be observed on the two cases: $\rho\neq Q$ and $\rho=Q$. Theorem \[rhoto1mle4\], which is on $\rho\neq Q$, has its leading term in $\Lambda_I(\rho^k)$ involved with global information of the manifold, while the leading term in Theorem \[rhotome4\], which corresponds to $\rho=Q$, only depends on the geometric information at the blowup point. Moreover, the error terms in Theorem \[rhoto1mle4\] and Theorem \[rhotome4\] respectively also indicate the different asymptotic behaviors of blowup solutions near the singularity. All these differences in the two cases will lead to separate strategies in the construction of bubbling solutions in forthcoming works.
Since Liouville systems and Toda systems share a lot of common features, it is informative to compare our main theorems with the ones for the $SU(3)$ Toda system. First, the location of the blowup point in Theorem \[locations\] is a critical point of a combination of $\log h_i$, $\rho_i$ and $\gamma$ ($\nabla_1 \gamma$ vanishes if the Riemann surface has constant curvature). However for the $SU(3)$ Toda system, the blowup point $p$ is a critical point of both $\log h_1$ and $\log h_2$, i.e. $p$ satisfies (see [@lwz]) $$\nabla h_1(p)=\nabla h_2(p)=0.$$ Second, for the $SU(3)$ Toda system, the convergence rate of $\rho_i^k-\rho_i$ is estimated to be $$\rho_i^k-\rho_i=(e^{\tilde c_i}b_i+o(1))\epsilon_k^2\log \epsilon_k^{-1},$$ where $b_i$ is the term in (\[12apr23e10\]). Nevertheless our result in (\[12apr23e10\]) is again a combination of the $b_i$s. The comparison of the results reflects some major difference between the Toda system and our Liouville system:
1. The dimension of kernel space of the linearized operator at an entire solution is $8$ for $SU(3)$ Toda system, and is $3$ for our Liouville system.
2. The set $\Gamma:=\{(\rho_1,...,\rho_n);\quad
\rho_i=\int_{\mathbb R^2}e^{u_i},\quad (u_1,..,u_n) \mbox{ is an entire solution}\,\, \}$ is only a point for $SU(3)$ Toda system, while for the Liouville system it is a $(n-1)$ dimensional manifold.
As far as the blowup analysis is concerned, our Liouville system has disadvantages in both respects, as the kernel space is too small and $\Gamma$ is too large. For a sequence of bubbling solutions, it is extremely difficult to pin-point suitable approximating solutions from $\Gamma$, because at the beginning, the local energy of bubbling solutions could be estimated in some rough way. This rough estimate of the local energy leads to a small perturbation of global bubbling solutions. This perturbation on global solutions, albeit small, has a non-negligible effect on the approximation of blowup solutions. This difficulty is particularly evident when we study bubbling solutions with multiple blowup points in [@linzhang4]. Therefore our method to obtain those sharp results is different from ones in Chen-Lin [@ChenLin1] for the mean field equations and Lin-Wei-Zhao [@lwz] for the $SU(3)$ Toda system (The methods in [@ChenLin1] and [@lwz] are similar).
The organization of the paper is as follows. In section two we first prove a uniqueness theorem for globally defined linearized Liouville systems. This result plays a central role for the delicate blowup analysis in sections three to five. The main idea of the proof uses a monotonicity property of solutions and we introduce a way to use maximum principles suitable for Liouville systems. In the second part of section two, we study the asymptotic behavior of global solutions to the Liouville system on $\mathbb R^2$ and obtain some Pohozaev identities. In section three and section four we obtain a sharp expansion result for blowup solutions around a blowup point. Then in section five, for local equations we use Pohozaev identity to determine the locations of blowup points. Then in section six we return to the equation on manifold and compute the leading term for $\rho^k\to \rho$ in both situations ($\rho\neq Q$ or $\rho=Q$) and complete the proofs of the main theorems.
[**Acknowledgement**]{} Part of the paper was finished when the second author was visiting Taida Institute for Mathematical Sciences (TIMS) in March 2011 and June 2012. He is very grateful to TIMS for their warm hospitality. He also would like to thank the National Science Foundation for the partial support (NSF-DMS-1027628).
Preliminary results on the Liouville systems
============================================
In this section we prove two theorems on the Liouville systems with the matrix $A$ satisfying $(H1)$. They are important for the blowup analysis and the computation of the leading terms of $\rho\to \Gamma_1$ in this paper and $\rho\to \Gamma_N$ in the forthcoming work [@linzhang4].
A uniqueness theorem for the linearized system of $n$ equations
---------------------------------------------------------------
In the first subsection we prove a uniqueness theorem for the linearized system of $n$ equations.
\[uniqlin\] Let $A$ be a matrix that satisfies (H1), $u=(u_1,..,u_n)$ be a radial solution of $$\left\{\begin{array}{ll}
-\Delta u_i=\sum_{j=1}^na_{ij}e^{u_j}\,\, \mbox{in}\,\, \mathbb R^2,\quad i\in I=\{1,..,n\}\\
\int_{\mathbb R^2}e^{u_i}<\infty.
\end{array}
\right.$$ Suppose $\phi=(\phi_1,..,\phi_n)$ satisfies $$-\Delta \phi_i=\sum_{j=1}^na_{ij}e^{u_j}\phi_j,\quad i\in I.$$
1. $$\label{1027e1}
|\phi_i(x)|\le C(1+|x|)^{\tau},\quad x\in \mathbb R^2,$$
for some $\tau\in (0,1)$ and $$\phi_i(0)=0,\quad i\in I.$$ Then there exist $c_1,c_2\in \mathbb R$ such that $$\phi_i(x)=c_1u_i'(r)\frac{x_1}r+c_2u_i'(r)\frac{x_2}r \quad \mbox{in}\quad \mathbb R^2, \quad i\in I.$$
2. If $
|\phi_i(x)|\le C$ for all $x\in \mathbb R^2$, then there exist $c_0,c_1,c_2\in \mathbb R$ such that $$\phi_i(x)=c_0(ru_i'(r)+2)+c_1u_i'(r)\frac{x_1}r+c_2u_i'(r)\frac{x_2}r,\quad \mathbb R^2 ,\quad i\in I.$$
3. If $\phi_i(x)=O(|x|^2)$ near $0$ and $|\phi_i(x)|\le C(1+|x|)^{2-\epsilon_0}$ for some $\epsilon_0>0$, then $\phi_i\equiv 0$.
Before the proof of Theorem \[uniqlin\] we first establish
\[1129lem1\] Let $A$ and $u$ be as in Theorem \[uniqlin\], let $\Phi=(\Phi_1,..,\Phi_n)$ solve $$\label{12e1}
\left\{\begin{array}{ll}
\Phi_i''(r)+\frac 1r\Phi_i'(r)-\frac 1{r^2}\Phi_i(r)+\sum_{j=1}^na_{ij}e^{u_j}\Phi_j=0,\quad 0<r<\infty, \\
\\
|\Phi_i(r)|\le Cr/(1+r)^{\epsilon_0} \quad \mbox{ for some }\epsilon_0\in (0,1), \quad \forall i\in I.
\end{array}
\right.$$ Then there exists a constant $C$ such that $\Phi_i(r)=Cu_i'(r)$ for $i\in I$.
[**Proof of Lemma \[1129lem1\]:**]{}
The proof is in two steps. First we show that under the assumption of $\Phi_i$ at infinity we have the following sharper decay estimate: $$\label{phidecay}
|\Phi_i(r)|\le Cr(1+r)^{-2},\quad 0<r<\infty,\quad i\in I.$$
Indeed, let $\bar \psi_i(r)=\Phi_i(r)/r$. By direct computation we see that $\bar \psi=(\bar \psi_1,...,\bar \psi_n)$ satisfies $$\label{12e2}
\bar \psi_i''(r)+\frac 3r\bar \psi_i'(r)+\sum_ja_{ij}e^{u_j}\bar \psi_j=0, \quad 0<r<\infty.$$ Clearly in order to show (\[phidecay\]) we only need to show $|\bar \psi_i(r)|\le Cr^{-2}$ for $r>1$ under the assumption that $|\bar \psi_i(r)|\le Cr^{-\epsilon_0}$ for $i\in I$ and $r>1$. Let $$\tilde \psi_i(t)=\bar \psi_i(e^t)\quad \mbox{ and }\quad \tilde u_i(t)=u_i(e^t)+2t,$$ it is easy to see that $\tilde \psi(t)=(\tilde \psi_1(t),..,\tilde
\psi_n(t))$ satisfies $$\label{psidc1}
\tilde \psi_i''(t)+2\tilde \psi_i'(t)+\sum_{j\in I}a_{ij}e^{\tilde u_j(t)}\tilde \psi_j(t)=0,\quad -\infty<t<\infty$$ and our goal is to show $$\label{psidc2}
\tilde \psi_i(t)=O(e^{-2t})$$ knowing $\tilde \psi_i(t)=O(e^{-\epsilon_0t})$ for $t>0$, $i\in I$. Set $$l_i=\frac 1{2\pi}\sum_{j\in I}a_{ij}\int_{\mathbb R^2}e^{u_j},\quad l=\min\{l_1,..,l_n\}.$$ By Lemma \[globalcpt\] below $l>2$. Let $$h_i(t)=-\sum_{j\in I}a_{ij}e^{\tilde u_j(t)}\tilde \psi_j(t)=O(e^{(2-l-\epsilon_0)t}),\quad t>0.$$ Then $$\tilde \psi_i(t)=C_0+C_1 e^{-2t}+\frac 12\int_0^th_i(s)ds-\frac 12e^{-2t}\int_0^te^{2s}h_i(s)ds.$$ Using the asymptotic rate of $h_i(t)$ at infinity we further have $$\tilde \psi_i(t)=(C_0+\frac 12\int_0^{\infty}h_i(s)ds)+C_1e^{-2t}+O(e^{(2-l-\epsilon_0)t}).$$ Since $\tilde \psi_i(t)$ tends to $0$ as $t$ goes to infinity we know $\tilde \psi_i(t)=O(e^{-2t})$ if $l+\epsilon_0\ge 4$, in which case (\[psidc2\]) is established. Otherwise we obtain $\tilde \psi_i(t)=O(e^{(2-l-\epsilon_0)t})$. In the latter case, we apply the same procedure to obtain a better decaying rate of $\tilde \psi_i(t)$ at infinity. After finite steps, (\[psidc2\]) is established.
In the second step we complete the proof of the Lemma \[1129lem1\]. By way of contradiction we suppose there is a solution $\Phi=(\Phi_1,..,\Phi_n)$ that satisfies (\[12e1\]) and $\Phi$ is not a multiple of $f=(u_1'(r),..,u_n'(r))$. Let $\psi_i^0=-u_i'(r)/r$, then clearly both $\psi^0=(\psi_1^0,..,\psi_n^0)$ and $\bar \psi=(\bar \psi_1,..,\bar \psi_n)$ satisfy (\[12e2\]). We verify by direct computation that $$\int_0^re^{u_j}\psi_j^0(\sigma)\sigma^3ds=-\int_0^r(e^{u_j})'\sigma^2d\sigma
=-e^{u_j(r)}r^2+2\int_0^re^{u_j}\sigma d\sigma>0.$$ Note that the last inequality is justified by $u'_i(r)<0$ for $r>0$ and $i\in I$. Also, since $e^{u_i}\le Cr^{-2-\delta}$ for some $\delta>0$ and $r>1$, $\displaystyle{\int_0^{\infty}e^{u_i}\psi_i^0(\sigma)\sigma^3d\sigma<\infty}$. Based on the computation above we set $$S=\bigg \{ t|\, \psi_i^t=\psi_i^0+t(\psi_i^0-\bar \psi_i) \, \mbox{and }\,
\int_0^re^{u_i}\psi_i^t(\sigma )\sigma^3d\sigma>0,\,\, \forall r>0,\,\, i\in I\, \bigg \}.$$ Let $\bar \psi=(\bar \psi_1,..,\bar \psi_n)$ and $\psi^0=(\psi_1^0,..,\psi_n^0)$. We first observe that $\psi_i^0(0)>0$. Suppose $\bar \psi \not\equiv \psi^0$, we can assume that $\bar \psi_1(0)\neq 0$ and $|\bar \psi_i(0)|<\psi_i^0(0)$ for all $i\in I$.
From the definition of $S$ we immediately see that $0\in S$. Moreover, since $|\bar \psi_i(r)|\le Cr^{-2}$ near infinity, we can choose $|t|<\delta$ with $\delta$ small so that all $|t|<\delta$ belong to $S$. Another immediate observation is that $S$ has a lower bound. Indeed, for $T$ sufficiently negative, $\psi_1^T(0)<0$, which is impossible to have $\int_0^re^{u_1}\psi_1^T(\sigma)\sigma^3d\sigma>0$ for $r$ small.
Let $T=\inf S$ and let $t_m\to T^+$. The sequence $\psi^{t_m}_i$ obviously converges to a function $\psi_i$, which is just $\psi_i^0+T(\psi_i^0-\bar \psi_i)$. $\psi=(\psi_1,..,\psi_n)$ satisfies the property $$\label{12e4}
\int_0^re^{u_i(\sigma)}\psi_i(\sigma)\sigma^3d\sigma\ge 0,\quad \forall r>0,\quad i=1,..,n.$$ On the other hand, from the behavior of $\bar \psi$ and $\psi^0$ at infinity (both are $O(r^{-2})$) we immediately observe that $$\int_0^{\infty}e^{u_i(r)}\psi_i(r)r^3dr<\infty.$$ In regard to (\[12e2\]) we have $$\label{12e3}
r^3\psi_i'(r)=-\int_0^r\sum_{j}a_{ij}e^{u_j}\psi_j\sigma^3 d\sigma, \quad 0<r<\infty.$$ From (\[12e4\]) and (\[12e3\]) we see that $\psi_i$ is non-increasing. Since we have known that $|\psi_i(r)|\le Cr^{-2}$ near infinity we have $$\label{12e6}
\psi_i(r)\ge 0, \quad \forall r>0, \quad i\in I.$$
It is not possible to have all $\psi_i(0)=0$ because this implies $\psi_i\equiv 0$, a contradiction to the assumption that $\bar \psi$ is not a multiple of $\psi^0$. Therefore without loss of generality we assume $\psi_1(0)>0$. Then we further claim that $\psi_i$ is strictly decreasing for all $i\in I$. Indeed, let $I_1=\{ j\in I| \, a_{1j}>0 \, \}$, for each $j\in I_1$ we use (\[12e6\]) and $\psi_1(0)>0$ to obtain $$r^3\psi_j'(r)\le -\int_0^ra_{1j}e^{u_1}\psi_1\sigma^3d\sigma<0, \quad 0<r<\infty.$$ Therefore for each $j\in I_1$, $\psi_j$ is strictly decreasing, which immediately implies that $\psi_j(0)>0$. We can further define $I_2=\{ i\in I| \, a_{ij}>0 \mbox{ for some }j\in I_1. \}$. Then the same argument shows that $\psi_i$ is strictly decreasing for each $i\in I_2$ as well. Since the matrix $A=(a_{ij})_{n\times n}$ is irreducible, this process exhausts all $i\in I$.
(\[12e3\]) yields $\psi_i'(r)\le -Cr^{-3}$ for $r>1$ and $i\in I$. Then by using $\lim_{r\to \infty}\psi_i(r)=0$ we further have $$\label{12e5}
\psi_i(r)\ge Cr^{-2},\quad r\ge 1.$$ Then it is easy to see that for $t=T-\epsilon$ with $\epsilon>0$ small, we also have $$\int_0^re^{u_i}\psi_i^t(\sigma)
\sigma^3d\sigma>0,\quad \mbox{for all } r>0,$$ a contradiction to the definition of $T$. Lemma \[1129lem1\] is proved. $\Box $
[**Proof of Theorem \[uniqlin\]:**]{} We first prove the third statement. The following function plays an important role: Let $f=(f_1,..,f_n)=(u_1',..,u_n')$. Then $$\label{85e1}
-\Delta f_i=\sum_ja_{ij}e^{u_j}f_j-\frac 1{r^2}f_i,\quad i=1,..,n.$$
Let $\phi^k=(\phi_1^k,..,\phi_n^k)$ be defined as $$\phi_i^k(r)=\frac 1{2\pi}\int_0^{2\pi}\phi_i(r\cos \theta,r\sin \theta)\cos k\theta d\theta,\quad i\in I.$$ Then $\phi^k$ satisfies $$\label{85e2}
-\Delta \phi_i^k=\sum_ja_{ij}e^{u_j}\phi_j^k-\frac{k^2}{r^2}\phi_i^k,\quad i\in I, \quad k=2,..$$ Clearly $\phi_i^k(r)=o(r)$ near $0$ and $\phi_i^k(r)=O(r^{2-\epsilon_0})$ at $\infty$. We claim that $$\label{11mar29e1}
\phi_i^k(r)\equiv 0,\quad \forall k\ge 2,\quad \mbox{provided that }\,\, \phi_i^k(r)=O(r^{k-1+\tau}), \,\, r>1,\,\, k\ge 2.$$ Note that the growth condition in (\[11mar29e1\]) is weaker than what is assumed in the assumption in Theorem \[uniqlin\].
The argument below also applies if $\phi$ is projected on $\sin k\theta$. First we show that $\phi_i^k=o(r^{-1})$ as $r\to \infty$. Indeed, using $\phi_j^k(x)\le C|x|^{k-1+\tau}$ we write $\sum_ja_{ij}e^{u_j}\phi^k_j$ as $O(r^{k-1+\tau-2-\delta_0})$ (for some $\delta_0>0$). Let $g(t)=\phi_i^k(e^t)$, then from (\[85e2\]) $g(t)$ satisfies $$g''(t)-k^2g(t)=h(t),\quad t\in \mathbb R$$ where $$\label{11mar29e2}
h(t)=O(e^{(k-1+\tau-\delta_0)t})\quad t>0.$$ Let $g_1(t)=e^{kt}$ and $g_2(t)=e^{-kt}$ be two fundamental solutions of the homogeneous equation, a general solution $g(t)$ is of the form $$g(t)=c_1g_1(t)+c_2g_2(t)-\frac{g_1(t)}{2k}\int_0^tg_2(s)h(s)ds+\frac {g_2(t)}{2k}\int_0^tg_1(s)h(s)ds$$ where $c_1,c_2$ are constants. Using (\[11mar29e2\]) in the above we obtain $$g(t)=c'_1g_1(t)+c'_2g_2(t)+O(e^{(k-1+\tau-\delta_0)t}),\quad \mbox{for }t>1$$ where $c'_1,c'_2$ are two constants. Since $g(t)=O(e^{(k-1+\tau )t})$ for $t\to \infty$, we see that $c'_1=0$ and therefore $g(t)=O(e^{(k-1+\tau-\delta_0)t})$ as $t\to \infty$. Equivalently $$\label{11mar29e3}
\phi_i^k(r)=O(r^{k-1+\tau-\delta_0}) \quad i\in I.$$ With (\[11mar29e3\]) we further obtain $$\sum_ja_{ij}e^{u_j}\phi^k_j=O(r^{k-1+\tau-2-2\delta_0}).$$ Consequently $\phi_i^k(x)=O(r^{k-1+\tau-2\delta_0})$. Keep doing this for finite steps we obtain that $\phi_i^k$ decays faster than $r^{-1}$ at infinity. The asymptotic theory of ODE can be similarly used to show that $\phi_i^k(r)=o(r)$ as $r\to 0$.
To get a contradiction, without loss of generality, we may assume that some of $\phi_i^k$, say $\phi_1^k(r)>0$ for some $r>0$ and $$\max_{\mathbb R^+}\bigg (\frac{\phi_1^k(r)}{f_1(r)}\bigg )
=\max_{1\le j\le n}\bigg (\max_{\mathbb R^+}\frac{\phi_j^k(r)}{f_j(r)}\bigg ).$$ By noting $\phi_1^k(r)=o(r)$ as $r\to 0$ and $\phi_1^k(r)=o(\frac 1r)$ as $r\to \infty$, $\phi_1^k/
f_1(r)$ attains its maximum at some point $r_0\in \mathbb R^+$. Let $w_1(r)=\phi_1^k(r)/f_1(r)$. By a direct computation, $w_1(r)$ satisfies $$\label{1019e1}
\Delta w_1+2\nabla w_1\cdot \frac{\nabla f_1}{f_1}+\frac{1-k^2}{r^2}w_1=\sum_{j=2}^na_{1j}e^{u_j}
\bigg (\frac{w_1f_j-\phi_j}{f_1}\bigg ).$$ Now we apply the maximum principle at $r=r_0$, and obtain $$\Delta w_1(r_0)\le 0,\quad \mbox{and}\quad \nabla w_1(r_0)=0.$$ Since $k>1$, (\[1019e1\]) yields $$\label{1019e2}
\sum_{j=2}^na_{1j}e^{u_j}\bigg (\frac{w_1f_j-\phi_j^k}{f_1}\bigg )(r_0)<0$$ because $w_1(r_0)>0$. On the other hand, for $j\ge 2$, $$w_1(r_0)f_j(r_0)-\phi_j^k(r_0)=f_j(r_0)(\frac{\phi_1^k(r_0)}{f_1(r_0)}-\frac{\phi_j^k(r_0)}{f_j(r_0)})\ge
0,$$ which obviously contradicts (\[1019e2\]). Therefore (\[11mar29e1\]) is established. When $k=1$, $\phi_i^1\equiv 0$ because by Lemma \[1129lem1\], $\phi_i^1(r)=Cu_i'(r)$. By the assumption $\phi_i^1(x)=O(|x|^2)$ near $0$, $C=0$. The third statement of Theorem \[uniqlin\] is established.
Again by Lemma \[1129lem1\] the first statement of Theorem \[uniqlin\] is established.
Finally, the second statement of Theorem \[uniqlin\] is an immediate consequence of Lemma 3.1 of [@linzhang1]. Theorem \[uniqlin\] is established. $\Box$
A Pohozaev identity for global solutions
----------------------------------------
\[globalcpt\] Let $u=(u_1,...,u_n)$ be an entire, radial solution of $$\left\{\begin{array}{ll}-\Delta u_i=\sum_{j=1}^na_{ij}e^{u_j}, \quad \mbox{in}\quad \mathbb R^2, \\
\int_{\mathbb R^2}e^{u_i}<\infty
\end{array}
\right.$$ where $A$ is a constant matrix that satisfies $(H1)$. Let $$c_i=u_i(0)+\frac 1{2\pi}\int_{\mathbb R^2}\log |\eta |(\sum_{j=1}^n a_{ij}e^{u_j(\eta)})d\eta,$$ $$\sigma_{i}=\frac 1{2\pi}\int_{\mathbb R^2}e^{u_i},\quad l_i=\sum_{j=1}^na_{ij}\sigma_{j},\,\, l=\min\{l_{1},...,l_{n}\}$$ and $$\sigma_{iR}=\frac 1{2\pi}\int_{B_R}e^{u_i}.$$ Then for some $\delta_0>0$ $$\label{12apr29e1}
e^{u_i(r)}=e^{c_i}r^{-l_i}(1+o(r^{-\delta_0})), \quad r>1,$$ $$\label{1105e1}
4\sum_{i\in I}\sigma_{iR}=\sum_{i,j\in I}a_{ij}\sigma_{iR}\sigma_{jR}+2\sum_{i\in I}e^{c_{i}}R^{2-l_{i}}+O(R^{2-l-\delta_0}).$$
[**Proof of Lemma \[globalcpt\]:**]{}
It is well known that $$\label{11may27e2}
u_i(x)=-\frac 1{2\pi}\int_{\mathbb R^2}\log |x-\eta |(\sum_{j=1}^n a_{ij}e^{u_j(\eta)})d\eta+c_i.$$ Indeed, let $w_i$ be the function defined by the right hand side of (\[11may27e2\]). Then $w_i-u_i$ is a harmonic function. Since they both have logarithmic growth at infinity, $w_i-u_i=c$. Evaluating both functions at $0$ we have $c=c_i$.
Clearly $$\label{12apr24e1}
u_i(x)+l_{i}\log |x|=-\frac 1{2\pi}\int_{\mathbb R^2}\log \frac{|x-\eta |}{|x|}\sum_ja_{ij}e^{u_j(\eta)}d\eta+c_{i}.$$ Using $\sum_j a_{ij}e^{u_j(r)}=O(r^{-2-\delta_0})$ for some $\delta_0>0$ and $r$ large, we obtain, by elementary estimates, $$u_i(r)=-l_i \log r+c_i+o(r^{-\delta_0}),$$ which leads to $$\label{11may30e1}
\sigma_{i}=\sigma_{iR}+\frac{e^{c_{i}}}{l_{i}-2}R^{2-l_{i}}+O(R^{2-l_{i}-\delta}).$$ We arrive at (\[1105e1\]) by using (\[11may30e1\]) in the Pohozaev identity for $\sigma$: $$4\sum_{i} \sigma_{i}=\sum_{i,j}a_{ij}\sigma_{i}\sigma_{j}.$$ Lemma \[globalcpt\] is established. $\Box$.
First order estimates {#blowup}
=====================
Let $(h_1^k,...,h_n^k)$ be a family of positive, $C^3$ functions on $B_1$ with a uniform bound on their positivity and $C^3$ norm: $$\label{hikn}
\frac 1C\le h_i^k(x)\le C,\quad \|h_i^k\|_{C^3(B_1)}\le C, \quad x\in B_1, \quad i=1,..,n.$$ In the next three sections we consider a sequence of locally defined, fully blown-up solutions $u^k=(u_1^k,...,u_n^k)$ and we shall derive their precise asymptotic behavior near their singularity and the precise location of their singularity. Here we abuse the notation $u^k=(u_1^k,...,u_n^k)$ and it is independent of the one used in the introduction.
Specifically we assume that $u^k$ satisfies the following equation in $B_1$, the unit ball: $$\label{uik}
-\Delta u_i^k=\sum_{j=1}^na_{ij}h_j^ke^{u_j^k},\quad i=1,..,n,\quad x\in B_1,$$ with $0$ being the only blowup point in $B_1$: $$\max_Ku_i^k\le C(K), \quad \forall K\subset\subset \bar B_1\setminus \{0\}, \,\, \mbox{ and } \max_{B_1}u_i^k\to \infty,$$ with bounded oscillation on $\partial B_1$: $$\label{88e1}
|u_i^k(x)-u_i^k(y)|\le C,\forall x,y\in \partial B_1,\quad C\mbox{ independent of }k$$ and uniformly bounded energy: $$\label{89e1}
\int_{B_1}e^{u_i^k}\le C,\quad C \mbox{ is independent of } k.$$ Finally we assume that $u^k$ is a fully blown-up sequence, which means when re-scaled according its maximum, $\{u_k\}$ converges to a system of $n$ equations: Let $u_1^k(0)=\max_{B_1}u_1^k$ and $\epsilon_k=e^{-\frac 12u_1^k(0)}$, and $$\label{vikdef}
v_i^k(y)=u_i^k(\epsilon_ky)-u_1^k(0),\quad y\in \Omega_k:=B(0,\epsilon_k^{-1}).$$ Then $v^k=(v_1^k,...,v_n^k)$ converges in $C^2_{loc}(\mathbb R^2)$ to $v=(v_1,..,v_n)$, which satisfies $$\label{817e1}
\left\{\begin{array}{ll}-\Delta v_i=\sum_{j}a_{ij}\tilde h_j(0)e^{v_j},\quad \mathbb R^2,\quad i=1,..,n\\
\\
\int_{\mathbb R^2}e^{v_i}<\infty,\quad i=1,..,n,\quad v_1(0)=0,
\end{array}
\right.$$ where $\tilde h_j(0)=\lim_{k\to \infty}h_j^k(0)$.
For the rest of the paper we set $$m_i:=\frac{1}{2\pi}\int_{\mathbb R^2}\sum_{j=1}^na_{ij} \tilde h_j(0)e^{v_j}>2,\quad i\in I$$ and $m=\min\{m_1,...,m_n\}$. In this section we derive a first order estimate for $v^k$ in $\Omega_k$. In [@linzhang1] the authors prove that there is a sequence of global radial solutions $U^k=(U_1^k,..,U_n^k)$ of (\[817e1\]) such that $$\label{11mar9e1}
|u_i^k(\epsilon_ky)-U_i^k(y)|\le C,\quad \mbox{for } |y|\le r_0\epsilon_k^{-1}.$$ From (\[11mar9e1\]) we have the following spherical Harnack inequality: $$\label{11mar9e3}
|u_i^k(\epsilon_ky)-u_i^k(\epsilon_k y')|\le C$$ for all $|y|=|y'|=r\le r_0\epsilon_k^{-1}$ and $C$ is a constant independent of $k,r$. (\[11mar9e3\]) will play an essential role in the first order estimate. To improve (\[11mar9e1\]) we introduce $\phi_i^k$ to be a harmonic function: $$\label{816e5}
\left\{\begin{array}{ll}
-\Delta \phi_i^k=0,\quad B_1,\\
\\
\phi_i^k=u_i^k-\frac 1{2\pi}\int_{\partial B_1}u_i^k,\quad \mbox{on}\quad \partial B_1.
\end{array}
\right.$$ Obviously $\phi_i^k(0)=0$ by the mean value theorem and $\phi_i^k$ is uniformly bounded on $B_{1/2}$ because of (\[88e1\]). Later in section 6, when the results in section 3,4,5 will be used to prove the main theorems, the function $\phi_i^k$ will be specified when we consider the system defined on Riemann surface.
Let $V^k=(V_1^k,..,V_n^k)$ be the radial solutions of $$\label{88e2}
\left\{\begin{array}{ll}-\Delta V_i^k=\sum_{j=1}^na_{ij}h_j^k(0)e^{V_j^k}\quad \mathbb R^2,\quad i\in I\\ \\
V_i^k(0)=v_i^k(0),\quad i\in I
\end{array}
\right.$$ where $v_i^k$ is defined in (\[817e1\]). It is easy to see that any radial solution $V$ of (\[88e2\]) exists for all $r>0$ and $e^{V_i}\in L^2(\mathbb R^2)$. The main result of this section is to prove that $V_i^k(y)+\phi_i^k(\epsilon_ky)$ is the first order approximation to $v_i^k(y)$.
\[thm2\] Let $A$, $u^k$, $h^k=(h_1^k,..,h_n^k)$ ,$\phi_i^k$ and $v^k$ be described as above. Then for any $\delta>0$, there exist $k_0(\delta)>1$ and $C$ independent of $k$ and $\delta$ such that for all $k\ge k_0$, $$\begin{aligned}
\label{11mar9e2}
&&|D^{\alpha}(v_i^k(y)-V_i^k(y)-\phi_i^k(\epsilon_ky))|\\
&\le &\left\{\begin{array}{ll}C\epsilon_k(1+|y|)^{3-m+\delta-|\alpha|},\quad m\le 3,\\
C\epsilon_k(1+|y|)^{\delta-l},\quad m>3,
\end{array}
\right.
\quad |y|<\epsilon_k^{-1}/2,\quad |\alpha|=0,1,2. \nonumber\end{aligned}$$
\[11def1\] $$\sigma_i^k=\frac 1{2\pi}\int_{\mathbb R^2} h_i^k(0)e^{V_i^k},\,\, m_i^k=\sum_{j=1}^n a_{ij}\sigma_j^k,\,\, m^k=\min \{m_1^k,..,m_n^k\}.$$ From Theorem \[thm2\] it is easy to see that $\lim_{k\to\infty}m_i^k=m_i$. Thus $m_i^k\ge 2+\delta_0$ for some $\delta_0>0$ independent of $k$.
To prove Theorem \[thm2\], we have $$\label{1020e1}
-\Delta (v_i^k(y)-\phi_i^k(\epsilon_ky))=\sum_ja_{ij}H_j^k(\epsilon_ky)e^{v_j^k(y)-\phi_j^k(\epsilon_ky)},\,\, \mbox{ in } \Omega_k\,\, \big( \mbox{ see (\ref{vikdef})} \big )$$ where $$\label{Hik}
H_i^k(\cdot)=h_i^k(\cdot)e^{\phi_i^k(\cdot)}.$$ Since $\phi_i^k(0)=0$ we have $H_i^k(0)=h_i^k(0)$. Also, the definition of $\phi_i^k$ implies that $v_i^k-\phi_i^k(\epsilon_k\cdot)$ is a constant on $\partial \Omega_k$.
To estimate the error term $w_i^k=v_i^k-\phi_i^k(\epsilon_k \cdot )-V_i^k$. We find $w_i^k$ satisfies $$\label{wik01}
\left\{\begin{array}{ll}
\Delta w_i^k(y)+\sum_ja_{ij}H_j^k(\epsilon_ky)e^{\xi_j^k}w_j^k=-\sum_ja_{ij}(H_j^k(\epsilon_ky)-H_j^k(0))e^{V_j^k}, \\ \\
w_i^k(0)=0,\quad i\in I,\quad \nabla w_1^k(0)=O(\epsilon_k),
\end{array}
\right.$$ where $\xi_i^k$ is defined by $$\label{1211e1}
e^{\xi_i^k}=\int_0^1e^{tv_i^k+(1-t)V_i^k}dt.$$
Since both $v^k$ and $V^k$ converge to $v$, $w_i^k=o(1)$ over any compact subset of $\mathbb R^2$. The first estimate of $w_i^k$ is the following
\[1020lem1\] $$\label{820e1}
w_i^k(y)=o(1)\log (1+|y|)+O(1),\quad \mbox{ for } y\in \Omega_k.$$
[**Proof**]{}: By (\[11mar9e3\]) $$|v_i^k(y)-\bar v_i^k(|y|)|\le C, \quad \forall y\in \Omega_k$$ where $\bar v_i^k(r)$ is the average of $v_i^k$ on $\partial B_r$: $$\bar v_i^k(r)=\frac{1}{2\pi r}\int_{\partial B_r}v_i^k.$$ Thus we have $e^{v_i^k(y)}=O(r^{-2-\delta_0})$ and $e^{V_i^k(y)}=O(r^{-2-\delta_0})$ where $r=|y|$ and $\delta_0>0$. Then $$r(\bar w_i^k)'(r)=\frac 1{2\pi }\bigg (\int_{B_r}\sum_ja_{ij}H_j^k(\epsilon_k\cdot)e^{v_j^k}
-\int_{B_r}\sum_ja_{ij}h_j^k(0)e^{V_j^k}\bigg )$$
It is easy to use the decay rate of $e^{v_i^k}$, $e^{V_i^k}$ and the closeness between $v_i^k$ and $V_i^k$ to obtain $$r(\bar w_i^k)'(r)=o(1),\quad r\ge 1.$$ Hence $\bar w_i^k(r)=o(1)\log r$ and (\[820e1\]) follows from this easily. Lemma \[1020lem1\] is established. $\Box$
The following estimate is immediately implied by Lemma \[1020lem1\]: $$e^{\xi_i^k(y)}\le C(1+|y|)^{-m+o(1)}\quad \mbox{ for } y\in \Omega_k=B(0,\epsilon_k^{-1}).$$
Before we derive further estimate for $w_i^k$ we establish a useful estimate for the Green’s function on $\Omega_k$ with respect to the Dirichlet boundary condition:
\[greenk\] Let $G(y,\eta)$ be the Green’s function with respect to Dirichlet boundary condition on $\Omega_k$. For $y\in \Omega_k$, let $$\begin{aligned}
\Sigma_1&=&\{\eta \in \Omega_k;\quad |\eta |<|y|/2 \quad \}\\
\Sigma_2&=&\{\eta \in \Omega_k;\quad |y-\eta |<|y|/2 \quad \}\\
\Sigma_3&=&\Omega_k\setminus (\Sigma_1\cup \Sigma_2).\end{aligned}$$ Then in addition for $|y|>2$, $$\label{1020e5}
|G(y,\eta)-G(0,\eta)|\le \left\{\begin{array}{ll}
C(\log |y|+|\log |\eta ||),\quad \eta\in \Sigma_1,\\
C(\log |y|+|\log |y-\eta ||),\quad \eta\in \Sigma_2,\\
C|y|/|\eta |,\quad \eta \in \Sigma_3.
\end{array}
\right.$$
[**Proof:**]{} The expression for $G(y,\eta)$ is $$G(y,\eta)=-\frac{1}{2\pi}\log |y-\eta |+\frac 1{2\pi}\log (\frac{|y|}{\epsilon_k^{-1}}|\frac{\epsilon_k^{-2}y}{|y|^2}-\eta |), \quad y,\eta\in \Omega_k.$$ In particular $$G(0,\eta)=-\frac{1}{2\pi}\log |\eta |+\frac 1{2\pi}\log \epsilon_k^{-1},\quad \eta\in \Omega_k.$$ Therefore we write $G(y,\eta)-G(0,\eta)$ as $$\label{1020e3}
G(y,\eta)-G(0,\eta)=\frac 1{2\pi}\log \frac{|\eta |}{|y-\eta |}
+\frac 1{2\pi}\log |\frac{y}{|y|}-\frac{|y|\eta}{\epsilon_k^{-2}}|.$$
The proof of (\[1020e5\]) for $\eta\in \Sigma_1$ is obvious. For $\eta \in \Sigma_2$, (\[1020e5\]) also obviously holds if either $|y|$ or $|\eta |$ is less than $\frac 78 \epsilon_k^{-1}$ because in this case $$\bigg |\log \big | \frac{y}{|y|}-\epsilon_k^2|y|\eta \big |\bigg |\le C.$$ Consequently $$\begin{aligned}
|G(y,\eta)-G(0,\eta)|&\le &C(\log |\eta |+\big |\log |y-\eta |\big |+C)\\
&\le &C(\log |\eta |+\big |\log |y-\eta |\big |).\end{aligned}$$ Therefore for $\eta\in \Sigma_2$ we only need to consider the case when $|y|,|\eta|>\frac 78\epsilon_k^{-1}$. In this case it is immediate to observe that $$\bigg |\log |\frac{y}{|y|}-\epsilon_k^2|y|\eta |\bigg |<C, \quad \mbox{ if }\,\, \angle (\frac{y}{|y}, \frac{\eta}{|\eta |})>\frac{\pi}8$$ where $\angle(\cdot,\cdot)$ is the angle between two unit vectors. Thus for $\eta\in \Sigma_2$ we only consider the situation when $|y|,|\eta|>\frac 78\epsilon_k^{-1}$, $\angle(\frac{y}{|y|},\frac{\eta}{|\eta |})<\frac{\pi}8$. For this case we estimate $G(y,\eta)-G(0,\eta)$ as follows: $$|G(y,\eta)-G(0,\eta)|\le |G(y,\eta)|+|G(0,\eta)|$$ $$|G(0,\eta)|\le C\log |y|.$$ $$\begin{aligned}
|G(y,\eta)|&\le &\frac 1{2\pi}\big |\log |y-\eta |\big |+\frac 1{2\pi}\log \frac 87+\frac 1{2\pi}\big |\log |\frac{\epsilon_k^{-2}y}{|y|^2}-\eta |\big |\\
&\le & C(\log |y|+\big |\log |y-\eta |\big |)\end{aligned}$$ where the last inequality holds because $$|y-\eta |\le |\frac{\epsilon_k^{-2}y}{|y|^2}-\eta |<C|y|,$$ which implies $$\big |\log |\frac{\epsilon_k^{-2}y}{|y|^2}-\eta | \big |\le C(\big |\log |y-\eta |\big |+\log |y|).$$ The second case of (\[1020e5\]) (when $\eta\in \Sigma_2$) is proved.
For $\eta\in \Sigma_3$, we first consider when $|\eta|>2|y|$. In this case $$|\log \frac{|\eta |}{|\eta -y|}|=|\log |\frac{\eta}{|\eta |}-\frac{y}{|\eta |}||\le C\frac{|y|}{|\eta |}.$$ For the second term, since $\eta,y\in \Omega_k$ and $|\eta |>2|y|$, we have $|y||\eta |<\frac 12 \epsilon_k^{-2}$, consequently $$\bigg |\log |\frac{y}{|y|}-\frac{|y|\eta }{\epsilon_k^{-2}}| \bigg |\le C|y||\eta |\epsilon_k^2\le C\frac{|y|}{|\eta |}.$$ So (\[1020e5\]) is proved in this case. Now we consider $\frac{|y|}2\le |\eta |\le 2|y|$ and $|\eta -y|\ge \frac{|y|}2$. For the first term we have $$\bigg |\log \frac{|\eta |}{|y-\eta |}\bigg |\le \log 4\le C\frac{|y|}{|\eta |}.$$ For the second term, we want to show $$\label{1027e5}
\bigg |\log |\frac{y}{|y|}-\frac{|y|\eta }{\epsilon_k^{-2}}| \bigg |\le C\le C\frac{|y|}{|\eta |}.$$ If either $|y|\le \frac{15}{16}\epsilon_k^{-1}$ or $|\eta |\le \frac{15}{16}\epsilon_k^{-1}$ we have $$|\frac{y}{|y|}-\frac{|y|\eta}{\epsilon_k^{-2}}|\ge \frac{1}{16},$$ therefore (\[1027e5\]) obviously holds. For $\frac{15}{16}\epsilon_k^{-1}<|y|, |\eta |\le \epsilon_k^{-1}$, using $|y-\eta |>\frac 12|y|$ we obtain easily $$\bigg |\frac{y}{|y|}-\frac{|y|\eta}{\epsilon_k^{-2}}\bigg |\ge \frac 38.$$ Therefore (\[1020e5\]) is proved in all cases. Lemma \[greenk\] is established. $\Box$
[**Proof of Theorem \[thm2\]:**]{} First we prove (\[11mar9e2\]) for $\alpha=0$. We consider the case $m\le 3$, the proof for the case $m>3$ is similar. By way of contradiction, we assume $$\Lambda_k:=\max_{y\in\Omega_k}\frac{\max_{i\in I}|w_i^k(y)|}{\epsilon_k(1+|y|)^{3+\delta-m}}\to \infty.$$ Suppose $\Lambda_k$ is attained at $y_k\in \bar \Omega_k$ for some $i_0\in I$. We thus define $$\bar w_i^k(y)=\frac{w_i^k(y)}{\Lambda_k\epsilon_k(1+|y_k|)^{3+\delta-m}}.$$ Here we require $\delta$ to be small so that $m-2-\delta>0$ (Thus $3-m+\delta<1$). It follows from the definition of $\Lambda_k$ that for $y\in \Omega_k$ $$\label{11apr20e1}
|\bar w_i^k(y)|=\frac{|w_i^k(y)}{\Lambda_k\delta_k(1+|y|)^{3+\delta-m}}\frac{(1+|y|)^{3+\delta-m}}{(1+|y_k|)^{3+\delta-m}}\le \frac{(1+|y|)^{3+\delta-m}}{(1+|y_k|)^{3+\delta-m}}.$$ The equation for $\bar w_i^k$ is $$\label{11apr20e2}
-\Delta \bar w_i^k(y)=\sum_ja_{ij}h_j^k(0)e^{\xi_j^k}\bar w_j^k+o(1)\frac{(1+|y|)^{1-m}}{(1+|y_k|)^{3+\delta-m}},
\quad \Omega_k$$ for $i\in I$. Here $\xi_i^k$ is given by (\[1211e1\]). $\xi_i^k$ converges to $v_i$ in $C^2_{loc}(\mathbb R^2)$. Besides, we also have $\bar w_i^k(0)=0$ for all $i$ and $\nabla \bar w_1^k(0)=o(1)$. If a subsequence of $y_k$ stays bounded, then along a subsequence $\bar w^k=(\bar w_1^k,..,\bar w_n^k)$ converges to $\bar w=(\bar w_1,...,\bar w_n)$ that satisfies $$\left\{\begin{array}{ll}
-\Delta \bar w_i=\sum_ja_{ij}h_j(0)e^{v_j}\bar w_j,\quad \mathbb R^2, \quad i\in I,\\
\bar w_i(0)=0,\quad \nabla \bar w_1(0)=0,\quad |\bar w_i(y)|\le C(1+|y|)^{3+\delta-m},\quad y\in \mathbb R^2.
\end{array}
\right.$$ Thanks to (1) of Theorem \[uniqlin\] $$\bar w_i(x)=c_1\frac{\partial v_i}{\partial x_1}+c_2\frac{\partial v_i}{\partial x_2}.$$ Since $\nabla \bar w_{1}(0)=0$ we have $c_1=c_2=0$, thus $\bar w_i\equiv 0$ for all $i$. On the other hand, the fact that $\bar w_{i_0}^k(y_k)=\pm 1$ for some $i_0\in I$ implies that $\bar w_{i_0}(\bar y)=\pm 1$ where $\bar y$ is the limit of $y_k$. This contradiction means that $y_k\to \infty$. Next we shall show a contradiction if $|y_k|\to \infty$. By the Green’s representation formula for $\bar w_i^k$, $$\bar w_i^k(y)=\int_{\Omega_k}G(y,\eta)(-\Delta \bar w_i^k(\eta))d\eta + \bar w_i^k|_{\partial \Omega_k}$$ where $\bar w_i^k|_{\partial \Omega_k}$ is the boundary value of $\bar w_i$ on $\partial \Omega_k$ ( which is a constant). From (\[11apr20e1\]) and (\[11apr20e2\]) we have $$|-\Delta \bar w_i^k(\eta)|\le \frac{C(1+|\eta |)^{3+\delta-2m}}{(1+|y_k|)^{3+\delta-m}}+
\frac{C(1+|\eta |)^{1-m+\delta}}{\Lambda_k(1+|y_k|)^{3+\delta-m}}.$$
Thus for some $i\in I$ we have $$\begin{aligned}
\label{11apr20e3}
&&1=|\bar w_i^k(y_k)-\bar w_i^k(0)|\\
&\le &C\int_{\Omega_k}|G(y_k,\eta)-G(0,\eta)|\bigg (\frac{(1+|\eta |)^{3+\delta-2m}}{(1+|y_k|)^{3+\delta-m}}
+\frac{(1+|\eta |)^{1-m+\delta}}{\Lambda_k(1+|y_k|)^{3+\delta-m}}\bigg ),\nonumber\end{aligned}$$ where the constant on the boundary is canceled out. To compute the right hand side of the above, we decompose the $\Omega_k$ as $\Omega_k=\Sigma_1\cup \Sigma_2\cup \Sigma_3$ as in Lemma \[greenk\]. Using (\[1020e5\]) we have $$\int_{\Sigma_1\cup \Sigma_2}|G(y_k,\eta)-G(0,\eta)|(1+|\eta |)^{3+\delta-2m}d\eta=O(1)(\log |y_k|)(1+|y_k|)^{(5+\delta-2m)_+}$$ where $$(1+|y_k|)^{\alpha_+}=\left\{\begin{array}{ll}(1+|y_k|)^{\alpha}, &\quad \alpha>0, \\
\log (1+|y_k|), &\quad \alpha=0, \\
1, &\quad \alpha<0.
\end{array}
\right.$$ $$\int_{\Sigma_3}|G(y_k,\eta)-G(0,\eta)|(1+|\eta |)^{3+\delta-2m}d\eta=O(1)(1+|y_k|)^{5+\delta-2m}.$$ Hence $$\int_{\Omega_k}|G(y_k, \eta)-G(0,\eta)|\frac{(1+|\eta |)^{3+\delta-2m}}{(1+|y_k|)^{3+\delta-m}}d\eta
=O(1)(1+|y_k|)^{2-m}.$$ Similarly we can compute the other term: $$\int_{\Omega_k}|G(y_k, \eta)-G(0, \eta)|\frac{(1+|\eta |)^{1-m+\delta}}{\Lambda_k(1+|y_k|)^{3+\delta-m}}d\eta=O(1)\Lambda_k^{-1}(\log (1+|y_k|))^{-\frac{\delta}2}.$$ By the computations above we see that the right hand side of (\[11apr20e3\]) is $o(1)$, a contradiction to the left hand side of (\[11apr20e3\]). Thus (\[11mar9e2\]) is established for $\alpha=0$. The estimates for $|\alpha|=1$ and $2$ follow easily by scaling and standard elliptic estimates. Therefore Theorem \[thm2\] is completely proved. $\Box$
Second order estimates
======================
In this section we improve the estimates in Theorem \[thm2\] for $m<4$ and $m=4$, respectively. Let $p_{i,k}$ be the maximum point of $v_i^k(\cdot)-\phi_i^k(\epsilon_k\cdot)$. The following lemma estimates the location of $p_{i,k}$.
\[pik\] $p_{i,k}=O(\epsilon_k),\quad i\in I. $
[**Proof :**]{} Applying Theorem \[thm2\] to $v_i^k-\phi_i^k(\delta \cdot )$ on $B_1$: $$\begin{aligned}
\label{11mar12e2}
&&D^{\alpha}(v_i^k(y)-\phi_i^k(\epsilon_ky))\\
&=&D^{\alpha}(V_i^k(|y|))+O(\epsilon_k),\quad |y|<1,\quad |\alpha |=0,1,2. \nonumber\end{aligned}$$ The equation for $V_i^k$ is $$(V_i^k)''(r)+\frac 1r(V_i^k)'(r)+\sum_{j=1}^na_{ij}H_j^k(0)e^{V_j^k(r)}=0,\quad r>0.$$ From $(V_i^k)'(0)=0$, we see $\lim_{r\to 0}(V_i^k)'(r)/r=(V_i^k)''(0)$. Thus $$\label{11apr4e2}
(V_i^k)''(0)=-\frac 12\sum_{j=1}^na_{ij}H_j^k(0)e^{V_j^k(0)}<-C$$ for some $C>0$ independent of $k$. Since $p_{i,k}$ is the maximum point of $v_i^k(\cdot)-\phi_i^k(\epsilon_k\cdot)$, we deduce from (\[11mar12e2\]) that $(V_i^k)'(|p_{i,k}|)=O(\epsilon_k)$, thus from (\[11apr4e2\]) we have $p_{i,k}=O(\epsilon_k)$. Lemma \[pik\] is established. $\Box$
The main result in this section is to find the $\epsilon_k$ approximation to $v_i^k(\cdot)-\phi_i^k(\epsilon_k\cdot)$. It is most convenient to write the expansion around one of the $p_{i,k}$s. We choose $p_{1,k}$ and shall use $\Phi^k=(\Phi_1^k,..,\Phi_n^k)$ to denote the projection of $v_i^k(\cdot)-\phi_i^k(\epsilon_k\cdot)$ onto $span\{\sin \theta, \cos\theta\}$. i.e. $$\label{11aug11e1}
\Phi^k_i(r\cos\theta,r\sin\theta)=\epsilon_k(G_{1,i}^k(r)\cos \theta+G_{2,i}^k(r)\sin\theta), \quad i\in I$$ with $G_{t,i}^k(r)$ ($t=1,2$) satisfying some ordinary differential equations to be specified later.
Set $v^{1,k}=(v^{1,k}_1,..,v^{1,k}_n)$ as $$\label{vi0kd}
v^{1,k}_i(\cdot)=v_i^k(\cdot+p_{1,k})-\phi_i^k(\epsilon_k(\cdot+p_{1,k}))$$ in $$\label{omegai0kd}
\Omega_{1,k}:=\{\eta;\,\, \eta+p_{1,k}\in \Omega_k\,\, \}.$$ Using $\nabla v_i^k(0)=O(\epsilon_k)$ (by Theorem \[thm2\]) and $\phi_i^k(0)=0$ we observe that $$\begin{aligned}
\label{11mar12e1}
&&v^{1,k}_i(0)=v_i^k(p_{1,k})-\phi_i^k(\epsilon_k p_{1,k})\\
&=&v_i^k(0)+\nabla v_i^k(0)\cdot p_{1,k}+O(|p_{1,k}|^2)+O(p_{1,k}\epsilon_k)\nonumber\\
&=&v_i^k(0)+O(\epsilon_k^2) \nonumber\end{aligned}$$ The equation that $v^{1,k}$ satisfies is (combining (\[11mar9e2\]) and (\[1020e1\])) $$\label{vpke}
\left\{\begin{array}{ll}
\Delta v_i^{1,k}+\sum_{j=1}^na_{ij}H_j^{1,k}(y)e^{v_j^{1,k}}=0,\quad \mbox{in}\,\, \Omega_{1,k}\\
\\
\nabla v_1^{1,k}(0)=0, \quad \nabla v_i^{1,k}(0)=O(\epsilon_k),\quad i=2,...,n
\end{array}
\right.$$ where $H^{1,k}=(H^{1,k}_1,...,H^{1,k}_n)$ is defined by (see (\[Hik\])) $$\label{hi0k}
H_i^{1,k}(\cdot)=H_i^k(\epsilon_k\cdot+\epsilon_kp_{1,k})=h_i^k(\epsilon_k\cdot +\epsilon_kp_{1,k})e^{\phi_i^k(\epsilon_k\cdot +\epsilon_kp_{1,k})}.$$ Trivially $$\label{12feb16e1}
H_i^{1,k}(0)=h_i^k(0)+O(\epsilon_k^2).$$
In the coordinate around $p_{1,k}$ and we seek to approximate $v_i^{1,k}$ in $\Omega_{1,k}$. The first term in the approximation of $v^{1,k}$ is $V^k$. Here we note that the domain $\Omega_{1,k}$ is shifted from the ball $\Omega_k$ by $p_{1,k}$.
We shall use five steps to establish an approximation of $v_i^{1,k}$ without distinguishing $m=4$ or not.
[**Step one:**]{}
Let $w^{1,k}=(w^{1,k}_1,..,w^{1,k}_n)$ be the difference between $v^k$ and $V^k$: $$w^{1,k}_i(y)=v^{1,k}_i(y)-V^{k}_i(|y|), \quad y\in \Omega_{1k}.$$ Taking the difference between (\[vpke\]) and (\[88e2\]), we have $$\Delta w_i^{1,k}+\sum_j a_{ij}H_j^{1,k}(y)e^{V_j^k+w_j^{1,k}}-\sum_j a_{ij}h_j^k(0)e^{V_j^k}=0,$$ which is $$\Delta w_i^{1,k}+\sum_j a_{ij}h_j^k(0)e^{V_j^k}(\frac{e^{w_j^{1,k}}H_j^{1,k}(y)}{h_j^k(0)}-1)=0.$$ Here we observe that the oscillation of $V_i^k $ on $\partial \Omega_k$ is $O(\epsilon_k^2)$. Indeed, recall that $\Omega_{1,k}$ is the shift of the large ball $\Omega_k$ by $p_{1,k}$. Let $y_1,y_2\in
\partial \Omega_{1,k}$, one can find $y_3$ such that $|y_3|=|y_2|$ and $|y_3-y_1|\le C\epsilon_k$. Since $(V_i^{k})'(r)\sim r^{-1}$ for $r>1$ and $|y_1|\sim \epsilon_k^{-1}$, we have $$\label{12jun19e1}
V_i^{,k}(y_1)-V_i^{k}(y_2)=V_i^{k}(y_1)-V_i^{k}(y_3)=O(\epsilon_k^2).$$
With (\[12jun19e1\]) we further write the equation for $w_i^{1,k}$ as $$\label{newwk}
\left\{\begin{array}{ll}
\Delta w_i^{1,k}+ \sum_ja_{ij}h_j^k(0)e^{V_j^{k}}w_j^{1,k} =E_i^k,\quad \Omega_{1,k}.\\
\\
w_i^{1,k}(0)=O(\epsilon_k^2),\,\, i\in I,\quad \bar w_i^{1,k}=O(\epsilon_k^2) \mbox{ on } \partial \Omega_{1,k},\\
\\
\nabla w_{1}^{1,k}(0)=0,\quad \nabla w_i^{1,k}(0)=O(\epsilon_k),\quad i\in I\setminus \{1\}
\end{array}
\right.$$ where $$\label{ewik}
E_i^k=-\sum_ja_{ij}h_j^k(0)e^{V_j^{k}}\bigg (\frac{H_j^{1,k}(y)}{h_j^k(0)}e^{w_j^{1,k}}-1-w_j^{1,k}\bigg ).$$ Similar to Theorem \[thm2\] we also have
\[wi0ki\] For any $\delta>0$, there exists $k_0(\delta)>1$ such that for some $C>0$ independent of $k$ and $\delta$, the following estimate holds for all $k\ge k_0$: $$\label{wi0k}
|w^{1,k}_i(y)|\le \left\{\begin{array}{ll}
C\epsilon_k(1+|y|)^{3-m+\delta},\,\, m\le 3,\\
C\epsilon_k(1+|y|)^{\delta},\,\, m> 3,
\end{array}
\right.
\quad y\in \Omega_{1,k}.$$
[**Proof:**]{} Using the definition of $v^{1,k}$ and Theorem \[thm2\] we have $$|v_i^{1,k}(y)-V_i^k(y+p_{1,k})|\le \left\{\begin{array}{ll}
C\epsilon_k (1+|y|)^{3-m+\delta},\quad m\le 3, \\
C\epsilon_k (1+|y|)^{\delta},\quad m>3.
\end{array}
\right.$$ On the other hand we clearly have $$|V_i^k(y)-V_i^k(y+p_{1,k})|\le C\epsilon_k(1+|y|)^{-1}$$ by mean value theorem and the estimate of $\nabla V_i^k$. Lemma \[wi0ki\] is established. $\Box$
Using Lemma \[wi0ki\] and (\[12feb16e1\]) we now rewrite $E_i^k$, clearly $$E_i^k=-\sum_ja_{ij}e^{V_j^{k}}(H_j^{1,k}(y)-h_j^k(0))+(H_j^k(y)-h_j^k(0))w_j^{1,k}+O((w_j^{1,k})^2).$$ By Lemma \[wi0ki\] and (\[12feb16e1\]), the last two terms are $O(\epsilon_k^2(1+|y|)^{2-m})$ regardless whether $m\ge 3$ or not. Thus $$\begin{aligned}
\label{ewik1}
E_i^k&=& -\sum_ja_{ij}e^{V_j^{k}}(H_j^{1,k}(y)-h_j^k(0))+O(\epsilon_k^2)(1+|y|)^{2-m}\\
&=&-\sum_ja_{ij}e^{V_j^{k}}(H_j^{1,k}(y)-H_j^{1,k}(0))+O(\epsilon_k^2)(1+|y|)^{2-m} \nonumber\end{aligned}$$ where in the last step we used (\[12feb16e1\]) again.
[**Step Two: Estimate of the radial part of $w^{1,k}$:**]{}
Let $g^{k,0}=(g_1^{k,0},...,g^{k,0}_n)$ be the radial part of $w^{1,k}$: $$g_i^{k,0}(r)=\frac 1{2\pi}\int_0^{2\pi}w_i^{1,k}(r\cos \theta, r\sin \theta)d\theta.$$ Due to the radial symmetry of $V_i^k$, $g^{k,0}$ satisfies $$\label{11apr8e1}
\left\{\begin{array}{ll}
L_i g^{k,0}
=-\frac{\epsilon_k^2}4\sum_ja_{ij}\Delta H_j^{1,k}(0)r^2e^{V_j^{k}}
+O(\epsilon_k^2)(1+r)^{\delta-m}\\
\\
g_i^{k,0}(0)=O(\epsilon_k^2),\quad i\in I, \quad \frac{d}{dr}g_{1}^{k,0}(0)=0.
\end{array}
\right.$$ where ( for simplicity we omit $k$ in $L_i$) $$L_ig^{k,0}=\frac{d^2}{dr^2}g_i^{k,0}+\frac 1r\frac{d}{dr}g_i^{k,0}+\sum_ja_{ij}h_j^{k}(0)e^{V^{k}_j}g_j^{k,0}$$ We claim that for $\delta>0$, there exists $k_0(\delta)>1$ such that for all $k\ge k_0$ $$\label{11mar29e4}
|g_i^{k,0}(r)|\le C\epsilon_k^2(1+r)^{4-m+\delta},\quad 0<r<\epsilon_k^{-1}$$ holds for some $C$ independent of $k$ and $\delta$. So, $g_i^{k,0}(r)$ can be discarded as an error term.
To prove (\[11mar29e4\]), we first observe that $m\le 4$ and by (\[11apr8e1\]) $$|L_ig^{k,0}|\le C\epsilon_k^2(1+r)^{2-m+\delta/2}.$$ Let $f^k=(f_1^k,..,f_n^k)$ be the solution of $$\left\{\begin{array}{ll}
\frac{d^2}{dr^2}f_i^k+\frac 1r\frac{d}{dr}f_i^k=L_ig^{k,0}, \\
\\
f_i^k(0)=\frac{d}{dr}f_i^k(0)=0.
\end{array}
\right.$$ Then elementary estimate shows $$\label{12feb13e1}
|f_i^k(r)|\le C\epsilon_k^2(1+r)^{4-m+\delta}.$$ If $m>\frac 52$, we claim $$\label{11apr5e1}
|g_i^{k,0}(r)-f_i^k(r)|\le C\epsilon_k^2(1+r)^{\delta}, \quad 0<r<\epsilon_k^{-1}.$$
Indeed, let $\bar g^{k}=g^{k,0}-f^{k}$, then clearly $$\left\{\begin{array}{ll}
L_i\bar g^k=F_i^k, \\
\\
\bar g_i^k(0)=O(\epsilon_k^2), \quad \frac{d}{dr}\bar g_{1}^k(0)=0,
\end{array}
\right.$$ where $$F_i^k:=-\sum_ja_{ij}h_j^{k}(0)e^{V^{k}_j}f_j^k=O(\epsilon_k^2)(1+r)^{4-2m+\delta}.$$
By considering $\frac{d^2}{dr^2}+\frac 1r\frac{d}{dr}$ as $\Delta $ in $\mathbb R^2$ and $\bar g^k$ a solution with boundary oscillation $0$ in $B(0,\epsilon_k^{-1})$, we obtain (\[11apr5e1\]) by the argument for Theorem \[thm2\]. Here we note that to apply Theorem \[uniqlin\], it is essential to require $6-2m+\delta<1$ (it holds if $m>5/2$), the estimate on the Green’s function in Lemma \[greenk\] and the condition $\frac{d}{dr}\bar g_{1}^k(0)=0$. Since $m\le 4$, $O(\epsilon_k^2)(1+r)^{\delta}$ is part of the error.
If $m\le \frac 52$ we apply the same ideas by adding more correction functions to $g^{k,0}$: Let $N$ satisfy $2+(2-m)N<1$, we add $N$ correcting functions to make the right hand side of the equation of the order $O(\epsilon_k^2(1+r)^{(2-m)N+\delta})$. Note that each correction can be discarded as an error in the sense that they are smaller than the right hand side of (\[11jun14e5\]). Using $2+2N-mN<1$ and the argument in the proof of Theorem \[thm2\] we obtain (\[11mar29e4\]).
[**Step Three: The projection on $\sin\theta$ and $\cos \theta$**]{}
In this step we consider the projection of $w^{1,k}$ over $\cos\theta$ and $\sin \theta$, respectively: $$\epsilon_kG_{1,i}^k(r)=\frac 1{2\pi}\int_0^{2\pi}w_i^{1,k}(r,\theta)\cos\theta d\theta, \quad
\epsilon_kG_{2,i}^k(r)=\frac 1{2\pi}\int_0^{2\pi}w_i^{1,k}(r,\theta)\sin\theta d\theta.$$ Let $$\label{1224e2}
\Phi_i^k=\epsilon_kG_{1,i}^k(r)\cos \theta +\epsilon_kG_{2,i}^k(r) \sin \theta$$ clearly $G_{1,i}^k$ and $G_{2,i}^k$ solve the following linear systems: For $0<r<\epsilon_k^{-1}$ and $t=1,2$ $$\begin{aligned}
\label{1030e1}
(\frac{d^2}{dr^2}+\frac 1r\frac{d}{dr}-\frac{1}{r^2})G^k_{t,i}+\sum_j a_{ij}h_j^k(0)e^{V_j^k}G^k_{t,j}\\
=-\sum_ja_{ij}\partial_tH_j^{1,k}(0)re^{V_j^{k}}
+O(\epsilon_k)(1+r)^{2-m}. \nonumber\end{aligned}$$
Then $\Phi^k$ solves $$\begin{aligned}
&&\Delta \Phi_i^k+\sum_ja_{ij}h_j^k(0)e^{V_j^{k}}\Phi_j^k\\
\label{1224e3}
&=&-\epsilon_k\sum_ja_{ij}(\partial_1H_j^{1,k}(0)y_1+\partial_2H_j^{1,k}(0)y_2)e^{V_j^{k}}
+O(\epsilon_k^2)(1+|y|)^{2-m}. \nonumber\end{aligned}$$
By the long behavior of $w^{1,k}$ (Lemma \[wi0ki\]) we have $$\label{11mar21e1}
|G_{1,i}^k(r)|+|G_{2,i}^k(r)|\le \left\{\begin{array}{ll}C(1+r)^{3-m+\delta}, \,\, m\le 3, \\
C(1+r)^{\delta}, \,\, m>3.
\end{array}
\right.$$ Note that $G_{t,i}^k(0)=0$ and $G_{t,i}^k(r)=O(r)$ near $0$.
[**Step Four: Projection of $w^{1,k}$ onto higher frequencies** ]{}
Let $$g^{k,l}=(g^{k,l}_1,...,g^{k,l}_n)$$ be the projection of $w^{1,k}$ on $\sin l\theta$. In this step we first establish a preliminary estimate for all these projections:
\[11jun14lem3\] There exist $l_0\ge 3$ and $C>0$ independent of $k,l$ such that $$|g_i^{k,l}(r)|\le C\epsilon_k^2r^2,\quad 0<r<\epsilon_k^{-1}, \quad \forall l\ge l_0.$$
[**Proof:**]{}
By (\[newwk\]), (\[ewik1\]) and Lemma \[wi0ki\] $g^{k,l}$ satisfies $$\label{11mar31e1}
\left\{\begin{array}{ll}
(\frac{d^2}{dr^2}+\frac{1}r\frac{d}{dr}-\frac{l^2}{r^2})g_i^{k,l}+\sum_ja_{ij}h_j^k(0)e^{V_j^{k}}g_j^{k,l}=h_{ikl}\\
\\
g_i^{k,l}(0)=0.
\end{array}
\right.$$ where, applying the Taylor expansion of $H_j^{i,k}$ up to the second order, $$\label{12mar9e1}
|h_{ikl}(r)|\le C\epsilon_k^2(1+r)^{2-m+\delta/2}$$ for some $C$ independent of $k$ and $l$. Thus $$\label{11apr5e9}
\Delta g_i^{k,l}-\frac{l^2}{r^2}g_i^{k,l}+\sum_ja_{ij}h_j^{k}(0)e^{V^{k}_j}g_j^{k,l}>-c_0\epsilon_k^2(1+r)^{2-m+\delta}$$ for some $c_0>0$. Let $$g(r)=\frac{r^2}4\int_r^{\infty}\frac{c_0}s(1+s)^{2-m+\delta}ds+c_0\frac{r^{-2}}4\int_0^rs^3(1+s)^{2-m+\delta}ds+r^2.$$ Then clearly $g(r)>0$ for $r>0$, $g$ solves $$\label{11apr11e11}
\left\{\begin{array}{ll}
g''+\frac 1r g'-\frac{4}{r^2}g(r)=-c_0r^2(1+r)^{-m+\delta}, \quad r>0, \\
\\
g(0)=g'(0)=0.
\end{array}
\right.$$ and $$g(r)=r^2+O(r^{4-m+\delta}), \quad r>1, \quad g(r)\le Cr^2\log (\frac 1r+1),\quad r\le 1.$$ Clearly by Lemma \[wi0ki\] $$\label{11apr5e7}
\epsilon_k^2g(\epsilon_k^{-1})>\max_{y\in \partial \Omega_{1,k}}|w^{1,k}_i(y)|\ge |g_i^{k,l}(\epsilon_k^{-1})|.$$ The reason that we include $r^2$ in the definition of $g(\cdot)$ is because by Lemma \[wi0ki\] we only know $w_i^{1,k}(x)=O(\epsilon_k (1+|x|)^{\delta})$ for $m>3$. Let $g^k=(g_1^k,..,g_n^k)=\epsilon_k^2(g,..,g)$, then it is easy to see that for $l_0$ sufficiently large and $l>l_0$ $$\begin{aligned}
\label{11apr5e8}
&&\Delta g_i^k-\frac{l^2}{r^2}g_i^k+\sum_ja_{ij}h_j^{k}(0)e^{V^{k}_j}g_j^k\\
&=&-\epsilon_k^2c_0(1+r)^{2-m+\delta}+\sum_ja_{ij}h_j^{k}(0)e^{V^{k}_j}g_j^k-\frac{l^2-4}{r^2}g_i^k \nonumber\\
&\le &-\epsilon_k^2c_0(1+r)^{2-m+\delta}.\nonumber\end{aligned}$$ To prove Lemma \[11jun14lem3\] it is enough to show $$\label{11jun15e1}
|g_i^{k,l}(r)|\le C g_i^k, \quad 0<r<\epsilon_k^{-1},\quad l\ge l_0.$$ with $C$ independent of $k$ and $l$. To this end, we shall use (\[11apr5e9\]) and the initial value $$\label{11apr5e10}
g_i^{k,l}(r)=o(r) \mbox{ near }0.$$ In the following we use the argument in the proof of Theorem \[uniqlin\]. Let $\psi_i^k=g_i^k-g_i^{k,l}$, by (\[11apr5e7\]),(\[11apr5e8\]),(\[11apr5e9\]) and (\[11apr5e10\]) $$\left\{\begin{array}{ll}
\Delta \psi_i^k-\frac{l^2}{r^2}\psi_i^k+\sum_ja_{ij}h_j^{k}(0)e^{V^{k}_j}\psi_j^k<0,\quad 0<r<\epsilon_k^{-1}, \\
\\
\psi_i^k(r)=o(r)\mbox{ near }0, \quad \psi_i^k(\epsilon_k^{-1})>0 \mbox{ for all } i.
\end{array}
\right.$$ Our goal is to prove $\psi_i^k(r)>0$ for all $0<r<\epsilon_k^{-1}$. To do this, let $f^k=(f^k_1,...,f^k_n)$ be the positive solution to the homogeneous system: $$\Delta f_i^k-\frac{1}{r^2}f_i^k+\sum_ja_{ij}h_j^{k}(0)e^{V^{k}_j}f_j^k=0, \quad r>0$$ such that $$f_i^k(r)\sim r, \mbox{ near } 0, \quad f^k_i(r)\sim \frac 1r \mbox{ near } \infty, \quad f_i^k(r)>0 \mbox{ for all } r>0.$$ If $\psi_i^k$ is not always non-negative we assume $$\min_{\mathbb R^+}\frac{\psi_1^k(r)}{f_1^k(r)}=\min_{1\le i\le n}\min_{\mathbb R^+}\frac{\psi_i^k(r)}{f_i^k(r)}<0.$$ Suppose the minimum is attained at $r_k$. By the behavior of $\psi_i^k$ and $f_i^k$, we have $0<r_k<\epsilon_k^{-1}$. Let $w_1^k=\psi_1^k/f_1^k$, then $w_1^k$ satisfies $$\Delta w_1^k+2\nabla w_1^k\cdot \frac{\nabla f_1^k}{f_1^k}+\frac{1-l^2}{r^2}w_1^k=
\sum_{j=2}^na_{1j}h_j^{k}(0)e^{V^{k}_j}(\frac{w_1^kf_j^k-\psi_j^k}{f_1^k}).$$ Evaluating both sides at $r_k$, the left hand side is strictly positive while the right hand side is non-positive by the definition of $w_1^k$. This contradiction proves (\[11jun15e1\]). Finally since $g(r)=O(r^2\log\frac 1r)$ near $0$, $g_i^{k,l}(r)=O(\epsilon_k^2r^2)$ near $0$ for $l\ge l_0\ge 3$. Lemma \[11jun14lem3\] is established. $\Box$
\[11jun14lem4\] Given $\delta>0$, there exist $C(\delta)>0$ independent of $k,l$ and $k_0(\delta)>1$ such that for $l\ge 3$ and $k\ge k_0$ $$\label{11may2e4}
|g^{k,l}_i(r)|\le C\epsilon_k^{(m-2-\delta)^+}(\epsilon_kr)^l+\frac{C}{l^2}\epsilon_k^2r^2(1+r)^{2-m+\delta},\quad r\le \frac 12 \epsilon_k^{-1}$$ where $$(m-2-\delta)^+=\left\{\begin{array}{ll}
m-2-\delta \quad \mbox{ if }\quad m\le 3 \\
1-\delta\quad \mbox{ if }\quad m>3.
\end{array}
\right.$$
[**Proof:**]{}
By Lemma \[11jun14lem3\], (\[11mar31e1\]) can be rewritten as as $$(\frac{d^2}{dr^2}+\frac 1r\frac{d}{dr}-\frac{l^2}{r^2})g_i^{k,l}=\bar h_{ikl},\quad
g_i^{k,l}(0)=0$$ where $$\label{11may2e2}
|\bar h_{ikl}(r)|\le C\epsilon_k^2(1+r)^{2-m+\delta},\quad k\ge k_0(\delta)$$ for some $C$ independent of $k$ and $l$. By standard ODE theory $$\label{11may2e1}
g_i^{k,l}(r)=c^i_{1kl}r^l+c^i_{2kl}r^{-l}-\frac{r^{l}}{2l}\int_r^{\epsilon_k^{-1}}s^{-l+1}\bar h_{ikl}(s)ds
-\frac{r^{-l}}{2l}\int_0^rs^{l+1}\bar h_{ikl}(s)ds.$$ When $r\to 0$, it is easy to see that the last two terms in (\[11may2e1\]) both tend to $0$. Thus $c^i_{2kl}=0$. Let $I_3$ and $I_4$ represent the last two terms in (\[11may2e1\]), respectively. By (\[11may2e2\]), we have $$\label{11may2e3}
|I_3(r)|+|I_4(r)|\le \frac{C\epsilon_k^2}{l^2}(1+r)^{4-m+\delta}$$ where $C$ is independent of $k$ and $l$. On the other hand, using the information of $w^{1,k}$ for $r\sim \epsilon_k^{-1}$, we know $$\label{12apr27e5}
|g^{k,l}_i(r)|\le C\epsilon_k^{(m-2-\delta)^+}, \quad r\sim \epsilon_k^{-1}.$$ In regard to (\[11may2e3\]) and (\[11may2e1\]) we have $$\label{12apr27e6}
|c^i_{1kl}|\le C\epsilon_k^{(m-2-\delta)^++l}.$$ Hence (\[11may2e4\]) follows immediately. Lemma \[11jun14lem4\] is established. $\Box$
Let $z^k=(z_1^k,...,z_n^k)$ be the projection of $w^{1,k}$ to $span\{ \sin l\theta, \cos l\theta, l\ge 2\}$, i.e. $$z_i^{k}=\sum_{l=2}^{\infty}\bigg (g_i^{k,l}(r)\sin l\theta + \tilde g_i^{k,l}(r)\cos l \theta \bigg )$$ where $\tilde g_i^{k,l}$ is the projection of $w_i^{1,k}$ on $\cos l \theta$. $\tilde g_i^{k,l}$ has similar estimates as that for $g_i^{k,l}$. Then Lemma \[11jun14lem4\] leads to $$\label{11jun17e1}
|z_i^k(y)|\le C\epsilon_k^2(1+|y|)^{4-m+\delta}, \quad |y|\le \frac 12\epsilon_k^{-1}$$ for $m\le 3$. However for $m>3$, Lemma \[11jun14lem4\] only gives $$\label{12jun20e5}
|z_i^k(y)|\le C\epsilon_k^{3-\delta}(1+|y|)^2+C\epsilon_k^2(1+|y|)^{4-m+\delta}, \quad |y|\le \frac 12\epsilon_k^{-1}.$$ In the following we shall get rid of the first term on the right hand side of (\[12jun20e5\]). To this end, we need to evaluate the value of $w^{1,k}$ on $\partial \Omega_{1,k}$.
\[wikbry\] (a) If $m<4$, then $$\label{11apr11e10}
|w_i^{1,k}(y)|\le C\epsilon_k^{m^k-2}\log \epsilon_k^{-1},\quad y\in \partial \Omega_{1,k}.$$ (b) If $m=4$ and $|m_i^k-4|\le C/\log \epsilon_k^{-1}$ for all $i$, then $$\label{11jun13e3}
|w_i^{1,k}(y)|\le C\epsilon_k^2(\log \epsilon_k)^2,\quad y\in \partial \Omega_{1,k}.$$
The assumption $|m_i^k-4|\le C/\log \epsilon_k^{-1}$ when $m=4$ is natural and will be justified in the proof of the main theorems in section 6. We also remark that in (\[11apr11e10\]) we use $\epsilon_k^{m^k-2}\log 1/\epsilon_k$ instead of the crude $\epsilon_k^{m-2-\delta}$ as before.
[**Proof:**]{}
The Green’s representation formula for $w^{1,k}_i$ gives (see (\[newwk\])) $$\begin{aligned}
O(\epsilon_k^2)=w_i^{1,k}(0)=\int_{\Omega_{1,k}}G(0,\eta)\big ( \sum_ja_{ij}h_j^{1,k}(0)e^{V_j^{k}}w_j^{1,k}(\eta)-E_i^k(\eta))d\eta \nonumber\\
-\int_{\partial \Omega_{1,k}}\partial_{\nu}G(0,\eta)w_i^{1,k}(\eta)dS_{\eta} \label{11jun13e5}\end{aligned}$$ where $G(\cdot, \cdot)$ is the Green’s function on $\Omega_{1,k}$ with respect to the Dirichlet boundary condition. We observe that $$\label{11jun13e6}
G(0,\eta)=-\frac 1{2\pi}\log |\eta |+\frac 1{2\pi} \log \epsilon_k^{-1}+O(\epsilon_k^2).$$ Indeed, since $\Omega_{1,k}$ is a translation of $\Omega_k$ by $p_{1,k}$ ( recall $p_{1,k}=O(\epsilon_k)$), the oscillation of $\log |\eta|$ on $\partial \Omega_{1,k}$ is $O(\epsilon_k^2)$. Thus the oscillation of the regular part of $G(0, \eta)$ is $O(\epsilon_k^2)$, which leads to (\[11jun13e6\]). On the other hand let $\bar w_i^{1,k}$ be the average of $w^{1,k}_i$ on $\partial \Omega_{1,k}$, using the fact that $v^{1,k}_i$ is constant on $\partial \Omega_{1,k}$ and $V_i^k$ has oscillation $O(\epsilon_k^2)$ on $\partial \Omega_{1,k}$ we have $$\label{11jun13e7}
w_i^{1,k}(y)=\bar w_i^{1,k}+O(\epsilon_k^2), \quad \forall y\in \partial \Omega_{1,k}.$$ Thus $$\label{11jun13e8}
-\int_{\partial \Omega_{1,k}}\partial_{\nu}G(0,\eta)w_i^{1,k}(\eta)dS_{\eta}
=\bar w_i^{1,k}+O(\epsilon_k^2).$$ By using (\[11jun13e6\]), (\[11jun13e7\]) and (\[11jun13e8\]) in (\[11jun13e5\]) we have $$-\bar w_i^{1,k}+O(\epsilon_k^2)=\int_{\Omega_{1,k}}(\frac 1{2\pi}\log \frac{\epsilon_k^{-1}}{|\eta |})\bigg (\sum_j a_{ij}h_j^{k}(0)e^{V_j^{k}}w_j^{1,k}(\eta)
-E_i^k(\eta)\bigg )d\eta$$ To evaluate the right hand side, we divide $\Omega_{1,k}$ into a symmetric part: $D_1:=B(0, \epsilon_k^{-1}-|p_{1,k}|)$ and a nonsymmetric part: $\Omega_{1,k}\setminus D_1$ and use $I_1$ and $I_2$ to represent the corresponding integrals on them. If $\eta\in \Omega_{1,k}\setminus D_1$, it is easy to see from $|p_{1,k}|=O(\epsilon_k)$ that $$\log (\frac{\epsilon_k^{-1}}{|\eta |})=O(\epsilon_k^2) \mbox{ and } | \Omega_{1,k}\setminus D_1|=O(1).$$ Moreover for $\eta\in \Omega_{1,k}\setminus D_1$, by (\[wi0k\]) and (\[ewik1\]) $$w_j^{1,k}(\eta)=O(\epsilon_k^{1-\epsilon}),
\quad E_i^k(\eta)=O(\epsilon_k^{m^k}).$$ Combining these facts we have $I_2=O(\epsilon_k^2)$.
To evaluate $I_1$, since $\log (\epsilon_k^{-1}/|\eta |)$ is radial, by symmetry only the projections of $w_j^{1,k}$ and $E_i^k$ onto $1$ remain. By (\[11mar29e4\]) and (\[ewik1\]) $$|I_1|\le C\log \epsilon_k^{-1}\int_{D_1}\epsilon_k^2(1+|\eta |)^{2-m^k}d\eta.$$ (Note that we use the fact that $e^{V_k(x)}\le C(1+|y|)^{-m^k}$ for $C$ independent of $k$.) Therefore if $m<4$, $I_1=O(\epsilon_k^{m^k-2})\log \epsilon_k^{-1}$. If $m=4$ and $|m_i^k-4|\le C\log \epsilon_k^{-1}$ we have $(1+r)^{-m^k}\le C(1+r)^{-4}$, thus by elementary computation $$|I_1|\le C\epsilon_k^2(\log \epsilon_k)^2.$$ Lemma \[wikbry\] is established. $\Box$
By Lemma \[wikbry\], (\[12apr27e5\]) can be replaced by $$|g^{k,l}_i(r)|\le C\epsilon_k^{m-2-\delta}, \quad r\sim \epsilon_k^{-1}.$$ Correspondingly, the estimate for $c_{1kl}^i$ becomes $$|c_{1kl}^i|\le C\epsilon_k^{m-2-\delta+l}.$$ Then it is easy to see that the first term in (\[12jun20e5\]) can be removed.
[**Step five:**]{}
For the projection on $span\{\sin \theta, \cos \theta\}$, we write (\[1030e1\]) as $$(\frac{d^2}{dr^2}+\frac 1r\frac{d}{dr}-\frac{1}{r^2})G_{1,i}^k=h(r),\quad 0<r<\epsilon_k^{-1}$$ where $h$ is the collection of other terms. By (\[11mar21e1\]) $$|h(r)|\le C(1+r)^{1-m+\delta}.$$ Then $$\label{12mar8e1}
G_{1,i}^k=c_{1k}r+\frac{c_{2k}}r-\frac r2\int_r^{\infty}h(s)ds-\frac{r^{-1}}2\int_0^r s^2 h(s)ds.$$ Since $G_{1,i}$ is bounded near $0$, $c_{2k}=0$. Using $G_{1,i}^k(\epsilon_k^{-1})=O(\epsilon_k^{m-3-\delta})$ we have $$|c_{1k}|\le C\epsilon_k^{m-2-\delta}.$$ Then it is easy to see from (\[12mar8e1\]) that $$\label{11jun23e1}
|G_{t,i}^k(r)|\le Cr(1+r)^{2-m+\delta}, \, t=1,2, \, \, \mbox{ when } m>3.$$ Similarly (\[11jun23e1\]) also holds for $G_{2,i}^k$.
Combining the results in the five steps we arrive at the following estimate without distinguishing $m<4$ or not.
\[cruexp\] Given $\delta>0$, there exist $C(\delta)>0$, $k_0(\delta)>1$ such that for $|y|\le \epsilon_k^{-1}/2$ and $|\alpha |=0,1$, the following holds for all $k\ge k_0$ $$\label{11jun14e5}
|D^{\alpha}(v_i^{1,k}(y)-V_i^k(y)-\Phi_i^k(y))|\le C\epsilon_k^2(1+|y|)^{4-m-|\alpha |+\delta}.$$ where $$\Phi_i^k(y)=\epsilon_k(G_{1,i}^k(r)\cos\theta+G_{2,i}^k(r)\sin \theta)$$ with $$\label{11jun14e6}
|G_{t,i}^k(r)|\le Cr(1+r)^{2-m+\delta}\quad t=1,2.$$
Note that the estimate for $|\alpha |=0$ follows directly from the five steps. The estimate for $|\alpha |=1$ follows from standard gradient estimate for elliptic equations.
Theorem \[cruexp\] does not distinguish $m<4$ or $m=4$. In the following we apply Theorem \[cruexp\] to obtain more accurate estimates for $m<4$ and $m=4$, respectively. Both results in the sequel (Theorem \[expthm1\] and Theorem \[expthm2\]) play a crucial role in determining the location of maximum points of bubbling solutions.
\[expthm1\] Suppose $m<4$, then for $|y|\le \epsilon_k^{-1}/2$ and $i\in I$ $$\begin{aligned}
\label{mle4e1}
&&|D^{\alpha}\big (v_i^k(y)-\phi_i^k(\epsilon_ky)-V_i^k(y-p_{1,k})-\Phi^k_i(y-p_{1,k})\big )|\\
&\le &
C\epsilon_k^2(1+|y|)^{4-m^k-l}\log (2+|y|)
\quad |\alpha |=0,1 \nonumber\end{aligned}$$ where $v_i^k,\phi_i^k,V_i^k,\Phi_i^k$ are defined by (\[vikdef\]), (\[816e5\]), (\[88e2\]) and (\[11aug11e1\]), respectively. Moreover $G_{t,i}^k$ ($t=1,2, i\in I$) satisfy $$\label{11jun9e1}
|G_{t,i}^k(r)|\le Cr(1+r)^{2-m^k}\quad 0<r<\epsilon_k^{-1}.$$
[**Proof:**]{}
We use the same notations as in the proof of Proposition \[cruexp\]. First we consider the radial part of $w^{1,k}$: Recall $m<4$. Using Proposition \[cruexp\] and (\[11apr8e1\]) we have $$\begin{aligned}
&&(\frac{d^2}{dr^2}+\frac 1r\frac{d}{dr})g_i^{k,0}\\
&=&-\sum_j a_{ij}h_j^k(0)e^{V_j^k}g_j^{k,0}-\frac{\epsilon_k^2}4\sum_ja_{ij}\Delta H_j^{1,k}(0)r^2e^{V_j^k}+O(\epsilon_k^2)(1+r)^{\epsilon-m}\\
&=&O(\epsilon_k^2)(1+r)^{2-m^k},\quad 0<r<\epsilon_k^{-1},\end{aligned}$$ and $$g_i^{k,0}(0)=O(\epsilon_k^2).$$
Multiplying $r$ on both sides of the equation and integrating, we obtain $$|g_i^{k,0}(r)|\le C\epsilon_k^2\log (2+r).$$
Next we consider the projection of $w^{1,k}$ on $\sin\theta $ and $\cos\theta$. Let $\omega^k=(\omega_1^k,...,\omega_n^k)$ be $$\omega_i^k=w^{1,k}_i-g^{k,0}_i.$$ Then $\omega^k$ satisfies $$\Delta \omega_i^k+\sum_{j=1}^n a_{ij}h^{k}_j(0)e^{V^{k}_j}\omega_j^k=E_{1,i}^k, \quad \Omega_{1,i}$$ where $E_{1,i}^k$ is the projection of the original right hand side on the subspace spanned by $\sin k\theta$ and $\cos ku$ ($k=1,2,..$). Since $g_i^{k,0}$ is a radial function, from the asymptotic behavior of $g^{k,0}$ it is easy to see that the oscillation of it on $\partial \Omega_{1,k}$ is $O(\epsilon_k^2)$, therefore the oscillation of $\omega_i^k$ on $\partial \Omega_{1,k}$ is $O(\epsilon_k^2)$. Let $\bar \omega_i^k$ be the average of $\omega_i^k$ on $\partial \Omega_{1,k}$, then by $\omega_i^k(0)=0$ we have $$0=\int_{\Omega_{1,k}}G(0,\eta)(\sum_{j=1}^n a_{ij}h^{k}_j(0)e^{V^{k}_j}\omega_j^k(\eta)-E_{1,i}^k(\eta))d\eta+\bar \omega_i^k+O(\epsilon_k^2).$$ Using (\[11jun13e6\]) in the equation above we have $\bar \omega_i^k=O(\epsilon_k^2)$, thus $$\label{12mar9e3}
\omega_i^k=O(\epsilon_k^2),\quad \mbox{ on } \quad \partial \Omega_{1,k}.$$ Since we have known that $ \epsilon_k G_{t,i}^k(\epsilon_k^{-1})=O(\epsilon_k^{m^k-2})$, we can improve the estimate of $\Phi^k$ using (\[12mar8e1\]). The estimate of $G_{t,i}^k$ now is $$|G_{t,i}^k(r)|\le Cr(1+r)^{2-m^k}, \quad r<\epsilon_k^{-1}$$ and $$|\frac{d}{dr}G_{t,i}^k|\le Cr(1+r)^{1-m^k},\quad 0<r<\epsilon_k^{-1}/2$$ which leads to $$\label{11jun16e1}
|\nabla \Phi^{k}_i(y)|\le C\epsilon_kr(1+r)^{1-m^k},\quad |y|<\epsilon_k^{-1}/2.$$ As far as the projection of $w^{1,k}$ on higher frequencies is concerned, since we have (\[12mar9e3\]), for $g^{k,l}$ we now have, instead of Lemma \[11jun14lem4\] $$\label{11jun16e2}
|g^{k,l}_i(r)|\le C\epsilon_k^{m^k-2}(\epsilon_k r)^{l}+\frac{C}{l^2}\epsilon_k^2r^2(1+r)^{2-m^k}, l\ge 2.$$ As before we let $z_i^k=w^{1,k}_i-\Phi_i^{k}$, then $z_i^k$ satisfies $$\left\{\begin{array}{ll}
\Delta z_i^k=O(\epsilon_k^2)|y|^2(1+|y|)^{-m^k},\quad \Omega_{1,k}\subset \mathbb R^2, \\
\\
z_i^k=O(\epsilon_k^2)\quad \mbox{ on }\quad \partial \Omega_{1,k}.
\end{array}
\right.$$ Because of (\[11jun16e2\]) we have $$|z_i^k(y)|\le C\epsilon_k^2(1+|y|)^{4-m^k}, \quad |y|\le \epsilon_k^{-1}.$$ By standard re-scaling method $$\label{11jun16e3}
|\nabla z_i^k(y)|\le C\epsilon_k^2(1+|y|)^{3-m^k},\quad |y|<\epsilon_k^{-1}/2.$$ We have established $$|D^{\alpha}(v_i^{1,k}(y)-V_i^{k}(y)-\Phi_i^{k}(y)|\le C\epsilon_k^2(1+|y|)^{4-m^k-|\alpha |}\log (2+|y|)$$ for $|\alpha |=0,1$ and $|y|\le \epsilon_k^{-1}/2$. Recall that $v_i^{1,k}$ is defined in (\[vi0kd\]). Instead of using the coordinate around $p_{1,k}$ we use the coordinate around the origin to obtain (\[mle4e1\]).
Theorem \[expthm1\] is established. $\Box$
\[expthm2\] If $m=4$ and $|m_i^k-4|\le C/\log \epsilon_k^{-1}$ for all $i\in I$, then we have, for $|y|\le \epsilon_k^{-1}/2$ and $i\in I$ $$\begin{aligned}
\label{m4e1}
&&| D^{\alpha}(v_i^k(y)-\phi_i^k(\epsilon_ky)-V_i^k(y-p_{1,k})-\Phi_i^{k}(y-p_{1,k})\big )|\\
&\le &
C\epsilon_k^2(1+|y|)^{-|\alpha |}(\log (2+|y|))^2.
\quad |\alpha |=0,1, \nonumber\end{aligned}$$ where $\Phi^{k}$ is of the form stated in (\[11aug11e1\]) with $G_{t,i}^k$ ($t=1,2$) satisfying $$\label{11jun9e3}
|G_{t,i}^k(r)|\le Cr(1+r)^{-2},\quad 0<r<\epsilon_k^{-1},\quad i\in I.$$
[**Proof:**]{}
For $m=4$ we use (\[11mar29e4\]) to write (\[11apr8e1\]) as $$\left\{\begin{array}{ll}
(\frac{d^2}{dr^2}+\frac 1r\frac{d}{dr})g_i^{k,0}=-\epsilon_k^2\sum_j a_{ij}\frac{\Delta H_j^{1,k}(0)}4 r^2e^{V_j^{k}}+O(\epsilon_k^2)(1+r)^{-4+\epsilon}, \\
\\
g_i^{k,0}(0)=O(\epsilon_k^2),\quad \frac{d}{dr}g_i^{k,0}(0)=0.
\end{array}
\right.$$ Note that $(1+r)^{-m^k}\le C(1+r)^{-4}$. Multiplying both sides by $r$ and integrating, we have $$g_i^{k,0}(r)=-\epsilon_k^2\sum_j a_{ij}\frac{\Delta H_j^{1,k}(0)}4\int_0^rt^3e^{V_j^{k}(t)}(\log r-\log t)dt+O(\epsilon_k^2)\log (1+r).$$ To evaluate the integral, we use (\[12apr24e1\]) to write $$H_i^{1,k}(0)e^{V_i^{k}(t)}=e^{c_i^k}t^{-m^k}+O(t^{-4-\epsilon_0}), \quad t>1$$ for some $\epsilon_0>0$. Under the assumption of Theorem \[expthm2\]: $|m_i^k-4|\le C/\log \epsilon_k^{-1}$ we have $t^{-m^k}=O(t^{-4})$, thus $$|g_i^{k,0}(r)|\le C\epsilon_k^2(\log (2+r))^2,\quad 0<r<\epsilon_k^{-1}.$$ The projection on higher frequencies has the same estimates as in the case for $m<4$. Specifically, let $\omega_i^k$ be the same as in Theorem \[expthm1\]. Then (\[12mar9e3\]) also holds. Correspondingly (\[11jun16e1\]), (\[11jun16e2\]) and (\[11jun16e3\]) still hold with $m^k=4$ by the same proof. Theorem \[expthm2\] is established. $\Box$
Location of the blowup points
=============================
In this section we determine the locations of the blowup points in Theorem \[expthm1\] and Theorem \[expthm2\].
\[location\] Under the same assumptions as in Theorem \[expthm1\] $$\label{10050404}
|\sum_i\bigg (\frac{\partial_l h_i^k(0)}{h_i^k(0)}+\partial_l \phi_i^k(0)\bigg )\sigma_i^k|
\le C\epsilon_k^{m^k-2},\quad l=1,2,$$ where $C$ is independent of $k$. On the other hand, under the assumptions in Theorem \[expthm2\] we have $$\label{11jun28e1}
|\sum_i\bigg (\frac{\partial_l h_i^k(0)}{h_i^k(0)}+\partial_l \phi_i^k(0)\bigg )\sigma_i^k|
\le C\epsilon_k^2\log \epsilon_k^{-1},\quad l=1,2,$$ where $\sigma_i^k$ is defined in Definition \[11def1\].
[**Proof of Theorem \[location\]:**]{} Recall $H_i^k(\epsilon_ky)=h_i^k(\epsilon_ky)e^{\phi_i^k(\epsilon_ky)}$ and $H_i^{1,k}$ is defined in (\[hi0k\]). Let $\tilde \Omega_k=B(0, \epsilon_k^{-1}/2)$, we use the following Pohozaev identity for the equation for $v^{1,k}$: For $\xi\in S^1$, $$\begin{aligned}
\label{10050406}
&&\qquad \sum_i\int_{\tilde \Omega_k}\partial_{\xi}H_i^{1,k}(y)e^{v_i^{1,k}(y)}\\
&=&\int_{\partial \tilde \Omega_k}(\xi\cdot \nu)\sum_ie^{v_i^{1,k}}H_i^{1,k}
+\sum_{ij}a^{ij}\bigg (\partial_{\nu}v_i^{1,k}\partial_{\xi}v_j^{1,k}
-\frac 12\nabla v_i^{1,k}\nabla v_j^{1,k}(\xi \cdot \nu)\bigg )\nonumber\end{aligned}$$ According to the definition of $H_i^{1,k}$ in (\[hi0k\]) $$\partial_{\xi}H_i^{1,k}(y)=\epsilon_k\partial_{\xi}H_i^k(0)+\sum_{l=1}^2\epsilon_k^2\partial_{\xi l}H_i^k(0)y^l+O(\epsilon_k^3(1+|y|)^2).$$ Using the expansion of $v_i^{1,k}$ in Proposition \[cruexp\] ( which holds for $m<4$ and $m=4$) we have $$\begin{aligned}
&&\int_{\tilde \Omega_k}\partial_{\xi}H_i^{1,k}(y)e^{v_i^{1,k}(y)}\\
&=&\int_{\tilde \Omega_k}(\epsilon_k\partial_{\xi}H_i^k(0)+\sum_{l=1}^2\epsilon_k^2\partial_{\xi l}H_i^k(0)y^l+O(\epsilon_k^3(1+|y|)^2))\\
&&\cdot \bigg (e^{V_i^{k}(y)}(1+\Phi_i^{k})+O(\epsilon_k^2)(1+|y|)^{4-2m+\epsilon} \bigg )dy.\end{aligned}$$ By symmetry we have $$\int_{\tilde \Omega_k}\sum_{l=1}^2\epsilon_k^2\partial_{\xi l}H_i^k(0)y^l e^{V_i^{k}(y)}
=\int_{\tilde \Omega_k} \epsilon_k\partial_{\xi}H_i^k(0) e^{V_i^{k}(y)} \Phi_i^k(y)dy=0.$$ Also by elementary estimates we have $$\int_{\tilde \Omega_k}
\big (\sum_{l=1}^2\epsilon_k^2\partial_{\xi l}H_i^k(0)y^l+O(\epsilon_k^3(1+|y|)^2) \big )
\big (e^{V_i^{k}(y)}\Phi_i^{k}+O(\epsilon_k^2)(1+|y|)^{4-2m+\epsilon} \big )=O(\epsilon_k^{m^k-1})$$ and $$\int_{\tilde \Omega_k}\epsilon_k\partial_{\xi}H_i^k(0)O(\epsilon_k^2)(1+|y|)^{4-2m+\epsilon}=O(\epsilon_k^{m^k-1}).$$ Thus $$\begin{aligned}
&&\int_{\tilde \Omega_k}\partial_{\xi}H_i^{1,k}(y)e^{v_i^{1,k}(y)}\\
&=&\int_{\tilde \Omega_k} \epsilon_k\partial_{\xi}H_i^k(0) e^{V_i^{k}(y)} + O(\epsilon_k^3(1+|y|)^2) e^{V_i^k(y)}.\end{aligned}$$
The only difference on whether $m<4$ or $m=4$ is on this term. For $m<4$, direct computation gives $$\begin{aligned}
&&\int_{\tilde \Omega_k}\partial_{\xi}H_i^{1,k}(y)e^{v_i^{1,k}(y)}\\
&=&\epsilon_k\partial_{\xi}H_i^k(0)\int_{\tilde \Omega_k}e^{V_i^{k}(y)}dy
+O(\epsilon_k^{m^k-1}).\\
&=&\epsilon_k\frac{\partial_{\xi}H_i^k(0)}{H_i^k(0)}\int_{\tilde \Omega_k} H_i^k(0)e^{V_i^{k}(y)}dy+O(\epsilon_k^{m^k-1})\\
&=&2\pi \epsilon_k \partial_{\xi}(\log H_i^k)(0)\sigma_i^k+O(\epsilon_k^{m^k-1}).\end{aligned}$$ On the other hand for $m=4$, using the closeness between $m_i^k$ and $4$, we have $$\begin{aligned}
&&\int_{\tilde \Omega_k}\partial_{\xi}H_i^{1,k}(y)e^{v_i^{1,k}(y)}\\
&=&2\pi \epsilon_k \partial_{\xi} (\log H_i^k)(0)\sigma_i^k+O(\epsilon_k^{3}\log\epsilon_k^{-1}).\end{aligned}$$
Next we estimate the right hand side of the Pohozaev identity. The key point in this part is that the only difference between Theorem \[expthm1\] and Theorem \[expthm2\] is on the radial part of $w_i^{1,k}$: $g_i^{k,0}$. The projections of $w_i^{1,k}$ on higher frequencies ($ \omega_i^k$ ) have the same estimates. i.e. (\[12mar9e3\])- (\[11jun16e3\]) hold for $m<4$ and $m=4$. The detail is as follows: First by the decay rate of $v_i^{1,k}$ one sees easily that $$\int_{\partial \tilde \Omega_k}(\xi\cdot \nu)\sum_ie^{v_i^{1,k}}H_i^{1,k}=O(\epsilon_k^{m^k-1}).$$ Before evaluating the remaining two terms we first observe that by (\[11jun16e1\]) and (\[11jun16e3\]) that $$\nabla v_i^{1,k}=\nabla V_i^{k}+\nabla g_i^{k,0}+O(\epsilon_k^{m^k-1}),\quad y\in \partial \tilde \Omega_k.$$ Moreover, by symmetry we have $$\int_{\partial \tilde \Omega_k}\partial_{\nu}V_i^{k}\partial_{\xi}g^{k,0}_j=\int_{\partial \tilde \Omega_k}\partial_{\xi}V_i^{k}\partial_{\nu}g^{k,0}_j=0.$$ Thus $$\begin{aligned}
&&\sum_{i,j}\int_{\partial \tilde \Omega_k}a^{ij}\partial_{\nu}v_i^{1,k}\partial_{\xi}v_j^{1,k} \\
&=&\sum_{i,j}\int_{\partial \tilde \Omega_k}a^{ij}\big ((\partial_{\nu}(V_i^{k}+g_i^{k,0})+O(\epsilon_k^{m^k-1})\big )
\big (\partial_{\xi}(V_j^{k}+g^{k,0}_j)+O(\epsilon_k^{m^k-1})\big )\\
&&=O(\epsilon_k^{m^k-1}).\end{aligned}$$
Similarly in the last integral: $$\begin{aligned}
&&\sum_{i,j}\int_{\partial \tilde \Omega_k}a^{ij}\nabla v_i^{1,k}\nabla v_j^{1,k} (\xi\cdot \nu)\\
&=&\sum_{i,j}\int_{\partial \tilde \Omega_k}a^{ij}(\frac{d}{dr}(V_i^{k}+g_i^{k,0})+O(\epsilon_k^{m^k-1}))
(\frac{d}{dr}(V_j^{k}+g_j^{k,0})+O(\epsilon_k^{m^k-1}))(\xi\cdot \nu)\\
&=&O(\epsilon_k^{m^k-1}).\end{aligned}$$ Theorem \[location\] is established. $\Box$
The leading term for $\rho^k\to \rho\in \Gamma_1$
==================================================
In this section we complete the proofs of Theorem \[rhoto1mle4\], Theorem \[rhotome4\] and Theorem \[locations\]. The notation $u^k$ refer to the one in the introduction. Let $$\label{11jun25e1}
\Theta_i^k=u_i^k-\log \int_M h_ie^{u_i^k}dV_g,\quad i\in I.$$ Then we have $$\label{1013e1}
-\Delta_g \Theta_i^k=\sum_{j=1}^na_{ij}\rho_j^k(h_je^{\Theta_j^k}-1).$$ where $$\label{1013e2}
\int_Mh_ie^{\Theta_i^k}dV_g=1,\quad Vol(M)=1.$$ $\Theta^k=(\Theta_1^k,..,\Theta_n^k)$ is a sequence of blow-up solutions. We use $p_k$ to denote the point where the maximum of $\Theta^k$ on $M$ is taken. Take the local coordinates around $p_k$. Then $ds^2$ has the form $e^{\psi(y_{p_k})}(dy_1^2+dy_2^2)$ where $$|\nabla\psi(0)|=0,\quad \psi(0)=0,\quad \Delta \psi=-2Ke^{\psi},$$ $K$ is the Gauss curvature. In local coordinates, (\[1013e1\]) becomes $$\label{1013e3}
-\Delta \Theta_i^k=\sum_{j=1}^na_{ij}\rho_j^ke^{\psi}(h_je^{\Theta_j^k}-1), \quad \mbox{in}\quad B_{\delta_0}$$ for some $\delta_0$ small. Let $f_i^k$ solve $$\label{1013e4}\left\{\begin{array}{ll}
\Delta f_i^k=\sum_{j=1}^na_{ij}\rho_j^ke^{\psi},\quad B_{\delta_0}\\
\\
f_i^k(0)=0,\quad \nabla f_i^k(0)=0.
\end{array}
\right.$$ and $f_i^k$ has bounded oscillation on $\partial B_{\delta_0}$. Set $$\tilde \Theta_i^k=\Theta_i^k(p_k+\cdot)-f_i^k\quad \mbox{and }\quad \tilde h_i^k=\rho_i^k h_i(p_k+\cdot)e^{\psi+f_i^k},$$ then $\tilde \Theta_i^k$ satisfies $$\label{1013e5}
-\Delta \tilde \Theta_i^k=\sum_ja_{ij}\tilde h_j^ke^{\tilde \Theta_j^k},\quad B_{\delta_0}.$$ Note that from the definition of $\tilde h_i^k$, $f_i^k$ and $\psi$, the following can be verified by direct computation: $$\label{11jun18e1}
\left\{\begin{array}{ll}
\tilde h_i^k(0)=\rho_i^k h_i(p_k), \quad \nabla \tilde h_i^k(0)=\rho_i^k \nabla h_i(p_k), \\
\\
\Delta \tilde h_i^k(0)=\rho_i^k \bigg (\Delta h_i(p_k)+h_i(p_k)(-2K(p_k)+\sum_{j=1}^n a_{ij}\rho_j^k)\bigg ).
\end{array}
\right.$$
Let $M_k=\max_{i\in I}\max_M\Theta_i^k$, $\epsilon_k=e^{-\frac 12 M_k}$, $$\label{11mar7e2}
\tilde v_i^k(y)=\tilde \Theta_i^k(\epsilon_ky)+2\log \epsilon_k=\tilde \Theta_i^k(\epsilon_ky)-M_k,$$ then we have $$\label{11jun11e1}
\Delta \tilde v_i^k(y)+\sum_{j=1}^n a_{ij} \tilde h_j^k(\epsilon_ky)e^{\tilde v_j^k(y)}=0,\quad |y|\le \delta_0\epsilon_k^{-1}.$$
The following lemma proves that the $\Theta^k$ is a fully blown up sequence.
\[ful\] Along a subsequence $(\tilde v_1^k,...,\tilde v_n^k)$ converge in $C^2_{loc}(\mathbb R^2)$ to a system of $n$ equations.
[**Proof:**]{} By way of contradiction we assume that only $l$ ($l<n$) components of $(\tilde v_1^k,..,\tilde v_n^k)$ converge to a system of $l$ equations. Without loss of generality we assume that the first $l$ components of $(\tilde v_1^k,..,\tilde v_n^k)$ converge to $(\tilde v_1,..,\tilde v_l)$ that satisfies $$-\Delta \tilde v_i=\sum_{j=1}^l a_{ij} \tilde h_j e^{\tilde v_j}, \quad i=1,..,l. \mbox{ in } \mathbb R^2$$ where $\tilde h_i=\rho_i h_i(p)$. Let $$\sigma_{i,v}=\frac 1{2\pi} \int_{\mathbb R^2} \tilde h_i e^{\tilde v_i}.$$ Then the entire solution $(\tilde v_1,...,\tilde v_l)$ with finite energy satisfies $$\sum_{j=1}^l a_{ij}\sigma_{j, v}>2.$$ Let $J=\{1,...,l\}$, then by Theorem C in [@linzhang2] $(\tilde v_1,...,\tilde v_l)$ also satisfies $$\frac 1{4\pi^2} \Lambda_J(\sigma_v):=4\sum_{i\in J}\sigma_{i,v}-\sum_{i,j\in J} a_{ij} \sigma_{i,v}\sigma_{j, v}=0.$$ By (\[1013e2\]) $$\frac{\rho_i^k}{2\pi}\ge \frac 1{2\pi}\int_{B(p_k,\delta)}\rho_i^k h_i e^{\Theta_i^k}dV_g
=\frac{1}{2\pi }\int_{B_{\delta}}\tilde h_i^k e^{\tilde \Theta_i^k}dx.$$ Thus by letting $k\to \infty$ we have $$\frac{\rho_i}{2\pi }\ge \sigma_{i,v},\quad i=1,..,l.$$ Let $s_i=\frac{\rho_i}{2\pi }-\sigma_{i,v}$, $i\in J$, then easy to see $$\label{12jun22e1}
\frac{1}{4\pi^2}(\Lambda_J(\rho)-\Lambda_J(\sigma_v))=-2\sum_{i\in J}(\sum_{j\in J} a_{ij}\sigma_{j,v}-2)s_i-\sum_{i,j}a_{ij}s_is_j\le 0.$$ Since $\Lambda_J(\sigma_v)=0$, (\[12jun22e1\]) is a violation of the definition of $\Gamma_1$ in the introduction. Lemma \[ful\] is established. $\Box$
Let $\phi^k=(\phi_1^k,...,\phi_n^k)$ be the harmonic function that takes $0$ at $0$ and that kills the oscillation of $\tilde \Theta^k$ on $\partial B_{\delta_0}$. The first term in the expansion of $\tilde v^k-\phi^k(\epsilon_k\cdot)$ is $U^k(\cdot -p_{1,k})$ where $U^k=(U_1^k,..,U_n^k)$ satisfies $$\label{11jun25e2}
\left\{\begin{array}{ll}
-\Delta U_i^k=\sum_ja_{ij}\tilde h_j^k(0)e^{U_j^k}\quad \mathbb R^2, \\
U_i^k(0)=\tilde v_i^k(0),\quad i\in I.
\end{array}
\right.$$ Moreover, $U_i^k+\log (\rho_i^k h_i(p_k))$ satisfies $$-\Delta (U_i^k +\log (\rho_i^k h_i(p_k)))=\sum_ja_{ij}e^{U_j^k+\log (\rho_j^k h_j(p_k))}\quad \mathbb R^2.$$ Set $$\label{12apr29e2}
\bar \sigma_i^k=\frac 1{2\pi}\int_{\mathbb R^2}\rho_i^k h_i(p_k) e^{U_i^k},\quad m_i^k=\sum_{j=1}^n a_{ij}\bar \sigma_j^k.$$ Correspondingly $$m^k=\min\{m_1^k,..,m_n^k\},\quad m=\lim_{k\to \infty} m^k.$$ Later we shall show that $m$ is equal to the one defined in (\[12apr23e1\]). In regard to Lemma \[globalcpt\] we set $c_i^k$ as $$\begin{aligned}
\label{11jun25e3}
c_i^k=U_i^k(0)+\log (\rho_i^k h_i(p_k))+\frac 1{2\pi}\int_{\mathbb R^2}\log |\eta |\sum_{j=1}^na_{ij}\tilde h_j^k(0)e^{U_j^k}d\eta
\\
=\Theta_i^k(p_k)-M_k+\log (\rho_i^k h_i(p_k))+\frac 1{2\pi}\int_{\mathbb R^2}\log |\eta |\sum_{j=1}^na_{ij}\rho_j^k h_j(p_k)e^{U_j^k}d\eta,
\nonumber\end{aligned}$$ and $$\label{12jun29e1}
c_i=\lim_{k\to \infty} c_i^k.$$ According to Proposition \[cruexp\], the expansion of $\tilde v_i^k$ can be written as $$\label{11jun17e8}
\tilde v_i^k(y)=U_i^k(y-p_{1,k})+\phi_i^k(\epsilon_ky)+\Phi_i^{k}(y-p_{1,k})+O(\epsilon_k^2)(1+|y|)^{4-m+\epsilon}$$ on $B(0,\delta_0\epsilon_k^{-1})$, where $p_{1,k}=O(\epsilon_k)$, $$\Phi_i^{k}(y)=\epsilon_k(G_{1,i}^k(r)\cos \theta +G_{2,i}^k(r)\sin \theta )$$ such that $$|G_{t,i}^k(r)|\le Cr(1+r)^{2-m^k},\quad 0<r<\delta_0 \epsilon_k^{-1},\quad t=1,2.$$
Here we note that the Green’s function is of the form $$G(x,\eta)=-\frac 1{2\pi}\log |x-\eta |\chi +\gamma(x,\eta)$$ where $\chi$ is a cut-off function such that $\chi\equiv 1$ in $B_{\epsilon_1}$ for some $\epsilon_1>0$ ($2\epsilon_1$ is less than the injectivity radius of $M$) and $\chi\equiv 0$ outside $B_{2\epsilon_1}$. In the sequel we always assume $\delta_0<\epsilon_1/4$.
The Green’s representation formula for $\Theta_i^k$ is $$\label{11may23e1}
\Theta_i^k(x)=\bar \Theta_i^k+\int_MG(x,\eta)(\sum_ja_{ij}\rho_j^kh_je^{\Theta_j^k})dV_g.$$
Next we claim that $$\begin{aligned}
\label{nov2e5}
&&|\bar \Theta_i^k+\frac{m_i^k-2}2M_k-c_i^k+\log (\rho_i^k h_i(p_k))+2\pi\gamma(p_k,p_k)m_i^k|\nonumber \\
&\le & C_0(\delta_0)\epsilon_k^{m-2-\delta}\end{aligned}$$ where $\delta>0$ is small positive number, $\bar \Theta_i^k$ is the average of $\Theta_i^k$ on $M$, $C_0(\cdot)$ is a positive function such that $C_0(\delta_0)\to \infty$ as $\delta_0\to 0$. To see this, let $x=p_k$ in (\[11may23e1\]), $$\begin{aligned}
\Theta_i^k(p_k)&=&\bar \Theta_i^k-\frac 1{2\pi}\int_{B(p_k,\delta_0)}\log |p_k-\eta |\sum_{j}a_{ij}\rho_j^k h_j e^{\Theta_j^k}dV_g \nonumber\\
&&+\int_{B(p_k,\delta_0)}\gamma(p_k, \eta)\sum_{j}a_{ij}\rho_j^k h_j e^{\Theta_j^k}dV_g \nonumber\\
\label{11mar7e0}
&&+\int_{M\setminus B(p_k,\delta_0)}G(p_k,\eta)\sum_ja_{ij}\rho_j^kh_je^{\Theta_j^k}dV_g.\end{aligned}$$ Observe that $$\label{11mar7e1}
\rho_j^kh_je^{\Theta_j^k}dV_g(\eta)=\rho_j^kh_je^{f_j^k+\psi}e^{\Theta_j^k-f_j^k}d\eta=\tilde h_j^ke^{\tilde \Theta_j^k}d\eta.$$ Then by the definition of $\epsilon_k$, $\tilde v^k$, $c_i^k$ we obtain $$\begin{aligned}
\label{11jun24e2}
&&-\frac 1{2\pi}\int_{B(p_k,\delta_0)}\log |p_k-\eta |\sum_ja_{ij}\rho_j^kh_je^{\Theta_j^k}dV_g\\
&=&-\frac 1{2\pi}\int_{B(0,\delta_0\epsilon_k^{-1})}(\log (\epsilon_k|\eta_1|))\sum_ja_{ij}(\tilde h_j^ke^{\phi_j^k})(\epsilon_k\eta_1)e^{\tilde v_j^k(\eta_1)-\phi_j^k(\epsilon_k\eta_1)}d\eta_1 \nonumber\\
&=&-\frac 1{2\pi}\int_{B(0,\delta_0\epsilon_k^{-1})}(\log \epsilon_k+\log |\eta_1|)\sum_ja_{ij}H_j^{1,k}(\eta)
e^{v_j^{1,k}(\eta_1)}
d\eta_1 \nonumber\end{aligned}$$ where, by the same kind of notations used in section \[blowup\] and section 4, $$v_i^{1,k}(\cdot)=\tilde v_i^k(p_{1,k}+\cdot)-\phi_i^k(\epsilon_kp_{1,k}+\epsilon_k\cdot)$$ and $$\label{11jun25e4}
H_i^{1,k}(\cdot )=\tilde h_i^k(\epsilon_k p_{1,k}+\epsilon_k \cdot )e^{\phi_i^k(\epsilon_kp_{1,k}+\epsilon_k \cdot)}.$$
By (\[11jun18e1\]) we have $$H_i^{1,k}(0)=\rho_i^k h_i(p_k)+O(\epsilon_k^2), \quad \nabla H_i^{1,k}(0)=\epsilon_k \rho_i^k \nabla h_i(p_k)+O(\epsilon_k^3)$$ and $$H_i^{1,k}(\eta)=\rho_i^k h_i(p_k)+ \epsilon_k \rho_i^k \nabla h_i(p_k)\cdot \eta +O(\epsilon_k^2)(1+|\eta |)^2.$$ On the other hand by Proposition \[cruexp\] $$e^{v_i^{1,k}(\eta_1)}=e^{U_i^{k}}(1+\Phi_i^{k}(\eta_1)+O(\epsilon_k^2)(1+|\eta_1|)^{4-m+\delta}).$$ Hence $$\begin{aligned}
\label{11jun24e1}
&&H_i^{1,k}(\eta)e^{v_i^{1,k}(\eta)}\\
&=& (\rho_i^k h_i(p_k)+\epsilon_k \rho_i^k \nabla h_i(p_k)\cdot \eta +\Phi_i^{k}(\eta))e^{U_i^k(\eta)}
+O(\epsilon_k^2)(1+|\eta |)^{2-m+\delta}. \nonumber\end{aligned}$$ Using (\[11jun24e1\]) in the evaluation of (\[11jun24e2\]) we have $$\begin{aligned}
&&-\frac 1{2\pi}\int_{B(p_k,\delta_0)}\log |p_k-\eta |\sum_ja_{ij}\rho_j^kh_je^{\Theta_j^k}dV_g\\
&=&\frac{M_k}2m_i^k+\Theta_i^k(p_k)-M_k-c_i^k+\log (\rho_i^k h_i(p_k))+C_0(\delta_0)(\epsilon_k^{m-2-\delta}).\end{aligned}$$
Similarly $$\int_{B(p_k,\delta_0)}\gamma(p_k,\eta)\sum_ja_{ij}\rho_j^kh_je^{\Theta_j^k}dV_g=\gamma(p_k,p_k)2\pi m_i^k+C_0(\delta_0)(\epsilon_k^{m-2-\delta}).$$ For the final term in (\[11mar7e0\]) by the following crude estimate established in [@linzhang2] $$\Theta_i^k(x)=-\frac{m_i^k-2}2 M_k +O(1), \quad x\in M\setminus B(p_k, \delta_0), \quad O(1)\to \infty \mbox{ if } \delta_0\to 0$$ it is easy to see $$\int_{M\setminus B(p_k,\delta_0)}G(p_k,\eta)\sum_j a_{ij} \rho_j^k h_j e^{\Theta_j^k} dV_g=C_0(\delta_0)O(\epsilon_k^{m-2-\delta}).$$ Thus, back to (\[11mar7e0\]), $$\begin{aligned}
\Theta_i^k(p_k)=\bar \Theta_i^k+\frac{M_k}2m_i^k+\Theta_i^k(p_k)-M_k-c_i^k+\log (\rho_i^k h_i(p_k))\\
+\gamma(p_k,p_k)2\pi m_i^k+O(\epsilon_k^{m-2-\delta}).\end{aligned}$$ Then (\[nov2e5\]) follows.
Next for $x\in M\setminus B(p_k,\delta_0)$, by (\[11may23e1\]) and standard estimates $$\label{11july6e3}
\Theta_i^k(x)=\bar \Theta_i^k+2\pi G(x,p_k)m_i^k+C_0(\delta_0)\epsilon_k^{m-2-\delta}.$$ Consequently by (\[11july6e3\]) and (\[nov2e5\]) $$\begin{aligned}
\label{1216e1}
e^{\Theta_i^k(x)}&=&e^{\bar \Theta_i^k+2\pi G(x,p_k)m_i^k}+E_{\delta_0},\quad x\in M\setminus B(p_k,\delta_0) \\
&=&\epsilon_k^{m_i^k-2}\frac{e^{c_i^k}}{\rho_i^k h_i(p_k)} e^{2\pi m_i^k (G(x,p_k)-\gamma(p_k,p_k))} +E_{\delta_0}. \nonumber\end{aligned}$$ where $$\label{11131e1}
|E_{\delta_0}|\le C_0(\delta_0)\epsilon_k^{m^k-2+\delta}.$$ In the sequel we shall always use $E_{\delta}$ to represent a term bounded by the right hand side of (\[11131e1\]).
Proof of Theorem \[rhoto1mle4\].
--------------------------------
In this case $m<4$. Since $\int_M h_ie^{\Theta_i^k}dV_g=1$ we write $$\rho_i^k=\int_{B(p_k,\delta_0)}\rho_i^kh_ie^{\Theta_i^k}dV_g+\int_{M\setminus B(p_k,\delta_0)}\rho_i^kh_ie^{\Theta_i^k}dV_g
=\rho_{ia}^k+\rho_{ib}^k.$$ By (\[11mar7e1\]), (\[11mar7e2\]) $$\rho_{ia}^k=\int_{B(0,\delta_0)}\tilde h_i^ke^{\tilde \Theta_i^k}d\eta=
\int_{B(0,\delta_0\epsilon_k^{-1})}\tilde h_i^k(\epsilon_ky)e^{\tilde v_i^k(y)}dy.$$ Let $$I_1=\{i\in I;\quad \lim_{k\to \infty}m_i^k=\lim_{k\to \infty} m^k.\quad \}.$$ Using the expansion of $\tilde v_i^k$ in (\[11jun17e8\]) we have (since $m<4$) $$\label{rhoia}
\rho_{ia}^k=\int_{B(0,\delta_0\epsilon_k^{-1})}\tilde h_i^k(0)e^{U_i^k(y)}dy+o(\delta_0)\epsilon_k^{m^k-2},\quad i\in I_1.$$ Now for $i\not \in I_1$ we have $$\label{rhoia1}
\rho_{ia}^k=\int_{B(0,\delta_0\epsilon_k^{-1})}\tilde h_i^k(0)e^{U_i^k(y)}dy+E_{\delta_0},\quad i\not \in I_1$$ and $$\label{rhobc}
|\rho_{ib}^k|=E_{\delta_0},\quad i\not \in I_1.$$ It is easy to see from (\[12apr29e2\]) and (\[11jun18e1\]) that the $m$ defined by (\[12apr29e2\]) is the same as the one in (\[12apr23e1\]).
Combining (\[rhoia\]), (\[rhoia1\]), (\[rhobc\]) and Lemma \[globalcpt\] we have $$\begin{aligned}
&&\sum_i \frac{4}{2\pi}\rho_{ia}^k-\sum_{ij}a_{ij}\frac{\rho_{ia}^k}{2\pi}\frac{\rho_{ja}^k}{2\pi}\\
&=&2\sum_{i\in I_1}e^{c_i^k}\delta_0^{2-m_i^k}\epsilon_k^{m_i^k-2}+o(\delta_0)\epsilon_k^{m^k-2}+E_{\delta_0}.\end{aligned}$$ Using (\[rhoia1\]), (\[rhobc\]) to change from $\rho_{ia}^k$ to $\rho_i^k$, $$\begin{aligned}
\label{1105e2}
&&\frac{4}{2\pi}\sum_i\rho_i^k =\sum_{ij}a_{ij}\frac{\rho_i^k}{2\pi}\frac{\rho_j^k}{2\pi}+2\sum_{i\in I_1}e^{c_i^k}\delta_0^{2-m_i^k}\epsilon_k^{m_i^k-2}\\
&&-2
\sum_{i\in I_1}(m_i^k-2)\frac{\rho_{ib}^k}{2\pi}+o(\delta_0)\epsilon_k^{m^k-2}+E_{\delta_0}. \nonumber\end{aligned}$$
Using (\[1216e1\]) we have
$$\begin{aligned}
\rho_{ib}^k&=&\int_{M\setminus B(p_k,\delta_0)}\rho_i^kh_ie^{\Theta_i^k}dV_g\\
&=&\epsilon_k^{m_i^k-2}\int_{M\setminus B(p_k,\delta_0)}\frac{h_i(x)e^{c_i^k}}{h_i(p_k)}e^{2\pi m_i^k(G(x,p_k)-\gamma(p_k,p_k))}dV_g
+E_{\delta_0},\quad i\in I_1.\end{aligned}$$
Combining terms we have $$\begin{aligned}
&&\frac{4}{2\pi}\sum_i\rho_i^k-\sum_{i,j}a_{ij}\frac{\rho_i^k}{2\pi}\frac{\rho_j^k}{2\pi} \nonumber\\
&=&2\sum_{i\in I_1}e^{c_i^k}\epsilon_k^{m_i^k-2}\bigg (\delta_0^{2-m_i^k}-\frac {m_i^k-2}{2\pi}\int_{M\setminus B(p_k,\delta_0)}
\frac{h_i(x)}{h_i(p_k)}e^{2\pi m_i^k(G(x,p_k)-\gamma(p_k,p_k))}dV_g \bigg )\nonumber\\
&&\qquad \qquad \qquad +o(\delta_0)\epsilon_k^{m^k-2}+E_{\delta_0}. \label{11jun22e1}\end{aligned}$$
We claim that for fixed $k$ the following limit exists: $$\label{leadterm1}
\lim_{\delta_0\to 0}\bigg (\delta_0^{2-m_i^k}-\frac {m_i^k-2}{2\pi}\int_{M\setminus B(p_k,\delta_0)}
e^{2\pi m_i^k(G(x,p_k)-\gamma(p_k,p_k))}\frac{h_i}{h_i(p_k)}dV_g \bigg )$$ Indeed, write the second integral as the sum of one integral over $B(p_k,\delta_1)\setminus B(p_k,\delta_0)$ and the other over $M\setminus B(p_k,\delta_1)$, where $\delta_1$ is chosen small enough so that in $B(p_k,\delta_1)$, $$\label{11mar9e5}
dV_g=e^{\psi}dx\quad \mbox{ where }\psi(x)=O(|x|^2)\mbox{ in }B_{\delta_1}.$$ In the local coordinate at $p_k$, it suffices to prove $$\lim_{\delta_0\to 0}\bigg (\delta_0^{2-m_i^k}-\frac {m_i^k-2}{2\pi}\int_{B_{\delta_1}\setminus B_{\delta_0}}
e^{2\pi m_i^k(G(p_k+x,p_k)-\gamma(p_k,p_k))}\frac{h_i(p_k+x)}{h_i(p_k)}dV_g(x) \bigg )$$ exists. Since $$G(p_k+x,p_k)=-\frac 1{2\pi}\log |x|+\gamma(p_k+x,p_k),$$ we use (\[11mar9e5\]) and the Taylor expansions of $h_i$ and $\gamma(p_k+x,p_k)-\gamma(p_k,p_k)$ to obtain $$\begin{aligned}
&&e^{2\pi m_i^k(G(p_k+x,p_k)-\gamma(p_k,p_k))}h_i(x+p_k)/h_i(p_k) dV_g\\
&=&|x|^{-m_i^k}(1+\sum_{l=1}^2c_lx^l+O(|x|^2)) dx.\end{aligned}$$ where $c_1,c_2$ are some constants. Observe that the terms with $c_1,c_2$ disappear in the integration and $m_i^k<4-\epsilon_1$ for some $\epsilon_1>0$ for all $i\in I_1$, thus the limits in (\[leadterm1\]) exists.
The following lemma states the closeness between $\rho^k$ to $\rho$ if $m<4$. It will be used to simplify the leading terms in the statements of main theorems.
\[close1\] Let $\rho^k\to \rho\in \Gamma_1$ ($\rho\neq Q$) such that all $\rho_i^k-\rho_i$ have the same sign, then $$\label{11jun22e2}
\rho_i^k-\rho_i=O(\epsilon_k^{m-2-\epsilon})\quad \mbox{ and }\quad m_i^k-m_i=O(\epsilon_k^{m-2-\epsilon}).$$
[**Proof of Lemma \[close1\]:**]{} Let $s_i^k=\rho_i-\rho_i^k$. Then all $s_i^k$ are of the same sign. By (\[11jun22e1\]) we have $$\sum_i (m_i-2)s_i^k=O(\epsilon_k^{m-2-\epsilon}).$$ Thus (\[11jun22e2\]) holds and Lemma \[close1\] is established. $\Box$
Using (\[11jun22e2\]) in (\[11jun22e1\]) we can rewrite the leading term as $$\begin{aligned}
&&\frac{4}{2\pi}\sum_i\rho_i^k-\sum_{i,j}a_{ij}\frac{\rho_i^k}{2\pi}\frac{\rho_j^k}{2\pi} \nonumber\\
&=&2\sum_{i\in I_1}e^{c_i^k}\epsilon_k^{m-2}\bigg (\delta_0^{2-m}-\frac {m-2}{2\pi}\int_{M\setminus B(p_k,\delta_0)}
\frac{h_i(x)}{h_i(p_k)}e^{2\pi m(G(x,p_k)-\gamma(p_k,p_k))}dV_g \bigg )\nonumber\\
&&\qquad +o(\delta_0)\epsilon_k^{m-2}+E_{\delta_0}.\end{aligned}$$ Since $c_i^k\to c_i$, the $c_i^k$ can be replaced by $c_i$ in (\[12jun29e1\]). Theorem \[rhoto1mle4\] is established. $\Box$
Next we establish the following lemma regardless of $m=4$ or not.
\[graest\] $$\label{graphik}
\nabla \phi_i^k(0)=2\pi m_i^k\nabla_1 \gamma(p_k,p_k)+O(\epsilon_k^{m-2-\epsilon}).$$ where $\nabla_1$ means the differentiation with respect to the first component.
[**Proof of Lemma \[graest\]:**]{}
First from (\[11july6e3\]), for $x\in \partial B(p_k,\delta_0)$ $$\begin{aligned}
&& \tilde \Theta_i^k(x)=\Theta_i^k(x)-f_i^k(x)\\
&=&\bar \Theta_i^k-m_i^k\log |x-p_k|+2\pi m_i^k \gamma(x,p_k)-f_i^k(x)+O(\epsilon_k^{m-2-\epsilon}).\end{aligned}$$ For the first derivative, by standard estimates $$|D^l(\tilde \Theta_i^k)(x)-D^l(-m_i^k\log |x-p_k|+2\pi m_i^k \gamma(x,p_k)-f_i^k(x))|\le C\epsilon_k^{m-2-\epsilon}, \quad l=0,1$$ for $x\in B(p_k, \delta_0)$. On the other hand, we recall that $\tilde v_i^k(y)=\tilde \Theta_i^k(\epsilon_ky)+2\log \epsilon_k$. According to Proposition \[cruexp\] $$|D^l(\tilde v_i^k(y)-U_i^k(y-p_{1,k})-\phi_i^k(\epsilon_k y)-\Phi_i^{k}(y-p_{1,k})|\le C\epsilon_k^2(1+|y|)^{4-m+\epsilon-l}$$ for $l=0,1$ and $|y|\le 2\delta_0\epsilon_k^{-1}$. Using the asymptotic behavior $U_i^k$ for $|y|\sim \epsilon_k^{-1}$ (\[12apr29e1\]) we have $$U_i^k(y-p_{1,k})+\log (\rho_i^kh_i(p_k))=-\frac{m_i^k}2M_k-m_i^k\log |x|+c_i^k+O(\epsilon_k^{m^k-2})$$ where $|x|=\epsilon_k |y|$. Thus by (\[nov2e5\]) for $l=0$ we have $$\label{11july7e11}
\phi_i^k(x)=2\pi m_i^k(\gamma(x,p_k)-\gamma(p_k,p_k))-f_i^k(x)+O(\epsilon_k^{2-\epsilon}), \quad x\in B(p_k,2\delta_0)\setminus B(p_k, \delta_0/2).$$ Thus the comparison on $l=1$ and (\[1013e4\]) yield (\[graphik\]). Lemma \[graest\] is established. $\Box$
Proof of Theorem \[rhotome4\]
-----------------------------
In this case $m=4$, we first give a rough estimate of $\rho_i^k-\rho_i$.
\[lemq4\] Let $\rho^k$ tend to $Q$ such that all $\rho_i^k-Q_i$ have the same sign. Then $$\label{11jun17e9}
m_i^k-4=O(\epsilon_k^{2-\epsilon}),\quad \rho_i^k-\rho_i=O(\epsilon_k^{2-\epsilon})\quad \forall i\in I.$$
[**Proof of Lemma \[lemq4\]:**]{} Since $\rho^k\to Q$, $m=4$, then all $m_i^k\to 4$. Recall that $\rho_i^k=2\pi\bar \sigma_i^k+O(\epsilon_k^{2-\epsilon})$, The rest of the proof is the same as that in the proof of Lemma \[close1\]. Lemma \[lemq4\] is established. $\Box$
By (\[11jun28e1\]) $$\sum_i(\partial_l \log \tilde h_i^k(0)+\partial_l \phi_i^k(p_k))\bar \sigma_i^k=O(\epsilon_k^2\log \epsilon_k^{-1}),$$ then by (\[11jun18e1\]) $$\label{locationm4}
\sum_{i=1}^n \bigg (\partial_l(\log h_i+\phi_i^k)(p_k) \bigg )\bar \sigma_i^k=O(\epsilon_k^2)\log \epsilon_k^{-1},\quad l=1,2.$$
The distance from $\rho^k$ to $\Gamma_1$ can be computed as follows: $$\rho_i^k=\int_{B(p_k,\delta_0)}\rho_i^k h_i e^{\Theta_i^k} dV_g +\int_{M\setminus B(p_k,\delta_0)} \rho_i^k h_i e^{\Theta_i^k} dV_g.$$ By (\[11jun17e9\]) and Theorem \[expthm2\] the second integral is $O(\epsilon_k^2)$, this is the same as the computation for the single equation [@ChenLin1]. Therefore $$\begin{aligned}
\label{11july5e1}
&&\rho_i^k= \int_{B(p_k,\delta_0)}\rho_i^k h_i e^{\Theta_i^k} dV_g+O(\epsilon_k^2)\\
&=&\int_{B(0,\delta_0)}\tilde h_i^k e^{\phi_i^k}e^{\tilde \Theta_i^k-\phi_i^k}d\eta+O(\epsilon_k^2)\nonumber\\
&=&\int_{B(0,\delta_0 \epsilon_k^{-1})}H_i^{1,k}(\eta)e^{v_i^{1,k}(\eta)}d\eta+O(\epsilon_k^2)\nonumber\\
&=&\int_{B(0,\delta_0 \epsilon_k^{-1})} \rho_i^k h_i(p_k) e^{U_i^k(\eta)}d\eta+\int_{B(0,\delta_0 \epsilon_k^{-1})}
\frac 14 \Delta H_i^{1,k}(0)|\eta |^2 e^{U_i^k(\eta)}d\eta \nonumber\\
&&\qquad +O(\epsilon_k^2).\nonumber\end{aligned}$$ The first integral on the right hand side of the above is $2\pi \bar \sigma_i^k+O(\epsilon_k^2)$. To evaluate the second term on the right hand side, we first use the definition of the $H_i^{1,k}$ in (\[11jun25e4\]), (\[11jun17e9\]), (\[11jun18e1\]) and (\[graphik\]) to have $$\begin{aligned}
\label{h1kdel}
\Delta H_i^{1,k}(0)=\rho_i^k h_i(p_k)\bigg (\frac{\Delta h_i(p_k)}{h_i(p_k)}-2K(p_k)+8\pi\\
+16\pi \frac{\nabla h_i(p_k)}{h_i(p_k)}\cdot \nabla_1\gamma(p_k,p_k)
+64\pi^2 |\nabla_1\gamma(p_k,p_k)|^2\bigg )\epsilon_k^2+O(\epsilon_k^{4-\epsilon}). \nonumber\end{aligned}$$ For $e^{U_i^k}$ we use the definition of $c_i^k$ in (\[11jun25e3\]) and (\[12apr29e1\]) to have $$\label{11jun25e8}
\rho_i^k h_i(p_k) e^{U_i^k(\eta)}=e^{c_i^k} |\eta |^{-4}+O(|\eta |^{-4-\delta_0}),\quad |\eta |>1$$ for some $\delta_0>0$ independent of $k$. Using (\[h1kdel\]) and (\[11jun25e8\]) in the computation of the second integral of $\rho_i^k$ we have $$\rho_i^k=2\pi \bar \sigma_i^k+2\pi \epsilon_k^2 \log \epsilon_k^{-1}
b_i^ke^{c_i^k}+O(\epsilon_k^2).$$ where $$\begin{aligned}
b_i^k=\frac 14\bigg (\frac{\Delta h_i(p_k)}{h_i(p_k)}-2K(p_k)+8\pi+16\pi \frac{\nabla h_i(p_k)}{h_i(p_k)}\nabla_1\gamma(p_k,p_k)
+64\pi^2|\nabla_1 \gamma(p_k,p_k)|^2\bigg )\\
=\frac 14\bigg (\Delta\, \log h_i(p_k)-2K(p_k)+8\pi
+|\nabla \log h_i(p_k)+8\pi \nabla_1\gamma(p_k,p_k)|^2\bigg )\end{aligned}$$ Consequently, $$\begin{aligned}
\label{11jun8e1}
&&4\sum_i\frac{\rho_i^k}{2\pi}-\sum_{i,j}a_{ij}\frac{\rho_i^k}{2\pi}\frac{\rho_j^k}{2\pi}\\
&=&\epsilon_k^2\log \epsilon_k^{-1}(\sum_i (4-2m_i^k)b_i^k e^{c_i^k})+O(\epsilon_k^2). \nonumber\\
&=&-4\epsilon_k^2\log \epsilon_k^{-1}\sum_i b_i^k e^{c_i^k} +O(\epsilon_k^2). \nonumber\end{aligned}$$ Theorem \[rhotome4\] is established. $\Box$
Even though there is a cut-off function in the definition of $\chi$, changing the domain of this cut-off function does not change $\nabla_1\gamma(p_k,p_k)$.
[**Proof of Theorem \[locations\]:**]{}
By going through the proof of Lemma \[graest\] using Theorem \[rhoto1mle4\] for $\rho\neq Q$ or Theorem \[rhotome4\] for $\rho=Q$ instead of Proposition \[cruexp\] one sees easily that $$\phi_i^k(x)=2\pi m_{i}^k (\gamma(p_k+x, p_k)-\gamma(p_k,p_k))
-c_{i}^k-f_{i}^k(x)+O(\epsilon_k^{m^k-2})$$ if $\rho\neq Q$. On the other hand $$\phi_{i}^k(x)=2\pi m_{i}^k (\gamma(p_k+x, p_k)-\gamma(p_k,p_k))
-c_{i}^k-f_{i}^k(x)+O(\epsilon_k^{2}\log \epsilon_k^{-1})$$ if $\rho=Q$. Correspondingly $$\label{11july13e5}
\nabla \phi_{i}^k(0)=2\pi m_{i}^k \nabla_1\gamma(p_k, p_k)+O(\epsilon_k^{m^k-2})$$ if $\rho\neq Q$ and $$\label{11july13e6}
\nabla \phi_{i}^k(0)=2\pi m_{i}^k \nabla_1\gamma (p_k, p_k)+O(\epsilon_k^{2}\log \epsilon_k^{-1})$$ if $\rho=Q$. Theorem \[location\] yields $$|\sum_i(\partial_l(\log h_{i}^k)(0)+\partial_l \phi_{i}^k(0))\sigma_{i}^k|\le C\epsilon_k^{m^k-2} \quad \mbox{ if } \rho\neq Q$$ and $$|\sum_i (\partial_l (\log h_{i}^k)(0)+\partial_l\phi_{i}^k(0))\sigma_{i}^k|\le C\epsilon_k^{2}\log \epsilon_k^{-1} \quad \mbox{ if } \rho= Q.$$ Also the proofs of Theorem \[rhoto1mle4\] and Theorem \[rhotome4\] give $$\rho_i^k-\rho_i=O(\epsilon_k^{m^k-2})=O(\epsilon_k^{m-2}), \quad m<4$$ and $$\rho_i^k-\rho_i=O(\epsilon_k^{m^k-2})\log \epsilon_k^{-1}=O(\epsilon_k^2\log \epsilon_k^{-1}), \quad m=4.$$ So $$\bar \sigma_i^k-\frac{\rho_i}{2\pi}=\left\{\begin{array}{ll} O(\epsilon_k^{m-2}) &\quad m<4,\\
O(\epsilon_k^2\log \epsilon_k^{-1})&\quad m=4.
\end{array}
\right.$$
By (\[11july13e5\]) and (\[11july13e6\]) we obtain (\[11july13e7\]) and (\[11july13e8\]) in Theorem \[locations\]. Theorem \[locations\] is established $\Box$
[99]{}
J. J. Aly, Thermodynamics of a two-dimensional self-gravitating system, Phy. Rev. A 49, No. 5, Part A (1994), 3771–3783.
D. Bartolucci, Compactness result for periodic multivortices in the electroweak theory, Nonlinear Anal., 53 (2003), 277–297.
D. Bartolucci and G. Tarantello, Liouville type equations with singular data and their applications to periodic multivortices for the electroweak theory, Comm. Math. Phys., 229 (2002), 3–47.
W. H. Bennet, Magnetically self-focusing streams, Phys. Rev. 45 (1934), 890–897.
L. A. Caffarelli, Y. Yang, Vortex condensation in the Chern-Simons Higgs model: an existence theorem, Comm. Math. Phys. 168 (1995), no. 2, 321–336.
E. Caglioti, P. L. Lions, C. Marchioro, C, M. Pulvirenti, A special class of stationary flows for two-dimensional Euler equations: a statistical mechanics description, Comm. Math. Phys. 143 (1992), no. 3, 501–525.
E. Caglioti, P. L. Lions, C. Marchioro, C, M. Pulvirenti, A special class of stationary flows for two-dimensional Euler equations: a statistical mechanics description. II, Comm. Math. Phys. 174 (1995), no. 2, 229–260.
S. A. Chang, C. C. Chen, C. S. Lin, Extremal functions for a mean field equation in two dimension. (English summary) Lectures on partial differential equations, 61–93, New Stud. Adv. Math., 2, Int. Press, Somerville, MA, 2003. S. Chanillo, M. K-H Kiessling, Rotational symmetry of solutions of some nonlinear problems in statistical mechanics and in geometry, Comm. Math. Phys. 160 (1994), no. 2, 217–238.
S. Chanillo, M. K-H Kiessling, Conformally invariant systems of nonlinear PDE of Liouville type, Geom. Funct. Anal. 5 (1995), no. 6, 924–947.
J. L. Chern, Z. Y. Chen and C. S. Lin, Uniqueness of topological solutions and the structure of solutions for the Chern-Simons with two Higgs particles, Comm. Math. Phys. 296 (2010), 323–351.
C. C. Chen, C. S. Lin, Sharp estimates for solutions of multi-bubbles in compact Riemann surfaces, Comm. Pure Appl. Math. 55 (2002), no. 6, 728–771.
C. C. Chen, C. S. Lin, Topological degree for a mean field equation on Riemann surfaces, Comm. Pure Appl. Math. 56 (2003), no. 12, 1667–1727.
C. C. Chen, C. S. Lin, Mean field equations of Liouville type with singular data: Sharper estimates, Discrete and continuous dynamic systems, Vol 28, No 3, 2010, 1237-1272.
C. C. Chen, C. S. Lin and G. Wang, Concentration phenomena of two-vortex solutions in a Chern-Simons model, Ann. Sc. Norm. Super. Pisa Cl. Sci. (5), 3 (2004), 367–397.
W. X. Chen, C. M. Li, Classification of solutions of some nonlinear elliptic equations. Duke Math. J. 63 (1991), no. 3, 615–622.
S. Childress and J. K. Percus, Nonlinear aspects of Chemotaxis, Math. Biosci. 56 (1981), 217–237.
M. Chipot, I. Shafrir, G. Wolansky, On the solutions of Liouville systems, J. Differential Equations 140 (1997), no. 1, 59–105.
M. Chipot, I. Shafrir, G. Wolansky, Erratum: “On the solutions of Liouville systems” \[J. Differential Equations 140 (1997), no. 1, 59–105; MR1473855 (98j:35053)\], J. Differential Equations 178 (2002), no. 2, 630.
P. Debye and E. Huckel, Zur Theorie der Electrolyte, Phys. Zft 24 (1923), 305–325.
G. Dunne, Self-dual Chern-Simons Theories, Lecture Notes in Physics, vol. m36, Berlin: Springer-Verlag, 1995.
J. Dziarmaga, Low energy dynamics of $[U(1)]^N$ Chern-Simons solitons and two dimensional nonlinear equations, Phys. Rev. D 49 (1994), 5469–5479.
L. Ferretti, S.B. Gudnason, K. Konishi, Non-Abelian vortices and monopoles in SO(N) theories, Nuclear Physics B, Volume 789, Issues 1-2, 21 January 2008, Pages 84-110.
J. Hong, Y. Kim and P. Y. Pac, Multivortex solutions of the Abelian Chern-Simons-Higgs theory, Phys. Rev. Letter 64 (1990), 2230–2233.
R. Jackiw and E. J. Weinberg, Selfdual Chern Simons vortices, Phys. Rev. Lett. 64 (1990), 2234–2237.
J. Jost, C. S. Lin and G. F. Wang, Analytic aspects of the Toda system II: bubbling behavior and existence of solutions, Comm. Pure Appl. Math. 59 (2006), no. 4, 526–558.
E. F. Keller and L. A. Segel, Traveling bands of Chemotactic Bacteria: A theoretical analysis, J. Theor. Biol. 30 (1971), 235–248.
M. K.-H. Kiessling, Statistical mechanics of classical particles with logarithmic interactions, Comm. Pure Appl. Math. 46 (1993), no. 1, 27–56.
M. K.-H. Kiessling, Symmetry results for finite-temperature, relativistic Thomas-Fermi equations, Commun. Math. Phys. 226 (2002), no. 3, 607–626.
M. K.-H. Kiessling and J. L. Lebowitz, Dissipative stationary Plasmas: Kinetic Modeling Bennet Pinch, and generalizations, Phys. Plasmas 1 (1994), 1841–1849.
C. Kim, C. Lee and B.-H. Lee, Schrödinger fields on the plane with $[U(1)]^N$ Chern-Simons interactions and generalized self-dual solitons, Phys. Rev. D (3) 48 (1993), 1821–1840.
Y. Y. Li, Harnack type inequality: the method of moving planes, Comm. Math. Phys. 200 (1999), no. 2, 421–444.
C. S. Lin, Topological degree for mean field equations on S2, Duke Math. J. 104 (2000), no. 3, 501–536.
C. S. Lin, Uniqueness of solutions to the mean field equations for the spherical Onsager vortex, Arch. Ration. Mech. Anal. 153 (2000), no. 2, 153–176.
C. S. Lin, J. C. Wei, D. Ye, Classifcation and nondegeneracy of $SU(n+1)$ Toda system, to appear on Invent. Math.
C. S. Lin, J. C. Wei, C. Zhao, Sharp estimates for fully bubbling solutions of a $SU(3)$ Toda system, to appear on GAFA.
C. S. Lin, L. Zhang, Profile of bubbling solutions to a Liouville system. Ann. Inst. H. Poincaré Anal. Non Linéaire 27 (2010), no. 1, 117–143.
C. S. Lin, L. Zhang, A topological degree counting for some Liouville systems of mean field equations, Comm. Pure Appl. Math. volume 64, Issue 4, pages 556–-590, April 2011.
C. S. Lin, L. Zhang, Sharp estimates for blowup solutions to general Liouville systems, Part II, work in progress.
M. Nolasco, G. Tarantello, Vortex condensates for the SU(3) Chern-Simons theory, Comm. Math. Phys. 213 (2000), no. 3, 599–639.
I. Rubinstein, Electro diffusion of Ions, SIAM, Stud. Appl. Math. 11 (1990).
J. Spruck, Y. Yang, Topological solutions in the self-dual Chern-Simons theory: existence and approximation, Ann. Inst. H. Poincare Anal. Non Lineaire 12 (1995), no. 1, 75–97.
G. Wolansky, On steady distributions of self-attracting clusters under friction and fluctuations, Arch. Rational Mech. Anal. 119 (1992), 355–391.
G. Wolansky, On the evolution of self-interacting clusters and applications to semi-linear equations with exponential nonlinearity, J. Anal. Math. 59 (1992), 251–272.
Y. Yang, Solitons in field theory and nonlinear analysis, Springer-Verlag, 2001.
L. Zhang, Blowup solutions of some nonlinear elliptic equations involving exponential nonlinearities. Comm. Math. Phys. 268 (2006), no. 1, 105–133.
L. Zhang, Asymptotic behavior of blowup solutions for elliptic equations with exponential nonlinearity and singular data. Commun. Contemp. Math. 11 (2009), no. 3, 395–411.
[^1]: Zhang is supported in part by NSF Grant 0900864 (1027628)
|
---
abstract: 'Dissipating of disorder quantum vortices in an annular two-dimensional Bose-Einstein condensate can form a macroscopic persistent flow of atoms. We propose a protocol to create persistent flow with high winding number based on a double concentric ring-shaped configuration. We find that a sudden geometric quench of the trap from single ring-shape into double concentric ring-shape will enhance the circulation flow in the outer ring-shaped region of the trap when the initial state of the condensate is with randomly distributed vortices of the same charge. The circulation flows that we created are with high stability and good uniformity free from topological excitations. Our study is promising for new atomtronic designing, and is also helpful for quantitatively understanding quantum tunneling and interacting quantum systems driven far from equilibrium.'
author:
- Xiyu Chen
- Tao Yang
- 'Wen-Li Yang'
- 'Wu-Ming Liu'
title: Turbulent cascade induced persistent current of cold atomic superfluids
---
Study of persistent flow of superfluid enables understanding of fundamental characteristics of superfluidity and may lead to applications in high-precision metrology and atomtronics[@PRL.95.143201; @PRL.103.140405; @nature.506.200]. Thanks to the technical development for achieving tailored trapping potential with arbitrary geometries, quantum transport experiments with quantum gases can be carried out in Bose-Einstein condensate (BEC) systems with weak interatomic interactions, which can be accurately described in the frame of the mean-field Gross-Pitaveskii (GP) equation[@RMP.73.307]. Two-dimensional (2D) annular BECs in ring-shaped traps[@PRL.95.143201; @PRA.74.023617; @PRL.99.260401; @NJP.10.043012], sustaining persistent currents with a quantized circulation around any closed path[@PRL.99.260401; @PRL.106.130401; @PRA.86.013629; @nature.506.200; @PRL.111.205301] with the corresponding phase winding of the macroscopic wave function being an integer multiple of $2\pi$, attract growing interest as an atomic analog of superconducting quantum interference device constituting a basic element for atomtronic circuits[@PRA.75.023615; @PRL.103.140405; @NJP.19.020201; @Jopt.18.093001], and are analogous to persistent currents in superconducting rings with a quantized magnetic flux.
The direct macroscopic circulation flow of superfluid BECs in annular traps can be achieved by rotating [a perturbing potential]{}[@PRA.58.580; @PRL.110.025302]. Although it is efficient, excitations in the condensate flow are hardly avoidable when the rotation is fast. Internal atomic state manipulation[@PRL.99.260401; @PRL.106.130401; @PRL.110.025301] is another way to achieve circulation states. However, the experimental setup is difficult for magnetically trapped condensates due to the coupling of different Zeeman substates. Generating circulation states by phase imprinting[@PRA.86.013629; @PRL.97.170406; @PRA.97.043615] is also hard to be applied in systems with complex structures. Matter-wave guiding of BECs over large distances was realized experimentally in a neutral-atom accelerator ring based on magnetic time-averaged adiabatic potentials which can result in accelerating of BECs with hypersonic velocities but nonuniform density[@nature.570.205].
Generally, adding energy into a system through transient stirring usually leads to more disorder. However, in 2D superfluids, vortex bending and tilting and Kelvin wave perturbations are suppressed. Hence, the well known inverse energy cascade[@PhysFluids.10.1417; @PhysFluids.11.671; @PhysFluids.12.233; @RepProgPhys.43.547; @PhysRep.362.1], i.e. the emergence of order from turbulence, is induced by the inhibition of energy dissipation at small length scales[@PRL.113.165302; @PRA.93.043614; @Science.364.1264; @Science.364.1267]. One of the most remarkable phenomena is the formation of Onsager vortex clusters[@Nuo.Cim.6.279], which can be used as a vortex thermometry[@PRL.120.034504]. Moreover, in toroidal BECs, a large-scale persistent flow of atoms can be induced by the decay of 2D quantum turbulence (2DQT) in form of disordered vortex distributions, indicating energy transport from small to large length scales[@PRL.111.235301].
Although the decay of turbulent cascade is studied extensively, the unpredictability inherent to turbulent systems is further confounded by physical properties such as boundaries and spatial dimensionality. It is of interest to investigate the dissipative dynamics of 2DQT with varying geometries. Moreover, by driving the systems far away from equilibrium through a sudden geometrical quench of the trapping potential, we may find out some novel nonequilibrium phenomena. In this letter, by designing a BEC system in [a double concentric ring-shaped trap]{}, we study quantitatively the quench dynamics of condensates with random vortex distributions and the emergence of persistent atom flows (currents) with high winding numbers in this novel geometry, which will facilitate the study of superfluidity and quantum tunneling in a large range of transport regimes.
![Proposed experimental configuration (a) and some typical density distributions of the condensate with the corresponding phase diagrams during the dynamics of the system (b)-(e). The trap is deformed suddenly from the initial single ring-shaped configuration to a double concentric ring-shaped configuration at the beginning of the dynamical process. []{data-label="fig1"}](fig1.pdf){width="50.00000%"}
At zero temperatures, the complex-valued mean field order parameter of a trapped 2D condensate with $N$ atoms satisfies the GP equation of the form $$\begin{aligned}
i\hbar\frac{\partial \psi}{\partial t}=\left(-\frac{\hbar^2}{2m}\nabla^2+V_{trap}+g_{2D}N\left|\psi\right|^2\right)\psi\text{,}\end{aligned}$$ where $g_{2D}$ is the 2D coupling constant and the order parameter $\psi$ is normalized to 1. To avoid the detrimental effects of density variation, our system is based on an experimentally accessible box potential[@PRL.110.200406; @nature.539.72] where the condensates are spatially uniform away from the boundaries. Then, an 2D annular BEC is prepared in a double-concentric ring-shaped box potential of the form $$\begin{aligned}
V_{trap}=\left\{
\begin{array}{rcl}
0 & &{(R_1< r < R_{2}) \,\bigcup\, (R_3 < r < R_{4})}\\
V_{0} & & elsewhere \end{array}\right.\end{aligned}$$ with $r=\sqrt{x^{2}+y^2}$. [This configuration can be realized in the optical-box trap formed by two hollow tube beam and two sheet beams as shown schematically in Fig.\[fig1\](a).]{} $R_j\ (j=1,3)$ and $R_k\ (k=2,4)$ being the inner and outer radius of the two concentric rings, respectively. The width of the potential barrier between the two rings is then $d=R_3-R_2$. The units of time and length are $t_0=1/\omega_z$ and $a_0=\sqrt{\hbar/m\omega_z}$, respectively. If not specified, we choose $V_0=1.43\hbar\omega_z$ and $d=1.25a_0$. In our study, the superfluid system contains a BEC of $N=10^5$ $^{87}$Rb atoms with $\omega_z=2\pi\times 700$ Hz.The injection of angular momentum by vortex imprinting is embedded in vortex distribution rather than the macroscopic annular superflow, which is the same as in Ref.[@PRL.111.235301]. The randomly distributed vortices can be seeded in the condensate cloud by employing the phase imprinting technology or optical obstacle stirring. The initial states of the system are obtained by the imaginary time evolution of the GP equation numerically. We note that the vortices can be seeded into the system either initially or during the dynamic process, which will not affect the intrinsic physics. However, the latter will excite stronger density wave oscillations in the system.
![ Typical density profiles of the evolution of the condensate initially prepared in a double concentric ring-shaped trap with corresponding phase diagrams and vortex distributions (a)-(c), and the time evolution of the winding number of the persistent current (d) with 20 initially seeded vortices.[]{data-label="fig2"}](fig2.pdf){width="50.00000%"}
![Typical density profiles of the evolution of the condensate initially prepared in a sing ring-shaped trap and then released in a double concentric ring-shaped trap with corresponding phase diagrams (a)-(c), and the time evolution of the averaged winding number of the persistent current for 64 realizations (d) with 20 initially seeded vortices.[]{data-label="fig3"}](fig3-1.pdf "fig:"){width="40.00000%"} ![Typical density profiles of the evolution of the condensate initially prepared in a sing ring-shaped trap and then released in a double concentric ring-shaped trap with corresponding phase diagrams (a)-(c), and the time evolution of the averaged winding number of the persistent current for 64 realizations (d) with 20 initially seeded vortices.[]{data-label="fig3"}](fig3-2.eps "fig:"){height="20.00000%" width="45.00000%"}
The properties of the system under geometry-quench processes away from equilibrium are less straightforward. To gain insight into the role of the quenched trapping potential, we load atoms in a single ring trap initially to obtain an annular condensate as shown in Fig.\[fig1\]b, and then tune the trapping potential into a double ring geometry as schematically shown in Fig.\[fig1\]a. Initially, there are no vortices and circular current in the system. After suddenly released in the double ring trap, atoms appears in the outer ring region through tunnelling across the barrier between the two rings with highly disordered density distribution as shown in Fig.\[fig1\]c. With increasing homogeneity of density distribution of the condensate in the outer ring trap, vortex-antivortex pairs excited as shown in Fig.\[fig1\]d without changing the total angular momentum of the system. Finally, the system reach a steady state with a few vortex pairs with symmetric distribution within the outer ring, while the density distribution of the condensate in the inner ring tends to be uniform as shown in Fig.\[fig1\]e. This quadrupole structure is generally unstable in a 2D disk-shaped condensate[@SRep.6.29066]. The generation of topological excitations is similar to that during the interfering and merging of BECs in a double well potential[@PRA.87.023603; @PRA.88.043602]. For condensate initially loaded in a double concentric ring trap, the ground state gives uniform density distribution in both inner and outer rings. Then, a number of vortices with the same charge, for example $s=+1$, but random positions, are excited in the inner ring region. Due to the narrow region of the ring-shaped condensate, the vortices are not far from the boundary (barrier potential between two rings). During the evolution of the system, there are ghost vortices with opposite charge emerging within the flimsy between the two parts of the condensate, whose locations are identified by the yellow points as shown in Fig.\[fig2\]a-\[fig2\]c. The number of the ghost vortices are about the same as that of the vortices. The same as previous study about 2DQT in an annular 2D quantum fluid with amount of vortices, persistent atom flow forms spontaneously in the inner ring as the number of vortices varies (decreases with respect to the initial vortex number). As shown in Fig.\[fig2\]d, the current in the outer ring-shaped condensate emerges following the appearance of the current of the inner condensate. They will finally approach steady values. The variation of the circular current accompanies with the creation and annihilation of vortices in the system. The average growth or decay of the current is quantized with integer numbers indicated by the steps shown in Fig.\[fig2\], which means the transition between different quantized persistent current states. In Ref.[@PRA.80.021601], decay of the atomic current through a barrier in a toroidal BEC was in the form of $2\pi$ sharp drops, leaving steps on the time evolution of the circulation (winding) number. During the time evolution, most of the ghost vortices are pinned within the barrier region and can hardly escape to the condensate cloud. But we did identify that at some stage this happened as shown in Fig.\[fig2\]b-\[fig2\]c. Moreover, the initially seeded vortices can hardly tunnel from the inner ring-shaped condensate into the outer one, and no vortices excited in the outer ring-shaped condensate comparing with the situation in Fig.\[fig1\].
![Variation of the winding number of the circulation flow with respect to the number of initially seeded vortex. []{data-label="fig4"}](fig4.pdf){height="20.00000%" width="45.00000%"}
Next, we set the initial state of the system to be a single ring-shaped condensate with randomly distributed vortices of the same charge, and then deform the trap into a double concentric ring-shaped geometry as what we have done for the condensate without the initial vortices in Fig.\[fig1\]. The condensate originally in the inner ring tunnels through the barrier and spread in the outer ring-shaped region of the trap as shown in Fig.\[fig3\]a. Formation of vortex pairs still occurs in the outer ring-shaped condensate during the tunneling process while the density distribution of the condensate is getting more and more uniform as shown in Figs.\[fig3\]b-\[fig3\]c, which is similar to the evolution of the initial state without vortices in the inner ring-shaped condensate (see Fig.\[fig1\](c)). However, different from the results shown in Fig.\[fig1\](e), all the initially seeded and dynamically excited vortices in the system decay eventually accompanied by the emergence of persistent currents in both inner and outer rings. Comparing with the result of the system without quenching of the trapping potential as shown in Fig.\[fig2\], the winding number of the current in the outer ring turns to be much larger, which means that the transfer of angular momentum from vortices to global atom flows is much more efficient in this case. The gray and red lines in Fig.\[fig3\]g are the time evolution of the winding number of the atom flows in the inner ring and the outer ring, respectively, with a random distribution of 20 initially seeded vortices. In Ref.[@PRL.110.025302], the circulation caused by a rotating barrier is only stable when the angular frequency of the barrier is small, while the system becomes dynamically unstable with high angular frequencies resulting from the excitations of vortices are in the condensate cloud. [The error bars show the [standard deviation]{} of 64 realizations for a given initial vortex distribution.]{} The circulation flow of the outer ring-shaped condensate reaches its steady state much faster than the inner ring-shaped condensate. In Fig.\[fig4\], we show that the average winding number of the outer ring-shaped condensate increases [linearly]{} with increasing initial number of the seeded vortices. For 30 initial vortices, the winding number of the circulation flow is -16, while that of the inner ring-shaped BEC is only -3.
We also investigate the effects of the thickness and hight of the middle flimsy barrier on the winding number of the circulation flow as shown in Fig.\[fig5\]. The number of the initial vortices is kept to be 20. [For $d$ larger than $5a_0$]{} or [$V_0$ larger than 6$\hbar\omega_z$]{}, the tunneling is strongly suppressed to prevent the emergence of the circulation flow in the outer ring-shaped trap. The winding number of the outer ring-shaped condensate does not vary monotonically with respect to $d$ and $V_0$. There exists a specific value of these parameters to obtain maximum persistent current of the circulation flow. The variations of the winding number with respect to both parameters satisfy Gaussian distributions.
![Variation of the winding number of the circulation flow with respect to the number of initially seeded vortex. []{data-label="fig5"}](fig5.pdf){height="20.00000%" width="45.00000%"}
[In conclusion, we propose a novel setup and a quench protocol which can be used to generate persistent currents of circular condensate flows with high winding numbers. For a double ring-shaped condensate, if we quench vortices with the same charge and random distributions, the circular condensate flows can emerge in both inner and outer ring-shaped condensate due to 2D turbulent cascade and quantum tunneling. The differences of the winding numbers of the two flows are not significant. If we quench the condensate with randomly distributed like vortices by deforming the trap from a single ring-shaped geometry into a concentric double ring-shaped one, the winding number of the outer ring shaped condensate will be much larger than that of the inner ring-shaped condensate flow. However, if the initial state of the system is without any vortices, the geometric quench itself cannot generate persistent atom flows in the system. The persistent currents generated in our system are with nearly uniform density distributions and free from vortex excitations. This method is much more efficient to create circulation flows of atoms than the technique of rotating barrier and may be used to design complex devices based on ultracold atoms. ]{}
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by the National Natural Science Foundation of China under grants Nos. 11775178, 11947301 and 61835013, National Key R&D Program of China under grants No. 2016YFA0301500, the Strategic Priority Research Program of the Chinese Academy of Sciences under grants Nos. XDB01020300 and XDB21030300, the Major Basic Research Program of Natural Science of Shaanxi Province under grants Nos. 2017KCT-12 and 2017ZDJC-32. This research is also supported by The Double First-class University Construction Project of Northwest University.
[37]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase 10.1103/PhysRevLett.95.143201) [****, ()](\doibase 10.1103/PhysRevLett.103.140405) [****, ()](\doibase
10.1038/nature12958) [****, ()](\doibase 10.1103/RevModPhys.73.307) [****, ()](\doibase
10.1103/PhysRevA.74.023617) [****, ()](\doibase
10.1103/PhysRevLett.99.260401) [****, ()](\doibase 10.1088/1367-2630/10/4/043012) [****, ()](\doibase
10.1103/PhysRevLett.106.130401) [****, ()](\doibase
10.1103/PhysRevA.86.013629) [****, ()](\doibase 10.1103/PhysRevLett.111.205301) [****, ()](\doibase 10.1103/PhysRevA.75.023615) [****, ()](\doibase 10.1088/1367-2630/aa5a6d) [****, ()](\doibase
10.1088/2040-8978/18/9/093001) [****, ()](\doibase 10.1103/PhysRevA.58.580) [****, ()](\doibase 10.1103/PhysRevLett.110.025302) [****, ()](\doibase 10.1103/PhysRevLett.110.025301) [****, ()](\doibase
10.1103/PhysRevLett.97.170406) [****, ()](\doibase
10.1103/PhysRevA.97.043615) [****, ()](\doibase 10.1038/s41586-019-1273-5) [****, ()](\doibase 10.1063/1.1762301) @noop [****, ]{} [****, ()](\doibase 10.1063/1.870297) [****, ()](\doibase 10.1088/0034-4885/43/5/001) [****, ()](\doibase https://doi.org/10.1016/S0370-1573(01)00064-3) [****, ()](\doibase 10.1103/PhysRevLett.113.165302) [****, ()](\doibase 10.1103/PhysRevA.93.043614) [****, ()](\doibase
10.1126/science.aat5718) [****, ()](\doibase 10.1126/science.aat5793) [****, ()](\doibase 10.1007/BF02780991) [****, ()](\doibase
10.1103/PhysRevLett.120.034504) [****, ()](\doibase
10.1103/PhysRevLett.111.235301) [****, ()](\doibase 10.1103/PhysRevLett.110.200406) [****, ()](\doibase 10.1038/nature20114) [****, ()](http://dx.doi.org/10.1038/srep29066) [****, ()](\doibase 10.1103/PhysRevA.87.023603) [****, ()](\doibase 10.1103/PhysRevA.88.043602) [****, ()](\doibase 10.1103/PhysRevA.80.021601)
|
---
abstract: 'A numerical approach to disordered 2D superconductors described by BCS mean field theory is outlined. The energy gap and the superfluid density at zero temperature and the quasiparticle density of states are studied. The method involves approximate self-consistent solutions of the Bogolubov-de$\,$Gennes equations on finite square lattices. Where comparison is possible, the results of standard analytic approaches to this problem are reproduced. Detailed modeling of impurity effects is practical using this approach. The [*range*]{} of the impurity potential is shown to be of [*quantitative importance*]{} in the case of strong potential scatterers. We discuss the implications for experiments, such as the rapid suppression of superconductivity by Zn doping in Copper-Oxide superconductors.'
address: ' Research Center in Superconductivity, University of Cambridge, Madingley Road, Cambridge CB3 0HE, United Kingdom '
author:
- 'T. Xiang and J. M. Wheatley'
title: ' Non-magnetic impurities in two dimensional superconductors '
---
epsf
Introduction
============
It is well established that chemical substitutions on copper oxide superconductors can qualitatively alter the properties of these materials in both their normal and superconducting states[@Exp]. In the normal state,impurity substitution can be used to test models of electronic transport. In the superconducting state, impurity scattering effects are sensitive to order parameter symmetry[@Theory] as well as other properties. It is well known that non-magnetic impurities are pair breakers in d-wave or other non-trivial pairing states with nodes of the energy gap on the Fermi surface. They produce a finite lifetime of the quasiparticles around the gap nodes and a finite density of states at low energy.
In order to parameterize impurity scattering, a physical picture for the normal state, the superconducting state and the impurity is required. In this paper we are motivated by Zn doping experiments in cuprates which we assume can be treated as a pure potential scatterer of fermions which obey Luttinger’s theorem. Since d-orbitals of Zn$^{++}$ are fully occupied, it is naively expected to behave as a non-magnetic impurity. However, electronic correlations may modify this picture somewhat. Nuclear Magnetic Resonance measurements by Mahajan and co-workers[@NMR] show that the copper spin correlations are severely modified on neighboring Cu sites, which could in principle give rise to effects analogous to the spin-flip scattering process of conventional magnetic impurities. As discussed recently by Borkowski and Hirschfeld[@BH], the unknown relative strength of spin flip and impurity scattering rates in Zn doped systems poses an obstacle to the quantitative analysis of experiments involving Zn even within the conventional BCS formalism.
Even if spin-flip scattering is negligible, the modification of copper spin correlations in the vicinity of an impurity site has the effect of increasing the range of an effective scattering potential. Information about the effective potential in $YBa_2(Cu_{1-x}Zn_x)_3O_{6.9}$ can be obtained from the residual sheet resistance estimated by extrapolation from the normal state. The residual resistance from two dimensional potential scattering is given by the approximate expression[@ziman] $$\Delta\rho = {1\over \pi} x
{\displaystyle h\over \displaystyle e^2}
\sum_l\sin^2(\delta_l-\delta_{l+1}).
\label{rr}$$ $\delta_l$ are the phase shifts which are constrained to satisfy the Friedel sum-rule $\Delta Z={2\over \pi}\sum_{l=-\infty}^\infty \delta_l$. $\Delta Z$ is the difference in the number of conduction electrons in the system without impurities and with one impurity. If we assume Zn removes one electron from the conduction band (i.e. $\Delta Z=-1$) and the effective potential is an impenetrable disc of radius $a$, we estimate that the residual sheet resistance is $\approx 0.6k\Omega$ for $x=0.01$ in the $l=0$ dominant scattering channel ($a\sim 0.54/k_f$), which is a factor three smaller than the experimental value $1.7k\Omega$[@RR]. This discrepancy may indicate appreciable phase shifts in one or more higher angular momentum channels. In fact the $l=2$ “near resonant channel", which requires a larger radius of the scattering potential ($a\sim 2.8/k_f$)[@phaseshifts] yields $\approx 1.8k\Omega$ for $x=0.01$. Recently Poilblanc, Scalapino, and Hanke[@Poilblanc] have investigated the effects of non-magnetic impurities in antiferromagnetically correlated systems. They find that the scatterings in $l\le 2$ channels are strong in that system.
The above discussion suggests that detailed structure of the impurity potential is important in making a quantitative study of Zn doped materials. As seen below, lattice effects may also be of quantitative importance for short coherence length superconductors. Recently, in a brief report[@Tao], we presented a numerical study of the disorder effect in two dimensional superconductors of various pairing symmetries in the limit of strong impurity potentials. We demonstrated that a short but finite range potential has a much stronger effect than an on-site (“$\delta$-function”) potential. In fact the finite range of the potential may be the primary reason for the rapid suppression of pairing correlations with impurity concentration in $YBa(Cu,Zn)O$. The importance of finite potential range has been pointed out recently by Balatsky et al[@Balatsky] in connection with non-universal behavior of the low frequency conductance in d-wave superconductors.
In this paper we present a more detailed study of the disorder effect in two dimensional superconductors. The model used here is a lattice BCS mean-field Hamiltonian with disorder defined by, $$H[\Delta_{\mathbf r\tau} ]
=-t\sum_{\langle {\mathbf r}{\mathbf r}^\prime \rangle \sigma}
c_{{\mathbf r}\sigma}^\dagger c_{{\mathbf r}^\prime\sigma}
+\sum_{{\mathbf r}{\mathbf \tau}}
(\Delta_{{\mathbf r}{\mathbf \tau}}c_{{\mathbf r}\uparrow}^\dagger
c_{{\mathbf r}+{\mathbf \tau}\downarrow}^\dagger +
h.c.)+\sum_{{\mathbf r}\sigma}(\sum_{{\mathbf r}_{i}}
V_{{\mathbf r}_{i},\mathbf r}-\mu )
c_{{\mathbf r}\sigma}^\dagger c_{{\mathbf r}\sigma} ,
\label{ee1}$$ where $\langle {\mathbf r}{\mathbf r}^\prime\rangle $ denotes nearest neighbors, $\mu$ is the chemical potential. $V_{{\mathbf r}_{i},\mathbf r}$ is a scattering potential of an impurity at ${\mathbf r}_i$. Lacking detailed knowledge of the scattering potential in high-$T_c$ cuprates, we assume a model potential form for it: $V_{{\mathbf r}_i,\mathbf r}$=$V_0\delta_{{\mathbf r}_i,\mathbf r}
+V_1(\delta_{{\mathbf r}_i-{\mathbf r} \pm \hat x}
+\delta_{{\mathbf r}_i-{\mathbf r} \pm \hat y}) $. When $V_1=0$, it is a $\delta$-function potential, otherwise it is finite ranged. $\Delta_{{\mathbf r}{\mathbf \tau}}$ is the superconducting gap parameter. For a given filling factor of electrons $n_e$, $\Delta_{\mathbf r\tau}$ and $\mu$ should be determined self-consistently from the relations $$\begin{aligned}
\Delta_{{\mathbf r}{\mathbf \tau}}&=&J\langle c_{{\mathbf r}
\uparrow}c_{{\mathbf r}+
{\mathbf \tau}\downarrow}\rangle _{\Delta_{\mathbf r\tau}},
\label{gapeq} \\
n_e&=&{1\over N}\sum_{{\mathbf r} \sigma} \langle c_{{\mathbf r}
\sigma}^{\dag} c_{{\mathbf r}\sigma}
\rangle _{\Delta_{\mathbf r\tau}},
\label{numbereq}\end{aligned}$$ where J is the coupling constant (assumed disorder independent), $N$ is the system size, and $\langle A \rangle _{\Delta_{\mathbf r\tau}}$ means the average of A in the ground state of $H[\Delta_{\mathbf r\tau}]$. Without disorder $\Delta_{{\mathbf r}{\mathbf \tau}}$ is independent of ${\mathbf r}$ and has a particular symmetry with respect to ${\mathbf \tau}$. In this paper only the on-site s-wave pairing state $\Delta_{\mathbf r\tau}
=\Delta \delta_{\mathbf \tau 0}$ and the d-wave pairing state $\Delta_{\mathbf r\tau}=\Delta (\delta_{{\mathbf \tau}\pm \hat x}-
\delta_{{\mathbf \tau}\pm \hat y})$ in two dimensions will be considered.
The Hamiltonian (\[ee1\]) is bilinear in fermion operators and can be diagonalized by solving a one-particle problem. $\Delta_{\mathbf r\tau}$ and $\mu$ should be determined self-consistently from Eqs. \[gapeq\] and \[numbereq\], which amount to solving the Bogolubov-de Gennes equations for the disordered superconductor. For simplicity in calculation, we perform a particle-hole transformation for the down-spin electrons, i.e. $c_{{\mathbf r}\downarrow}
\longleftrightarrow c_{{\mathbf r}\downarrow}^\dagger$, and re-express (\[ee1\]) as $$\begin{aligned}
H^\prime [\Delta_{\mathbf r\tau} ]& =&
-t\sum_{\langle {\mathbf r}{\mathbf r}^\prime \rangle }
(c_{{\mathbf r}\uparrow}^\dagger c_{{\mathbf r}^\prime\uparrow}
- c_{{\mathbf r}\downarrow}^\dagger
c_{{\mathbf r}^\prime\downarrow})
+\sum_{{\mathbf r}{\mathbf \tau}}
(\Delta_{{\mathbf r}{\mathbf \tau}}
c_{{\mathbf r}\uparrow}^\dagger
c_{{\mathbf r}+{\mathbf \tau}\downarrow} +
h.c.)\nonumber\\
&&+\sum_{{\mathbf r}}(\sum_{{\mathbf r}_i}
V_{{\mathbf r}_i,\mathbf r}-\mu )
( c_{{\mathbf r}\uparrow}^\dagger c_{{\mathbf r}\uparrow}
-c_{{\mathbf r}\downarrow}^\dagger
c_{{\mathbf r}\downarrow})+const.
\label{ee2}\end{aligned}$$ $H^\prime$ has the usual tight binding model form, but the hopping constant, the chemical potential, and the impurity potential have opposite signs for the up spin electrons and the down spin electrons[@half]. In the remainder of the paper we set the hopping constant t=1.
In Sec. II, the numerical method and essential approximation is outlined. Results for the disorder dependence of the zero temperature gap and superfluid density $\rho_s$ as well as the quasiparticle density of states $\rho$ are presented. In Sec. III, a concluding remark is given.
Numerical method and results
============================
Gap parameter and superfluid density
------------------------------------
In the presence of disorder, the energy gap is space dependent. We determine $\Delta_{\mathbf r\tau}$ by iteratively solving $H^\prime$ with the self-consistent conditions. We start from an initial gap function $\Delta_{\mathbf r\tau}$ with a certain pairing symmetry. After diagonalizing $H^\prime$, we find a new gap function $\Delta_{\mathbf r\tau}$ from the self-consistent equations and then use it as input to repeat the above process until the self-consistent conditions are satisfied. This is a strict self-consistent iterative process. However, since the gap function at every site needs be adjusted to satisfy the self-consistent equations, this is excessively time consuming when an average over a large number of impurity configurations is required. We shall however perform this strict self-consistent iteration procedure only for studying the properties of a single impurity system and for checking the accuracy of the approximation used in many-impurity cases.
As mentioned above, to solve the self-consistent equations, the Hamiltonian needs be exactly diagonalized. This can be done, however, only on small lattices in the presence of disorder. For a superconductor, a characteristic length scale is the superconducting correlation length $\xi\sim {\hbar}v_F/\pi\Delta$. If $\xi$ is larger than the dimension of the system, the finite size effect is large and the analysis of the disorder effect may be subtle. To avoid this situation, we shall limit our calculations only to cases where $\xi$ is much smaller than the dimension of the system.
The existence of a finite superfluid density $\rho_s$ is a defining property of superconductors. Experimentally $\rho_s$ is determined from the microwave measurement of the penetration depth. $\rho_s$ on a finite lattice can be evaluated directly from the current-current correlation function, using the eigenfunctions obtained from the iteration procedure described above, [@super] $${\rho_s \over 4} =\langle -K_x\rangle
-\Lambda_{xx}(q_x=0,q_y\rightarrow 0,\omega =0)$$ where $\langle A\rangle =Tr(Ae^{-\beta H})/Tre^{-\beta H}$, $\langle K_x\rangle $ is the kinetic energy along x-direction, and $$\begin{aligned}
\Lambda_{xx}({\mathbf q},\omega )&=& {i\over V}\int_{-\infty}^tdt'
e^{i\omega (t-t')}\langle
[J_x^p(-{\mathbf q},t),J_x^p({\mathbf q},t')]\rangle
\nonumber\\
&=&{1\over V}\sum_{n_1\not= n_2}{A_{n_1,n_2}(-q)A_{n_2,n_1}(q)
\over \omega +i\delta +E_{n_1}-E_{n_2}}
(\langle c_{n_1}^\dagger c_{n_1}\rangle
- \langle c_{n_2}^\dagger c_{n_2}\rangle )\end{aligned}$$ with $c_n = \sum_{{\mathbf r}\sigma }
\phi_n({\mathbf r},\sigma ) c_{{\mathbf r}\sigma }$ the quasiparticle operator and $$A_{n,m}({\mathbf q})= \sum_{{\mathbf r}\sigma}
e^{i{\mathbf q}\cdot {\mathbf r}}(
\phi_n({{\mathbf r}+x\sigma }) \phi_m({{\mathbf r}\sigma })
-\phi_n({{\mathbf r}\sigma }) \phi_m({{\mathbf r}+x\sigma })).$$
Before studying disordered systems, we consider a one-impurity system. We first consider the change of the gap function induced by an impurity. Fig. \[fig1\] shows the self-consistent energy gap $\Delta_{\mathbf r \tau}$ for a s-wave superconducting state with one impurity on a 21$\times$21 lattice. Since the scattering potential is short ranged, the gap function changes only in the vicinity of the impurity. Far away from the impurity, $\Delta_{\mathbf r\tau}$ approaches to the value of the energy gap without disorder. In a region with a length scale of the range of scattering potential around the impurity site $\Delta_{\mathbf r}$ is largely reduced due to the strong suppression of the probability of an electron hopping to this region by the impurity scattering. Beyond this region, there exists a relatively larger region with a length scale comparable with the superconducting correlation length $\xi$ where a weak oscillation of $\Delta_{\mathbf r\tau}$ in space is observed. This oscillation is due to the interplay between the impurity scattering and the superconducting correlations. Because of the lattice effect, $\Delta_{\mathbf r\tau}$ is not isotropic in space. It is relatively strong along the diagonal direction. Along other directions, it is too weak to be resolved from the figure. For the d-wave state, similar results have been found. But the oscillation of $\Delta_{\mathbf r\tau}$ along two axes seems more apparent in this case.
For finite impurity doping systems where many impurity configurations are used in the disorder average we shall approximate $\Delta_{{\mathbf r}{\mathbf \tau}}$ in (\[ee1\]) by the average of the pairing correlation functions $\langle c_{{\mathbf r}\downarrow}
c_{{\mathbf r}+{\mathbf \tau}\uparrow}\rangle$ in space, ${\bar\Delta}_{\mathbf \tau}$, so that only a simplified self-consistent equation, $$\bar{\Delta}_{\tau}
={1\over N}\sum_{\mathbf r}J\langle c_{{\mathbf r}\downarrow}
c_{{\mathbf r}+{\mathbf \tau}\uparrow}\rangle_{{\bar\Delta}_\tau}$$ needs be solved. This approximation is to ignore the fluctuation of $\Delta_{\mathbf r\tau}$ (but not $\langle c_{{\mathbf r}\downarrow}
c_{{\mathbf r}+{\mathbf \tau}\uparrow}\rangle$) in space.
The errors resulting from the above approximation can be found by directly comparing the results obtained with and without the approximation. We have calculated the errors for the pairing amplitude for several arbitrarily chosen configurations of impurities in both strong and weak scattering potential limits. For all the cases we have studied, we find that the relative errors in the average energy gap are small compared with the combined errors produced by the disorder average or the finite size effect. For example, for a randomly chosen system of 3 impurities on a 14$\times$14 lattice, the relative errors in the average energy gap are 0.1% (0.6%) in a weak potential $V_0=2$ and $V_1=0$ and 2%(2%) in a strong potential $V_0=20$ and $V_1=0$ for the s-wave (d-wave) pairing state. As the impurity concentration increases, the error increases slightly. The errors for the local pairing correlation functions $\langle c_{{\mathbf r}\downarrow}
c_{{\mathbf r}+{\mathbf \tau}\uparrow}\rangle$ are larger than that for the average gap, but still small. Fig. \[fig2\], as an example, shows the relative error pattern for the local correlation function $\langle c_{{\mathbf r}\downarrow}
c_{{\mathbf r}+{\mathbf \tau}\uparrow}\rangle$ for a s-wave pairing state. We find that the errors for $\langle c_{{\mathbf r}\downarrow}
c_{{\mathbf r}+{\mathbf \tau}\uparrow}\rangle$ are largest (about 5%) at the impurity sites. Clearly $\langle c_{{\mathbf r}\downarrow}
c_{{\mathbf r}+{\mathbf \tau}\uparrow}\rangle$ is overestimated in the vicinity of impurities and underestimated far away from the impurities. Nevertheless it is encouraging that the error made in neglecting off-diagonal disorder is small[@lambert].
Using the above approximation, we have evaluated the average gaps $\bar\Delta$ for both s- and d-wave pairing states in different scattering potentials. For all the cases we have studied, we find that $\bar\Delta$ decays almost linearly with x for small x. Fig. \[fig3\] shows ${\bar\Delta}_\tau$ as a function of the impurity concentration x for both pairing states in a strong $\delta$-function scattering potential $V_0=20$ on a 14$\times$14 lattice. All four curve shown in the figure are nearly parallel to each other although their values at x=0, i.e. $\bar\Delta (0)$, are quite different. This remarkable behavior indicates that the reduction of $\Delta$ by disorder ($d\Delta / dx$) is determined only by the scattering potential to a first approximation and is independent of the pairing symmetries and the values of the energy gaps for pure samples. This nearly universal property of the energy gap could be of use in future studies of disordered superconductors. For weak $\delta$-function potentials or finite but short ranged potentials, similar results for $\bar\Delta$ have been found. But for a weak $\delta$-function potential the slope of the decay of $\bar\Delta$ with x becomes smaller.
Our calculations for $\bar\Delta$ and $\rho_s$ have been done mostly on lattices with size ranged from 10$\times$10 to 18$\times$18 sites. In general, $\bar\Delta$ and $\rho_s$ are size dependent. However, for the quantities which are physically more interesting, the relative energy gap $\bar\Delta (x)/\Delta (0)$ and the relative superfluid density $\rho_s (x)/\rho_s (0)$, the finite size effect is small. A comparison for $\bar\Delta (x)/\Delta (0)$ and $\rho_s (x)/\rho_s (0)$ for the s- and d-wave states with a $\delta$-function potential $V_0=8$ on lattices of 10$\times$10, 14$\times$14, and 18$\times$18 sites is given in Fig. \[fig4\]. Similar results have been found for other impurity potentials. In general, we find that the finite size effect for $\rho_s$ is larger than that for ${\bar\Delta}_\tau$ when ${\bar\Delta}_\tau$ is small (i.e. large $\xi$). Physically this is because that $\Delta_{\mathbf r\tau}$ is determined only by the local pair correlation function while $\rho_s$ is determined by the current-current correlation function at long wavelengths which in turn is determined by the properties of the eigenfunction on the whole lattice.
Fig. \[fig5\] shows $\bar{\Delta} (x)/\bar{\Delta} (0)$ and $\rho_s (x)/\rho_s (0)$ as functions of $x$ for the s- and d-wave pairing states in three potentials. For a weak $\delta$-function potential, $\bar{\Delta}
(x)/\bar{\Delta} (0)$ and $\rho_s (x)/\rho_s (0)$ decrease very slowly with $x$ and the difference between the s- and d-wave pairing states is also small. With increasing $V_0$, the difference between these two pairing states increases. The disorder has stronger effect on the d-wave state than the s-wave state; in particularly, $\rho_s$ falls much faster in the d-wave state than in the s-wave state. This property may be useful for distinguishing a d-wave pairing state from a s-wave pairing state in the unitary scattering limit from an experimental point of view. But if an off-site scattering potential (i.e. $V_1$ term) is present, this difference will be eventually reduced. The disorder effect is clearly strongly enhanced in a finite range potential than in a $\delta$-function potential.
Fig. \[fig6\] shows $\bar{\Delta} (V_0,V_1)/\bar{\Delta} (0,0)$ as functions of $V_0$ for the s- and d-wave pairing states with $x$=0.02. The qualitative behaviors of $\bar{\Delta} (V_0,V_1)/\bar{\Delta} (0,0)$ are similar for all the cases shown in the figure. $\bar{\Delta} (V_0,V_1)/\bar{\Delta} (0,0)$ decreases with $V_0$ for small $V_0$, but soon becomes saturated when $V_0$ surpasses the band width. $d\bar{\Delta}/dx$ for the s- and d-wave states can be very large depending on the value of $V_1$.
Density of states
-----------------
Now we consider the effect of disorder on the density of states of quasiparticles. We calculate the density of states using a recursion method[@recursion]. This method addresses the local density of states of an infinite lattice. In this method, the density of states $\rho (E)$ is obtained from the imaginary part of the one-particle Green’s function $G(E)$. $$\rho (E)=\lim_{\epsilon\longrightarrow 0}-{1\over \pi} {\hbox {Im}}
G(E+i\epsilon).$$
Given a starting state $|0\rangle$, the recursion method is defined by recurrence relations $$H|0\rangle = a_0 |0\rangle + b_1 |1\rangle \label{rec1}$$ and $$H|n\rangle = b_n |n-1\rangle +a_n |n\rangle + b_{n+1} |n+1\rangle
\qquad (n >0), \label{rec2}$$ where $\{ |n\rangle \}$ is a set of normalized bases generated automatically from equations (\[rec1\]) and (\[rec2\]). From the a’s and b’s generated, $G(E)$ can be expressed in a continued fraction form $$G(E)={1\over\displaystyle (E-a_0)-{\displaystyle b_1^2\over
\displaystyle (E-a_1)-{\displaystyle b_2^2\over
\displaystyle (E-a_2)- \cdot\cdot\cdot}}}.$$ In real calculation, this continued fraction is truncated at a certain step and the remainder of the continued fraction is replaced by a parameter which is determined such that the error is minimized. We truncate the continued fraction at a step when the difference between the result obtained at that step and that with 5 more steps is smaller than the error demanded. In using the recursion method, the values of $\Delta_{\mathbf r\tau}$ and $\mu$ obtained previously on finite lattices will be used. When the impurity concentration is finite, the approximation $\Delta_{\mathbf r\tau}={\bar\Delta}_\tau$ is assumed. For one impurity system, the strict self-consistent solution for $\Delta_{\mathbf r\tau}$ on a small lattice around the impurity is used, while for the rest part of the lattice $\Delta_{\mathbf r\tau}$ are approximated by the average value of $\Delta_{\mathbf r\tau}$ on the edges of the small lattice.
As discussed by Byers et al[@byers; @choi] the densities of states in the vicinity of the impurity is in principle measurable via the spatial variation of the tunneling conductance around an impurity with a scanning-tunneling-microscope study of the surface of a superconductor. In continuum space, the local density of states (or the tunneling conductance) around an impurity has been calculated by Byers et al[@byers] for both the s- and d-wave pairing states and by Choi[@choi] for the d-wave pairing state. They find that the density of states for a give energy oscillates in space and depends strongly on the anisotropy of the gap parameter. When the energy is larger than $\Delta$, the oscillation is largest in the directions of the gap maxima and smallest in the directions of the gap minima. In lattice space, however, we find that their results are partly altered. Fig. \[fig7\] shows the impurity induced density of states as a function of distance from an impurity along two directions for the s- and d-wave pairing states. (Here the fully self-consistent isolated impurity result for the gap function on a 21 $\times$ 21 lattice shown in Fig. \[fig1\] is used as an input.) The density of states oscillates in space with an energy and direction dependent wavelength in agreement with the results of Ref. [@byers]. In the s-wave case, the oscillation along the diagonal direction is much larger than that along the x-axis direction, in contrast to the isotropic s-wave pairing state in continuum space. Since the energy gaps are the same on the Fermi surface for a s-wave states, this difference is purely a lattice effect. For the d-wave pairing state, the oscillations along two directions are not so different as shown in Refs. [@byers; @choi]. The impurity induced density of states decays slightly faster along the x-axis direction than along the diagonal direction.
To compute the density of states with finite doping of impurities, we have used the results of the average gap $\bar\Delta$ obtained previously. The fluctuation of $\Delta_{\mathbf r\tau}$ in space is ignored in this calculation. Fig. \[fig8\] shows the density of states for the s- and d-wave pairing states with different potentials and x. The main results are summarized as follows:
\(a) For s-wave pairing with weak potential or strong potential with very small x, $\rho$ has no qualitative change with respect to the case without disorder. In particularly, the energy gap still exists and is hardly changed by disorder in agreement with Anderson theorem[@pwa].
\(b) For the d-wave pairing state with weak scattering potential, the change of $\rho$ with respect to the case without disorder is small. But $\rho$ at the Fermi energy $E_F$ becomes finite, in consistent with the non-magnetic impurity scattering theory in the Born scattering limit[@GK].
\(c) For the d-wave pairing state with very strong onsite potential, $\rho$ shows a peak around $E_F$. This results agrees very well with the self-consistent t-matrix theory for the d-wave superconductor in the unitary scattering limit[@PEAK]. On the other hand it also shows that the approximations made in the self-consistent t-matrix theory, such as ignoring the vertex corrections and the energy dependence of the self-energy, are valid for the d-wave state.
\(d) For the d-wave pairing state with a strong on-site potential or both pairing states with a finite range potential, $\rho$ at $E_F$ grows quickly with increasing x and becomes comparable with the average density of states at some critical x. For the s-wave pairing state with a strong $\delta$-function potential, a finite gap remains when x is smaller than a critical value x$_c$ within numerical errors. When x$>$x$_c$, the gap vanishes (but $\rho_s$ and ${\bar\Delta}$ are non-zero), $\rho$ at $E_f$ is small and increases slowly with increasing x.
\(e) All singularities of $\rho$ are suppressed by disorder average in these calculations. We have not found any evidence for the singular behavior predicted recently by Nersesyan et al[@ntw] within numerical error.
Conclusion
==========
We have discussed a straightforward numerical technique which allows detailed effects of various impurity potentials on superconductors to be investigated with a BCS mean field framework. In particular we evaluated the gap parameter, the superfluid density, and the density of states for the s-wave and d-wave superconducting states with non-magnetic impurities, as functions of impurity concentrations and scattering potentials. For one impurity systems, the local density of states induced by impurity oscillates in space, in agreement with known analytic results. For s-wave pairing, the energy gap and the density of states are hardly affected by weak disorder (i.e. either weak scatterers or dilute strong scatterers), consistent with Anderson theorem. In dilute impurity limit, our results agree well with the self-consistent t-matrix theory in both Born and unitary scattering limits, and in both s- and d- pairing states. For the d-wave pairing state, the density of states at $E_F$ becomes finite even for weak scattering potential, consistent with the Born scattering theory of the d-wave superconductor. For the d-wave pairing state with strong on-site potential, the density of states is in good agreement with the self-consistent t-matrix theory for the d-wave superconductor in the unitary limit. For strong scatterers, the energy gap of the s-wave state disappears beyond a critical doping level which is sensitive to the range of the impurity potential. A finite range potential is shown to have a stronger effect than a short range potential in either pairing state. For Zn doped YBaCuO, experiments find that $T_c$ varies almost linearly with $x$ and drops about 25% for 2% Zn doping[@TC]. If we assume the change of $T_c$ is equivalent to the change of $\bar{\Delta}$ at zero temperature, we find that $d\bar{\Delta}/dx$ in a $\delta$-function potential is too small to fit quantitatively with experiments even in unitary scattering limit. However for a finite range potential, no such difficulty exists.
Acknowledgement
===============
We wish to thank J. Loram, J. Cooper, and P. Hirschfeld for helpful communications.
Gang Xiao, F. H. Streitz, A. Gavrin, Y. W. Du, and C. L. Chien, Phys. Rev. B [**35**]{}, 8782 (1987); Y. Maeno, T. Tomita, M. Kyogoku, S. Awaji, Y. Aoki, K. Hoshino, A. Minami, and T. Fujita, Nature [**328**]{}, 512 (1987); K. Ishida, Y. Kitaoka, T. Yoshitomi, N. Ogatta, T. Kamino, and K. Asayama, Physica C [**179**]{}, 29 (1991).
For recent discussions, see W. H. Hardy, D. A. Bonn, D. C. Morgan, R. Liang, and K Zhang, Phys. Rev. Lett. [**70**]{}, 3999 (1993); P. J. Hirschfeld and N. Goldenfeld, Phys. Rev. B [**48**]{}, 4219 (1993); P. A. Lee, Phys. Rev. Lett. [**71**]{}, 1887 (1993).
A. V. Mahajan, H. Alloul, G. Collin, and J. F. Marucco, Phys. Rev. Lett. [**72**]{}, 3100 (94).
L. S. Borkowski and P. J. Hirschfeld, preprint.
A derivation for the impurity residual resistivity in 3D can be found from G. D. Mahan, “Many-particle physics" (2nd edition, 1990 Plenum Press, New York), Chapter 7.
T. R. Chien, Z. Z. Wang, and N. P. Ong, Phys. Rev. Lett. [**67**]{}, 2088 (1991).
It is worth noting some features of this 2D scattering problem. Using the impenetrable disc model with radius $a$ we find that $k_f a=0.542$ satisfies the sum rule with near unitary scattering in the $l=0$ channel; the corresponding phase shifts are $\delta_0 = -0.75 {(\pi / 2)}$, $\delta_1 =-0.12{(\pi / 2)}$, $\delta_2 = -0.005{(\pi / 2)}$ etc. The next values of $k_f a$ compatible with the sum rule are $1.52$ and $2.787$. The latter value leads to scattering phase shifts $\delta_0 = -0.25 {(\pi / 2)}$, $\delta_1 = 0.64{\pi / 2}$, $\delta_2 = -0.69{(\pi / 2)}$, $\delta_3 =-0.26{(\pi / 2)}$ and $\delta_4 = -0.06{(\pi / 2)}$; they are appreciable in all channels up to $l=2$. This set of phase shifts leads to the best agreement with the $YBCO$ data.
D. Poilblanc, D. J. Scalapino, and W. Hanke, Phys. Rev. Lett. [**72**]{}, 884 (1994).
T. Xiang and J.M. Wheatley, Physica C, to be published.
A. V. Balatsky, A. Rosengren, and B. L. Altshuler, Phys. Rev. Lett. [**73**]{} 720 (1994).
The results presented here are for a half-filled band. For the simple tight-binding model used here, this has a van Hove singularity at the Fermi surface in the normal state. This fact plays no role in our $T=0$ calculations however. To see this note that (a) in the absence of disorder, the superconducting gap pushes the van Hove singularity far from the Fermi surface, (b) in the presence of disorder, all singularities in the density of states are rapidly suppressed after the disorder average, even in the normal state. The chemical potential $\mu$ must be determined self-consistently along with $\Delta$ even at half filling, because particle-hole symmetry is destroyed by disorder.
D. J. Scalapino, S. R. White, and S. C. Zhang, Phys. Rev. Lett. [**68**]{}, 2830 (1992).
For a discussion of the effect of off-diagonal disorder, see for example, V. C. Hui and C. J. Lambert, J. Phys.: Condens. Matter [**5**]{} 697 (1993).
R. Haydock, V. Heine, and M. J. Kelly, J. Phys. C [**5**]{}, 2845 (1972).
J. M. Byers, M. E. Flatte, and D. J. Scalapino, Phys. Rev. Lett. [**71**]{}, 3363 (1993).
C. H. Choi, Phys. Rev. B [**50**]{}, 3491 (1994).
P. W. Anderson, J. Phys. Chem. Solids, [**11**]{}, 26 (1959); L. P. Gorkov, Sov. Phys. JETP [**10**]{}, 998 (1960).
L. P. Gorkov and P. A. Kalugin, Pis’ma Zh. Eksp. Teor. Fiz. [**41**]{} 208 (1985) \[JETP Lett. [**41**]{} 253 (1985)\].
P. J. Hirschfeld, P. Wolfle, and D. Einzel, Phys. Rev. B [**37**]{}, 83 (1988).
A. A. Nersesyan, A. M. Tsvelik, and F. Wenger, Phys. Rev. Lett. [**72**]{}, 2628 (1994).
A. Janossy, J. R. Cooper, L. C. Brunel, and A. Carrington, Phys. Rev. B [**50**]{}, 3442 (1994); J. R. Cooper, private communication.
|
\#1[[$\backslash$\#1]{}]{}
In order to understand the mechanism of high-temperature superconductivity in doped cuprates, a central issue has been the evolution of the electronic structure with hole doping near the filling-control metal-insulator transition (MIT). In spite of extensive photoemission studies, it still remains unclear how the electronic structure evolves, especially, between underdoped metal and antiferromagnetic insulator. For a systematic study of the doping dependence near the MIT, La$_{2-x}$Sr$_x$CuO$_4$ (LSCO) is a suitable system. It has the simplest crystal structure with single CuO$_2$ layers and the hole concentration in the CuO$_2$ plane is well controlled over a wide range and uniquely determined by the Sr concentration $x$ (and small oxygen non-stoichiometry). So far photoemission studies of high-$T_{c}$ cuprates were concentrated on Bi$_2$Sr$_2$CaCu$_2$O$_8$ (BSCCO) and YBa$_2$Cu$_3$O$_{7-y}$ (YBCO) systems. With the LSCO system, one can investigate the electronic structure of the CuO$_2$ plane continuously from the heavily overdoped limit ($x\sim0.35$) to the undoped insulator ($x=0$) in a single system.
Recently, in underdoped cuprates a “normal-state gap” behavior above $T_c$ has been observed by angle-resolved photoemission spectroscopy (ARPES) in BSCCO [@ARPES] and a “spin-gap behavior” by NMR in YBCO [@Yasuoka]. The magnitude of the normal-state gap is of the same order as the superconducting gap at optimal doping. Meanwhile, underdoped cuprates have characteristic temperatures which are considerably [*higher*]{} than $T_{c}$ in the uniform magnetic susceptibility [@chi], the electronic specific heat [@Loram], the Hall coefficient [@Hwang] and the electrical resistivity [@chi]. All these characteristic temperatures show similar behaviors in LSCO: they increase from $\sim$300 K at optimal doping $x\sim0.15$ to $\sim$600 K at $x\sim0.1$ for the LSCO system, suggesting a pseudogap-type electronic structure. In addition, it can be reconciled only if a pseudogap is opened at the chemical potential that both the electronic specific heat coefficient $\gamma$ [@Loram; @Momono] and the chemical potential shift with doping [@chempot] are suppressed towards the MIT. To obtain a full picture of the evolution of those “gaps,” it is also necessary to know the total density of state (DOS), which is most directly observed by angle-integrated photoemission spectroscopy (AIPES). In the present study, we have performed high-resolution AIPES measurements and made a systematic study of the doping dependence of the electronic structure of the LSCO system, focusing on the evolution of a pseudogap, in a wide hole concentration range from $x=0$ (undoped) to $x=0.3$ (heavily overdoped).
High-quality single crystals of LSCO were grown by the traveling-solvent floating-zone method. The samples had $T_c$’s of 18, 34, 32, 21 and 0 K for $x = 0.074, 0.13, 0.175, 0.203$ and 0.30, respectively, with transition widths of about 0.5 K. All the samples were annealed to make the oxygen content stoichiometric. The $x=0$ sample was annealed in a reducing atmosphere (100 Torr O$_2$ at 800$^\circ$C) and confirmed that the Néel temperature was higher than 250 K, meaning that the hole concentration was less than 0.004 [@CYchen]. Photoemission measurements were carried out using the He [I]{} line ($h\nu=21.4$ eV) with an overall energy resolution of $\sim22$ meV. The base pressure in the spectrometer was in the $10^{-11}$ Torr range. Clean surfaces were obtained by [*in situ*]{} scraping with a diamond file in every 40 minutes. In order to minimize the degradation of sample surfaces, the measurements were performed at $T\sim18$ K. Only $x=0$ sample was measured at $\sim70$ K to avoid charging effect, but still a slight charging effect (at most $\sim10$ meV) could not be eliminated. Energies were carefully calibrated using Au evaporated on each sample so that its uncertainty was about 1 meV. The spectra were corrected for the He [ I]{}$^\ast$ satellite assuming that He [I]{}$^\ast$/He [I]{} ratio is constant in all the spectra.
Figure \[spectra\] shows photoemission spectra for various compositions. The whole valence-band spectra given in the inset show no trace of a hump at $\sim-9$ eV, indicating the high sample quality and the cleanliness of the sample surfaces. The spectra have been normalized to the maximum intensity of the valence band since the area of the valence band was hard to determine accurately due to the experimentally ambiguous background. As $x$ decreases, the intensity at the chemical potential $\mu$ decreases and disappears in the insulating phase ($x=0$). While in the overdoped region ($x > 0.2$) the spectra show an ordinary metallic Fermi edge, in the underdoped region ($x < 0.15$) the spectra show a pseudogap-type lineshape around $\mu$ in the sense that the spectral intensity gradually diminishes towards $\mu$ from somewhat below $\mu$ as marked by vertical bars in Fig. \[spectra\]. Here the vertical bars indicate the points of maximum curvature. It appears that a rather large pseudogap develops with decreasing $x$, with its width and depth increasing towards $x=0$. Note that the energy scale of the DOS suppression is as large as the order of $\sim0.1$ eV (“high-energy pseudogap”) and therefore that it does not correspond to the “normal-state gap” which has a magnitude similar to the superconducting gap ($\sim25$ meV for BSCCO) and has been observed in the ARPES spectra of underdoped BSCCO (“low-energy pseudogap”) [@ARPES]. In view of the energy scale, the marked feature in the AIPES spectra may rather correspond to the broad feature at ($\pi$,0) in the ARPES spectra of underdoped BSCCO . Unfortunately, since the present spectra were taken at a low temperature ($T\sim18$ K), one cannot judge whether the low-energy normal-state gap is opened or not in the underdoped LSCO.
In order to determine the spectral DOS at $\mu$, $\rho(\mu)$, and the precise position of the leading edge, the spectra near $\mu$ ($>-0.1$ eV) have been fitted to a linear DOS multiplied by the Fermi-Dirac distribution function convoluted with a Gaussian of the instrumental resolution, as shown in Fig. \[fitting\]. The obtained $\rho(\mu)$ are shown in Fig. \[Allin1\] (a), where error bars include uncertainties in the normalization procedure due to subtle changes in the valence-band lineshape. In Fig. \[Allin1\] (a), $\rho(\mu)$ is compared with the specific heat coefficient $\gamma$ [@Momono] and the Pauli-paramagnetic component $\chi_s^c$ of the spin susceptibility [@chi]. The three quantities, $\rho(\mu)$, $\gamma$ and $\chi_s^c$, show quite similar $x$-dependences: for $x>0.2$, with decreasing $x$ they slowly increase or remain nearly constant, take a maximum around $x=0.2$ and then decrease towards $x\sim0$ for $x<0.2$. A similar behavior was predicted by Hubbard-model calculations [@Duffy] although the absolute value of the calculated $\rho(\mu)$ was much higher than the observed one. Using the quasiparticle (QP) density at $\mu$, $N^\ast(\mu)$, obtained from $\gamma$ \[$N^\ast(\mu) =
\gamma/(\frac{1}{3}\pi^2 k_B^2)$\], and the unrenormalized DOS at $\mu$, $N_b(\mu)$, calculated by band theory [@LSCObandcalc], one can deduce the mass enhancement factor $m^\ast/m_b\equiv
N^\ast(\mu)/N_b(\mu)$ and the renormalization factor $Z \equiv
\rho(\mu) / N^\ast(\mu)$ as shown in Fig. \[Allin1\] (b). Figures \[Allin1\] (a) and (b) imply that the decrease in $m^\ast$ is driven by the decrease in $\rho(\mu)$ associated with the development of the high-energy pseudogap. The renormalization factor (or equivalently the QP spectral weight) $Z$ also decreases towards $x=0$, but it is not clear whether $Z$ goes to zero or remains finite as $x \rightarrow 0$ because of experimental uncertainties. It should be noted that the QP mass is enhanced as $x$ decreases for $x>0.2$, i.e., in the region where the pseudogap is absent. The behavior for $x>0.2$ is similar to that of the typical Mott MIT system La$_{1-x}$Sr$_x$TiO$_3$, which shows a QP mass ($\propto m^*
\propto\gamma$) enhancement with decreasing $x$ until the boundary of the antiferromagnetic (AF) phase is reached [@TiOspecheat].
The energy of the high-energy pseudogap $\Delta_{PG}$ was defined by the binding energy of the point of maximum curvature determined by taking the second derivatives (as marked by the vertical bars in Fig. \[spectra\]). The $\Delta_{PG}$ values are plotted in Fig. \[Allin1\] (c), which shows that $\Delta_{PG}$ follows the $x$-dependence of the characteristic temperatures $T_{\chi}$ [@chi] and $T_H$ [@Hwang]. Here $T_{\chi}$ and $T_H$ are temperatures at which the magnetic susceptibility $\chi$ takes a maximum and below which the Hall coefficient $R_H$ increases, respectively. It has also been reported that the temperature $T_{\rho}$ below which the electrical resistivity $\rho$ deviates downward from the linear-$T$ behavior follows $T_{\chi}$ and $T_{H}$ [@chi]. The similar $x$-dependences of $\Delta_{PG}$, $T_{\chi}$, $T_H$ and $T_{\rho}$ imply that these temperatures are closely related to the development of the high-energy pseudogap. We may therefore refer to those characteristic temperatures as the “pseudogap temperature” $T_{PG}$. We then find $\Delta_{PG}/k_BT_{PG}\simeq3$, indicating the interaction is in the strong coupling regime.
In the low-energy region, the position of the leading-edge midpoint obtained from the fitting is shown in Fig. \[Allin1\] (d). In going from the overdoped region to the optimum doping, the edge is shifted downward by $\sim 4$ meV. If we attribute the shift to the opening of a superconducting gap, we obtain the ratio $2\Delta/k_BT_c\sim2$, which is smaller than the typical value ($4-6$) deduced from ARPES of BSCCO and YBCO [@Dessau]. However, if the superconducting gap is anisotropic as in $d$-wave pairing, it is understandable that the angle-integrated spectra fitted to the simple step function give a smaller leading-edge shift than the ARPES data.
In the underdoped regime, since $\rho(\mu)$ is small and the slope of DOS at $\mu$, $\partial\rho(\omega)/\partial\omega|_{\omega=\mu}$, is steep, the Fermi edge is obscured at high temperatures and the Fermi-Dirac distribution lose its meaning. The crossover temperature for such a disappearance of the Fermi edge is thus given by $T_F =
(1/\pi k_B) \rho(\omega)/\frac{\partial\rho(\omega)}
{\partial\omega}|_{\omega=\mu}$ \[Fig. \[Allin1\] (e)\] and may be called the “coherence temperature” of the doped holes. $T_F$ may also be understood as the “Fermi temperature” of the doped holes because if one linearly extrapolates the DOS $\rho(\omega)$ beyond $\mu$, then a hole pocket has the Fermi energy of $\varepsilon_F
\equiv \pi k_BT_F$. For such a hole pocket, transport at high temperatures $T\gg T_F$ would be dominated by the incoherent charge dynamics of the doped holes [@Jaklik]. Figure \[Allin1\] (e) shows that $T_F$ becomes very low ($<100$K) in the underdoped regime, indicating that the thermodynamic and transport properties behave as those in “incoherent metals.” Note that $T_F$ is considerably lower than $T_{PG}$ and even lower than $J/\pi k_B$ in the underdoped region. Under such a condition, the kinetic energy gain of the doped holes alone may be insufficient to destroy the AF order, and alternatively the disappearance of the long-range order may be attributed to strong quantum fluctuations characteristic of two-dimensional systems. Figure \[Allin1\] (e) also suggests that for small $x$, $T_F$ scales with $x^2$ rather than $x$. According to the hyperscaling hypothesis of MIT [@Imada], the critical behavior of $T_F$ and chemical potential shift $\Delta\mu$ near MIT is given by $T_F\propto x^{z/d}$ and $\Delta\mu \propto x^{z/d}$, where $z$ is the dynamical exponent of the MIT and $d$ is the spatial dimension ($d=2$). Thus the present result ($T_F\propto x^2$) implies $z=4$ and is consistent with the observed suppression of the chemical potential shift ($\Delta\mu \propto x^{2}$) in underdoped LSCO [@chempot]. This $z$ value is distinctly different from that of an ordinary metal-to-band insulator transitions, where $z=2$.
Finally, let us discuss the microscopic origin of the pseudogap behavior. Since the energy scale of the present pseudogap is of the order of the super-exchange energy $J\sim0.1$ eV, it is tempting to associate the pseudogap with the development of AF correlations or short-range AF order in the underdoped region [@Schrieffer; @Pines]. In this scenario, when the temperature is lowered below $T_{PG}$, the AF correlation length increases and consequently the DOS shows a pseudogap reminiscent of the AF band gap [@Preuss]. The observed pseudogap may also be related with the short-range stripe order because it also originates from AF correlations. It has been stressed that LSCO is close to the instability of stripe order even in the metallic phase [@Emery; @Tranquada]. Since $T_F \ll T_{PG}\sim
J/\pi k_B$ in the underdoped region, the system cannot be regarded as a normal Fermi liquid which is weakly perturbed by AF correlation but rather as an AF state perturbed by the motion of doped holes. As an alternative scenario, a pseudogap may be produced by preformed Cooper pairs, which lose their coherence above $T_{c}$ but still keep local pairing, [@Trivedi; @Kivelson; @Imada2], or by spinon pairing which already occurs above $T_c$ [@Fukuyama]. However, interactions which lead to such pairing would have only small energies; that is, the energy scale of such a pseudogap should be of the same order as that of the superconducting gap ($\Delta_{SC}= 10 -20$ meV) at least in the slightly underdoped region. Probably, such a low-energy pseudogap corresponds to the spin gap in NMR of YBCO, the normal-state gap in ARPES of BSCCO and the gap observed by tunneling spectroscopy and is related with the drop in the electric resistivity just above $T_c$ [@Tunnel]. On the other hand, the high-energy pseudogap observed in the present work corresponds to the anomalies in the magnetic susceptibility, the Hall effect and the specific heat. Then the question may arise whether the high-energy pseudogap and the low-energy normal-state gap or spin gap are interrelated or not. Further studies are necessary to resolve this interesting issue, which is directly related to the mechanism of superconductivity in the doped cuprates.
In conclusion, we have found a large pseudogap on the energy scale of 0.1 eV in the DOS of La$_{2-x}$Sr$_x$CuO$_4$ by AIPES. The presented spectra have given us a clear view of the evolution of the pseudogap from the overdoped metal to the undoped insulator. Its evolution with decreasing $x$ is represented by the increase of the pseudogap energy $\Delta_{PG}$ and the suppression of the spectral DOS $\rho(\mu)$ as well as the QP density at $\mu$ near the AF insulating phase. Consequently, the coherence temperature $T_F$ is much smaller than $\Delta_{PG}$ in the underdoped region, suggesting that an incoherent metallic state may dominate the thermodynamic and transport properties of the underdoped cuprates. The origin of the present large pseudogap is most likely due to AF correlations or short-range AF order, while its relation to the smaller “normal-state gap” or “spin gap” remains an open question and should be addressed in future studies.
We would like to thank M. Imada and Z.-X. Shen for helpful discussions. This work is supported by a Grant-in-Aid for Scientific Research from the Ministry of Education, Science, Sports and Culture, the New Energy and Industrial Technology Development Organization (NEDO) and Special Promotion Funds of the Science and Technology Agency, Japan.
Present address: Joint Research Center for Atom Technology, National Institute for Advanced Interdisciplinary Research, 1-1-4 Higashi, Tsukuba, Ibaraki 305, Japan. Present address: The Institute of Physical and Chemical Research (RIKEN), SPring-8, Kamigori-cho, Hyogo 678-12, Japan.
D. S. Marshall [*et al.*]{}, Phys. Rev. Lett. [**76**]{}, 4841 (1996); H. Ding [*et al.*]{}, Nature (London) [**382**]{}, 51 (1996).
H. Yasuoka, T. Imai, and T. Shimizu, [*in Strong Correlation and Superconductivity*]{} edited by H. Fukuyama, S. Maekawa, and A. P. Malozemof (Springer-Verlag, Berlin, 1989) p. 254; M. Takigawa [*et al.*]{}, Phys. Rev. B [**43**]{}, 247 (1991).
T. Nakano [*et al.*]{}, Phys. Rev. B [**49**]{}, 16000 (1994).
J. W. Loram [*et al.*]{}, Physica (Amsterdam) [ **235C**]{}, 134 (1994)
H. Y. Hwang [*et al.*]{}, Phys. Rev. Lett. [**72**]{}, 2636 (1994); T. Nishikawa, J. Takeda, and M. Sato, J. Phys. Soc. Jpn. [**63**]{}, 1441 (1994).
N. Momono [*et al.*]{}, Physica (Amsterdam) [**233C**]{}, 395 (1994).
A. Ino [*et al.*]{}, Phys. Rev. Lett. [**79**]{}, 2101 (1997).
C. Y. Chen [*et al.*]{}, Phys. Rev. B [**43**]{}, 392 (1991).
Z.-X. Shen and J. R. Schrieffer, Phys. Rev. Lett. [**78**]{}, 1771 (1997).
D. Duffy [*et al.*]{}, Phys. Rev. B [**56**]{}, 5597 (1997).
L. F. Mattheiss, Phys. Rev. Lett. [**58**]{}, 1028 (1987).
Y. Tokura [*et al.*]{}, Phys. Rev. Lett. [**70**]{}, 2126 (1993).
Z.-X. Shen [*et al.*]{}, Phys. Rev. Lett. [ **70**]{}, 1553 (1993).
J. Jaklič and P. Prelovšek, Phys. Rev. B [ **52**]{}, 6903 (1995).
M. Imada, J. Phys. Soc. Jpn. [**64**]{}, 2954 (1995).
J. R. Schrieffer and A. P. Kampf, J. Phys. Chem. Solids [**56**]{}, 1673 (1995)
J. Schmalian, D. Pines and B. Stojkovic, Phys. Rev. Lett. [**80**]{}, 3839 (1998).
R. Preuss [*et al.*]{}, Phys. Rev. Lett. [**79**]{}, 1122 (1997).
M. I. Salkola, V. J. Emery, and S. A. Kivelson, Phys. Rev. Lett. [**77**]{}, 155 (1996).
J. M. Tranquada [*et al.*]{}, Nature (London) [**375**]{} 561 (1995).
S. Doniach and M. Inui, Phys. Rev. B [**41**]{}, 6668 (1990); N. Trivedi and M. Randeria, Phys. Rev. Lett. [**75**]{}, 312 (1995).
V. J. Emery and S. A. Kivelson, Nature (London) [**374**]{}, 434 (1995).
M. Imada, J. Phys. Soc. Jpn [**62**]{}, 1105 (1993).
T. Tanamoto, H. Kohno, and H. Fukuyama, J. Phys. Soc. Jpn. [**61**]{}, 1886 (1992).
N. Momono [*et al.*]{}, J. Phys. Chem. Solids, in press.
|
---
address:
- 'Copenhagen University, Mathematics Department, Universitetsparken 5, 2100 Copenhagen, Denmark. E-mail [grubbmath.ku.dk]{}'
- 'Institut für Mathematik, Universität Hannover, Welfengarten 1, 30167 Hannover, Germany.'
author:
- Gerd Grubb and Elmar Schrohe
title: 'Traces and quasi-traces on the Boutet de Monvel algebra'
---
amstex =1 1200 ‘=11 @ ‘==23.5truecm =16.5truecm
\#1[ ]{} \#1[{[\#1]{}}]{}
E-mail [schrohemath.uni-hannover.de]{} Canonical trace — Nonlocal invariant — Pseudodifferential boundary value problems — Boutet de Monvel calculus — Asymptotic resolvent trace expansions 58J42, 35S15
We construct an analogue of Kontsevich and Vishik’s canonical trace for pseudodifferential boundary value problems in the Boutet de Monvel calculus on compact manifolds with boundary. For an operator $A$ in the calculus (of class zero), and an auxiliary operator $B$, formed of the Dirichlet realization of a strongly elliptic second-order differential operator and an elliptic operator on the boundary, we consider the coefficient $C_0(A,B)$ of $(-\lambda)^{-N}$ in the asymptotic expansion of the resolvent trace $\operatorname{Tr}(A(B-\lambda)^{-N})$ (with $N$ large) in powers and log-powers of $\lambda$. This coefficient identifies with the zero-power coefficient in the Laurent series for the zeta function $\operatorname{Tr}(AB^{-s})$ at $s=0$, when $B$ is invertible. We show that $C_0(A,B)$ is in general a quasi-trace, in the sense that it vanishes on commutators $[A,A']$ modulo local terms, and has a specific value independent of the auxiliary operator, modulo local terms. The local “errors” vanish when $A$ is a singular Green operator of noninteger order, or of integer order with a certain parity; then $C_0(A,B)$ is a trace of $A$. They do not in general vanish when the interior ps.d.o. part of $A$ is nontrivial.
[*Keywords:*]{} Canonical trace — Nonlocal invariant — Pseudodifferential boundary value problems — Boutet de Monvel calculus — Asymptotic resolvent trace
[*Math. classification:*]{} 58J42 – 35S15
1\. Introduction
In their work on the geometry of determinants of elliptic operators, Kontsevich and Vishik [@KV] introduced the [*canonical trace*]{}, a novel functional $\operatorname{TR} A$ defined for a class of pseudodifferential operators ($\psi$do’s) $A$ on a closed $n$-dimensional manifold $X$. What makes it remarkable is that it extends the standard operator trace, and hereby complements the noncommutative residue.
The noncommutative residue $\operatorname{res}A$ was discovered by Wodzicki [@W] and, independently, by Guillemin [@Gu]; it is a trace (i.e., a nontrivial linear functional which vanishes on commutators) on the full algebra ${\Cal A}$ of all classical pseudodifferential operators on $X$. The noncommutative residue turns out to be the only trace on $\Cal A$ with this property, up to multiples. Moreover, the value of $\operatorname{res}A$ is determined from finitely many terms in the asymptotic expansion of the symbol of $A$ (in fact, from the component of homogeneity $-n$ in this expansion); we call such functionals “locally determined”.
It vanishes on operators of noninteger order or of order $<-n$, and it is here that Kontsevich and Vishik’s functional takes over: $\TR A$ is well-defined and nontrivial when $A$ is of noninteger order or of order $<-n$, and it vanishes on commutators $[A,A']$ of these types. It is “global” in that the value of $\TR A$ depends on the full operator and cannot be determined from the terms in the asymptotic expansion of the symbol. The canonical trace equals the standard trace when the order is $<-n$; it is moreover defined on integer-order operators with certain parity properties, see below.
In this article we shall generalize the canonical trace to pseudodifferential boundary value problems in the calculus of Boutet de Monvel [@B].
The original approach of [@KV] is based on studies of generalized zeta functions $\zeta (A,P,s)=\Tr(AP^{-s})$ and certain regularizations of them, where $P$ is an auxiliary elliptic operator. This does not extend easily to the situation of manifolds with boundary; in the case without boundary, $P^{-s}$ is a classical $\psi $do again (by Seeley [@S]), whereas complex powers of elliptic boundary problems fall outside the calculus of Boutet de Monvel. However, the introduction of $\TR A$ can instead be based on trace expansions of heat operators $ Ae^{-tP}$ (Lesch [@L]) or resolvents $A(P-\lambda )^{-N}$ (Grubb [@G4]), where the latter admit a direct generalization to manifolds with boundary (Grubb and Schrohe [@GSc]).
Therefore, let us explain the functional of [@KV] from the resolvent point of view: Let $A$ be a classical $\psi $do of order $\nu$; its symbol has an expansion in local coordinates $a\sim
\sum_{j\ge 0}a_{\nu -j}$ with $a_{\nu -j}(x,\xi )$ smooth in $(x,\xi )$ and homogeneous of degree $\nu -j$ in $\xi $ for $|\xi |\ge 1$. When $P$ is an auxiliary elliptic operator of integer order $m>0$ with no principal symbol eigenvalues on $\Bbb R_-$, a calculation in local coordinates shows that the operator family $A(P-\lambda )^{-N}$ has for $N>(n+\nu )/m$ a trace expansion $$\Tr (A(P-\lambda )^{-N})
\sim
\sum_{j\ge 0}\tilde c_{j}(-\lambda )^{\frac {n+\nu -j}m-N}
+\sum_{k\ge 0}
\bigl(\tilde c'_{k}\log (-\lambda )+\tilde
c''_{k}\bigr)(-\lambda )^{-k-N},\tag 1.1$$ for $\lambda \to \infty$ in a small sector around $\Bbb R_-$, cf. Grubb and Seeley [@GS1]. Here each $\tilde c_j$ (and each $\tilde c_k'$) comes from a specific homogeneous term in the symbol of $A(P-\lambda)^{-N}$, whereas the $\tilde c''_k$ depend on the full symbol. So the coefficients $\tilde c_j$ and $\tilde c_k'$ depend each on a finite set of homogeneous terms in the symbols of $A$ and $P$; we call such coefficients ‘locally determined’ (or ‘local’), while the $\tilde
c''_k$ are called ‘global’. When $\nu \notin
\Bbb Z$, the $\tilde c'_k$ vanish. When $\nu \in\Bbb Z$ and $(j-n-\nu
)/m$ is an integer $k\ge 0$, both $\tilde c_j$ and $\tilde c''_k$ contribute to the power $(-\lambda )^{-k-N}$; their sum is independent of the choice of local coordinates, whereas the splitting in $\tilde c_j$ and $\tilde c''_k$ depends in a well-defined way on the symbol structure in local coordinates (see or [@G4; @Th.; @1.3]).
In (1.1), $m\cdot \tilde c_0'= \operatorname{res}A$, depending solely on $A$, cf. [@W], [@Gu]. Moreover, cf. [@KV], [@G4], $$\operatorname{TR} A= \tilde c''_0\tag1.2$$ in the following four cases:
“(1)” $\nu <-n$,
“(2)” $\nu \notin\Bbb Z$,
“(3)” $\nu \in\Bbb Z$, $A$ is even-even and $n$ is odd,
“(4)” $\nu \in\Bbb Z$, $A$ is even-odd and $n$ is even; in the cases (3) and (4), $P$ is taken to be even-even with $m$ even. In all the cases (1)–(4), $\operatorname{res}A=0$, there is no contribution to $(-\lambda)^{-N}$ from the first sum in (1.1), and $$\tilde c''_0=\int\slint \tr a(x,\xi )\,\d\xi dx,\tag1.3$$ where $\tslint \tr a(x,\xi )\,\d\xi $ is a finite part integral (explained in detail in Section 3 below).
When $\nu \in\Bbb Z $, we say that $A$ or $a$ has [*even-even*]{} alternating parity (in short: is even-even), when the symbols with even (resp. odd) degree $\nu -j$ are even (resp. odd) in $\xi $: $$a_{\nu -j}(x,-\xi )=(-1)^{\nu -j}a_{\nu -j}(x,\xi ) \text{ for }|\xi
|\ge 1,\tag1.4$$ and the derivatives in $x$ and $\xi $ likewise have this property. $A$ and $a$ are said to have [*even-odd*]{} alternating parity in the reversed situation, where the symbols with even (resp. odd) degree $\nu -j$ are odd (resp. even) in $\xi $: $$a_{\nu -j}(x,-\xi )=(-1)^{\nu -j-1}a_{\nu -j}(x,\xi ) \text{ for
}|\xi |\ge 1,\tag1.5$$ etc. These properties are preserved under coordinate changes. For brevity, we shall say that $A$ (or $a$) [*has a parity that fits with the dimension $n$*]{}, when (3) or (4) holds.
[@KV] only considered the cases (1)–(3); the operators satisfying (3) were called odd-class operators. (4) was included in [@G4].
The relations can also be formulated in terms of the generalized zeta function $\zeta (A,P,s)=\Tr(AP^{-s})$; here we assume $P$ invertible for simplicity. It is known from [@W], [@Gu] that $\zeta
(A,P,s)$ extends meromorphically from large $\operatorname{Re}s$ across $\operatorname{Re}s=0$ with a simple pole at $0$, $$\zeta (A,P,s)=C_{-1}(A,P)s^{-1}+C_0(A,P)+O(s) \text{ for }s\to 0;\tag1.6$$ this follows also by translating (1.1) to a statement on the pole structure of $\zeta (A,P,s)$ (as e.g. in ). Here$$\aligned
C_{-1}(A,P)&=\tilde c'_0=\tfrac 1m \operatorname{res}A,\\
C_0(A,P)&=\tilde c_{\nu +n}+\tilde c''_0,
\endaligned\tag1.7$$ with $\tilde c_{\nu +n}$ defined as 0 if $\nu <-n$ or $\nu
\notin\Bbb Z$. In the cases (1)–(4), $\tilde c_{\nu +n}$ and $\operatorname{res}A$ vanish (for any choice of local coordinates), so that$$\operatorname{TR}A=
\tilde c''_0 =C_0(A,P)
=\zeta (A,P,0).\tag1.8$$
The coefficient $C_0(A,P)$ has an interest also when none of the conditions (1)–(4) is satisfied. Then it has the properties:$$C_0(A,P)-C_0(A,P') \text{ and }C_0([A,A'],P)\text{ are locally
determined}.\tag1.9$$ Moreover, $C_0(A,P)$ can be written as (1.3) plus local terms. Functionals with properties as in (1.9) for a system of auxiliary elliptic operators $P$ will be called quasi-traces.
By use of more functional calculus, defining $\log P$, one can moreover show that the two expressions in (1.9) are noncommutative residues of suitable combinations of the given operators and $\log P$; cf. [@KV] and Okikiolu [@O] for $C_0(A,P)-C_0(A,P')$, cf. Melrose and Nistor [@MN] for $C_0([A,A'],P)$. (In works of Melrose et al., notation such as $\widehat{\operatorname{Tr}}(A)$ or $\Tr_P(A)$ is used for $C_0(A,P)$; it is called a regularized trace there.)
In order to extend the definition of the canonical trace to operators in the calculus of Boutet de Monvel [@B] on an $n$-dimensional compact $C^\infty$-manifold $X$ with boundary $\partial X=X'$, we shall rely on resolvent expansions analogous to (1.1). To this end we choose an auxiliary operator $P_{1,\operatorname{D}}$, which is the (invertible) Dirichlet realization of a strongly elliptic principally scalar second-order differential operator.
It was shown in [@GSc] that when $A=P_++G$ is a pseudodifferential boundary operator ($\psi$dbo) in this calculus of order $\nu \in\Bbb Z$ with $G$ of class 0, then there is a trace expansion for $N>(n+\nu )/2$: $$\Tr(A(P_{1,\operatorname{D}}-\lambda )^{-N})\sim \sum_{j\ge
0}\!\tilde c_{j}(-\lambda)
^{\frac{n+\nu - j}{2}-N}
+\sum_{k\ge
0}(\tilde c'_k\log (-\lambda )+\tilde c''_k)(-\lambda )
^{-\frac k2-N}.
\tag 1.10$$ The proof depends on a reduction to parameter-dependent $\psi $do’s on $X'$, where [@GS1] can be used; it is easily generalized to cases where $A=G$ of noninteger order $\nu $ and class 0. When $\nu
\notin\Bbb Z$, the $\tilde c'_k$ vanish.
As explained e.g. in [@GS2] this yields the expansions: $$\align
\Tr(Ae^{-tP_{1,\operatorname{D}}})\sim
\sum_{j\ge 0}
c_{j} t
^{\frac{j-n-\nu }{2}}
+\sum_{k\ge 0}(- c'_k\log t + c''_k)t
^{\frac k2},\tag1.11\\
\Gamma (s)\Tr
(AP_{1,\operatorname{D}}^{-s})\sim \sum_{j\ge 0}\frac{
c_j }
{s+\frac{j-n-\nu }{2}}
+\sum_{k\ge 0}\Bigl(\frac{ c'_k}{(s+\frac k2)^2} +\frac{
c''_k}
{s+\frac k2}\Bigr) .\tag1.12
\endalign$$ In (1.10), $\lambda \to\infty $ in a sector of $\Bbb C$; in (1.11), $t\to 0+$; (1.12) describes the pole structure of the meromorphic extension of $\Gamma (s)\Tr
(AP_{1,\operatorname{D}}^{-s})$ to $\Bbb C$. The coefficients $c_j, c_k', c''_k$ are proportional to the coefficients $\tilde c_j, \tilde c_k', \tilde
c''_k$ in (1.10) by universal constants; in particular,$$c'_0=\tilde c'_0,\quad c''_0=\tilde c''_0, \quad c_{\nu +n}=\tilde
c_{\nu +n},\tag1.13$$ where $c_{\nu +n}$ and $\tilde c_{\nu +n}$ are taken as 0 when $\nu
<-n$ or $\nu \notin\Bbb Z$. The coefficients $c_j$ and $c_k'$ are locally determined, whereas the $c''_k$ are global.
Relation (1.12) implies that the generalized zeta function $\zeta
(A,P_{1,\operatorname{D}},s)=\Tr (AP^{-s}_{1,\operatorname{D}})$ extends meromorphically with $$\gathered
\zeta
(A,P_{1,\operatorname{D}},s)=C_{-1}(A,P_{1,\operatorname{D}})s^{-1}
+C_0(A,P_{1,\operatorname{D}})+O(s) \text{ for }s\to 0,\\
C_{-1}(A,P_{1,\operatorname{D}})=\tilde c'_0,\quad
C_0(A,P_{1,\operatorname{D}})=\tilde c_{\nu +n}+\tilde c''_0.
\endgathered\tag1.14$$ In [@FGLS], an analogue of Wodzicki’s noncommutative residue was introduced by Fedosov, Golse, Leichtnam, and Schrohe for operators in Boutet de Monvel’s calculus. In [@GSc] we were able to show that the coefficient $\tilde
c_0'=C_{-1}(A,P_{1,\operatorname D})$ in the above expansions satisfies the relation $$\tilde c_0'=\tfrac12 \res A.\tag1.15$$
We shall presently investigate $C_0(A,P_{1,\operatorname{D}})$ as a candidate for a canonical trace. We establish quasi-trace properties as in (1.9), with formulas for the “value modulo local terms”. Moreover, we extract some cases where $C_0(A,P_{1,\operatorname{D}})$ is independent of $P_{1,\operatorname{D}}$ and vanishes on commutators $[A,A']$, so that it is a trace. Here we show in particular that $C_0([A,A'],P_{1,\operatorname{D}})=0$ whenever $A$ and $A'$ are singular Green operators of class zero and orders $\nu$ and $\nu'$ with $\nu+\nu'<1-n$ or $\nu+\nu'\notin \Bbb Z$; the same is true when $\nu+\nu'$ is an integer and a certain parity holds, provided we narrow down slightly the class of admissible auxiliary operators $P_1$. When $A=P_++G$ and $A'=P'_++G'$ both have nontrivial pseudodifferential part (and hence the orders are integer), however, one cannot hope for more than the quasi-trace property, see Remark 4.2 for a discussion of this point.
In contrast with the boundaryless case, the powers $P_{1,\operatorname{D}}^s$ are far from belonging to the $\psi $dbo calculus when $s\notin \Bbb Z$, so we do not expect to find residue formulas as in [@KV], [@O], [@MN].
The quasi-trace property has an interest in itself, for situations where one has some control over the local terms. Our formulas for $C_0(P_+,P_{1,\operatorname{D}})$ in Theorem 4.1 and for $C_0(G,P_{1,\operatorname{D}})$ in Theorem 3.6 moreover show how the values are related to formulas for the boundaryless manifolds $\widetilde X$ resp. $X'$.
2\. Preliminaries
Let $X$ be an $n$-dimensional compact $C^\infty $ manifold with boundary $\partial X=X'$, $X$ and $X'$ provided with $C^\infty $ vector bundles $E$ and $E'$. We can assume that $X$ is smoothly imbedded in an $n$-dimensional manifold $\widetilde X$ without boundary, provided with a vector bundle $\widetilde E$ such that $E=\widetilde E|_X$. We consider the algebra of (one-step) polyhomogeneous operators in the calculus of Boutet de Monvel $$\pmatrix P_++G & K\\ \quad&\quad \\T&S
\endpmatrix
\:
\matrix
C^\infty (E)&&C^\infty(E)\\
\oplus&\longrightarrow&\oplus\\
C^\infty (E')&& C^\infty (E')
\endaligned.
\tag2.1$$ Here $P$ is a pseudodifferential operator satisfying the transmission condition. The subscript ‘$+$’ indicates that we are taking the truncation of $P$ to $X$, i.e., the operator given by extending a function in $C^\infty (E)$ by zero to a function in $L^2(\widetilde X,\widetilde E)$, applying $P$, and restricting the result to $X^\circ$; the transmission property assures that this gives an element of $C^\infty (E)$. Moreover, $G$ is a singular Green operator (s.g.o.), $K$ a Poisson operator, $T$ a trace operator, and $S$ a pseudodifferential operator on $X'$. We assume that all are of order $\nu$. Details on the calculus can e.g. be found in [@G2].
As a first observation we note that for classical (one-step polyhomogeneous) $\psi $do’s $P$ of order $\nu $ on $\widetilde X$, the calculus of Boutet de Monvel requires $\nu \in\Bbb Z$ in order for the transmission condition to be satisfied at $\partial X$. The definition of the operators $G$, $K$ and $T$ (and, of course, $S$), however, extends readily to noninteger $\nu
$, so one can also ask for an extension of the canonical trace to operators of the form $$\pmatrix G & K\\ T&S
\endpmatrix ,\tag2.2$$ with pseudodifferential part equal to zero and all elements of order $\nu \in \Bbb R$.
We next fix an auxiliary operator: We let $P_1$ be a second-order strongly elliptic differential operator in $\widetilde E$ with scalar principal symbol, and denote by $P_{1,D}$ its Dirichlet realization on $X$. By possibly shifting the operator by a constant, we can assume that $P_1$ and $P_{1,\operatorname{D}}$ have positive lower bound, so that the resolvents $Q_\lambda =(P_1-\lambda )^{-1}$ and $R_\lambda =(P_{1,\operatorname{D}}-\lambda )^{-1}$ are defined for $\lambda $ in a region$$\Lambda =\{ \,\lambda \in\Bbb C\mid \arg\lambda \in [\tfrac \pi
2-\varepsilon ,\tfrac{3\pi }2+\varepsilon] \text{ or }|\lambda |\le
r(\varepsilon )\,\}\tag2.3$$ with suitable $\varepsilon>0$ and $r(\varepsilon)>0$. Moreover, we take a strongly elliptic second-order pseudodifferential operator $S_1$ on $C^\infty(X',E')$ and set $
B=\pmatrix P_{1,D}& 0\\ 0&S_1\endpmatrix.
$ Then $$\aligned
&\Tr \Bigg(\pmatrix P_++G & K\\ T&S
\endpmatrix(B-\lambda)^{-N}\Bigg)\\
&\qquad\qquad= \Tr_X \left((P_++G)(P_{1,D}-\lambda)^{-N}\right)
+\Tr_{X'}\left(S(S_1-\lambda)^{-N}\right)\endaligned\tag2.4$$ is well-defined for $N>(n+\nu)/2$. The behavior of $\Tr_{X'}\left(S(S_1-\lambda)^{-N}\right)$ is well-known from the theory of $\psi $do’s on manifolds without boundary, so we shall restrict the attention to $$\Tr_X \left((P_++G)(P_{1,D}-\lambda)^{-N}\right)\text{ if $\nu
\in\Bbb Z$},\text{ and }~~ \Tr_X
\left(G(P_{1,D}-\lambda)^{-N}\right)\text{ if $\nu
\in\Bbb R\setminus \Bbb Z$}.\tag2.5$$ We henceforth denote $P_++G=A$.
A natural candidate for an extension of the canonical trace is the functional $A\mapsto C_0(A,P_{1,\operatorname D})$ which associates with the $\psi$dbo $A=P_++G$ the coefficient of $(-\lambda )^{-N}$ in (1.10), equal to the coefficient of $s^0$ in the Laurent expansion of $\zeta(A,P_{1,\operatorname D},s)$ at $s=0$ (1.14). We shall show that this is indeed a good choice at least in two cases:
\(i) $A=G$ is a singular Green operator of noninteger order $\nu$ and class zero with vanishing pseudodifferential part;
\(ii) $A=P_++G$ is of integer order $\nu$ with a pseudodifferential operator $P$ of normal order $0$ and $G$ of class zero (we say that a $\psi
$do symbol has normal order $d$ when it is $O(\xi _n^{d})$ at the boundary).
There are two conditions involved, namely on the class of the singular Green part and on the normal order of the pseudodifferential part in (ii). They are both natural, and related:
If $G$ is effectively of class $r>0$, i.e., if it can be written in the form $G=G_0+\sum_{j=0}^{r-1} K_j\gamma_j$, where $G_0$ is a singular Green operator of class zero, $\gamma_ju=(\partial _n^ju)|_{X'}$, and the $K_j$ are Poisson operators with $K_{r-1}\ne 0$, then $G$ will not be bounded on $L^2(E)$, much less of trace class, even if its order is arbitrarily low, cf. e.g. . As we are looking for a functional coinciding with the standard trace for operators of sufficiently low order, we shall exclude operators of positive class. An expansion analogous to (1.10) still exists, with the sum over $k\ge 0$ replaced by a sum over $k\ge k_0$ with a possibly negative starting index $k_0$. In this case, however, the coefficient $c'_0$ will not in general coincide with the noncommutative residue, cf. [@GSc; @Remark; @1.2].
For operators $P$ having the transmission property, we find that $C_0(P_+, P_{1,\operatorname{D}})$ does indeed have a value modulo local terms that is independent of $P_1$ modulo local terms. The commutator $[P_+,P_+']$ of two pseudodifferential operators $P$ and $P'$ truncated to $X$, however, will be of the form $P_+''+G''$, where, in general, the singular Green operator $G''$ will be of class $r>0$ unless both $P$ and $P'$ are of normal order $\le 0$, cf. [@G2; @Section; @2.6]. Hence we can only hope to show the commutator property for this class.
Recall (cf. e.g. [@G2; @Lemma; @1.3.1]) that any $\psi $do $P$ having the transmission property at $X'$ can be written as a sum $$P=P^{(1)}+P^{(2)}+P^{(3)},\tag 2.6$$ where $P^{(1)}$ is a differential operator, $P^{(2)}$ is a $\psi $do whose symbol has normal order $-1$, and the symbol of $P^{(3)}$ has, near $X'$, a large power of $x_n$ as a factor to the left or right.
For a [*differential*]{} operator $P^{(1)}$, the expansion in (1.10) (with $A=P^{(1)}$) is valid without the second series over $k$; all terms are locally determined. Therefore $C_{-1}(P^{(1)},P_{1,\operatorname{D}})$ vanishes, and $C_0(P^{(1)},P_{1,\operatorname{D}})$ is local in this case, and so is $C_0(P^{(1)},P_{1,\operatorname{D}})-C_0(P^{(1)},P_{2,\operatorname{D}})$ for another choice of auxiliary operator $P_2$. Thus $C_0(P^{(1)},P_{1,\operatorname{D}})$ satisfies the first part of the requirement for being a quasi-trace (with value zero modulo local terms). As mentioned above, differential operators of positive normal order will be left out from the commutator considerations. So will operators of type $P^{(3)}$, since the structure is not preserved under differentiation of symbols.
At this point we can state a part of the results we will show in this paper:
Let $A=P_++G$ and $A'=P'_++G'$ be of orders $\nu $ resp. $\nu
'$, with $P$ resp. $P'$ vanishing if $\nu $ resp. $\nu ' \notin \Bbb
Z$, and with $G$ and $G'$ of class $0$. Let $P_1$ and $P_2$ be two auxiliary operators as explained above. Then
[(i)]{} $$C_{0}(A,P_{1,\operatorname{D}})-C_{0}(A,P_{2,\operatorname{D}})\text{
is locally determined};\tag2.7$$ it vanishes if $\nu <-n$ or $\nu \notin \Bbb Z$, and otherwise depends solely on the terms of the first $\nu +n+1$ homogeneity degrees in the symbols of $P_++G $, $P_1$ and $P_2$.
[(ii)]{} If $\nu $ and $\nu '\in\Bbb Z$, assume that $P$ and $P'$ have normal order $0$; if $\nu $ or $\nu '\in \Bbb R\setminus \Bbb Z
$, assume that $P$ and $P'$ are zero. Then $$C_{0}([A,A'],P_{1,\operatorname{D}})\text{
is locally determined};\tag2.8$$ it vanishes if $\nu +\nu '<-n$ or $\nu +\nu '\notin \Bbb Z$, and otherwise depends solely on the terms of the first $\nu +\nu '+n+1$ homogeneity degrees in the symbols of $P_++G $, $P'_++G'$ and $P_1$.
(Actually, the number of homogeneous terms entering from $G$ is one step lower; see the statements in Section 3.)
The operators of the form $A=P_++G$ of integer order with $G$ of class zero and $P$ of normal order $\le 0$ form an algebra; Theorem 2.1 shows that $C_0(A,P_{1,\operatorname{D}})$ is a quasi-trace on this algebra as well as on the singular Green operators of class zero.
Similarly as in [@GSc], our analysis is based on precise information about the resolvent. We therefore recall the structure of $R_\lambda =(P_{1,\operatorname{D}}-\lambda)^{-1}$: The Dirichlet problem $$(P_1-\lambda )u=f\text{ on }X,\quad \gamma _0u=\varphi \text{ on
}X',\tag2.9$$ is solved by a a row matrix $$\pmatrix P_1-\lambda \\ \gamma _0
\endpmatrix^{-1}=
\pmatrix R_\lambda & K_\lambda
\endpmatrix = \pmatrix Q_{\lambda ,+}+G_\lambda & K_\lambda
\endpmatrix ,\text{ with }G_\lambda =- K_\lambda \gamma
_0Q_{\lambda ,+}.\tag2.10$$ This is seen as follows: We denote by $K_\lambda $ the Poisson operator solving the semi-homogeneous problem with $f=0$, i.e., $K_\lambda \varphi$ is the solution $u$ of the equations $$(P_1-\lambda )u=0\text{ on }X,\quad \gamma _0u=\varphi \text{ on }X'.$$ The resolvent $R_\lambda $ of $P_{1,\operatorname{D}}$ should solve the other semi-homogeneous problem, with $\varphi =0$. With $Q_{\lambda ,+}$ denoting the truncation to $X$ of the resolvent $Q_\lambda $ of $P_1$ on $\widetilde X$, one can easily check that $R_\lambda =Q_{\lambda ,+}- K_\lambda \gamma_0Q_{\lambda ,+}$ solves the problem; it has the singular Green operator part $G_\lambda =- K_\lambda \gamma_0Q_{\lambda ,+}$.
We shall also study powers of the resolvent and write them in the form $$R^N_\lambda =(Q_{\lambda ,+}+G_\lambda)^{N}=(Q^N_\lambda
)_++G^{(N)}_\lambda
.\tag 2.11$$ Note that $$R_\lambda ^{N}=\tfrac 1{(N-1)!}\partial _\lambda
^{N-1}R_\lambda ,\quad Q_{\lambda ,+}^{N}=\tfrac 1{(N-1)!}\partial
_\lambda
^{N-1}Q^N_{\lambda ,+},\quad G^{(N)}_\lambda =\tfrac
1{(N-1)!}\partial _\lambda ^{N-1}G_\lambda ;
\tag 2.12$$ this allows us to replace calculations for powers of $R_\lambda $ by calculations for $\lambda $-derivatives.
We shall now start with the proof of Theorem 2.1. The case where $A=G$ is a singular Green operator of class zero and the pseudodifferential part vanishes can be attacked more directly and will be studied first.
3\. Traces on the algebra of singular Green operators
Let $G$ be a singular Green operator in $E$ of order $\nu \in\Bbb R$ and class 0, and consider $GR^N_\lambda = GQ^N_{\lambda ,+}+GG^{(N)}_\lambda$. In [@GSc] we introduced the auxiliary variable $\mu =(-\lambda )^{\frac12}$ and phrased the results in terms of $\mu$ instead of $\lambda $. In the present paper we shall keep $\lambda $ as an index, but will often use $\mu $ in symbol computations. Some formulas from [@GSc] and some immediate consequences are collected in Appendices A and B, to which we refer in the following.
As mentioned already, the proof in [@GSc] of the expansion (1.10) for $A=G$ is straightforwardly modified to allow $\nu \notin\Bbb Z$. For an analysis of the coefficients, we apply a partition of unity $1=\sum_{i=1}^{i_0}\theta _i$ subordinate to a cover of $X$ by coordinate patches $U _j$, $j=1,\dots,j_0$, with trivializations $\psi _j\:E|_{U_j}\to V_j\times
\Bbb C^{\operatorname{dim}E}$, $V_j\subset\subset \crnp$, $U_j\cap
\partial X$ mapped into $\partial \crnp$, such that any two of the functions $\theta _{i_1}$ and $\theta _{i_2}$ are supported in one of the coordinate patches $U _{j(i_1,i_2)}$. Replacing $G$ by $\sum_{i_1,i_2\le i_0}\theta _{i_1}G\theta _{i_2}$, we find from $GQ^N_{\lambda ,+}$ and $GG^{(N)}_\lambda
$ a system of terms $$G=\sum_{i_1,i_2\le i_0}(\theta _{i_1}G\theta _{i_2}Q^N_{\lambda ,+}+
\theta _{i_1}G\theta
_{i_2}G^{(N)}_\lambda ),\tag3.1$$ where $\theta _{i_1}$ and $\theta _{i_2}$ are supported either in an interior coordinate patch (with closure in the interior) or a patch meeting the boundary, and we can study each term in the local coordinates. Since $\theta _{i_2}G^{(N)}_\lambda $ is strongly polyhomogeneous of order $-\infty $ when $\theta _{i_2}$ is supported in the interior, its kernel is smooth and is $O(\lambda ^{-M})$ for any $M$. Hence the terms $\theta _{i_1}G\theta_{i_2}G^{(N)}_\lambda $ that are supported in the interior have traces that are $O(\lambda ^{-M})$ for any $M$ and can be disregarded in the following. For the terms $\theta _{i_1}G\theta _{i_2}Q^N_{\lambda ,+}$ supported in the interior, $\theta _{i_1}G\theta _{i_2}$ is of order $-\infty $; they will be covered by Lemma 3.1 below, which gives: $$\Tr(\theta _{i_1}G\theta
_{i_2}Q^N_{\lambda ,+})=
\Tr(\theta _{i_1}G\theta
_{i_2})(-\lambda )^{-N}+O(\lambda ^{-N-1/2}).\tag3.2$$ For the terms supported in patches meeting the boundary we can use the results of [@GSc] in an accurate way.
Let us first recall some elements of the basic symbol calculus we use, namely the calculus of parameter-dependent $\psi $do’s introduced in [@GS1]. It will be used both in the $(n-1)$-dimensional setting, relevant for operators on $X'$, and in the $n$-dimensional setting, relevant for operators on $\widetilde X$.
For $m\in\Bbb R$, the usual pseudodifferential symbol space $S^m(\Bbb
R^{n'}\stimes\Bbb R^{n'})$ consists of the functions $s(x,\xi )\in
C^\infty (\Bbb R^{n'}\stimes\Bbb R^{n'})$ satisfying estimates$$|D_{x}^\beta D_{\xi }^\alpha s(x,\xi )|\le C_{\alpha ,\beta
}\ang{\xi }^{m-|\alpha |},$$ for all $\alpha ,\beta \in \Bbb N^{n'}$. (We denote $(1+|\xi
|^2)^{\frac12}=\ang\xi $, $\{0,1,2,\dots\}=\Bbb N$.) The parameter-dependent version we shall use here is $S^{m,d,s}(\Bbb
R^{n'}\stimes\Bbb R^{n'},\Gamma )$, for $m\in\Bbb
R$, $d$ and $s\in\Bbb Z$, and with $\Gamma $ denoting a sector in $\Bbb C\setminus \{0\}$; it was introduced in [@G3], where a detailed account can be found. Here $S^{m,d,s}(\Bbb
R^{n'}\stimes\Bbb R^{n'},\rp )$ is defined as the space of $C^\infty
$ functions $f(x,\xi ,\mu )$ on $\Bbb R^{n'}\stimes\Bbb
R^{n'}\stimes\rp $ such that, with $z=1/\mu $,$$\multline
\partial _z^j[z^{d}|(\xi ,1/z)|^{-s}f(x,\xi ,1/z)]\in
S^{m+j}(\Bbb R^{n'}\stimes \Bbb R^{n'}),\\
\text{with symbol estimates uniform in $z$ for }z\le 1.
\endmultline$$ For more general $\Gamma $, the estimates have to hold on each ray in $\Gamma $, uniformly in closed subsectors of $\Gamma $. When the symbols moreover are holomorphic in $\mu \in\Gamma ^\circ$ (just for $|(\mu ,\xi
')|\ge \varepsilon $ with some $\varepsilon $ depending on the closed subsector), we speak of holomorphic symbols. The indication $\Bbb
R^{n'}\stimes \Bbb R^{n'}$ (and $\Gamma $) will often be omitted. We say that $f(x,\xi
,\mu )$ is (weakly) polyhomogeneous in $S^{m,d,s}(\Gamma )$, when there is a sequence of symbols $f_{j}$ in $S^{m-j,d,s}(\Gamma )$, homogeneous in $(\xi ,\mu )$ of degree $m-j+d+s$ for $|\xi |\ge 1$, such that $f-\sum_{j<J}f_j$ is in $S^{m-J,d,s}(\Gamma )$ for all $J\in\Bbb N$ (strongly polyhomogeneous if the homogeneity holds for $|(\xi ,\mu )|\ge 1$, with appropriate remainder estimates).
These spaces $S^{m,d,s}$ are a generalization of the spaces $S^{m,d}$ introduced in [@GS1] (in fact, $S^{m,d}(\Gamma )$ equals the space of holomorphic symbols in $S^{m,d,0}(\Gamma )$); they satisfy $$\aligned
S^{m,d,s}&\subset S^{m+s,d,0}\cap
S^{m,d+s,0}\text{ if }s\le 0,\\
S^{m,d,s}&\subset S^{m+s,d,0}+
S^{m,d+s,0}\text{ if }s\ge 0,
\endaligned\tag3.3$$ by [@GS1; @Lemma; @1.13]. The notation with the third upper index spares us from keeping track of such intersections and sums in general calculations. (It was used systematically in [@G3], [@G5], [@G6].)
When $f\in S^{m,d,0}$, there is an expansion in decreasing powers of $\mu $ starting with $\mu ^d$: $$f(x,\xi ,\mu )=\sum_{0\le k<N}\mu ^{d-k}f_{(k)}(x,\xi)+O(\mu
^{d-N}\ang\xi ^{m+N}),\quad \text{any $N$};\tag3.4$$ in fact the $N$’th remainder is in $S^{m+N,d-N,0}$. Such expansions play an important role in the proof of the following fact: When $F=\operatorname{OP}(f(x,\xi ,\mu ))$ has a polyhomogeneous symbol in $S^{m,d,0}$ such that all terms are integrable with respect to $\xi $, then by , the kernel of $F$ has an asymptotic expansion on the diagonal: $$K(F,x,x,\mu )\sim \sum_{j\in\Bbb N}c_j(x)\mu ^{m+d+n'-j}+\sum_{k\in\Bbb
N}(c'_k(x)\log\mu +c''_k(x))\mu ^{d-k};\tag 3.5$$ here the coefficients $c_j(x)$ and $c'_k(x)$ with $k=-m+j-n'$ are local, determined from the strictly homogeneous version of the $j$’th homogeneous term in the symbol of $F$, whereas the $c''_k(x)$ depend on the full operator (are global). If $m\notin\Bbb Z$, all $c'_k$ are zero. When $F$ is a $\psi $do family on $\widetilde X$ (or on $X'$), the trace expansion for $F$ is found by establishing expansions (3.5) for components of $F$ in local coordinate systems, carrying them back to $\widetilde X$ and integrating the fibrewise traces over $\widetilde X$; this leads to a similar expansion of $\Tr F$: $$\Tr(F )\sim \sum_{j\in\Bbb N}c_j\mu ^{m+d+n'-j}+\sum_{k\in\Bbb
N}(c'_k\log\mu +c''_k)\mu ^{d-k}.\tag 3.6$$ Of course, such a calculation depends on the choice of partitions of unity and local coordinates, but it is useful for qualitative information. The final result (3.6) is independent of the choices, in the sense that the collected coefficient of each power or log-power is so (for $k=j-m-n'\in\Bbb N$, $c_j+c''_k$ is independent of the choices).
This is the basis for the trace expansions we shall show. For purely $\psi $do terms on $X$, the expansions will be established relative to $\widetilde X$, but the integration of the kernels will be restricted to $X$. For all the other types of operators, the main idea is to reduce the considerations to $\psi $do’s over the boundary $X'$. In all cases the point is to show that the resulting symbols belong to suitable $S^{m,d,0}$-spaces such that (3.5) can be established.
Since we are particularly interested in the nonlocal coefficient $c_0''$, we have to be very careful with remainders and smoothing terms. Recall from that when an operator $F$ with symbol in $S^{m,d,0}$ is of order $-\infty $ (i.e., lies in $S^{-\infty ,d,0}$), then it has a $C^\infty $ kernel $K(F,x,y,\mu )$ with an expansion in $C^\infty $ terms$$K(F,x,y,\mu )=\sum_{0\le l<L}K_l(x,y)\mu ^{d-l}+K'_L(x,y,\mu ),\tag3.7$$ for all $L$, with $K'_L$ and its $(x,y)$-derivatives being $O(\ang \mu
^{d-L})$. Moreover, since $\partial _z^j(z^d F)=(-\mu ^2\partial _\mu )^j(\mu
^{-d}F)$ has symbol in $S^{-\infty ,0,0}$ for all $j$, there are expansions $$K((-\mu ^2\partial _\mu )^j(\mu
^{-d}F),x,y,\mu )=\sum_{0\le l<L}K^{(j)}_l(x,y)\mu ^{-l}+K^{(j)\prime}_L(x,y,\mu ),\tag3.8$$ for all $L$, with $K^{(j)\prime}_L$ and its $(x,y)$-derivatives being $O(\ang \mu
^{-L})$. Conversely, these kernel properties characterize the operators $F$ with symbol in $S^{-\infty ,d,0}$ (as is seen straightforwardly from the definition).
It follows from (3.3) that $S^{-\infty ,d,s}=S^{-\infty ,d+s,0}$. In the proof that a polyhomogeneous symbol lies in a symbol space $S^{m,d,s}$ one shows the symbol properties for the homogeneous terms and treats the remainders by establishing the kernel expansionss (3.8) (with $d$ replaced by $d+s$) with smoother coefficients and a higher number of terms, the higher the index of the remainder is taken.
The sector $\Gamma $ used in the present work is$$\Gamma =\{\,\mu \in\Bbb C\setminus \{0\}\mid |\arg \mu |\le\tfrac \pi
4+\tfrac
\varepsilon 2
\,\},\tag3.9$$ for some $\varepsilon>0$, since $\lambda =-\mu ^2$ runs in $\Lambda $ defined in (2.3). [*All symbols to be considered in this paper will be holomorphic on $\Gamma ^\circ$*]{}; this fact will not be mentioned explicitly each time.
One of the technical points in [@GSc] was that in compositions of two operators, the right-hand factor is taken in $y$-form and the resulting symbol is considered in $(x,y)$-form so that one does not need to consider the asymptotic expansion in the usual symbol composition formula (cf. e.g. (3.31) below). However, the passage from one of these forms to another can induce low-order errors. These have to be dealt with carefully in the present situation, where nonlocal contributions are in focus. Also for this reason, the following analysis of the effect of smoothing operators is important.
Smoothing parts of $Q_\lambda $ and $G_\lambda $ are easy to deal with, since they have smooth kernels that are $O(\lambda ^{-M})$ for any $M$. For compositions containing smoothing parts of $G$, we recall that an s.g.o. of order $-\infty $ is simply an operator with $C^\infty $ kernel on $X$. So is $P_+$ when $P$ is of order $-\infty $. Such operators enter as follows (for $G^{\pm}$, cf. (A.14)):
Let $G$ be an operator with $C^\infty $ kernel, compactly supported in $\crnp\stimes\crnp$. Then the traces of $GQ^N_{\lambda ,+}$, $GG^{(N)}_\lambda $ and $GG^\pm(Q^N_\lambda )$ have expansions $\sum_{j\ge 0}c''_j\mu ^{-2N-j}$, where $$\alignedat2
\text{\rm (i)}&&\quad
c''_0&= \Tr G \text{ in the expansion of }\Tr(GQ^N_{\lambda ,+}),\\
\text{\rm (ii)}&&\quad c''_0&= 0 \text{ in the expansions of
$GG^{(N)}_\lambda $ and
$GG^\pm(Q^N_\lambda )$.}
\endaligned \tag3.10$$
First consider the cases in (ii); here we use that the symbol of the $\lambda
$-dependent factor has the structure described in Lemma A.4 with $\mu^2=-\lambda$. Consider a term $G_{J,j,j'}=\operatorname{OPG}(g_{J,j,j'}(x',\xi ',\xi
_n,\eta _n,\mu ))$ with $g_{J,j,j'}=s_{J,j,j'}(\kappa +i\xi
_n)^{-j}(\kappa '-i\eta _n)^{-j'}$ ($\kappa $ and $\kappa '$ taking values $\kappa ^+$ or $\kappa ^-$), $s_{J,j,j'}\in
S^{0,0,j+j'-1-2N-J}$; its kernel is $$\aligned
K(G_{J,j,j'},x,y,\mu )
&=
\int e^{i (x'-y')\cdot \xi '+ix_n\xi
_n-iy_n\eta _n}g_{J,j,j'}(x',\xi ',\xi _n,\eta _n,\mu )\,\d\xi \d\eta
_n\\
&= c_jc_{j'}
\int e^{i (x'-y')\cdot \xi '}s_{J,j,j'}(x',\xi ',\mu
)x_n^{j-1}y_n^{j'-1}e^{-\kappa x_n-\kappa 'y_n}\,\d\xi ',
\endaligned$$ since $\Cal F^{-1}_{\xi _n\to x_n}(\kappa +i\xi
_n)^{-j}=c_jx_n^{j-1}e^{-\kappa x_n}$. Now $$\aligned
K(GG_{J,j,j'},x,y,\mu )&=\int_{\rnp}K(G,x,z)K(G_{J,j,j'}, z,y,\mu )\,dz
,\\
\Tr (GG_{J,j,j'})&=\int_{\rnp}\tr\int_{\rnp}K(G,x,z)K(G_{J,j,j'}, z,x,\mu
)\,dzdx .\endaligned$$ To evaluate the latter, we insert a Taylor expansion of $K(G,x,z)$ w.r.t.$(x_n,z_n)$:$$K(G,x,z)=\sum_{m,m'<M}b_{m,m'}(x',z')x_n^mz_n^{m'} +x_n^Mz_n^Mb'_M(x,z),$$ and use that $$\aligned
\int_0^\infty x_n^{m+j-1}e^{-\kappa x_n}\,dx_n&=c'_{m+j}\kappa
^{-m-j},\\
|\int_0^\infty f(x_n)x_n^{M+j-1}e^{-\kappa x_n}\,dx_n|&\le
\sup |f(x_n)e^{-\kappa x_n/2}\int_0^\infty x_n^{M+j-1}e^{-\kappa
x_n/2}\,dx_n|\\
&\le
C\kappa
^{-M-j}.
\endaligned$$ This gives when integrations in $(x_n,z_n)$ are performed: $$\multline
\Tr (GG_{J,j,j'})=\sum_{m,m'<M}c_{jj'mm'}\\
\int\tr
b_{m,m'}(x',z')\kappa ^{-m-j'}(\kappa ')^{-m'-j}e^{i(z'-x')\cdot \xi
'}s_{J,j,j'}(
z',\xi ',\mu )\,\d \xi 'dz'dx'\\
\quad+\int O(\ang{(\xi ',\mu )} ^{-2M-1-2N-J})\,\d\xi '.
\endmultline$$ Each term in the sum over $m,m'$ is the trace of a $\psi $do $S_{J,j,j',m,m'}$ on $\Bbb
R^{n-1}$ composed of the operator with kernel $c_{jj'mm'}b_{m,m'}(x',y')$ (smooth compactly supported) and the operator with symbol $$\kappa (x',\xi ',\mu )^{-m-j'}\kappa
'(x',\xi ',\mu )^{-m'-j}s_{J,j,j'}(
x',\xi ',\mu )\in S^{0,0,-m-m'-1-2N-J};$$ so $S_{J,j,j',m,m'}$ has symbol in $S^{-\infty
,0,-m-m'-1-2N-J}=S^{-\infty ,-m-m'-1-2N-J,0}$ and consequently a trace expansion as in (3.7) with $d\le-2N-1$ in all cases. The remainder is $O(\mu ^{-2M-1-2N-J+n-1})$. Since $M$ can be taken arbitrarily large, this shows (ii).
For (i), let $\widetilde G$ be an operator on $\Bbb R^n$ with compactly supported $C^\infty $ kernel such that $G=r^+\widetilde G
e^+$ (its kernel can be constructed by extending the kernel of $G$ smoothly across $x_n=0$ and $y_n=0$). Then $$GQ^N_{\lambda ,+}=\widetilde G_+Q^N_{\lambda ,+}=(\widetilde
GQ^N_{\lambda })_+-G^+(\widetilde G)G^-(Q^N_{\lambda }).$$ Here $G^+(\widetilde G)G^-(Q^N_{\lambda })$ is of the type we considered above, having a trace expansion with $d\le -2N-1$. For $(\widetilde
GQ^N_{\lambda })_+$ we use the result known from [@GS1] (cf. also ) that the kernel of $\widetilde GQ^N_{\lambda }$ has an expansion on the diagonal $$K(\widetilde GQ^N_{\lambda },x,x,\mu )\sim \sum_{j\ge 0}c''_j(x)\mu
^{-2N-j},\text{ with }c''_0(x)=K(\widetilde G,x,x ).\tag 3.11$$ The expansion of $\Tr(\widetilde GQ^N_\lambda )_+$ is found by integrating the matrix trace of (3.11) over $\rnp$, where $K(\widetilde G,x,x )=K( G,x,x )$, so (i) follows, with $c''_0=\int_{\rnp}\tr K( G,x,x )\, dx=\Tr G$.
As already mentioned, this lemma takes care of the interior parts of $GQ^N_{\lambda ,+}$, cf. (3.1), (3.2).
For the parts $\theta
_{i_1}G\theta
_{i_2}G^{(N)}_\lambda $ at the boundary, we can use Proposition B.3 directly; it shows that they contribute only locally to the coefficient of $\lambda^{-N}$.
It remains to consider terms $\theta _{i_1}G\theta
_{i_2}Q^N_{\lambda ,+}$ in patches intersecting the boundary. Here we can draw on the fine analysis of such terms made in [@GSc] to find the contribution to $c'_0$ in (1.12); it can be used also to isolate the contribution to $c''_0$ modulo local terms. This will give not only a proof of (2.7) but also a “value” (modulo local terms) of the constant.
To simplify the presentation, we can unify the treatment of the coordinate patches by spreading out the images $V_{j(i_1,i_2)}$ of the sets $U_{j(i_1,i_2)}$ by linear translations in the $x'$-variable, to obtain sets $V'_{i_1,i_2}$ with positive distance (in the $x'$-direction) from one another. Then the sum of the localized operators $\theta _{i_1}G\theta _{i_2}$ acts in $\crnp\times \Bbb C^{\operatorname{dim}E}$; we shall denote it by $G$ again or, if a distinction from the original $G$ is needed, by $\underline
G$. $R_{\lambda }$ is likewise considered in these coordinates and may be denoted $\underline R_\lambda $ if needed for precision.
Recall that when $G$ is defined on $\crnp$ from a symbol $g(x',\xi ',\xi
_n,\eta _n)$ with Laguerre expansion (B.15), then the operator $\tr_nG$ is defined as the $\psi $do on $\Bbb
R^{n-1}$ with symbol $$\tr_ng(x',\xi ')=\int g(x',\xi ',\xi _n,\xi
_n)\,\d\xi _n=\sum_{j\in\Bbb N}d_{jj}(x',\xi ').$$ Recall also (cf. e.g. [@G1]) that a singular Green operator on $\overline {\Bbb
R}^n_+$ with compact $x'$-support is trace-class, when its order is $<-n+1$. Then $\tr_nG$ is trace-class on $\Bbb R^{n-1}$ and $\Tr _{\Bbb R^n} G=\Tr_{\Bbb R^{n-1}}(\tr_n G)$.
We shall now apply the result of Proposition B.5; that $\tr_n(GQ^N_{\lambda ,+})=\widetilde S_0+\widetilde S_1$, with $\widetilde
S_0=\op'(\tr_ng(x',\xi
')\alpha ^{(N)}(x',\xi ',\mu ))$ (cf. (A.12)) and $\widetilde S_1$ having symbol in $S^{\nu +1,-2N-1,0}\cap S^{\nu -2N,0,0}$. This result was used in [@GSc] to pinpoint the first logarithmic coefficient $\tilde c_0'$ in (1.10), using that only $\widetilde S_0$ contributes to it. But the fact that $\widetilde
S_1$ has $d$-index $-1-2N$ also implies that it contributes only locally to $\tilde c_0''$, so nonlocal contributions to $\tilde c''_0$ come entirely from $\widetilde S_0$.
In the localized situation, denote $\tr_nG=\ttilde G$ and denote its symbol $\tr_ng=\ttilde g$, expanded in homogeneous terms $\ttilde g\sim\sum_{j\ge 0}\ttilde g_{\nu -j}$.
We shall use the notion of a regularized integral (or finite part integral) $\tslint f(\xi
)\,\d\xi $ as in Lesch [@L], [@G4], for polyhomogeneous functions $f(\xi )$ of $\xi \in\Bbb R^{n'}$ (it will be used with $n'=n-1$ or $n$): $\tslint f( \xi )\,\d\xi $ is the constant term in the expansion of $\int_{|\xi|\le\mu }f(\xi) \d\xi$ into powers of $\mu$ and $\log \mu$. In more detail:
When $f( \xi )$ is integrable in $\xi $, $\tslint f( \xi )\,\d\xi $ is the usual integral $\int_{\Bbb
R^{n'}}f( \xi )\,\d\xi $. When $f_{\nu-j}( \xi )$ is homogeneous of order $\nu -j$ in $\xi $ for $|\xi |\ge 1$, then $$\int_{|\xi |\le \mu }f_{\nu -j}\,\d\xi =\int_{|\xi |\le 1 }
f_{\nu -j}\,\d\xi +\int_{1\le |\xi |\le \mu }f_{\nu -j}\,\d\xi ,\tag3.12$$ where a calculation in polar coordinates gives $$\aligned
\int_{1\le |\xi |\le \mu }&f_{\nu -j}( \xi)\,\d\xi
=\int_1^\mu r^{\nu -j +n'-1}\,dr
\int_{|\xi |=1}f_{\nu -j} ( \xi )\,\d S(\xi )\\
&=\cases
c_{j}(\frac{\mu ^{\nu -j +n'}}{\nu -j+n'}-\frac1{\nu
-j+n'})&\text{ if }\nu -j+n'\ne 0,\\
c_{j}\log \mu &\text{ if }\nu -j+n'= 0;\endcases\\
c_{j}&= \int_{|\xi |=1}f_{\nu -j}( \xi
)\,\d S(\xi ).\endaligned\tag3.13$$ Here, $$\slint f_{\nu -j}( \xi )\,\d\xi =\int_{|\xi |\le 1}f_{\nu -j}\,\d\xi
+\cases -\frac1{\nu
-j+n'}c_j&\text{ if }\nu -j+n'\ne 0,\\
0 &\text{ if }\nu -j+n'= 0;\endcases\tag3.14$$ this is consistent with the integrable case. (Such calculations were basic in the proof of .) More generally, when $f(x, \xi )$ is a classical symbol of order $\nu $, $f(x, \xi )\sim \sum_{j\in\Bbb N}f_{\nu -j}(x,\xi
)$, then there is an asymptotic expansion for $\mu \to\infty $, $$\int _{|\xi |\le \mu }f(x,\xi )\,\d\xi \sim \sum_{j\in\Bbb N, j\ne
\nu +n'}f _j(x)\mu ^{n'+\nu -j}+f '_0(x)\log \mu +f ''_0(x),
\tag3.15$$ and we set $$\slint f(x,\xi )\,\d\xi =f ''_0(x).\tag3.16$$ (This is related to Hadamard’s definition of the finite part — [*partie finie*]{} — of certain integrals .) In view of (3.12)–(3.14), we have the precise formula: $$\multline
\slint f(x,\xi )\,\d\xi
=\sum_{0\le j\le \nu +n'}\bigl(\int_{|\xi|\le 1}
f_{\nu -j}(x,\xi )\,\d\xi -
\tfrac{1-\delta _{\nu +n',j}}{\nu +n'-j}
\int_{|\xi |=1}f_{\nu -j}\,\d S(\xi )\bigr)\\
+\int_{\Bbb R^{n'}}
(f-\sum_{j \le \nu +n'}f_{\nu -j})\,\d\xi ,
\endmultline\tag3.17$$ where $\delta _{r,s}$ is the Kronecker delta. (One can replace the sum over $j\le \nu +n'$ by the sum over $j\le J$ for any $J\ge \nu +n'$.)
Let $F=\op (f(x,\xi))$ be the $\psi$do associated with the above symbol. It was shown by Lesch [@L; @Section; @5] that the density $\omega (F)=\tslint f(x,\xi)\,\d\xi dx$ is invariant under the change of the symbol induced by diffeomorphisms of open sets, when $\nu\not\in\Bbb Z$ or $\nu<-n'$. Moreover, it was observed in [@G4] that the proof of [@L] extends to the cases where $\nu \in \Bbb Z$ and $f$ has a parity that fits with the dimension $n'$, cf. (1.4)–(1.5)ff. So in all these cases, $\omega (F)$ depends only on $F$, not on the representation of its symbol in local coordinates. As a consequence, we can consider this density also for $\psi$do’s having the mentioned properties on a manifold $M$ of dimension $n'$. Note that when $M$ is compact, $\int_M\tslint \tr
f(x,\xi)\,\d\xi dx=\Tr(F)$ if $\nu<-n'$.
The notation is used in the following with $n'=n-1$. It was shown in (by working out explicitly in this case) that when $f(x',\xi ')$ is the symbol of a classical $\psi $do $F$ on $\Bbb
R^{n-1}$, and $S=\operatorname{OP}'(s(x',\xi '))$ is an auxiliary second-order uniformly elliptic operator with no principal eigenvalues on $\Bbb
R_-$, then the kernels of $F(S+\mu
^2)^{-N}$ and $\operatorname{OP}'(f(x',\xi ')(s(x',\xi
')+\mu ^2)^{-N})$ with $N>(\nu +n-1)/2$ have diagonal expansions of the form$$K(x',x',\mu )\sim \sum_{j\ge 0}b_j(x')\mu ^{n-1+\nu
-j-2N}+\sum_{k\ge 0}(b'_k(x')\log \mu +b''_k(x'))\mu
^{-k-2N}, \tag3.18$$ where the $b'_k$ are 0 if $\nu \notin\Bbb Z$, and $$\aligned
b_{\nu +n-1}(x')+b''_0(x')&=\slint f(x',\xi ')\,\d\xi '+\text{ local
terms, if $\nu $ is integer $\ge 1-n$},\\
b''_0(x')&=\slint f(x',\xi ')\,\d\xi ' \text{ if $\nu <1-n$ or }
\nu \notin \Bbb Z.\endaligned
\tag3.19$$ Moreover, if $\nu \in\Bbb Z$ and $f$ has a parity that fits with the dimension $n-1$, then $b_{\nu +n-1}(x')=0$ and the second formula in (3.19) holds. (Furthermore, the sum over $k$ in (3.18) for the operators $F(S+\mu
^2)^{-N}$ and $\operatorname{OP}'(f(s+\mu ^2)^{-N})$ skips the terms where $k/2\notin \Bbb Z$, but we need to refer to the general expansion below. In parity cases, the $b_j$ are zero for $\nu +n-1-j$ even, and the $b'_k$ are zero.) What we shall show now is that $\alpha ^{(N)}$ plays to a large extent the same role as $(s+\mu ^2)^{-N}$. We shall deal with the special cases where $\nu <1-n$ or $\nu \notin \Bbb Z$ here, whereas parity cases will be discussed later, around Theorem 3.15.
One has for $\widetilde
S_0=\operatorname{OP}'(\ttilde g\alpha ^{(N)})$, when $N>(n-1+\nu
)/2$: The kernel of $\widetilde S_0$ has an expansion on the diagonal: $$K(\widetilde S_0,x',x',\mu )
\sim \sum_{j\ge 0} a_{j}(x')\mu ^{{n-1+\nu - j}-2N}+\sum_{k\ge
0}(a'_k(x')\log \mu + a''_k(x'))\mu
^{- k-2N},\tag 3.20$$ where the terms $a_j(x')$, and $a'_k(x')$ for $k=\nu +n-1-j$, depend on the first $j+1$ homogeneous terms in the symbols $\ttilde g(x',\xi ')$ and $\alpha ^{(N)}(x',\xi ',\mu )$; the $a''_k(x')$ are global.
If $\nu <-n+1$ or $\nu \in \Bbb R\setminus \Bbb Z$, then $$a''_0(x')=\slint
\ttilde g(x',\xi ')\,\d\xi ',
\tag3.21$$ whereas if $\nu $ is an integer $\ge -n+1$,$$a_{\nu +n-1}(x')+a''_0(x')=\slint
\ttilde g(x',\xi ')\,\d\xi '+\text{ local terms},
\tag3.22$$ where the local terms depend on the first $\nu +n$ homogeneous terms in the symbols of $\ttilde g$ and $\alpha ^{(N)}$. The $a'_k$ vanish if $\nu \notin\Bbb Z$.
The existence of the expansion is assured by , since $\widetilde S_0$ has a weakly polyhomogeneous symbol in $S^{\nu ,-2N,0}\cap S^{\nu -2N,0,0}$, so the main point is to prove the formulas for $a''_0$ resp. $a_{\nu +n-1}+a''_0$. Recall that $$K(\widetilde S_0,x',x',\mu )=\int_{\Bbb R^{n-1}}\ttilde g(x',\xi
')\alpha ^{(N)} (x',\xi ',\mu )\,\d\xi '.$$ First consider $\ttilde g'=\ttilde g-\sum_{j\le
\nu +n-1}\ttilde g_{\nu -j}$, it is of order $\nu -J<-n+1$ where $J=\max\{n+[\nu ],0\}$, hence integrable in $\xi '$. Since$$\alpha ^{(N)}=\mu ^{-2N}+\alpha ^{(N)}_1(x',\xi ',\mu )$$ with $\alpha ^{(N)}_1\in S^{0,-2N,0}\cap S^{1,-2N-1,0}$ (cf. (A.13)ff.), $\ttilde g'\alpha ^{(N)}_1$ is in $S^{\nu -J,-2N,0}\cap
S^{1+\nu -J,-2N-1,0}$, and we can write $$K(\operatorname{OP}'(\ttilde g'\alpha ^{(N)} ),x',x',\mu )=\mu
^{-2N}\int_{\Bbb R^{n-1}}\ttilde
g'\,\d\xi '+\int_{\Bbb R^{n-1}}\ttilde
g'\alpha ^{(N)}_1\,\d\xi '.\tag3.23$$ By , the last integral has an expansion as in (3.18) with $\nu $ replaced by $\nu -J$ and the sum over $k$ starting with $k=1$. This expansion has no term with $\mu
^{-2N}$, so the only contribution from $\ttilde g'$ to the coefficient of $\mu
^{-2N}$ is $\int \ttilde g' \,\d\xi =\tslint \ttilde g' \,\d\xi$. This ends the proof in the case $\nu <-n+1$, where $\ttilde g=\ttilde
g'$.
For the homogeneous terms $\ttilde g_{\nu -j}$, we use an analysis as in : Let $\mu \in\Bbb R_+$ and write the contribution from $\ttilde g_{\nu -j}$ as $$\int_{|\xi '|\ge \mu }\ttilde g_{\nu -j}(x',\xi
')\alpha ^{(N)} (x',\xi ',\mu )\,\d\xi '+\int_{|\xi '|\le \mu }\ttilde
g_{\nu -j}(x',\xi
')\alpha ^{(N)} (x',\xi ',\mu )\,\d\xi '.$$ The first integral gives a local term $\tilde a_j(x')\mu ^{n-1+\nu
-j-2N}$, since the integrand is homogeneous in $(\xi ',\mu )$ of degree $\nu -j-2N$. Note that when $\nu \in\Bbb R\setminus \Bbb Z$, the power cannot be ${-2N}$. The second integral may be written$$\int_{|\xi '|\le \mu }\ttilde g_{\nu -j}\alpha ^{(N)} \,\d\xi '=\mu
^{-2N}\int_{|\xi '|\le \mu }\ttilde g_{\nu -j}(x',\xi
')\,\d\xi '+\int_{|\xi '|\le \mu }\ttilde g_{\nu -j}\alpha^{(N)}_1
\,\d\xi '.\tag3.24$$ Here $\int_{|\xi '|\le \mu }\ttilde g_{\nu -j}\,\d\xi '$ is calculated as in (3.12)–(3.14), contributing $\tslint \ttilde g_{\nu -j}\,\d\xi '$ to the coefficient of $\mu
^{-2N}$. For the last integral in (3.24) we observe that since $\ttilde
g'\alpha ^{(N)}_1$ is in $S^{\nu -J,-2N,0}\cap
S^{1+\nu -J,-2N-1,0}$, with $d$-index $\le -2N-1$, the consideration of this integral in the proof of shows that it contributes only locally to the coefficient of $\mu ^{-2N}$, and not at all if $\nu \notin\Bbb Z$.
Similar statements hold when $\ttilde g$ is given in $(x',y')$-form; then $\ttilde g(x',\xi ')$ is replaced by $\ttilde g(x',x',\xi ')$ in [(3.21)]{} and [(3.22)]{}. The extension to this case is carried out as in [@G4; @Remark; @1.4].
In view of the information on general compositions recalled before the theorem, we have in particular:
Let $S=\operatorname{OP}'(s(x',\xi '))$ be a second-order uniformly elliptic operator on $\Bbb R^{n-1}$ with no principal eigenvalues on $\Bbb
R_-$. Comparing the expansion [(3.18)]{} for the kernel of $\ttilde G(S+\mu ^2)^{-N}$ with the expansion [(3.20)]{} for the kernel of $\widetilde S_0$, one has that $$\aligned
a''_0(x')&=b''_0(x')+\text{ local
terms, if $\nu $ is integer $\ge 1-n$},\\
a''_0(x')&=b''_0(x')\text{ if $\nu <1-n$ or }
\nu \notin \Bbb Z.\endaligned
\tag3.25$$
We then find:
Let $G$ be a singular Green operator on $X$ of order $\nu \in\Bbb R$ and class $0$.
$1^\circ$ Let $g(x',\xi ,\eta _n)$ be the symbol of the operator in a localization to $\crnp$ as described above. Then $$C_0(G,P_{1,\operatorname{D}})=
\int\slint
\tr (\tr_ng)(x',\xi ')\,\d\xi 'dx'+\text{ local
terms};\tag 3.26$$ it holds with vanishing local terms if $\nu <-n+1$ or $\nu \in\Bbb R\setminus \Bbb Z$.
$2^\circ$ Let $P_1$ and $P_2$ be two choices of auxiliary strongly elliptic operators. Then $$C_0(G,P_{1,\operatorname{D}})- C_0(G,P_{2,\operatorname{D}})
\tag3.27$$ is locally determined; it vanishes if $\nu <-n+1$ or $\nu \in \Bbb
R\setminus\Bbb Z$.
Consider $G$ and $Q^N_{\lambda ,+}$ carried over to $\crnp$. Here we have $$\Tr(GQ^N_{\lambda ,+})=\Tr_{\Bbb
R^{n-1}}\tr_n(GQ^N_{\lambda ,+})=
\Tr_{\Bbb R^{n-1}}\widetilde S_{0}+\Tr_{\Bbb R^{n-1}}\widetilde S_{1},
\tag3.28$$ as in Proposition B.5. Since $\widetilde S_{1}$ has symbol in $S^{\nu
+1,-2N-1,0}\cap S^{\nu -2N,0,0}$, it has a diagonal kernel expansion as in (3.18) with the sum over $k$ starting with $k=1$, so the coefficient of $\mu ^{-2N}$ is locally determined (vanishing if $\nu \in\Bbb
R\setminus\Bbb Z $ or $\nu <1-n$). For $\widetilde S_{0}$, we have the diagonal kernel expansion from Theorem 3.3. We integrate the matrix traces in $x'$, and we add on $\Tr(GG^{(N)}_\lambda )$, which is described in Proposition B.3. This gives a trace expansion as in (1.10), with coefficient of $(-\lambda )^{-N}=\mu ^{-2N}$ equal to $$\int \slint \tr \tr_ng\,\d\xi 'dx'+\text{
local terms}$$ (no local terms if $\nu \in\Bbb
R\setminus \Bbb Z $ or $\nu <1-n$); this shows $1^\circ$.
Now if $P_1$ is replaced by $P_2$ in these calculations, $1^\circ$ implies that the contribution to the coefficient of $\mu ^{-2N}$ is modified only in the local terms, vanishing if $\nu <-n+1$ or $\nu \in \Bbb
R\setminus\Bbb Z$; this shows $2^\circ$.
\(a) Another choice of local coordinates and partition of unity will give another decomposition in (3.1), but this will modify the term of a given order in the localized symbol only by terms of the same and higher order.
\(b) When $\nu \notin \Bbb Z$ or $\nu<-n+1$, or $\nu \in\Bbb Z$ and $\tr_n g$ has a parity that fits with the dimension $n-1$, then $\int \tslint \tr
\tr_ng\,\d\xi 'dx'$ will be invariant under such choices, by Remark 3.2.
\(c) In the localized situation, we find by integration from Corollary 3.5 that $$C_0( G, P_{1,\operatorname{D}})=C_0(\tr_n
G, S)
+\text{ local terms,}\tag 3.29$$ for any auxiliary operator $S$ as described there. This reduces the calculation of $C_0$ on $\rnp$ to a calculation of a $C_0$ on $\Bbb
R^{n-1}$, modulo local terms (vanishing if $\nu <1-n$ or $\nu \notin \Bbb Z$).
We have hereby obtained (2.7) in the case $A=G$. Now we turn to (2.8), the commutation property of $C_0(G,P_{1,\operatorname{D}})$. We shall give a proof that reduces it to the commutation property for closed manifolds (cf. e.g. [@G4]), using a variant of the preceding analysis.
Here we want to consider $G$ given in a different form in the localized situation, namely in the form $$G=\sum_{j,k\in\Bbb N}\Phi _jC_{jk}\Phi ^*
_k\tag 3.30$$ explained in Lemma B.6. When $G=\sum_{j,k\in\Bbb N}\Phi _jC_{jk}\Phi ^*
_k$, then $$\Tr_{\rnp}(GQ^N_{\lambda ,+})=\Tr_{\rnp}(\sum_{j,k\in\Bbb N}\Phi
_jC_{jk}\Phi ^*
_k Q^N_{\lambda ,+})=\sum_{j,k\in\Bbb N}\Tr_{\Bbb R^{n-1}}(C_{jk}\Phi ^*
_k Q^N_{\lambda ,+}\Phi _j),$$ since the sum over $j,k$ is rapidly decreasing in the relevant symbol- and operator norms. Here the $C_{jk}$ and $\Phi ^*
_k Q^N_{\lambda ,+}\Phi _j$ are $\psi $do’s on $\Bbb R^{n-1}$; to use their properties, we have to investigate $\Phi ^* _k Q^N_{\lambda ,+}\Phi _j$.
It was a typical feature of the calculations in [@GSc] that we used the passage from $x'$-form to $y'$-form and vice versa in the compositions in such a way that the case of $\psi $do symbols over the boundary was reached before it was necessary to verify asymptotic composition formulas such as $$a(x',\xi ',t)\circ b(x',\xi ',s)\sim \sum_{\alpha \in \Bbb
N^{n-1}}\tfrac{1}{\alpha !}D _{\xi'} ^\alpha
a(x',\xi ',t)\partial^\alpha _{x '}b(x',\xi ',s),\tag3.31$$ with parameters $t$ and $s$. In the $S^{m,d,s}$-spaces over the boundary, the composition formulas work as usual (), whereas the mixture of $\mu $-dependence and $\mu $-independence makes it harder to see what goes on in compositions of the $\psi $dbo symbols reaching into the interior of $X$. Here we rely heavily on the special rational structure of the symbols entering in $Q_\lambda $. (The compositions are not covered by [@G3].)
In compositions containing $\Phi ^*
_k Q^N_{\lambda ,+}\Phi _j$, we have to deal with terms of the form$$\operatorname{OPT}(\bar{\hat\varphi
}_k(\xi _n,\sigma (\xi ') )) \op (q_{-2-J}(x,\xi ,\mu
))_+\operatorname{OPK}(\hat\varphi _j(\xi _n,\sigma (\xi ') ))$$ (and their $\lambda $-derivatives), which do require asymptotic expansions. In fact, when $q_{2-J}$ is independent of $x_n$, the two factors to the right give a relatively simple Poisson symbol depending on $x'$, but then in the composition with the trace symbol to the left (depending on $\xi '$ through $\sigma (\xi
')$), there will be an infinite expansion as in (3.31), where remainder estimates have to be shown. We shall treat this by going back to the original proof of the pseudodifferential composition rule by Taylor expansion, keeping track of the properties of the terms, in particular the remainder, by use of the exact formulas. A composition to the left with $C_{jk}=\op'(c_{jk}(x',\xi '))$ does not make the expression harder to deal with, since $\op'(c_{jk}(x',\xi
'))\operatorname{OPT}(\bar{\hat\varphi
}_k(\xi _n,\sigma )) =\operatorname{OPT}(c_{j,k}(x',\xi ')\bar{\hat\varphi
}_k(\xi _n,\sigma ))$, so we include that in the following.
Let $C_{jk}=\op'(c_{jk}(x',\xi '))$ be of order $\nu $ and let $$S_{J,j,k}=\op'(c_{jk}(x',\xi '))\operatorname{OPT}(\bar{\hat\varphi
}_k(\xi _n,\sigma )) \op (q_{-2-J}(x,\xi ,\mu
))_+\operatorname{OPK}(\hat\varphi _j(\xi _n,\sigma )).\tag 3.32$$ For each $J\in\Bbb N$, $S_{J,j,k}$ is a $\psi $do with symbol in $S^{\nu ,0,-2-J}$, the $(N-1)$’st $\lambda $-derivatives of the symbol lying in $S^{\nu ,0,-2N-J}$.
In view of the expansion in (A.5), $S_{J,j,k}=\sum_{J/2+1\le m\le
2J+1}S_{J,j,k,m}$, where $$\aligned
S_{J,j,k,m}&=C_{jk}\Phi _k^*\op \Bigl(\frac{r_{J,m}(x,\xi
)}{(p_{1,2}+\mu ^2)^{m}}\Bigr)_+\Phi _j\\
&=\operatorname{OPT}(c_{jk}(x',\xi ')\bar{\hat\varphi
}_k(\xi _n,\sigma )) \op \Bigl(\frac{r_{J,m}(x,\xi
)}{(p_{1,2}+\mu ^2)^{m}}\Bigr)_+\operatorname{OPK}(\hat\varphi _j(\xi
_n,\sigma )),
\endaligned\tag 3.33$$ and we have to show the property for each of the terms in the expansion. The function $r_{J,m}$ is a polynomial in $\xi $ of degree $2m-J-2\ge 0$. Let us drop the indexations on $r_{J,m}$ and $p_{1,2}$, and simply write the fraction as $r(p+\mu
^2)^{-m}=q$.
We first consider the case where these symbols are independent of $x_n$. Here one has immediately: $$\operatorname{OP}(q(x',\xi ,\mu
))_+\operatorname{OPK}(\hat\varphi _j(\xi _n,\sigma
))=\operatorname{OPK}(b(x',\xi ,\mu )),\quad b=h^+_{\xi _n}(q\hat\varphi
_j).$$ To prepare for the composition of this Poisson operator with the trace operator $\operatorname{OPT}(c_{jk}\bar{\hat\varphi
}_k)$ to the left, we apply the usual procedure for changing an operator from $x'$-form to $y'$-form: $$\multline
\operatorname{OPK}(b(x',\xi ,\mu ))v=\int e^{i(x'-y')\cdot \xi
'+ix_n\xi _n}b(x',\xi ,\mu )v(y')\,dy'\d \xi \\
=\int e^{i(x'-y')\cdot \xi
'+ix_n\xi _n}\sum_{|\alpha |<M}
\tfrac 1{\alpha !}(x'-y')^\alpha
\partial _{x'}^\alpha b(y',\xi ,\mu )v(y')\,dy'\d \xi \\
+\int e^{i(x'-y')\cdot \xi
'+ix_n\xi _n}\sum_{|\alpha |=M}\tfrac {M(x'-y')^\alpha}{\alpha
!}\int_0^1(1-h)^{M-1}
\partial _{x'}^\alpha b(y'+(x'-y')h,\xi ,\mu )v(y')\,dhdy'\d \xi \\
=\sum_{|\alpha |<M}\operatorname{OPK}(b_\alpha (y',\xi ,\mu
))v+\operatorname{OPK}(b_M(x',y',\xi ,\mu ))v;
\endmultline$$ here we have replaced $(x'-y')^\alpha e^{i(x'-y')\cdot\xi '}$ by $D_{\xi '}^\alpha e^{i(x'-y')\cdot\xi '}$ and integrated by parts w.r.t. $\xi '$ (in the oscillatory integrals), and set $$\aligned
b_\alpha (y',\xi ,\mu )&=\tfrac 1{\alpha !}\overline
D_{\xi '}^\alpha
\partial _{y'}^\alpha b(y',\xi ,\mu ),\\
b_M(x',y',\xi ,\mu )&=\sum_{|\alpha |=M}\tfrac M{\alpha !}\overline
D_{\xi '}^\alpha
\int_0^1(1-h)^{M-1}
\partial _{x'}^\alpha b(y'+(x'-y')h,\xi ,\mu )\,dh.
\endaligned\tag3.34$$
Now we compose with $\operatorname{OPT}(c_{jk}\bar{\hat\varphi
}_k)$ in front. For each term in the sum over $|\alpha |<M$, this gives:$$\gather
\operatorname{OPT}(c_{jk}\bar{\hat\varphi
}_k)\operatorname{OPK}(b_\alpha )=\operatorname{OP}'(s_\alpha
(x',y',\xi ',\mu )),\text{ where}\\
s_\alpha (x',y',\xi ',\mu )=\int c_{jk}\bar{\hat\varphi }_k\tfrac 1{\alpha
!}\overline
D_{\xi '}^\alpha
\partial _{y'}^\alpha h^+_{\xi _n}(q\hat\varphi _j)\,\d \xi _n
=\int c_{jk}\bar{\hat\varphi }_k\tfrac 1{\alpha !}\overline
D_{\xi '}^\alpha
\partial _{y'}^\alpha (q(y',\xi ,\mu )\hat\varphi _j)\,\d \xi _n.
\endgather$$ In the last step it is used that $\overline D_{\xi '}^{\alpha
}\partial _{y'}^\alpha $ commutes with $h^+_{\xi _n}$ and that $\overline
D_{\xi '}^\alpha
\partial _{y'}^\alpha (q\hat\varphi _j)$ is a sum of terms with a similar structure as $q\hat\varphi _j$, so that one can remove $h^+$ in the calculation of the plus-integral, as in (B.5). More precisely,$$\gathered
\overline
D_{\xi '}^\alpha
\partial _{y'}^\alpha (q\hat\varphi _j)=\sum_{\gamma \le\alpha
}\tbinom\alpha \gamma \overline
D_{\xi '}^{\alpha-\gamma }
\partial _{y'}^\alpha q\overline D_{\xi '}^\gamma \hat\varphi _j,
\text{ where}\\
\overline
D_{\xi '}^{\alpha-\gamma }
\partial _{y'}^\alpha q=\sum_{m\le m'\le m+|\alpha -\gamma |}\frac
{r_{m'}(y',\xi )}{(p+\mu ^2)^{m'}},\quad
\overline D_{\xi '}^\gamma \hat\varphi _j=\sum_{|j'-j|\le |\gamma
|,j'\ge 0}r'_{j'}(\xi ')\hat\varphi _{j'};
\endgathered\tag3.35$$ here the $r_{m'}(y',\xi )$ are polynomials in $\xi $ of degree $\le 2m'-J-2-|\alpha -\gamma |$ (vanishing if this number is $<0$) and the $r'_{j'}(\xi ')$ are in $S^{-|\gamma |}$ (found from (B.2)). It follows that$$\overline
D_{\xi '}^\alpha
\partial _{y'}^\alpha (q(y',\xi ,\mu )\hat\varphi _j(\xi _n,\sigma ))=\sum
\frac {r_{\alpha
,m',j'}(y',\xi )}{(p+\mu ^2)^{m'}}\hat\varphi _{j'},\tag3.36$$ where the sum is over $m\le m'\le m+|\alpha |$, $|j-j'|\le |\alpha |$, $j'\ge 0$, and the $r_{\alpha ,m',j'}(y',\xi )$ are in $S^{2m'-J-2-|\alpha |}(\Bbb R^{n-1},\Bbb R^n)$; they are polynomials in $\xi
_n$ of degree $\le 2m'-J-2$, vanishing if this number is $<0$. Each $r_{\alpha ,m',j'}(p+\mu ^2)^{-m'}$ is decomposed in simple fractions with poles $\pm i\kappa ^\pm$ as in (A.7): $$\frac{r_{\alpha ,m',j'}(y',\xi )}{(p(y',\xi )+\mu
^2)^{m'}}= \sum_{1\le m''\le m'}\frac{r^+_{\alpha ,m',j',m''}(y',\xi ',\mu
)}{(\kappa ^+(y',\xi ',\mu )+ i\xi
_n)^{m''}}+\sum_{1\le m''\le m'}\frac{r^-_{\alpha ,m',j',m''}(y',\xi ',\mu
)}{(\kappa ^-(y',\xi ',\mu )- i\xi
_n)^{m''}};\tag3.37$$ here the numerators are of order $m''-|\alpha |-J-2$, lying in $S^{-|\alpha |,0,m''-J-2}$ (since the $\mu $-independent factors coming from derivatives of $\hat\varphi _j$ have order down to $ -|\alpha |$), and each $\lambda $-differentiation lowers the third upper index by $2$.
We can now apply Lemma B.4 to the compositions of $c_{jk}\bar{\hat \varphi }_k$ with (3.37) (recalling from (B.1) that the normalized Laguerre functions have a factor $(2\sigma )^{\frac12}$). This gives $\psi $do symbols in $(x',y')$-form on $\Bbb R^{n-1}$, lying in $S^{\nu -|\alpha |,0,m''-J-2}\cdot
S^{1,0,-m''-1}\subset S^{\nu +1-|\alpha |,0,-J-3}$ if $j'\ne k$, and in $S^{\nu -|\alpha |,0,-J-2}$ if $j'=k$.
Summing over the indices we find that $s_\alpha (y',\xi ',\mu )\in
S^{\nu -|\alpha |,0,-J-2}$, and each $\lambda $-differentiation lowers the third upper index by $2$.
From the sequence of symbols $s_\alpha $, one constructs a symbol $s\in S^{\nu ,0,-J-2}$ having the asymptotic expansion $s\sim
\sum_{\alpha \in\Bbb N^{n-1}}s_\alpha $ in the usual way. There remains the question of how well this represents a symbol of the operator given in (3.33). This will be dealt with by a consideration of the remainder after $M$ terms, using the formula in (3.34).
We shall use the strategy explained after (3.7). In order to be able to distinguish between the various covariables, we shall write here explicitly $\hat\varphi (\xi _n,[\xi '])$ and $\hat\varphi (\xi _n,[\eta '])$ instead of $\hat\varphi (\xi _n,\sigma )$. Consider, for large $M$, $$\multline
\operatorname{OPT}(c_{jk}\bar{\hat\varphi
}_k)\operatorname{OPK}(b_M )v(x')\\
=\int_{\Bbb R^{4n-3}}
e^{i(x'-z')\cdot \xi '+i(z'-y')\cdot \eta '}c_{jk}(x',\xi
')\bar{\hat\varphi }_k(\xi
_n,[\xi ']) h^+_{\xi _n}\bigl[\sum_{|\alpha |=M}\tfrac M{\alpha
!}\overline
D_{\eta '}^\alpha\\
\int_0^1(1-h)^{M-1}
\partial _{z'}^\alpha q(y'+(z'-y')h,\eta ',\xi _n,\mu )
\hat\varphi _j(\xi
_n,[\eta '])\bigr]v(y')\,dh\d
\xi _ndy'\d\eta 'dz'\d\xi '\\
=\int_{\Bbb R^{4n-3}}
e^{i(x'-z')\cdot \xi '+i(z'-y')\cdot \eta '}c_{jk}\bar{\hat\varphi }_k(\xi
_n,[\xi '])
\sum_{|\alpha |=M}\tfrac M{\alpha !}\overline
D_{\eta '}^\alpha
\int_0^1(1-h)^{M-1}\\
\times\partial _{z'}^\alpha q(y'+(z'-y')h,\eta ',\xi _n,\mu )
\hat\varphi _j(\xi
_n,[\eta '])v(y')\,dh\d
\xi _ndy'\d\eta 'dz'\d\xi ';
\endmultline\tag3.38$$ here we have again commuted $\overline D_{\eta '}^{\alpha
}\partial _{z'}^\alpha $ and $h^+_{\xi _n}$, and removed the latter because of the special structure as rational functions. The kernel of this operator is $$\aligned
&K_M(x',y ',\mu )=\int_{\Bbb R^{3n-2}}
e^{i(x'-z')\cdot\xi '+i(z'-y')\cdot \eta '}c_{jk}\bar{\hat\varphi }_k(\xi
_n,[\xi'])\sum_{|\alpha |=M}\tfrac M{\alpha !}\overline
D_{\eta '}^\alpha\\
&\times \int_0^1(1-h)^{M-1}
\partial _{z'}^\alpha q(y'+(z'-y')h,\eta ',\xi _n,\mu )
\hat\varphi _j(\xi
_n,[\eta '])\,dh\d\xi _nd\eta 'dz'\d\xi '.
\endaligned\tag3.39$$
Performing the differentiations $\overline D_{\eta '}^{\alpha
}\partial _{z'}^\alpha $, one gets a sum of similar expressions, now with order $M$ steps lower in the Poisson operator. Consider one term in such an expression, it is of the form (3.39) with $q$ replaced by a fraction $r'(p+\mu ^2)^{-m'}$ as in (3.36), with $r'$ in $S^{2m'-J-2-M}$: $$\aligned
&\operatorname{OPT}(c_{jk}\bar{\hat\varphi}_k)
\operatorname{OPK}(b')v(x')
\\&\quad=\int_{\Bbb R^{4n-3}}
e^{i(x'-z')\cdot \xi '+i(z'-y')\cdot \eta '}c_{jk}\bar{\hat\varphi }_k(\xi
_n,[\xi '])\int_0^1(1-h)^{M-1}\\
&\quad\times
\frac{r'(y'+(z'-y')h,\eta ',\xi _n)}{(p(y'+(z'-y')h,\eta ',\xi _n)+\mu ^2)^{m'}}\hat\varphi _j(\xi
_n,[\eta '])v(y')\,dh\d
\xi _ndy'\d\eta 'dz'\d\xi '.
\endaligned\tag3.40$$
Insert in (3.40) the expansion of the denominator from Lemma A.1: $$(p+\mu ^2)^{-m'}=\sum_{0\le i <L}c_{m',i }\mu ^{-2m'-2i }p^i
+\mu ^{-2m'-2L}p'_L.\tag 3.41$$ The composed operators resulting from the terms in the sum over $i $ are of the form $\mu ^{-2m'-2i }S_i $, where $S_i $ is a $\mu
$-independent $\psi $do on $\Bbb R^{n-1}$ of order $\le\nu+2m'+2i -J-2-M$ (with $i <L$). These operators belong to the calculus. Taking $M$ large in comparison with $2L$ we can assure that their kernels have any given finite degree of smoothness.
The contribution from $\mu ^{-2m'-2L}p'_L$ equals $\mu ^{-2m'-2L} S'$, where $S'$ has a form as in (3.40) with $(p+\mu
^2)^{-m'}$ replaced by $p'_L$; it is a $\psi $do on $\Bbb R^{n-1}$ of order $\le\nu+2m'+2L -J-2-M$. The estimates (A.2) imply that the entries satisfy all symbol estimates [*with uniform estimates in* ]{}$\mu \in\Gamma $ for $|\mu |\ge 1$. Then $ S'$ maps $$S'\: H^{s}(\Bbb R^{n-1})\to H^{s-\nu-2m'-2L+J+2+M}(\Bbb R^{n-1})
, \text{ any }s\in\Bbb R,$$ with uniform estimates in $\mu \in\Gamma $ for $|\mu |\ge 1$, since the norms of such mappings are estimated in terms of the symbol seminorms. Taking $M$ large in comparison with $2L$, we can achieve that the kernel of $S'$ is a smooth as we want, with uniform bounds on the $(x',y')$-derivatives.
It follows that $$S_{J,j,k,m}-\op'\bigl(s(x',y',\xi ',\mu
)\bigr)=S_{J,j,k,m}-\op'\bigl(\sum_{|\alpha
|<M}s_\alpha \bigr)-\op'\bigl(s-\sum_{|\alpha
|<M}s_\alpha \bigr)\tag3.42$$ has a kernel expansion as in (3.7) (for $d=-J-2$), where $L$ and the degree of smoothness can be as high as we want, since $M$ can be taken arbitrarily large. Since differentiation in $\lambda $ lowers the $d$-index by two steps, and $\partial _\lambda =-2\mu \partial _\mu $, the information required for (3.8) is likewise found in this way. This shows that $S_{J,j,k,m}$ is a $\psi $do with symbol $\sim s$ in $S^{\nu ,0,-J-2}$. Its $\lambda
$-derivatives are treated in a similar way.
This shows the proposition in the case where the symbols are independent of $x_n$. Now let $q_{-2-J}$ depend on $x_n$, and consider again (3.32). By a Taylor expansion in $x_n$, $q_{-2-J}$ is of the form$$\multline
q_{-2-J}(x,\xi ,\mu )=\sum_{0\le r<M}\tfrac 1{r!}x_n^r\partial
_{x_n}^rq_{-2-J}(x',0,\xi ,\mu )\\
+\tfrac1{(M-1)!}{x_n^M}\int_0^1(1-h)^{M-1}
\partial _{x_n}^{M}q_{-2-J}(x',hx_n,\xi ,\mu )\,dh.
\endmultline\tag 3.43$$ For the terms in the sum over $r$, we observe that the factor $x_n^r$ goes together with the trace operator to the left $\operatorname{OPT}(c_{jk}(x',\xi ')\bar{\hat\varphi }_k(\xi
_n,\sigma ))$ and replaces its symbol by $c_{jk} D_{\xi
_n}^r\bar{\hat\varphi
}_k(\xi _n,\sigma )$ \[G2, (2.4.14)\]. This lowers the degree — but not the $d$-index — by $r$ steps; more precisely it gives $c_{jk}[\xi
']^{-r}$ times a linear combination of adjacent Laguerre functions, cf. (B.2). The resulting composition is of the type we have dealt with above, giving a symbol in $S^{\nu -r,0,-2-J}$.
The last term in (3.43) is a linear combination of terms of the form $$x_n^M\int_0^1(1-h)^{M-1}\partial
_{x_n}^{M}\bigl(\frac{r(x',hx_n,\xi )}{(p(x',hx_n,\xi )+\mu
^2)^m}\bigr)\,dh,$$ with $r$ polynomial in $\xi $ of order $2m-2-J\ge 0$; consider one such term. Insert the expansion (3.41) of $(p+\mu ^2)^{-m}$. The composed operators resulting from the terms in the sum over $i $ are of the form $\mu ^{-2m-2i }S_i
$ with $S_i $ of order $\le \nu +2m+2i -2-J-M$ (with $i <L$). The operators belong to the calculus, with more smoothness of the kernels, the larger $M$ is taken in comparison with $2L$. Finally, the contribution from $p'_L$ is of the form $\mu
^{-2m-2L}\widetilde S$, with $$\gathered
\widetilde S=\operatorname{OPT}(c_{jk}D_{\xi _n}^M
\bar{\hat\varphi}_k(\xi _n,\sigma))
\operatorname{OP}( \tilde q_{L}(x,\xi ,\mu))_+
\operatorname{OPK}(\hat\varphi _j(\xi_n,\sigma )),\\
\tilde q_{L}(x,\xi ,\mu )=
\tfrac1{(M-1)!}\int_0^1(1-h)^{M-1}\partial
_{x_n}^{M}\bigl({r(x',hx_n,\xi )}p'_{L}(x',hx_n,\xi ,\mu
)\bigr)\,dh.\endgathered\tag 3.44$$ The estimates (A.2) imply that for all indices:$$|D_{x}^\beta D_\xi ^\alpha \tilde q_{L}(x,\xi ,\mu )|\le
C'_{\alpha ,\beta }\ang\xi ^{2m-2-J+2L-|\alpha |},\tag 3.45$$ with $C'_{\alpha ,\beta }$ independent of $\mu \in\Gamma $ for $|\mu |\ge
1$. Then the resulting operator $\widetilde S$ maps $H^{s}(\Bbb R^{n-1})$ to $ H^{s-\nu-2m+2+J-2L+M+1}(\Bbb R^{n-1})$ for any $s\in\Bbb R$, with uniform estimates in $\mu \in\Gamma $ for $|\mu |\ge 1$. Taking $L$ large, and $M$ large in comparison with $2L$, we can obtain that the resulting kernel is a smooth as we want, with coefficient $\mu ^{-2m-2L}$ of as low order as we want.
As in the preceding considerations, this allows us to conclude that the whole composed operator belongs to the calculus, with symbol properties as asserted.
In the above proof, one can moreover keep track of the size of symbol seminorms in their dependence on $j$ and $k$. It follows from Lemma B.4 and the results on $\xi '$-derivatives and $\xi _n$-derivatives we used, that for each $J$, the symbol seminorms resulting from calculations with Laguerre funtions are only polynomially increasing in $j$ and $k$. Thus if the symbol seminorms of the $c_{jk}$ are rapidly decreasing in $j$ and $k$, as they are when the $c_{jk}$ come from an s.g.o. as in Lemma B.6, summations in $j$ and $k$ of operators $S_{J,j,k}$ will converge in the relevant symbol seminorms.
Note that for $J\ge 1$, the $(N-1)$’st derivatives of the symbols of the operators $S_{J,j,k}$ lie in $S^{\nu ,0,-2N-1}\subset S^{\nu
-2N-1,0,0}\cap S^{\nu ,-2N-1,0}$ (cf. (3.3)); hence in their diagonal kernel expansions as in (3.18), the sum over $k$ starts with $k\ge 1$, so that they contribute only locally to the coefficient of $\mu ^{-2N}$. Thus only the terms with $J=0$, $$\aligned
\tfrac{\partial _\lambda ^{N-1}}{(N-1)!}S_{0,j,k}&=
\tfrac{\partial _\lambda ^{N-1}}{(N-1)!}C_{jk}\Phi
^*_k\op((p_{1,2}+\mu ^2)^{-1})_+\Phi _j\\
&=C_{jk}\Phi
^*_k\op((p_{1,2}+\mu ^2)^{-N})_+\Phi _j
\endaligned \tag3.46$$ can contribute nonlocally to the coefficient of $\mu ^{-2N}$. We can use the analysis in the proof of Proposition 3.8 to distinguish the contributing part still further:
We have for each $N\ge 1$ that$$C_{jk}\Phi
^*_k\op((p_{1,2}+\mu ^2)^{-N})_+\Phi _j
=\op'\big(\delta _{jk}c_{jk}(x',\xi ')\alpha ^{(N)}(y',\xi ',\mu
)\big)+\widetilde S_{0,j,k,N},\tag3.47$$ where $\alpha ^{(N)}$ is as defined in [(A.12)]{} and $\widetilde
S_{0,j,k,N}$ has symbol in $S^{\nu +1,0,-2N-1}$.
First let $q$ be independent of $x_n$, and consider how $q_{-2}=(p+\mu ^2)^{-1}$ (with $p=p_{1,2}$) enters in the proof of Proposition 3.8, for $N=1$. Whenever we apply a differentiation $\partial _{x'}^\alpha $ with $|\alpha |>0$ to $q$, it produces a sum of rational functions with denominators $(p+\mu ^2)^{-m}$ with $m\ge 2$ and $\mu
$-independent numerators, so only the undifferentiated term $q_{-2}$ retains the power $-1$. Therefore the resulting terms in (3.35) with $|\alpha |>0$ have denominators $(p+\mu ^2)^{-m'}$ with $m'\ge 2$. Then when we decompose in simple fractions and apply Lemma B.4, we get symbols lying in $S^{\nu +1,0,-3}$. There remains the term$$c_{jk}(x',\xi ')\int\bar{\hat\varphi }_k(\xi _n,\sigma )(p(y',\xi
',\xi _n)+\mu ^2)^{-1}\hat\varphi _j(\xi _n,\sigma )\,\d\xi _n.$$ Here Lemma B.4 shows that it gives a symbol in $S^{\nu +1,0,-3}$ when $j\ne k$, whereas it gives $c_{jk}(x',\xi ')\alpha ^{(1)}(y',\xi
',\mu )\in S^{\nu ,0,-2}$ when $j= k$ (note that the norming factors $(2\sigma )^\frac12$ in the Laguerre functions eliminate the division by $2\sigma $ in (B.8)).
This shows the claim for $N=1$, and in view of the information on $\lambda $-derivatives, it follows for general $N$.
When $q$ depends on $x_n$, we consider the Taylor expansion (3.43). Again, any derivative $\partial _{x_n}^r$ with $r>0$ produces rational functions with $p+\mu ^2$ in powers $\ge 2$ in the denominators. So we find that the Taylor terms, except for the first one, lead to symbols with lower third index. The first term is dealt with above.
Consider a localized situation, where $G$ is given in the form [(3.30)]{} and the symbols $c_{jk}(x',\xi ')$ have compact $x'$-support. Define the diagonal sum$$C=\sum_{j\in\Bbb N}C_{jj}, \text{ with symbol }c(x',\xi
')=\sum_{j\in\Bbb N}c_{jj}(x',\xi ').\tag 3.48$$ Then $$\Tr _{\Bbb R^n_+}(GQ^N_{\lambda ,+})=\Tr_{\Bbb
R^{n-1}}\op'(c(x',\xi ')\alpha ^{(N)}(x',\xi ',\mu ))+\Tr_{\Bbb
R^{n-1}}\widetilde S',\tag 3.49$$ where $\widetilde S'$ has symbol in $S^{\nu +1,0,-2N-1}$. It follows that for $N>(\nu
+n-1)/2$, the trace has an expansion (with all $c'_k=0$ if $\nu
\notin\Bbb Z$): $$\Tr_{\rnp}(GQ^N_{\lambda ,+})\sim \sum_{j\ge 1}\! c_{j}(-\lambda)
^{\frac{n+\nu - j}{2}-N}
+\sum_{k\ge
0}( c'_k\log (-\lambda )+ c''_k)(-\lambda )
^{-\frac k2-N},
\tag 3.50$$ where $$c''_0=\int \slint \tr c(x',\xi ')\,\d\xi 'dx'+\text{
local terms}\tag3.51$$ (with vanishing local terms if $\nu \in\Bbb
R\setminus \Bbb Z $ or $\nu <1-n$).
For each $j,k$, we conclude from Proposition 3.10 that $$\aligned
\Tr _{\Bbb R^n_+}(\Phi _jC_{jk}\Phi ^*_kQ^N_{\lambda ,+})&=
\Tr _{\Bbb R^{n-1}}(C_{jk}\Phi ^*_kQ^N_{\lambda ,+}\Phi _{j})
\\&=\Tr_{\Bbb R^{n-1}}\op'(\delta _{jk}c_{jk}(x',\xi ')\alpha
^{(N)}(y',\xi ',\mu
))+\Tr_{\Bbb R^{n-1}}\widetilde S'_{0,j,k,N}
\\ &=\Tr_{\Bbb
R^{n-1}}\op'(\delta _{jk}c_{jk}(x',\xi ')\alpha ^{(N)}(x',\xi ',\mu
))+\Tr_{\Bbb R^{n-1}}\widetilde S'_{0,j,k,N}.
\endaligned$$ with pseudodifferential operators $\tilde S'_{0,j,k,N}$ having symbols in $S^{\nu+1,0,-2N-1}$. For the last equality it is used that the trace is calulated from the diagonal values of the kernel, where $y'$ is taken equal to $x'$. The formula (3.49) follows by summation over $j$ and $k\in\Bbb N$, using that $(C_{jk})_{j,k\in\Bbb N}$ is rapidly decreasing in $j$ and $k$ and that the symbol seminorms of the composed operators are polynomially controlled in $j$ and $k$, as indicated in Remark 3.9.
Since $c_{jj}\alpha^{(N)}\in S^{\nu-2N,0,0}\cap S^{\nu,-2N,0}$, the trace expansion follows from (3.5) with the statement (3.51) as in the proof of Theorem 3.3. (We have replaced $j$ by $j+1$ in (3.18) and inserted $\mu =(-\lambda )^{\frac12}$, to facilitate comparison with (1.10).)
We have as an immediate corollary (using Proposition B.3 for $GG^{(N)}_\lambda $):
Let $G$ be a singular Green operator on $X$ of order $\nu \in\Bbb R$ and class $0$. Consider a localization $\underline G$ to $\crnp$ as described after Lemma [3.1]{}, write$$\underline G=\sum_{j,k\in\Bbb N}\Phi _j\underline C_{jk}\Phi ^*
_k,\quad \underline C= \sum_{j\in\Bbb N}\underline C_{jj},\tag3.52$$ with $\underline C=\op'(c(x',\xi '))$, and carry $\underline C$ back to an operator $C$ on $X'$ by use of the considered coordinate mappings. Then $$C_0(G,P_{1,\operatorname{D}})=\int\slint
\tr c(x',\xi ')\,\d\xi ' dx'
+\text{ local terms,}\tag 3.53$$ with vanishing local terms if $\nu <-n+1$ or $\nu \in\Bbb R\setminus \Bbb
Z$. In particular, we have for the localized operators as well as for the operators on $X,X'$: $$\aligned
C_0(\underline G,\underline P_{1,\operatorname{D}})&=C_0(\underline
C,\underline S)+\text{ local terms,}\\
C_0(G,P_{1,\operatorname{D}})&=C_0(C,S)+\text{ local terms,}
\endaligned\tag 3.54$$ for any auxiliary elliptic second-order operator $S$ on $X'$ with no principal symbol eigenvalues on $\Bbb R_-$.
This can be used to show the commutation property for $C_0(G,P_{1,\operatorname{D}})$:
When $G$ and $G'$ are singular Green operators of orders $\nu$ resp. $\nu'$ and class 0, then $C_0([G,G'],P_{1,\operatorname{D}})$ is locally determined; it vanishes if $\nu +\nu '\notin\Bbb Z$ or $\nu +\nu '<1-n$.
Expand the localized versions in series with Laguerre operators, $$\underline G=\sum_{j,k\in\Bbb N}\Phi _j\underline C_{jk}\Phi ^*
_k,\quad
\underline G'=\sum_{j,k\in\Bbb N}\Phi _j\underline C'_{jk}\Phi ^*_k.$$ In view of (B.13),$$\aligned
[\underline G,\underline G']&=\sum_{j,k,l,m}\Phi _j \underline C_{jk} \Phi
_k^*\Phi _l
\underline C'_{lm} \Phi _m^* - \sum_{j,k,l,m}\Phi _j \underline C'_{jk}
\Phi
_k^*\Phi _l \underline C_{lm} \Phi _m^*\\
&=\sum_{j,k,m}\Phi _j (\underline C_{jk} \underline
C'_{km}-\underline C'_{jk} \underline C_{km})
\Phi _m ^* ,
\endaligned$$ and the diagonal sum associated with this operator as in (3.52) is: $$\sum_{j,k} (\underline C_{jk} \underline C'_{kj}-\underline C'_{jk}
\underline C_{kj})=\sum_{j,k} \underline C_{jk} \underline
C'_{kj}-\sum_{j,k}\underline C'_{kj}
\underline C_{jk}=\sum_{j,k} [\underline C_{jk}, \underline C'_{kj}].$$ This series converges in the relevant symbol seminorms, in view of the rapid decrease of the $\underline C_{jk}$ and $\underline C'_{jk}$ for $j,k\to\infty $. Then it follows from Theorem 3.12 that $$C_0([\underline G,\underline G'],\underline
P_{1,\operatorname{D}})=C_0(\sum_{j,k}[\underline C_{jk},
\underline C'_{kj}],\underline S)+ \text{ local terms }.$$ Now we use the fact known for closed manifolds (cf. e.g. [@G4]) that each $C_0([\underline C_{jk},\underline C'_{kj}],\underline S)$ is local, defined from specific homogeneous terms in the symbols; then so is the sum. In particular, they vanish in the cases $\nu
+\nu '\notin\Bbb Z$ or $\nu +\nu '<1-n$; then so does the sum. The resulting statements carry back to the manifold situation. Hence $C_0([G,G'], P_{1,\operatorname{D}})$ is local.
It may be remarked that when $G=\sum _{j,k\in\Bbb N}\Phi
_jC_{jk}\Phi ^*_k$ on $\crnp$ with $C_{jk}=\op'(c_{jk}(x',\xi '))$, then the symbol of $G$ has the same principal part as $\sum_{j,k\in\Bbb N}\hat\varphi _j(\xi _n,\sigma
)c_{jk}(x',\xi ')\bar{\hat\varphi }_k(\eta _n,\sigma )$, but need not equal the full symbol. Likewise, the normal trace $\tr_n G$ has the same principal symbol as the diagonal sum $\sum_{j\in\Bbb N}C_{jj}$, but not always the same full symbol. In any case, the above considerations show that they give the same contribution to $C_0(G,P_{1,\operatorname{D}})$ (modulo local terms, if $\nu \in\Bbb Z$, $\nu \ge -n$).
This ends the proof that $C_0(G,P_{1,\operatorname{D}})$ is a quasi-trace on s.g.o.s. We can also describe cases where it is a trace; more about that below.
The local terms, we have talked about so far, are defined from the symbols on $X$. But in fact, it is only the behavior in an arbitrarily small neighborhood of $X'$ that enters:
Assume that a normal coordinate $x_n$ has been chosen such that a suitable neighborhood $X_1$ of $X'$ in $X$ is represented by the product $X_1=X'\times [0,1]$, with $x\in X_1$ written as $x=(x',x_n)$, $x'\in X'$ and $x_n\in [0,1]$. We can also assume that $E|_{X_1}$ is the lifting of $E|_{X'}$. Let $\chi (t)$ be a $C^\infty $ function on $\crp$ that is 1 for $t\le 1$ and $0$ for $t\ge 2$, and, for $\varepsilon <\frac12$, let $$\chi _\varepsilon (x)=\cases \chi (x_n/\varepsilon )\text{ for }x\in
X_1,\\
0\text{ for }x\in X\setminus X_1.\endcases\tag3.55$$ Then we can write $$G=G_{\operatorname{b}}+G_{\operatorname{i}}; \quad
G_{\operatorname{b}}=\chi _{\varepsilon /4}G\chi _{\varepsilon
/4},\tag3.56$$ here $G_{\operatorname{b}}$ is supported in $X_{\varepsilon
/2}=X'\times [0,\varepsilon /2]$, and $G_{\operatorname{i}}$ is of order $-\infty $ (in particular it is trace-class). The auxiliary local coordinates can be chosen such that the variable $x_n$ is preserved for the patches intersecting with $X_1$; then $\tr_n
G_{\operatorname{b}}$ has a meaning as a $\psi $do on $X'$.
[(i)]{} With the decomposition defined above, $$C_0(G,P_{1,\operatorname{D}})=C_0(G_{\operatorname{b}},P_{1,D})+\Tr_X
G_{\operatorname{i}}+ \text{ local terms},\tag 3.57$$ where the local terms depend on the symbol of $P_1$ and the first $\nu +n$ symbols of $G$ on $X_\varepsilon $ only.
[(ii)]{} If $\nu \in \Bbb R\setminus \Bbb Z$ or $\nu <-n+1$, then $$C_0(G,P_{1,\operatorname{D}})=\operatorname{TR}_{X'}(\tr_n
G_{\operatorname{b}})+\Tr_X
G_{\operatorname{i}}.\tag 3.58$$ For such values of $\nu $, $C_0(G,P_{1,\operatorname{D}})$ is a canonical trace on the operators $G$: $$C_0(G,P_{1,\operatorname{D}})=\TR G,\tag3.59$$ in the sense that it satisfies:
“(1)” $C_0(G, P_{1,\operatorname{D}})$ is independent of $P_1$ if $\nu \notin\Bbb Z$ or $\nu <1-n$;
“(2)” $C_0([G,G'],P_{1,\operatorname{D}})=0$ if $\nu +\nu
'\notin\Bbb Z$ or $\nu +\nu '<1-n$.
[(iii)]{} If $\nu < 1-n$, $$\TR G=\Tr_X G.\tag 3.60$$
For (i), note that since $\chi _{\varepsilon /4}=\chi _{\varepsilon /2}\chi _{\varepsilon /4}
$ $$\Tr(G_{\operatorname{b}}R_\lambda ^N)=\Tr(\chi _{\varepsilon /2}\chi
_{\varepsilon /4}G\chi _{\varepsilon /2}R_\lambda ^N)=
\Tr(\chi _{\varepsilon /4}G\chi _{\varepsilon
/2}R_\lambda ^N\chi _{\varepsilon /2}),$$ which shows that only the behavior of $ R_\lambda $ on $X_\varepsilon $ and the behavior of $G$ on $X_{\varepsilon /2}$ enter in the calculation of $C_0(G_{\operatorname{b}},
P_{1,\operatorname{D}})$.
As noted in the preceding theorems, the formulas are exact without unspecified local terms, when $\nu $ is as in (ii). Moreover, $C_0(G_{\operatorname{b}},
P_{1,\operatorname{D}})=C_0(\tr_nG_{\operatorname{b}},S)$ then by (3.29), for the allowed auxiliary elliptic operators $S$. For the manifold $X'$ without boundary we have from [@KV], [@L], [@G4] that $C_0(\tr_n
G_{\operatorname{b}},S)=\TR
(\tr_n G_{\operatorname{b}})$ for such $\nu $. This shows (3.58). The statements in (1) and (2) have been shown further above.
For (iii), we note that when $\nu <1-n$, $G_{\operatorname{b}}$ is trace-class with continuous kernel, and $$\Tr G_{\operatorname{b}}=\Tr _{X_\varepsilon
}G_{\operatorname{b}}=\int_{X'}\int_0^\varepsilon
K(G_{\operatorname{b}}, x,x)\,dx_ndx'=\Tr_{X'}(\tr_n
G_{\operatorname{b}}).\tag3.61$$ Then $C_0( G,P_{1,\operatorname{D}})=\Tr G_{\operatorname{b}}+\Tr
G_{\operatorname{i}}=\Tr G$.
If $G_{\operatorname{b}}$ is written in the local coordinates in the composition form (3.52), one can show similar formulas with $C$ instead of $\tr_nG_{\operatorname{b}}$.
When $\nu $ is integer, it may be so that $\tr_n
G_{\operatorname{b}}$ or $C$ has a parity that fits with the dimension $n-1$, cf. (1.4)–(1.5)ff. One could then try to show a canonical trace property of $C_0(G,P_{1,\operatorname{D}})$ as in the corresponding situation for operators given directly on the closed manifold $X'$ as in [@G4]. This is further supported by the fact that since $\kappa ^+(x',-\xi ',\mu )=\kappa ^{-}(x',\xi ',\mu
)$, $\alpha ^{(1)}$ is even-even, cf. (A.12). However, there are a lot of terms that give local contributions; not only lower order terms in compositions but notably all the $c_{lm}(x',\xi ')$ with $l\ne m$ in the symbol $\sum_{l,m\in\Bbb N}c_{lm}\hat\varphi
_l\bar{\hat\varphi }_m$ of $G$; cf. Lemma B.4: For $l>m$, the resulting symbols involve $\kappa ^+$, for $l<m$ $\kappa ^-$, on the principal level. A high degree of symmetry seems required, or we can restrict the auxiliary operators to a nicer type:
Consider $G$ as defined on the cylinder $X'\times \crp$ (after being cut down to a neighborhood $X_\varepsilon $ of $X'$). Referring to this fixed choice of normal coordinate, we assume that the symbol of $P_1$ is not only even-even with respect to $\xi $ (as any differential operator is), but it is so with respect to $\xi '$. This holds when there are no terms with $D_{x_n}$ times a first-order operator in $x'$, for example when$$P_1=D_{x_n}^2+P'_1,\tag3.62$$ where $P'_1$ is a second-order elliptic differential operator on $X'$. We can take $P'_1$ with positive principal symbol, to assure strong ellipticity of $P_1$ (a little more generality could be allowed).
Assume that $G=\opg(\sum_{j,k\in\Bbb N}c_{jk}\hat\varphi
_j\bar{\hat\varphi }_k)$ (expressed in local coordinates on $X'$, using $x_n$ as the normal coordinate), with the $c_{jk}(x',\xi ')$ all being even-even or all being even-odd. Then when $GQ^N_{\lambda ,+}$ and $GG^{(N)}_\lambda $ are calculated, we find that $$\tr_n GQ^N_{\lambda ,+}\text{ and }\tr_n GG^{(N)}_\lambda$$ are even-even of order $\nu -2N$, resp. even-odd of order $\nu
-2N$. An application of the proof of to these two $\psi $do’s on $X'$ gives that in the even-even case, the local contributions to the coefficient of $\mu ^0$ and to $\log \mu $ vanish if $n-1$ is odd, and in the even-odd case, the local contributions to the coefficient of $\mu ^0$ and to $\log \mu $ vanish if $n-1$ is even. So indeed, restricting the auxiliary operators to those of the form (3.62), we have a canonical trace in suitable parity cases. We have shown:
Let $G$ be given on the cylinder $X'\times \crp$ with points $(x',x_n)$ and let $P_1$, in addition to being strongly elliptic with scalar principal symbol, be of the form [(3.62)]{}. Let $G$ be of order $\nu \in\Bbb Z$, with symbol $\sum_{j,k\in\Bbb
N}c_{jk}(x',\xi ')\hat\varphi
_j(\xi _n,\sigma )\bar{\hat\varphi }_k(\eta _n,\sigma )$ (for a choice of local coordinates on $X'$, with $x_n$ as normal coordinate). Then $$C_0(G,P_{1,\operatorname{D}})=\TR (\tr_n G)\tag3.63$$ holds in the following cases:
[(i)]{} All the $c_{jk}$ are even-even and $n$ is even.
[(ii)]{} All the $c_{jk}$ are even-odd and $n$ is odd.
In this sense, we may say that s.g.o.s have the canonical trace $$\TR G=\TR_{X'}(\tr_n G),\tag3.64$$in the above cases (i) and (ii). One can also show that $C_0([G,G'], P_{1,\operatorname{D}})=0$ when, in the case $n$ even, the $c_{jk}$ and $c'_{jk}$ are either all even-even or all even-odd; in the case $n$ odd it holds when all the $c_{jk}$ are even-even and all the $c'_{jk}$ are even-odd.
If $G$ is given in the form (3.52), the statements hold with $\tr_n G$ replaced by $C=\sum_{j\in\Bbb N}C_{jj}$.
One could also introduce a parity concept directly for s.g.o. symbols $g$, but it involves some unpractical shifts: Recall that in the polyhomogeneous expansion $g\sim\sum_{l\ge 0}g_{\nu -1-l}$, the homogeneity degree of the $l$’th term is $\nu -1-l$, one step lower than the order. Since the $\varphi _j$ depend on $\xi '$ through $\sigma (\xi ')=[\xi ']$, the parity statements on the $c_{jk}$ correspond to the opposite parity in $\xi '$ for $g$, relative to the degrees. — Observe moreover that the parity property depends on the choice of normal coordinate (cf. the general transformation rule [@G2; @(2.4.62)ff.]).
For an example where the result may be of interest, let us mention that the singular Green part of the solution operator for the Dirichlet or Neumann problem for a strongly elliptic differential operator of the form (3.62) has even-even parity for the coefficients $c_{jk}$ in the Laguerre expansion of the symbol. (This is seen by calculations related to those in Remark 4.2 below, for fixed $\lambda $.) Then it has a canonical trace if $n$ is even.
If we merely have that $\tr_n g$ has a parity that fits the dimension $n-1$ (instead of all $c_{jk}$ having it), then there holds at least that $\TR (\tr_n G)=C_0(\tr_n G,S)$ is well-defined independently of $S$ when $n$ is even, resp. odd, by the result for closed manifolds ([@KV], [@G4]) applied to $X'$. This number can then be viewed as the nonlocal part of $C_0(G,P_{1,\operatorname{D}})$, in an explicit way.
The application of [@GS1; @pf.; @of; @Th.; @2.1] to $\tr_n GR^N_\lambda $ shows even more, namely that in the parity cases (i) and (ii) in Theorem 3.15, all the coefficients $\tilde c_j$ with $\nu +n-1-j$ even and all the log-coefficients $\tilde c'_k$ with $k$ even vanish in (1.10) (cf. also ). This holds also for the pointwise kernel expansions as in (3.18).
Finally, let us mention another useful application of the method leading to Theorem 3.13:
When $K$ is a Poisson operator of order $\nu
$ and $T$ is a trace operator of class $0$ and order $\nu '$, then, for auxiliary operators $P_{1,\operatorname{D}}$ and $S$ as above, $$C_0(KT,P_{1,\operatorname{D}})=C_0(TK,S) +\text{ local terms};\tag 3.65$$ the local terms vanishing if $\nu +\nu '<1-n$ or $\notin\Bbb Z$.
Applying a partition of unity, we can reduce to the case where the operators are given on $\rnp$ (with compact $x'$-support). Here we write (cf. Lemma B.7) $$K=\sum_{j\in\Bbb N}\Phi _jC_j,\quad T=\sum_{k\in\Bbb N}C'_k\Phi _k^*,$$ with rapidly decreasing sequences $(C_j)_{j\in\Bbb N}$ and $(C'_k)_{k\in\Bbb N}$ of $\psi $do’s on $\Bbb R^{n-1}$ of order $\nu
-\frac12$ resp. $\nu '+\frac12$ (and with compact $x'$-support). By (3.54) and (B.13),$$\aligned
C_0(KT, P_{1,\operatorname{D}})&=C_0(\sum_{j,k\in \Bbb N}\Phi
_jC_jC'_k\Phi _k^*, P_{1,\operatorname{D}})\\
&=C_0(\sum_{j}C_jC'_j,S)+\text{ local terms};\\
C_0(TK,S)&=C_0(\sum_{j,k\in\Bbb N}C'_k\Phi _k^*\Phi _jC_j,S)
=C_0(\sum_{j}C'_jC_j,S).
\endaligned$$ The latter equals $C_0(\sum_{j}C_jC'_j,S)$ modulo local terms, by the commutativity property for $\psi $do’s on $\Bbb R^{n-1}$ and the rapid decrease of the $C_j$ and $C'_j$. This shows the identity (3.65), and the vanishing statement follows in the usual way.
4\. The integer order case with nonvanishing $\psi $do part
Consider now the case where $A$ equals $P_++G$, of order $\nu \in\Bbb
Z$. Here $C_0(P_++G,P_{1,\operatorname{D}})=C_0(P_+,P_{1,\operatorname{D}})
+C_0(G,P_{1,\operatorname{D}})$, where $C_0(G,P_{1,\operatorname{D}})$ has already been analyzed. For the determination of $C_0(P_+,P_{1,\operatorname{D}})$, we rewrite as follows: $$\aligned
\Tr(P_+R_\lambda ^N)&=\Tr(P_+Q_{\lambda ,+} ^N)+\Tr(P_+G_\lambda
^{(N)})\\
&=\Tr((PQ^N_{\lambda })_+)-\Tr(G^+(P)G^-(Q^N_{\lambda }))+\Tr(P_+G_\lambda
^{(N)}).
\endaligned\tag4.1$$ It follows from the results of [@GSc] recalled in Appendix B that the last two terms have expansions as in Proposition B.3, contributing only locally to $C_0(P_+,P_{1,\operatorname{D}})$. For the first term, we find the desired information by considering $PQ^N_\lambda $ on $\widetilde X$. As in Section 3 (see the details after Lemma 3.1), we can use local coordinates and a subordinate parition of unity, now covering all of $\widetilde X$ with open sets $U_i$, such that subsets of $X$ resp. $\widetilde X\setminus X$ are mapped to subsets of $\crnp$ resp. $\rnm$, the intersections of the $U_i$ with $X'$ being mapped into $\partial \crnp$. Again we can replace the images $V_{j(i_1,i_2)}$ by sets $V'_{i_1,i_2}$ with positive distance from one another (in the $x'$-direction), so that we can refer to one localized operator. By and the more detailed information in , the kernel of $PQ^N_\lambda $ in the localized situation has a diagonal expansion:$$K( PQ^N_\lambda ,x,x)
\sim
\sum_{ j\in \Bbb N } \tilde c_{ j}(x)(-\lambda ) ^{\frac{\nu +n -j}2-N}+
\sum_{k\in \Bbb N}\bigl( \tilde c'_{ k}(x)\log (-\lambda ) +\tilde c''_{
k}(x)\bigr)(-\lambda ) ^{-k-N},
\tag4.2$$ where $$c_{\nu +n}(x)+c''_0(x)=\slint p(x,\xi )\,\d\xi +\text{ local
terms}.\tag4.3$$ From the sets mapped into $\crnp$, we find $\Tr_{X}((PQ^N_\lambda )_+)$ by integrating over ${\Bbb R^n_+}$. Then (4.2) implies $$\Tr_X((PQ^N_\lambda )_+)\sim \sum_{ j\in \Bbb N } \tilde c_{ j,+}(-\lambda
) ^{\frac{\nu +n -j}2-N}+
\sum_{k\in \Bbb N}\bigl( \tilde c'_{ k,+}\log (-\lambda ) +\tilde c''_{
k,+}\bigr)(-\lambda ) ^{-k-N},
\tag4.4$$ where the coefficients are obtained from the coefficients in (4.2) by integration in $x$. In particular, in view of (4.3),$$c_{\nu +n,+}+c''_{0,+}=\int_{\Bbb R^n_+} \slint \tr p(x,\xi )\,\d\xi dx
+\text{ local terms.}\tag4.5$$ Here $c_{\nu +n,+}$ is also locally determined, so we can conclude:
Let $P$ be of order $\nu \in\Bbb Z$, having the transmission property at $X'$. In the localized situation, $$C_0(P_+,P_{1,\operatorname{D}})=\int_{\Bbb R^n_+} \slint \tr p(x,\xi
)\,\d\xi
dx+\text{ local terms.}\tag4.6$$ When $P_1$ and $P_2$ are two choices of auxiliary operator, then $C_0(P_+,P_{1,\operatorname{D}})-C_0(P_+,P_{2,\operatorname{D}})$ is locally determined.
It only remains to establish the last statement: It holds, since the global term $\int_{\Bbb R^n_+} \tslint \tr p(x,\xi )\,\d\xi dx$ cancels out in the calculation of the difference in local coordinates, leaving only locally determined terms.
As usual, we can ask for cases where the local terms vanish. They do so for $\nu <-n$, where $$C_0(P_+,P_{1,\operatorname{D}})=\Tr P_+$$ is easily seen. For $\nu \ge -n$, one can analyze cases with parity properties, but they will in general not have vanishing local contributions, as the following considerations show: Let $P_+=I$, the simplest possible choice, and let $P_1$ be of the simple form (3.62) with $P'_1$ selfadjoint positive. We have to expand$$\Tr(IR^N_\lambda )=\Tr (Q^N_{\lambda ,+})+ \Tr (G^{(N)}_\lambda ).$$ Since $Q_\lambda $ is even-even,$$\Tr (Q^N_{\lambda ,+})\sim\sum_{j'\in\Bbb N}\tilde c_{2j',+}(-\lambda
)^{\frac n2-j'-N},$$ which contributes to the power $(-\lambda )^{-N}$ when $n$ is [*even*]{}. On the cylinder $X'\times \crp$, the Dirichlet s.g.o. $G_\lambda $ can be explicitly constructed (cf. (2.10)) to be$$\gathered
G_\lambda =-K_{A_\lambda }(2A_\lambda )^{-1}T_{A_\lambda },\text{ where}\\
A_\lambda =(P'_1-\lambda )^{\frac12},\; K_{A_\lambda }\:
v(x')\mapsto e^{-x_nA_\lambda }v,\;T_{A_\lambda }\:u(x)\mapsto
\int_0^\infty e^{-x_nA_\lambda }u(x',x_n)\,dx_n;
\endgathered$$ with notation as in [@G6]. Then $\tr_n G_\lambda =-(2A_\lambda
)^{-2}=-\frac14(P'_1-\lambda )^{-1}$. This is a resolvent on $X'$, and the $N$’th derivative has a trace expansion $$\Tr (G^{(N)}_\lambda )=\Tr_{X'}(\tr_n G_\lambda ^{(N)})
\sim\sum_{j'\in\Bbb N}a_{2j'}(-\lambda
)^{\frac {n-1}2-j'-N},$$ which contributes to the power $(-\lambda )^{-N}$ when $n$ is [*odd*]{}.
However, in cases where $p$ has a parity that fits with the dimension $n$, one can at least observe that the integral $\int_{\Bbb R^n_+} \tslint \tr p(x,\xi )\,\d\xi dx$ has a coordinate invariant meaning (cf. Remark 3.2), so it gives an explicit value that can be considered as the nonlocal part of $C_0(P_+,P_{1,\operatorname{D}})$.
Theorem 4.1 assures the validity of (2.7) for $C_0(P_++G,P_{1,\operatorname{D}})$. In the proof of (2.8), we restrict the attention to operators $P$ of normal order $\le 0$.
For (2.8), it would be possible to refer again to Laguerre expansions (as in the considerations of s.g.o.s in Section 3), using that $P$ on the symbol level acts like a Toeplitz operator. More precisely, if one writes $p(x',0,\xi )=\sum_{j\in\Bbb Z}a_j(x',\xi ')\hat\psi _j(\xi _n,\sigma
)$ with $$\hat\psi _j(\xi _n,\sigma )=\frac{(\sigma -i\xi _n)^j}{(\sigma +i\xi
_n)^{j}},\tag4.7$$ then in the one-dimensional calculus based on Laguerre expansion in $L_2(\rp)$, $p(x',0,\xi ',D_n)_+$ acts like the Toeplitz operator with matrix $(a_{j-k})_{j,k\in\Bbb N}$, cf. . With some effort, the localness of $C_0([P_+,G'],P_{1,\operatorname{D}})$ and $C_0([P_+,P'_+],P_{1,\operatorname{D}})$ can be proved by use of the localness of $C_0$ on commutators over $X'$, but the interpretations and composition rules for the involved operators are not altogether simple to deal with. Instead we shall rely on repeated commutator techniques.
In order to study the commutator of a pseudodifferential operator $P_+$ of order $\nu \in\Bbb Z$ with an operator $A'=P'_++G'$ of order $\nu '\in \Bbb Z$, we write:$$\Tr([P_+,A']R_\lambda ^N)=
\Tr([P_+,A']Q_{\lambda ,+} ^N)+\Tr([P_+,A']G_\lambda^{(N)}).\tag4.8$$ We know from Proposition B.3 that the second term contributes only locally to $C_0$, and go on with an analysis the first term: $$\aligned
\Tr([P_+,A']&Q_{\lambda ,+} ^N)=\Tr(A'Q_{\lambda ,+}
^NP_+)-\Tr(A'P_+Q_{\lambda ,+} ^N)\\
&=\Tr(A'[Q_{\lambda } ^N,P]_+)-\Tr(A'G^+(Q_{\lambda } ^N)G^-(P))+
\Tr(A'G^+(P)G^-(Q_{\lambda } ^N))\\
&=\Tr(A'[Q_{\lambda } ^N,P]_+)-\Tr(G^-(P)A'G^+(Q_{\lambda } ^N))+
\Tr(A'G^+(P)G^-(Q_{\lambda } ^N)).
\endaligned\tag4.9$$ Again, the last two terms contribute only locally to $C_0$, by Proposition B.3. It is used that our hypotheses on $P$ assure that the s.g.o.s $G^{\pm}(P)$ are of class 0.
Now consider the remaining term $\Tr(A'[Q_{\lambda } ^N,P]_+)$. The expression $[Q_{\lambda } ^N,P]_+$ is not very convenient in the boundary calculus, as a mixture of strongly polyhomogeneous parameter-dependent and arbitrary parameter-independent interior operators. However, we shall now show that it can be reduced to an expression with $\lambda $-dependent factors to the right only, with better decrease in $\lambda $ than $Q_\lambda $, plus a manageable remainder term.
Let $P$ be of order $\nu $ and normal order $\le 0$. For $r>0$, let $P_{(r)}$ denote the $r$’th commutator of $P$ with $P_1$: $$P_{(1)}=[P,P_1],\;
P_{(2)}=[[P,P_1],P_1],\;\dots,\;P_{(r)}=[\cdots[[P,P_1],P_1]\dots,P_1];\tag4.10$$ it is of order $\nu +r$ and of normal order $\le \min\{\nu +r,2r\}$).
For any $M>0$, $[Q_\lambda ,P]$ may be written as$$[Q_\lambda ,P]=P_{(1)}Q_\lambda ^2+P_{(2)}Q_\lambda
^3+\dots+P_{(M)}Q_\lambda ^{M+1}+Q_\lambda P_{(M+1)}Q_\lambda
^{M+1}.\tag4.11$$
Since $P_1$ has scalar principal symbol and is of order 2, $[P,P_1]$ is of order $\nu +1$. Since $P$ is of normal order 0 and $P_1$ is of normal order 2, $[P,P_1]$ is of normal order 2 (e.g., $\partial _{x_1}p_{1,2}$ is generally so). Similarly, every subsequent commutation lifts the order by 1 and the normal order by 2, so the $r$’th commutator is of order $\nu +r$ and normal order $\le 2r$. Since it has the transmission property, it also has normal order $\le \nu +r$. This explains the first statement. The second statement follows by successive applications of the following calculation: $$\aligned
[Q_\lambda ,P_{(r)}]&=Q_\lambda P_{(r)}-P_{(r)}Q_\lambda
=Q_\lambda P_{(r)}(P_1-\lambda )Q_\lambda -Q_\lambda (P_1-\lambda
)P_{(r)}Q_\lambda \\
&=Q_\lambda [P_{(r)},P_1]Q_\lambda=Q_\lambda P_{(r+1)}Q_\lambda =
P_{(r+1)}Q_\lambda^2+ [Q_\lambda ,P_{(r+1)}]Q_\lambda;
\endaligned$$ hence $$\aligned
[Q_\lambda ,P]&= P_{(1)}Q_\lambda ^2+[Q_\lambda ,P_{(1)}]Q_\lambda
=\dots \\
&=P_{(1)}Q_\lambda ^2+P_{(2)}Q_\lambda
^3+\dots+P_{(M)}Q_\lambda ^{M+1}+[Q_\lambda ,P_{(M)}]Q_\lambda ^{M}\\
&=P_{(1)}Q_\lambda ^2+P_{(2)}Q_\lambda
^3+\dots+P_{(M)}Q_\lambda ^{M+1}+Q_\lambda P_{(M+1)}Q_\lambda
^{M+1}.\quad\square
\endaligned$$
Let $P$ and $P'$ be $\psi $do’s of order $\nu $ resp. $\nu '\in\Bbb Z$ and normal order $\le 0$, and let $G'$ be an s.g.o. of order $\nu '$ and class $0$; denote $P'_++G'=A'$. For $N>(\nu +\nu '+n)/2$, $\Tr(A'[Q_\lambda ^N,P]_+)$ has an expansion$$\Tr(A'[Q_\lambda ^N,P]_+)
\sim\sum_{j\ge
0}\!a_{j}(-\lambda)
^{\frac{n+\nu +\nu '- j}{2}-N}
+\sum_{k\ge
1}(a'_k\log (-\lambda )+a''_k)(-\lambda )
^{-\frac k2-N}.\tag 4.12$$
We know from the preceding calculation that there is an expansion as in (1.10) with $\nu $ replaced by $\nu +\nu '$; the point is to show that the series in $k$ starts with a lower power than $-N$.
Recalling that $Q_\lambda ^N=\tfrac 1{(N-1)!}\partial _\lambda
^{N-1}Q_\lambda $, we find from (4.11): $$\multline
[Q^N_\lambda ,P]=c_1P_{(1)}Q_\lambda ^{N+1}+c_2P_{(2)}Q_\lambda
^{N+2}+\dots+c_MP_{(M)}Q_\lambda ^{N+M}\\
+\sum_{1\le l\le N}c'_{l}Q_\lambda ^{l}P_{(M+1)}Q_\lambda
^{N+M+1-l}.\endmultline\tag4.13$$ In order to establish the expansion (4.12) we shall study the terms $$P'_+(P_{(r)}Q^{(N+r)}_\lambda)_+ \
\text{ and }
\ G'(P_{(r)}Q^{(N+r)}_\lambda)_+; \quad 1\le r\le M; \tag4.14$$ as well as $$A'(Q_\lambda ^{l}P_{(M+1)}Q_\lambda ^{N+M+1-l})_+ ;\quad
1\le l\le N. \tag4.15$$ The operator $Q_\lambda ^{l}P_{(M+1)}Q_\lambda ^{N+M+1-l}$ is a $\psi$do on $\widetilde X$ with a symbol in $S^{\nu+M+1,0,-2N-2M-2}$. The compositions in (4.15) will therefore be trace-class on $L^2(X)$ when $\nu -M-2N-1\le 0$, $\nu '+\nu -M-2N-1<-n$, and their trace will for any given $M'$ be $O(\lambda^{-M'})$ for sufficiently large $M$. Our assertion will thus be true for these terms.
We turn to the terms in (4.14), considered in local coordinates. In view of (2.6), we can decompose $P_{(r)}= P_{(r)}'+P_{(r)}''$, where $P_{(r)}''$ is a differential operator of order $\le \nu+r$ and normal order $\le \min\{\nu +r,2r\}$, and $P_{(r)}'$ has the property that $G^+(P'_{(r)})$ is of class 0, as a sum of an operator of normal order 0 and an operator with a factor $x_n^{M'}$ to the right, $M'\ge 2r$. Then $$(P_{(r)}Q_\lambda ^{N+r})_+
=P'_{(r),+}Q_{\lambda ,+}^{N+r}+G^+(P'_{(r)})G^-(Q_{\lambda
}^{N+r})+(P''_{(r)}Q_\lambda ^{N+r})_+. \tag4.16$$
Note that $$P''_{(r)}=\sum_{0\le j\le 2r}S_j(x,D_{x'})D_{x_n}^j$$ with tangential differential operators $S_j$ of order $\nu +r-j$. The symbol of $Q^{N+r}_\lambda $ has the structure described in (A.10), with typical term $r_{N+r,J,m}(p_{1,2}+\mu ^2)^{-m}$ where $J+2N+2r\le 2m\le 4J+2N+2r$ and $r_{N+r,J,m}$ is polynomial in $\xi $ of degree $\le 2m-2N-2r-J$. When this is multiplied by $\xi _n^{2r}$ and we redefine $J+2r=J'$, we get a term with numerator of order $\le 2m-2N-J'$ and denominator $(p_{1,2}+\mu ^2)^{m}$, $J'+2N\le 2m\le 4J'+2N-6r\le 4J'+2N$; this is of the type in the symbol expansion of $Q^N_\lambda $. Similarly, multiplication by $\xi _n^{j}$ for $j\le 2r$ gives terms of the types in the symbol expansions of $Q^{N'}_\lambda $ with $N'\ge N$. The point of this analysis is that we can write $$G^-(P''_{(r)}Q^{N+r}_\lambda )=\sum_{0\le j\le
2r}S_j(x,D_{x'})G^-(D_{x_n}^jQ^{N+r}_\lambda ),
\tag4.17$$ where the factors $G^-(D_{x_n}^{j}Q^{N+r}_\lambda )$ have structures like the $G^-(Q^{N'}_\lambda )$, $N'\ge N$, so that Proposition B.3 can be applied.
The composition of $A'$ with the terms in (4.16) gives a number of terms: $$\aligned
\text{(i) }&P'_+P'_{(r),+}Q_{\lambda ,+}^{N+r}=(P'P'_{(r)})_+Q_{\lambda
,+}^{N+r} -
G^+(P')G^-(P'_{(r)})Q_{\lambda ,+}^{N+r}\\
&=(P'P'_{(r)}Q_{\lambda
}^{N+r} )_+
-G^+(P'P'_{(r)})G^-(Q_{\lambda ,+}^{N+r})-G^+(P')G^-(P'_{(r)})Q_{\lambda
,+}^{N+r} ,\\
\text{(ii) }&G'P'_{(r),+}Q_{\lambda ,+}^{N+r},\\
\text{(iii) }&A'G^+(P'_{(r)})G^-(Q_{\lambda
}^{N+r}),\\
\text{(iv) }&P'_+(P''_{(r)}Q_\lambda ^{N+r})_+=
(P'P''_{(r)}Q_\lambda ^{N+r})_+-
G^+(P')G^-(P''_{(r)}Q_\lambda ^{N+r}),\\
\text{(v) }&G'(P''_{(r)}Q_\lambda ^{N+r})_+.
\endaligned$$ To the terms in (i) and (iii) with the s.g.o.-factor $G^-(Q_\lambda
^{r+N})$ we can apply Proposition B.3 with $\nu ,N$ replaced by $\nu +\nu ',N+r$; here the series in $k$ starts with the power $(-\lambda )^{-\frac12-N-r}$.
The term $G^+(P')G^-(P''_{(r)}Q_\lambda ^{N+r})$ in (iv) with the s.g.o.-factor $G^-(P''_{(r)}Q_\lambda ^{r+N})$ is written in view of (4.17) as $$G^+(P')G^-(P''_{(r)}Q_\lambda ^{N+r})=
\sum_{0\le j\le
2r}G^+(P')S_j(x,D_{x'})G^-(D_{x_n}^jQ^{N+r}_\lambda) ,$$ where each $G^+(P')S_j(x,D_{x'})$ is of class 0. Proposition B.3 applies and gives contributions as in (4.12).
For the terms in (i) and (ii) of the form of an s.g.o. composed with $Q^{r+N}_{\lambda ,+}$, we apply Proposition B.5. Since $r\ge 1$, the series in $k$ has only powers $\le -1-N$ of $\lambda $, so the contribution to the coefficient of $(-\lambda )^{-N}$ is local.
For (v), we rewrite $G'(P''_{(r)}Q_\lambda ^{N+r})_+=(G'P''_{(r),+})Q_{\lambda ,+}
^{N+r}$, noting that the leftover term for the product is zero, since $P''_{(r)}$ is a differential operator. Now $G'P''_{(r),+}$ is an s.g.o. of order $\nu'+\nu+r$ and class $\le\min\{\nu+r, 2r\}$. It therefore has a representation $$G_0 + \sum_{0\le j\le 2r-1} K_j\gamma_j$$ with an s.g.o. $G_0$ of order $\nu+\nu'+r$ and class 0 and Poisson operators $K_j$ of order $\nu+\nu'+r-j$. The composition $G_0Q_{\lambda ,+}^{N+r}$ is like (ii); Proposition B.5 applies to give an expansion (4.12).
Next we decompose $K_j$ according to Lemma B.7: $K_j=\sum_{l\ge 0}\Phi_lC_{jl}$ with a rapidly decreasing sequence $(C_{jl})_l$ in $S^{\nu+\nu'+r-j+\frac12}$. Then $$\Tr(K_j\gamma_jQ^{N+r}_{\lambda,+})=
\sum_{l\ge 0}
\Tr_{\Bbb R^{n-1}}(\gamma_0\partial^j_{x_n}
Q^{N+r}_{\lambda ,+}\Phi_lC_{jl}).$$ The composition $\gamma_0\partial_{x_n}^jQ^{N+r}_{\lambda ,+}\Phi_l$ is a $\psi$do on $\Bbb R^{n-1}$ with the symbol $$\sum_{j_1+j_2=j}\binom{j}{j_1}
\int^+
(i\xi_n)^{j_1}\,\partial^{j_2}_{x_n}q^{\circ(N+r)}(x',0,\xi,\mu)\,
\hat\varphi_l(\xi_n,\sigma)\d\xi_n.\tag 4.18$$ We know from (A.8)–(A.11) that $(i\xi_n)^{j_1}\partial_{x_n}^{j_2}q^{\circ(N+r)}_{-2(N+r)-J}$ is a finite sum of terms of the form $\xi_n^{j_1}r_{J,m}(p_{12}+\mu^2)^{-m}$ with a polynomial $r_{J,m}$ of degree $\le 2m-2(N+r)-J$. Decomposing it into simple fractions, we obtain a sum of terms of the form $r^\pm_{J,k}(\kappa^\pm\pm i\xi_n)^{-k}$ with $r^\pm_{J,k}(x',\xi',\mu)$ strongly homogeneous in $(\xi',\mu)$ of degree $\le j+k-2(N+r)-J$. Inserting this into (4.18) and using Lemma B.1, we see that $\gamma_jQ^{N+r}_{\lambda ,+}\Phi_l$ has its symbol in $S^{\frac12,0,-2N-2r+j}$; the symbol seminorms grow at most polynomially in $l$. Since $j<2r$, we then reach the expansion (4.12).
The remaining $\psi $do terms in (i) and (iv), $(P'P'_{(r)}Q_{\lambda
}^{N+r} )_+$ and $(P'P''_{(r)}Q_\lambda ^{N+r})_+$, add up to $(P'P_{(r)}Q_\lambda ^{N+r})_+$, which is treated as in Theorem 4.1, using the pointwise kernel expansion on $\widetilde X$. Here $P'P_{(r)}Q_\lambda ^{r+N}$ has symbol in $S^{\nu '+\nu +r,
0,-2r-2N}\subset S^{\nu '+\nu -r-2N,0,0}$$\cap S^{\nu '+\nu +r,
-2r-2N,0}$ with $d$-index $\le -2-2N$ since $r\ge 1$, so we get an expansion contributing only locally to the coefficient of $(-\lambda )^{-N}$.
We now have all the ingredients for the proof of:
Let $P$ and $P'$ be $\psi $do’s of order $\nu $ resp. $\nu '\in\Bbb Z$ and normal order $0$, and let $G$ and $G'$ be singular Green operators of order $\nu $ resp. $\nu '$ and class $0$. Then $C_0([P_++G,P'_++G'],P_{1,\operatorname{D}})$ is locally determined.
We have that $$[P_++G,P'_++G']=[P_++G,P'_+]+[P_+,G']+[G,G'].$$ The last term was shown in Theorem 3.13 to contribute locally to $C_0$. For the two other terms we have the analysis above, through (4.8), (4.9) and Proposition 4.4, showing that they contribute only locally to $C_0$.
This completes the proof that (2.8) holds also for operators of type $A=P_++G$, so $C_0(P_++G,P_{1,\operatorname{D}})$ is indeed a quasi-trace on such operators. In addition, we have found the interesting information that the contribution from $P_+$ can be traced back to a pointwise defined contribution from $P$ over $\widetilde X$, and that the contribution from $G$ can be traced back to an interior trace contribution of order $-\infty $ plus a contribution from the normal trace on $X'$; here both $\widetilde X$ and $X'$ are compact manifolds without boundary.
Appendix A. The structure of the auxiliary operators
We here recall the symbol formulas established and used in [@GSc], and some useful consequences. The parameter-dependent entries were indexed by $\mu =(-\lambda )^{\frac12}$ in [@GSc]; we simply replace this here by indexation by $\lambda $, although $\mu $ can still appear as a variable.
Throughout this paper, we denote by $[\xi ']$ a positive $C^\infty $ function of $\xi
'$ that coincides with $|\xi '|$ for $|\xi '|\ge 1$. It will often be denoted $\sigma (\xi ')$ or just $\sigma $.
The principal (second-order) symbol of $P_1$ is denoted $p_{1,2}$, so the principal symbol of $Q_\lambda $ is $q_{-2}=(p_{1,2}-\lambda)^{-1}
=(p_{1,2}+\mu ^2)^{-1}$.
We assume $P_1$ to be strongly elliptic; this means that the principal symbol $p_{1,2}(x,\xi )$ has positive real part when $\xi \ne 0$. Since $|\operatorname{Im}p_{1,2}(x,\xi )|\le |p_{1,2}(x,\xi )|\le
C\operatorname{Re}p_{1,2}(x,\xi )$, there is a sector $\Gamma $ such that $p_{1,2}(x,\xi )+\mu ^2\ne 0$ when $\mu \in \Gamma \cup \{0\}$, $(\xi
,\mu
)\ne (0,0)$; cf. also (3.9).
The following observation will be useful:
Let $p (x,\xi )$ be a uniformly strongly elliptic homogeneous second-order differential operator symbol on $\Bbb R^n$, with $p-\lambda =p+\mu ^2$ invertible for $\mu \in\Gamma $. Then for any $m$, $L>0$,$$\aligned
(p +\mu ^2)^{-m}&=
\sum_{0\le j <L}c_{m,j } \mu ^{-2m-2j } p^{j }+
\mu ^{-2m-2L } p'_L(x,\xi ,\mu ), \text{ where}\\
p'_L(x,\xi ,\mu )&=\sum_{0\le k\le m}c_{m,L,k}
p ^{L+k}(p +\mu ^2)^{-k}. \endaligned\tag A.1$$ Here one has for all indices:$$|\partial _\lambda ^N\partial _x^\beta \partial _\xi ^\alpha
p'_L(x,\xi ,\mu )|\le C_{\alpha ,\beta ,N
}\ang\xi ^{2L-|\alpha |-2N},\tag A.2$$ with constants $C_{\alpha ,\beta ,N}$ independent of $\mu \in\Gamma $ for $|\mu |\ge 1$.
One has for $b\in \Bbb C\setminus\{1\}$, $L\ge 1$, $$\aligned (1-b)^{-m}=&\frac1{(m-1)!}\partial_b^{m-1}\frac1{1-b}=
\frac1{(m-1)!}\partial_b^{m-1}\Bigl(\sum_{0\le k <L+m-1}b^k
+\frac{b^{L+m-1}}{1-b}\Bigr)
\\
=&\sum_{0\le j <L}c_j b^{j }+\sum_{1\le l\le m}
\frac{c'_lb^{L+l-1}}{(1-b)^{l}}.
\endaligned\tag A.3$$ An application to $1-b=1+p /\mu ^2$ gives$$\aligned
(p +\mu ^2)^{-m}&=\mu ^{-2m}(1+p /\mu ^2)^{-m}\\
&=\mu ^{-2m}\bigl(\sum_{0\le j <L}c_j \mu ^{-2j
}p ^j +\sum_{1\le l\le m}c'_l\mu
^{-2L-2l+2}p ^{L+l-1}
(1+p /\mu ^2)^{-l}\bigr)\\
&=\sum_{0\le j <L}c_j \mu ^{-2m-2j
}p ^j +\mu ^{-2m-2L+2}\sum_{1\le l\le m}c'_lp ^{L+l-1}
(p +\mu ^2)^{-l}.
\endaligned$$ Insertion of $\mu ^2(p +\mu ^2)^{-1}=1-p (p +\mu ^2)^{-1}$ in the last sum leads to (A.1). Since $p $ is polynomial of degree 2, $(p +\mu ^2)^{-1}$ is $C^\infty $ and homogeneous in $(\xi ,\mu
)$ of degree $-2$ for $(\xi ,\mu )\ne 0$, and $\partial
_\lambda ^N(p-\lambda )^{-l}=c_{l,N}(p-\lambda )^{-l-N},$ $$\multline
|\partial _\lambda ^N\partial _x^\beta
\partial _\xi ^\alpha p'_L(x,\xi ,\mu )|
\le C'_{\alpha ,\beta ,N}\sum_{0\le k\le m,\alpha '\le\alpha }
(\partial _\xi ^{\alpha '}p^{L+k})
|(\xi ,\mu )|^{-2k-2N-|\alpha -\alpha '|}\\
\le C_{\alpha ,\beta,N}
\ang\xi ^{2L-|\alpha |-2N}\endmultline$$ for $|\mu |\ge 1$, showing (A.2).
Let $\xi =(\xi ',\xi _n)$ in local coordinates at the boundary, then for $(\xi ',\mu )\ne (0,0)$, the strong ellipticity implies that the polynomial in $\xi _n$ $$p_{1,2}(x',0,\xi ',\xi _n)+\mu ^2=a(x')\xi _n^2+b(x',\xi ')\xi
_n+c(x',\xi ')+\mu ^2\tag A.4$$ has two roots $\varrho _1(x',\xi ',\mu )$ and $\varrho _2(x',\xi ',\mu )$ in $\Bbb C\setminus \Bbb R$. When $\mu \in\crp$, one of the roots, say $\varrho _1$, lies in $\Bbb C_+$ and the other, $\varrho
_2$, in $\Bbb C_-$. For, $a\xi _n^2+b\xi _n+c+\mu ^2$ can be carried into $\operatorname{Re}a\,\xi _n^2+\operatorname{Re}b\,\xi
_n+\operatorname{Re}c+\mu ^2$ by a homotopy that preserves the property of having positive real part, and for the latter polynomial, the roots are placed in this way; they depend continuously on the polynomial, hence cannot cross the real axis. The placement of the roots is also preserved when $\mu $ is moved to a general element of $\Gamma $. Thus we can denote the roots $\pm i\kappa ^{\pm}(x',\xi ',\mu )$, where $\kappa ^\pm$ have positive real part; they depend smoothly on $(x',\xi ',\mu )$ for $(\xi ',\mu )\ne (0,0)$. They are homogeneous of degree 1 in $(\xi ',\mu )$ and bounded away from $\Bbb R$ for $|(\xi ',\mu
)|=1$, $\mu \in \Gamma \cup \{0\}$, so in fact they take values in a sector $\{|\arg z|\le \frac \pi 2-\delta \}$ for some $\delta >0$.
We recall from (with some small precisions):
The symbol of $Q_\lambda $ has the following form in local coordinates, for $\mu \in \Gamma $: $$\aligned
q(x,\xi ,\mu )&\sim \sum_{l\in\Bbb N}q_{-2-l}(x,\xi ,\mu ),\text{ with}\\
q_{-2}(x,\xi ,\mu )&=(p_{1,2}(x,\xi )+\mu ^2)^{-1}I,\\
q_{-2-J}(x,\xi ,\mu )&=\sum_{J/2+1\le m\le 2J+1}\frac{r_{J,m}(x,\xi
)}{(p_{1,2}(x,\xi )+\mu ^2)^{m}},\text{ for }J\ge 0;\endaligned\tag A.5$$ here the $r_{J,m}$ are $(\operatorname {dim} E\times \operatorname{dim} E)$-matrices of homogeneous polynomials in $\xi $ of degree $2m-2-J$ with smooth coefficients, and the remainders $q^{\prime}_{-2-M}=q-\sum_{0\le
J<M}q_{-2-J}$ satisfy estimates for all indices $\partial _{x}^\beta \partial _\xi
^\alpha
\partial _\mu ^kq^{\prime}_{-2-M}=O(\ang{\xi ',\mu
}^{-2-M-|\alpha |-k})$, for $|\mu |\ge 1$, $\mu $ in closed subsectors of $\Gamma $.
Concerning $q(x',0,\xi )$, we have: Writing $$\aligned
p_{1,2}(x',0,\xi )+\mu ^2&=a(x')(\xi _n-i\kappa ^+(x',\xi ',\mu ))
(\xi _n+i\kappa ^-(x',\xi ',\mu ))\\
&=
a(x')(\kappa ^+(x',\xi ',\mu )+i\xi _n)
(\kappa ^-(x',\xi ',\mu )-i\xi _n),\endaligned\tag A.6$$ we can decompose each term in simple fractions (at $x_n=0$): $$\aligned
\frac{r_{J,m}(x',0,\xi )}{(p_{1,2}(x',\xi )+\mu
^2)^{m}}&=h^+\frac{r_{J,m}}{(p_{1,2}+\mu
^2)^{m}}+h^-\frac{r_{J,m}}{(p_{1,2}+\mu ^2)^{m}},\\
h^\pm \frac{r_{J,m}}{(p_{1,2}+\mu ^2)^{m}}&=
\sum_{1\le j\le m}\frac{r^\pm_{J,m,j}(x',\xi ',\mu
)}{(\kappa ^\pm(x',\xi ',\mu )\pm i\xi
_n)^{j}},\endaligned\tag A.7$$ where the $r_{J,m,j}^\pm(x',\xi ',\mu )$ are strongly homogeneous of degree $j-J-2$ in $(\xi ',\mu )$. This gives a decomposition of $q(x',0,\xi ',\mu )$: $$q=h^+q+h^-q=q^{+}+q^{-},\quad
q^{\pm}(x',\xi ,\mu )\sim\sum_{
J\ge 0}q^{ \pm}_{-2-J}(x',\xi ,\mu ),$$with terms as in [(A.7)]{}.
The structure of the normal derivatives is similar: We have $$\partial _{x_n}^lq_{-2-J}(x,\xi ,\mu )
=\sum_{1+J/2\le m\le 2J+1+l}
\frac{r^{l}_{J,m}(x,\xi)}{(p_{1,2}(x,\xi )+\mu ^2)^{m}},$$ for all $l$, with homogeneous polynomials $r^{l}_{J,m}(x,\xi )$ of degree $2m-2-J$ in $\xi $. At $x_n=0$ we can decompose as before, obtaining $$\aligned
\quad \partial _{x_n}^lq=\partial _{x_n}^lq^++\partial _{x_n}^lq^{-},&
\quad
\partial _{x_n}^lq^{\pm}(x',\xi ,\mu )
\sim\sum_{J\ge 0}
\partial _{x_n}^lq^{\pm}_{-2-J}(x',\xi ,\mu ),\\
\partial _{x_n}^lq_{-2-J}(x',\xi ,\mu )
&=\partial _{x_n}^lq^{+}_{-2-J}(x',\xi ,\mu
)+\partial _{x_n}^lq^{-}_{-2-J}(x',\xi ,\mu ),\\
\partial _{x_n}^lq^{\pm}_{-2-J}(x',\xi ,\mu )
&=\sum_{1\le j\le 2J+1+l}
\frac{r^{l,\pm}_{J,j}(x',\xi ',\mu )}
{(\kappa ^\pm(x',\xi ',\mu )\pm i\xi_n)^{j}},
\endaligned\tag A.8$$ where the numerators $r^{l,\pm}_{J,j}(x',\xi ',\mu )$ are strongly homogeneous of degree $j-J-2$.
It is useful to observe that the $r^{\pm}_{J,m,j}$ as well as $\kappa ^{\pm}$ are in fact functions of $(x',\xi
',\lambda )$, $\lambda =-\mu ^2$, strongly quasihomogeneous in $(\xi
',\lambda )$ with weight $(1,2)$ in the following sense: We say that $r(x',\xi ',\lambda )$ is $(1,2)$-homogeneous in $(\xi ',\lambda )$ of degree $d$, when $$r(x',t\xi ',t^2\lambda )=t^{d}r(x',\xi ',\lambda );\tag A.9$$ it is strongly so if (A.9) holds for $|\xi '|+|\lambda |\ge \varepsilon $, weakly so if it holds for $|\xi '|\ge \varepsilon $. Then since $r(x',\xi ',\lambda )=t^{-d}r(x',t\xi ',t^2\lambda )$, it is readily checked that $\partial _\lambda ^Nr$ is $(1,2)$-homogeneous of degree $d-2N$. Thus for the symbols that depend on $\mu $ through $\lambda $ in this way, differentiation with respect to $\lambda $ lowers the homogeneity degree in $(\xi ',\mu )$ by two steps, preserving strong homogeneity. The estimates of $\lambda $-derivatives of the remainders $q^{\prime}_{-2-M}$ likewise improve by two steps for each derivative.
These considerations of $\lambda $-derivatives, playing on the strong quasi-homogeneity of symbols coming from $R_\lambda $, will replace the calculations for higher powers of $R_\lambda $ used in [@GSc], cf. (2.12). (The calculus in [@GH] is set up to handle the anisotropic homogeneity in terms of $\xi '$ and $\lambda $ and could give further information on higher terms in the full trace expansions, but for the discussion of the leading nonlocal term we carry out here, the calculus of [@GS1] will suffice.) For convenience, we recall explicitly the structure of the formulas for symbols connected with higher powers of $Q_\lambda $, denoting the symbol of $Q_\lambda ^N=\frac{\partial _\lambda^{N-1}}{(N-1)!}Q_\lambda $ by $q^{\circ N}$: $$\aligned
q^{\circ N}(x,\xi ,\mu )
&\sim \sum_{l\in\Bbb N}q^{\circ N}_{-2N-l}(x,\xi ,\mu ),
\text{ with }q^{\circ N}_{-2N}=(p_{1,2}+\mu ^2)^{-N},\\
q^{\circ N}_{-2N-J}(x,\xi ,\mu )&=\sum_{J/2+N\le m\le 2J+N}
\frac{r_{N,J,m}(x,\xi)}{(p_{1,2}(x,\xi )+\mu ^2)^{m}},
\text{ for }J\ge 0,
\endaligned\tag A.10$$ where the $r_{N,J,m}$ are homogeneous polynomials in $\xi $ of degree $2m-2N-J$ with smooth coefficients. Moreover for $x=(x',0)$, $$\aligned
q^{\circ N}&=h^+q^{\circ N}+h^-q^{\circ N}=q^{\circ N,+}+q^{\circ N,-},\\
q^{\circ N,\pm}(x',\xi ,\mu )&\sim\sum_{
J\ge 0}q^{\circ N, \pm}_{-2-J}(x',\xi ,\mu ),\\
q^{\circ N, \pm}_{-2-J}(x',\xi ,\mu )&=\sum_{1\le j\le
2J+N}\frac{r^\pm_{N,J,j}(x',\xi ',\mu
)}{(\kappa ^\pm(x',\xi ',\mu )\pm i\xi
_n)^{j}};
\endaligned\tag
A.11$$ here $r_{N,J,j}^\pm(x',\xi ',\mu )$ are strongly homogeneous of degree $j-J-2N$ in $(\xi ',\mu )$.
For $\partial_{x_n}^l q^{\circ N}(x,\xi ,\mu )$ and $\partial_{x_n}^l q^{\circ N,\pm}(x',\xi ,\mu )$, we obtain corresponding results: Their structure is as in (A.10) and (A.11), except that now the summation will be over the sets $J/2+N\le m\le 2J+N+l$ and $1\le j\le 2J+N+l$.
The following symbols derived from the principal symbol of $q$ at $x_n=0$ played an important role in : $$\aligned
\alpha ^{(1)}(x',\xi ',\mu )&=[h^+q_{-2}(x',0,\xi ,\mu )]_{\xi _n=-i\sigma
}+
[h^-q_{-2}(x',0,\xi ,\mu )]_{\xi _n=i\sigma }\\
&=q^+_{-2}(x',\xi ',-i\sigma ,\mu )+q^-_{-2}(x',\xi ',i\sigma ,\mu
)\\
&=\frac 1{(\kappa ^++\kappa ^-)(\kappa ^++\sigma )}+\frac 1{(\kappa
^++\kappa ^-)(\kappa ^-+\sigma )},\\
\alpha ^{(N)}(x',\xi ',\mu )&=[h^+q^{\circ N}_{-2N}(x',0,\xi ,\mu )]_{\xi
_n=-i\sigma }+
[h^-q^{\circ N}_{-2N}(x',0,\xi ,\mu )]_{\xi _n=i\sigma }\\
&=\tfrac{\partial _\lambda ^{N-1}}{(N-1)!}\alpha ^{(1)}(x',\xi ',\mu );
\endaligned\tag A.12$$ here $\sigma =[\xi ']$, and $\alpha ^{(N)}$ is weakly polyhomogeneous in $S^{-2N,0,0}\cap S^{0,-2N,0}$ (in fact in $S^{0,0,-2N}$, see below). The crucial information established in [@GSc; @Lemma; @5.5] was that the coefficient of $\mu
^{-2N}$ in the expansion in powers of $\mu $ (as in (3.4)) is 1: $$\alpha ^{(N)}(x',\xi ',\mu )=\mu ^{-2N}+
\alpha ^{(N)}_1(x',\xi ',\mu )\text{ with } \alpha ^{(N)}_1(x',\xi ',\mu
)\in S^{1,-2N-1,0};
\tag A.13$$ the remainder $\alpha _1^{(N)}$ is also in $S^{0,-2N,0}$ since $\alpha ^{(N)}$ and $\mu ^{-2N}$ are so.
The symbols $(\kappa ^\pm+\sigma )^{-1}$ are weakly polyhomogeneous, belong to $S^{0,0,-1}$ and have $N$’th $\lambda $-derivatives in $S^{0,0,-1-2N}$. For each $N$, $\alpha ^{(N)}$ is weakly polyhomogeneous lying in $S^{0,0,-2N}$.
First note that $\kappa ^\pm$ and $(\kappa ^\pm)^{-1}$ are strongly polyhomogeneous in $(\xi ',\mu )$ of degree $1$ resp. $-1$, so they lie in $S^{0,0,1}$ resp. $S^{0,0,-1}$. As noted above, they are strongly $(1,2)$-homogeneous in $(\xi ',\lambda )$ of degree 1 resp. $-1$; hence the $N$’th $\lambda $-derivatives lie in $S^{0,0,1-2N}$ resp. $S^{0,0,-1-2N}$. Similarly, since $\kappa ^+$ and $\kappa ^-$ lie in a proper subsector of $\{z\in\Bbb C\mid \operatorname{Re}z>0\}$, $|\kappa ^++\kappa ^-|\ge \operatorname{Re}(\kappa ^++\kappa ^-)\ge
c_0|(\xi ',\mu )|$ with $c_0>0$, so also $(\kappa ^++\kappa ^-)^{-1}$ belongs to $S^{0,0,-1}$, with $N$’th $\lambda $-derivatives in $S^{0,0,-1-2N}$.
It was observed in [@GSc] that $(\kappa ^\pm+\sigma )^{-1}\in
S^{-1,0}\cap
S^{0,-1}$; we now recall the proof showing how it also assures that $(\kappa ^\pm+\sigma )^{-1}\in S^{0,0,-1}$:
Let $\kappa $ stand for $\kappa ^+$ or $\kappa ^-$. Write $(\kappa +\sigma
)^{-1}=(\kappa )^{-1}(1+\sigma /\kappa )^{-1}$. We know already that $(\kappa
)^{-1}\in S^{0,0,-1}$, so it remains to show that $(1+\sigma /\kappa
)^{-1}$ is in $S^{0,0,0}$. Since $\sigma \in S^{1,0,0}$ and $(\kappa )^{-1}\in
S^{0,0,-1}\subset S^{-1,0,0}$, $\sigma /\kappa $ is in $S^{0,0,0}$, and so is $1+\sigma
/\kappa $. Moreover, it is bounded in $(\xi ',\mu )$ (for $\mu \in\Gamma $, $|\mu |\ge 1$). Since $\operatorname{Re}\kappa >0$, $\sigma <\sigma +\operatorname{Re}\kappa \le |\sigma +\kappa |$, so also the inverse $(1+\sigma
/\kappa )^{-1}=\kappa /(\kappa +\sigma )=1-\sigma /(\sigma
+\kappa )$ is bounded. Then (just the beginning of the proof) shows that the inverse $(1+\sigma
/\kappa )^{-1}$ does belong to $S^{0,0,0}$.
For the $\lambda $-derivatives, we observe that $\partial _\lambda (\kappa +\sigma )^{-1}=
- (\kappa +\sigma )^{-2}\partial _\lambda \kappa \in
S^{0,0,-3}$ by the composition rules, and hence by successive application, $\partial _\lambda ^k(\kappa +\sigma )^{-1}\in S^{0,0,-1-2k}$ for all $k$.
Using these informations, it is now seen from the form of $\alpha
^{(1)}$ in (A.12) that it lies in $S^{0,0,-2}$ with $k$’th $\lambda $-derivatives in $S^{0,0,-2-2k}$, so $\alpha ^{(N)}\in S^{0,0,-2N}$.
Besides $G^{(N)}_\lambda $ (cf. (2.11)), we also need to consider the s.g.o.s $G^{\pm}(Q^N)$, which arise from compositions such as $$P_+Q^N_{\lambda ,+}=(PQ^N_\lambda )_+-G^+(P)G^-(Q^N_\lambda ).\tag A.14$$ Their symbols have the following structure:
The operators $G^{(N)}_\lambda $ and $G^{\pm}(Q^N)$ have symbols of the form$$\aligned
g(x',\xi ,\eta _n,\mu )&\sim
\sum_{J\in\Bbb N}g_{-2N-1-J}(x',\xi ,\eta _n,\mu ),\text{ with }\\
g_{-2N-1-J}(x',\xi ,\eta _n,\mu )&=\sum_{ \Sb j\ge 1,\,j'\ge 1\\
j+j'\le 2J+N+1\endSb}\tfrac{s_{J,j,j',N}(x',\xi ',\mu )}
{(\kappa +i\xi
_n)^{j}(\kappa '-i\eta
_n)^{j'}};
\endaligned\tag A.15$$ here $(\kappa ,\kappa ')$ equals $(\kappa ^+,\kappa ^-)$ for $G^{(N)}_\lambda $, $(\kappa ^+,\kappa ^+)$ for $G^{+}(Q^N)$, and $(\kappa ^-,\kappa ^-)$ for $G^{-}(Q^N)$. The numerators $s_{J,j,j',N}(x',\xi ',\mu )$ are strongly homogeneous in $(\xi ',\mu )$ of degree $j+j'-J-2N-1$. They are in fact functions of $(x',\xi
',\lambda )$, $\lambda =-\mu ^2$, strongly $(1,2)$-homogeneous in $(\xi
',\lambda )$ of the indicated degrees, such that differentiation with respect to $\lambda $ gives a strongly $(1,2)$-homogeneous symbol of $2$ steps lower degree.
For the remainders $g'_{-2N-1-M}=g-\sum_{0\le
J<M}g_{-2N-1-J}$, the sup-norms in $\xi _n$ resp. $(\xi
_n,\eta _n)$ are bounded by $\ang{\xi ',\mu }$ in powers $-2N-1-M$, decreasing by $|\alpha |$ for differentiations in $(\xi ,\eta _n)$ of order $\alpha $, and by $2$ for each differentiation in $\lambda $ (no change of the power for differentiations in $x'$).
The formulas for the symbols of $G^{(N)}_\lambda $ and $G^\pm(Q^N_\lambda )$ come from . The remainder estimates hold since the symbols are strongly polyhomogeneous in $(\xi ,\eta _n,\mu )$ so that standard estimates hold when $|\mu |$ is considered as an extra cotangent variable (on each ray). The strong (1,2)-polyhomogeneity $(\xi
,\eta _n,\lambda )$ assures the statements on $\lambda$-derivatives.
Appendix B. Formulas for compositions with Laguerre functions
Recall the formulas for the (Fourier transformed) Laguerre functions we use in expansions of parameter-independent operators: $$\hat\varphi '_k(\xi _n,\sigma )=\frac{(\sigma -i\xi _n)^k}{(\sigma
+i\xi _n)^{k+1}},\quad\hat\varphi _k(\xi _n,\sigma )
=(2\sigma )^{\frac12}\frac{(\sigma -i\xi _n)^k}{(\sigma
+i\xi _n)^{k+1}};\tag B.1$$ A priori, $\sigma$ here might be any positive number, but we will take $\sigma=[\xi']$. The $\hat\varphi _k$ are the [*normalized*]{} Laguerre functions, which for $k\in\Bbb Z$ form an orthonormal basis of $L_2(\Bbb R)$. Differentiations in $\xi '$ and $\xi _n$ follow the rules$$\aligned
\partial_{\xi_j}\hat\varphi_k(\xi_n,\sigma)
&=\big(k\hat\varphi_{k-1}-
\hat\varphi_{k}
-(k+1)
\hat\varphi_{k+1}\big)(2\sigma)^{-1}\partial_{\xi_j}\sigma ,\quad
j<n,\\
\partial_{\xi_n}\hat\varphi_k(\xi_n,\sigma)
&=-i\big(k\hat\varphi_{k-1}+
(2k+1)\hat\varphi_{k}
+(k+1)
\hat\varphi_{k+1}\big)(2\sigma)^{-1} ;
\endaligned
\tag B.2$$ here $\partial _{\xi _j}\sigma =\xi _j\sigma ^{-1}$ for $|\xi '|\ge 1$.
We have from [@GSc] the following formulas for $\xi _n$-compositions of Laguerre funtions and rational functions involving $\kappa ^\pm$:
For all $m\ge 0$ and $j\ge 1$: $$\int \frac{(\sigma \pm i\xi _n)^m}{(\sigma \mp i\xi _n)^{m+1}}\frac{1}
{(\kappa ^\pm(x',\xi ',\mu )\pm i\xi
_n)^{j}}\,\d\xi _n=\sum_{m'\ge 0,\, |m'-m|<j}(\kappa^\pm) ^{1-j}
a^\pm_{jm'}\frac{(\sigma -\kappa ^\pm)^{m'}}{(\sigma
+\kappa ^\pm
)^{m'+1}},
\tag B.3$$ with universal constants $a^\pm_{jm'}$ that are $O(m^j)$ for fixed $j$. The resulting expressions are weakly polyhomogeneous $\psi $do symbols belonging to $S^{0,0,-j}(\Bbb R^{n-1}\times\Bbb R^{n-1},\Gamma )$, with $\partial
_\lambda ^N$ of the symbols lying in $S^{0,0,-j-2N}(\Bbb
R^{n-1}\times\Bbb R^{n-1},\Gamma )$.
The formulas are shown in [@GSc; @Lemma; @3.2]. The statements on symbol classes follow since $\sigma =[\xi ']\in S^1\subset
S^{1,0,0}$, $\kappa ^{\pm}\in S^{0,0,1}$ with $N$’th $\lambda
$-derivatives in $S^{0,0,1-2N}$, and $(\sigma
+\kappa ^{\pm})^{-1}\in S^{0,0,-1}$ with $N$’th $\lambda
$-derivatives in $S^{0,0,-1-2N}$, by Lemma A.3. For the derivatives of the composed expressions one can use the formulas [@GSc; @(3.21)] for the derivatives of $(\sigma -\kappa
^\pm)^m(\sigma +\kappa ^\pm)^{-m}$.
These formulas enter in calculations of compositions such as $GG^{(N)}_\lambda $ and $GG^{\pm}(Q^N_\lambda )$, where the symbol of $G$ is Laguerre expanded and the rational structure of the symbols of $G^{(N)}_\lambda
$ and $G^\pm(Q^N_\lambda )$ (A.15) is used.
The symbol $$s_{j,j',m}=
\int \frac{(\sigma -i\xi _n)^m}{(\sigma
+i\xi _n)^{m+1}}\frac
{1}{(\kappa ^++ i\xi
_n)^{j}(\kappa ^--
i\xi
_n)^{j'}}\,\d\xi _n$$ satisfies for $m\ge 0$, $j$ and $j'\ge 1$: $$s_{j,j',m}=\sum_{\Sb |m-m'|\le j''< j'\\ m'\ge
0\endSb}b_{jj'j''m'}(\kappa ^-)^{-j''}\frac{(\sigma -\kappa
^-)^{m'}}{(\sigma +\kappa ^-)^{m'+1}}(\kappa ^++\kappa ^-)^{-j-j'+1+j''},
\tag B.4$$ where the $b_{jj'j''m'}$ are universal constants that are $O(m^{j'})$ for fixed $j,j'$. This is a weakly polyhomogeneous symbol in $S^{0,0,-j-j'}$, with $N$’th $\lambda $-derivatives in $S^{0,0,-j-j'-2N}$, for all $N$. There is a similar formula for $m\le 0$, with $j$ and $j'$ interchanged, $\kappa ^+$ and $\kappa ^-$ interchanged.
The formulas were shown in [@GSc; @Lemma; @4.2], and the symbols are analyzed as in the preceding proof.
These formulas enter in calculations of compositions such as $P_+G^{(N)}_\lambda $ and $P_+G^\pm(Q^N_\lambda )$, where the symbol of $P$ is Laguerre expanded and the rational structure of the s.g.o. is used.
Composition formulas involving $\psi $do’s on the interior contain $h^+_{\xi _n}$- and $h^-_{\xi _n}$-projections. We recall from [@GSc; @(3.9); @(4.8)] that the projections can be removed in certain integrals, e.g.: When $\tilde q$ is a rational function of $\xi
_n$ of the form $r(x',\xi ',\mu )(\kappa ^{\pm}\pm i\xi _n)^{-j}$, then$$\multline
\int \bar{\hat\varphi }_k(\xi _n,\sigma ) h^+_{\xi _n}[\tilde
q(x',\xi ,\mu )\hat\varphi _l(\xi _n,\sigma )]\,\d \xi _n
=\int \bar{\hat\varphi }_k(\xi _n,\sigma ) \tilde
q(x',\xi ,\mu )\hat\varphi _l(\xi _n,\sigma )\,\d \xi _n\\
=\int h^-_{\xi _n}[\bar{\hat\varphi }_k(\xi _n,\sigma ) \tilde
q(x',\xi ,\mu )]\hat\varphi _l(\xi _n,\sigma )\,\d \xi _n.
\endmultline\tag B.5$$ To see this, note that the integrand in all three expressions is $O(\xi _n^{-2})$ for $|\xi _n|\to \infty $ in $\Bbb C$ and meromorphic with no real poles, so that the integral can be replaced by the integral over a large contour in $\Bbb C_+$ (the “plus-integral”) or the integral over a large contour in $\Bbb C_-$. The first equality holds since $\bar{\hat \varphi }_kh^- [\tilde
q\hat\varphi _l]$ is meromorphic with no poles in $\Bbb
C_+$ and is $O(\xi _n^{-2})$ for $|\xi _n|\to \infty $ in $\Bbb C$, hence contributes with 0. The second equality holds since $h^+[\bar{\hat \varphi }_k\tilde
q]\hat\varphi _l$ is meromorphic with no poles in $\Bbb C_-$, and is $O(\xi _n^{-2})$ for $|\xi _n|\to \infty $ in $\Bbb C$, hence contributes with 0.
The above formulas were used in to show that the trace expansions of terms where $G$ or $P_+$ is composed with one of the parameter-dependent singular Green operators [*do not*]{} contribute to the residue coefficient $\tilde c_0'$ in (1.10). It was in fact shown that they give expansions where the summation over $k$ as in (1.10) starts with $k\ge 1$, so they do not contribute to the nonlocal coefficient $\tilde c''_0$ either. This is important for our present study and will therefore be formulated explicitly:
When $G$ is of class $0$ and order $\nu
\in\Bbb R$, or $P$ is of order $\nu \in\Bbb Z$, the operators $GG^{(N)}_\lambda $, $GG^{\pm}(Q^N_\lambda )$, $P_+G^{(N)}_\lambda $ and $P_+G^{\pm}(Q^N_\lambda )$ have trace expansions of the form $$\sum_{j\ge
1}\!a_{j}(-\lambda)
^{\frac{n+\nu - j}{2}-N}
+\sum_{k\ge
1}(a'_k\log (-\lambda )+a''_k)(-\lambda )
^{-\frac k2-N}.\tag B.6$$ For the compositions with $G$ in front, this is essentially the content of , see in particular (3.20) there. We have replaced the index $l$ there by $j=l+1$ in the first sum, $k=l+1$ in the second sum, to facilitate the comparison with (1.10); recall also that $\mu =(-\lambda )^{\frac12}$. The arguments there extend immediately to noninteger $\nu $.
For the compositions with $P_+$ in front, the statement is covered by the calculations in , see in particular (4.3) there (which contains neither nonlocal nor logarithmic terms, since $p'_{(l)}$ is a differential operator) and (4.14) there.
In each case, the result is found by showing that in local coordinates, the symbol of $\tr_n$ of the operator contains so many negative powers of $\kappa ^{\pm}$ and $\kappa ^{\pm}+\sigma $ that it is in $S^{\nu +1,-2N-1,0}\cap S^{\nu -2N,0,0}$, so that $d=-2N-1$ in (3.6).
When $\nu \notin\Bbb Z$, the log-coefficients $a'_k$ vanish in (B.6), since the degrees of the homogeneous symbols are noninteger.
The conclusions of the proposition hold also if the symbol $q$ of $Q_\lambda $ is replaced by one of its $(x,\xi )$-derivatives, since they have a similar structure.
The next lemma is used in calculations of compositions of the type $GQ^N_{\lambda ,+}$.
Let $$s^{\pm}_{j,l,m}(y',\xi ',\mu )=\int \frac
{(\sigma - i\xi
_n)^{l}}{(\sigma + i\xi
_n)^{l+1}}\frac{(\sigma +
i\xi
_n)^{m}}{(\sigma -
i\xi
_n)^{m+1}}\frac1{(\kappa
^{\pm}\pm i\xi _n)^j}\,\d\xi _n.
\tag B.7$$ One has for $l,m\in\Bbb Z$, $j\ge 1$: $$\alignedat2
&\text{For $m<l$, }&&s^+_{j,l,m}=0,\quad
s^-_{j,l,m}=\sum_{ |l-m-1-m'|<j,\, m'\ge
0}(\kappa^-) ^{1-j}b'_{jm'}\frac{(\sigma -\kappa ^-)^{m'}}{(\sigma +\kappa
^-
)^{m'+2}},\\
&\text{For $m=l$, }&&s^+_{j,l,l}=\frac1{(\kappa ^++\sigma )^j 2\sigma
},\quad s^-_{j,l,l}=\frac1{(\kappa ^-+\sigma )^j 2\sigma },\\
&\text{For $m>l$, }&&s^+_{j,l,m}=\sum_{ |m-l-1-m'|<j,\, m'\ge
0}(\kappa ^+) ^{1-j}b_{jm'}\frac{(\sigma -\kappa
^+)^{m'}}{(\sigma +\kappa ^+
)^{m'+2}},\quad s^-_{j,l,m}=0,
\endalignedat\tag B.8$$ with $b_{jm'}$ and $b'_{jm'}$ being $O(l^jm^j)$ for fixed $j$.
When $m\ne l$, the resulting symbols are in $S^{0,0,-j-1}$ having $N$’th $\lambda $-derivatives in $S^{0,0,-j-1-2N}$; for $m=l$, they are in $S^{-1,0,-j}$, with $N$’th $\lambda $-derivatives in $S^{1,0,-j-2N}$.
It follows (cf(A.12)) that$$\tr_n (\hat\varphi _l(\xi _n,\sigma )h^+(q^{\circ N}_{-2N}(x',\xi
',\eta _n,\mu
)\bar{\hat\varphi }_m(\eta _n,\sigma )))=\cases \alpha
^{(N)}(x',\xi ',\mu )&\text{ if }l=m,\\
s_{lm}(x',\xi ',\mu )&\text{ if }l\ne m,\endcases\tag B.9$$ where $s_{lm}\in S^{1,0,-2N-1}$, the symbol seminorms being polynomially bounded in $l$ and $m$.
The formulas in (B.8) were shown in [@GSc; @Lemma; @5.2], and the symbols are analyzed as in Lemma B.1.
Formula (B.9) is deduced from (B.7)–(B.8) using (B.5) and noting that (B.7) contains non-normalized Laguerre functions $\hat\varphi '_l$ and $\bar{\hat\varphi }'_m$ so that we get an extra factor $2\sigma $ from the normalized Laguerre functions. The statements for $l\ne m$ follow straightforwardly from the descriptions in (B.8). The statements for $l=m$ were proved in ; they can be verified very simply by doing the calculation for $N=1$ and passing via $\lambda $-derivatives to the general case.
The important point in this lemma is that a symbol contributing to $\tilde c'_0$ and $\tilde c''_0$ does appear, but with a special form that allows further clarification (using the information (A.13)).
We summarize some results from [@GSc] on compositions of the form $GQ^N_{\lambda,+} $ in the following statement:
Consider $GQ^N_{\lambda
,+}$ in a localization to $\Bbb R^n_+$. Here $$\tr_n(GQ^N_{\lambda ,+})=\widetilde S_0+\widetilde S_1\text{ with }
\widetilde
S_0=\op'(\tr_ng(x',\xi
')\alpha ^{(N)}(x',\xi ',\mu )),
\tag B.10$$ where $\widetilde S_0$ and $\widetilde S_1$ are $\psi $do’s on $\Bbb R^{n-1}$ with symbols in $S^{\nu ,-2N,0}\cap S^{\nu -2N,0,0}$ resp. $S^{\nu +1,-2N-1,0}\cap S^{\nu -2N,0,0}$. The traces $\Tr_{\Bbb R^{n-1}}(\widetilde S_0)$ and $\Tr_{\Bbb R^{n-1}}(\widetilde S_1)$ have expansions $$\sum_{j\ge
1}\!a_{j}(-\lambda)
^{\frac{n+\nu - j}{2}-N}
+\sum_{k\ge
0}(a'_k\log (-\lambda )+a''_k)(-\lambda )
^{-\frac k2-N},\tag B.11$$ where the sum over $k$ starts at $k=1$ for $\widetilde S_1$, and the value of $a'_0$ can be determined more precisely for $\widetilde S_0$.
This is proved in , see in particular Prop. 5.3, 5.4, and Sect. 5.b there.
We have in fact in view of Lemma A.3 that $\widetilde S_0\in
S^{\nu ,0,-2N}$, and also the information on $\widetilde S_1$ can be upgraded to $S^{\nu +1,0,-2N-1}$ by a closer analysis (as in Section 3 in this paper, using methods from Proposition 3.8 to handle the case where $q$ depends on $x_n$).
Let us introduce a notation for the Poisson and trace operators on $\rnp$ (Laguerre operators) defined from Laguerre functions: $$\Phi _j=\opk(\hat{\varphi}_j(\xi_n,\sigma)),\quad \Phi ^*
_k=\opt(\bar{\hat{\varphi}}_k(\xi_n,\sigma)).\tag B.12$$ Here $\Phi _j$ maps $L_2(\Bbb R^{n-1})$ continuously (in fact isometrically) into $L_2(\rnp)$, and its adjoint is $\operatorname{OPT}(\bar{\hat\varphi }_j)$. Moreover, because of the orthonormality of the $\hat\varphi _j$, $$\Phi ^*_j\Phi
_k=\delta _{jk}I,\tag B.13$$ where $I$ is the identity operator on functions on $\Bbb R^{n-1}$.
A singular Green operator $G$ on $\rnp$ of order $\nu$ and class 0 can be written in the form $$G=\sum_{j,k\in\Bbb N} \Phi _j C_{jk}
\Phi ^*_k
,\quad C_{jk}=\op'(c_{jk}(x',\xi')) , \tag B.14$$ with a rapidly decreasing sequence $(c_{jk})_{j,k\in\Bbb N}$ in $S^{\nu}$.
The symbol $g$ of $G$ has a Laguerre expansion: $$g(x',\xi',\xi_n,\eta_n) = \sum_{j,k\in\Bbb N}
d_{jk}(x',\xi'){\hat\varphi}_j(\xi_n,\sigma)
\bar{{\hat\varphi}}_k(\eta_n,\sigma),
\tag B.15$$ with $(d_{jk})_{j,k\in\Bbb N}$ rapidly decreasing in $S^{\nu}$, i.e., the relevant symbol seminorms on $\ang k^{N}\ang
j^{N'}d_{jk}$ are bounded in $j,k$ for any $N,N'\in\Bbb N$. Then $$G=\sum_{k\in\Bbb N} \opk \big(\sum_{j\in\Bbb N}
d_{jk}(x',\xi'){\hat\varphi}_j(\xi_n,\sigma)\big)\circ
\opt\big(\bar{\hat\varphi}_k(\eta_n,\sigma)\big).$$ Each Poisson operator $\opk(\sum_{j\in\Bbb N}
d_{jk}(x',\xi'){\hat\varphi}_j(\xi_n,\sigma))$ has a symbol in $y'$-form, $l_k(y',\xi)$. Since the $d_{jk}$ are rapidly decreasing in $S^{\nu}$, the sequence $(l_k)$ is rapidly decreasing in the topology of Poisson symbols of order $\nu+\frac12$. We now expand $l_k$ in a Laguerre series: $$l_k(y',\xi) = \sum_{j\in\Bbb N} c_{jk}(y',\xi')\hat\varphi_j(\xi_n,\sigma)
,$$ and conclude from [@G2; @Lemma; @2.2.1] that $(c_{jk})$ is rapidly decreasing in $S^{\nu}$. Since $$\opk(l_k) = \sum_{j\in\Bbb N}
\opk(\hat\varphi_j(\xi_n,\sigma))\op'(c_{jk}(x',\xi')),$$ we obtain $$G=\sum_{j,k\in\Bbb
N}\opk\big(\hat\varphi_j(\xi_n,\sigma)\big)\op'\big(c_{jk}(x',\xi')\big)
\opt\big(\bar{\hat\varphi}_k(\eta_n,\sigma)\big),$$ which shows the assertion.
The last part of the proof shows the following useful result:
A Poisson operator $K$ of order $\nu+\frac12$ on $\overline{\Bbb R}^{n}_+$ can be written $K=\sum_{j\ge 0}\Phi_jC_j$ with a rapidly decreasing sequence $(C_j)$ of $\psi$do’s with symbols in $S^\nu$.
Elmar Schrohe was partially supported by the European Research and Training Network ‘Geometric Analysis’ (Contract HPRN-CT-1999-00118).
\[\] L. Boutet de MonvelBoundary problems for pseudo-differential operatorsActa Math.1261971 11–51
\[\]B. V. Fedosov, F. Golse, E. Leichtnam, E. Schrohe The noncommutative residue for manifolds with boundary J. Funct. Anal. 1421996 1–31
\[\] G. Grubb Singular Green operators and their spectral asymptotics Duke Math. J. 51 1984 477–528
\[\] Functional calculus of pseudodifferential boundary problems, Progress in Math. vol. 65, Second Edition BirkhäuserBoston 1996 first edition issued 1986
\[\] A weakly polyhomogeneous calculus for pseudodifferential boundary problems J. Funct. Anal. 184 2001 19–76
\[\] A resolvent approach to traces and determinants AMS Contemp. Math. Proc. 366 2005 67–93
\[\] Spectral boundary conditions for generalizations of Laplace and Dirac operators Comm. Math. Phys. 2003 240 243–280
\[\] Ann. Global Anal. Geom. 24 2003 1–51
\[\] G. Grubb and L. Hansen Complex powers of resolvents of pseudodifferential operators 2002 Comm. Part. Diff. Eq. 27 2333–2361
\[\]G. Grubb and E. Schrohe Trace expansions and the noncommutative residue for manifolds with boundary J. Reine Angew. Math. (Crelle’s Journal) 2001 167–207 536
\[\]G. Grubb and R. Seeley Weakly parametric pseudodifferential operators and Atiyah-Patodi-Singer boundary problems Invent. Math. 121 1995 481–529
\[\] Zeta and eta functions for Atiyah-Patodi-Singer operatorsJ. Geom. Anal.1996 6 31–77
\[\]V. Guillemin A new proof of Weyl’s formula on the asymptotic distribution of eigenvalues Adv. Math. 102 1985 184–201
\[\]Le problème de Cauchy et les équations aux dérivées partielles linéaires hyperboliques J. Hadamard Hermann Paris 1932
\[\] M. Kontsevich and S. Vishik Geometry of determinants of elliptic operators Functional Analysis on the Eve of the 21’st Century (Rutgers Conference in honor of I. M. Gelfand 1993), Vol. I S. Gindikin et al. Progr. Math. 131, Birkhäuser Boston 1995 173–197
\[\] M. Lesch On the noncommutative residue for pseudodifferential operators with log-polyhomogeneous symbols Ann. Global Anal. Geom. 1999 17 151–187
\[\] R. Melrose and V. Nistor Homology of pseudodifferential operators I. Manifolds with boundary manuscript, arXiv:funct-an/9606005
\[\] K. Okikiolu The multiplicative anomaly for determinants of elliptic operators Duke Math. J. 79 1995 723–750
\[\] R. T. SeeleyComplex powers of an elliptic operator Amer. Math. Soc. Proc. Symp. Pure Math.101967288–307
\[\] M. Wodzicki Local invariants of spectral asymmetry Invent. Math. 1984 75 143-178
|
---
abstract: 'We present results from new deep HST/ACS photometry of I Zw 18, the most metal-poor blue compact dwarf galaxy in the nearby universe. It has been previously argued that this is a very young system that started forming stars only $\lesssim 500$ Myr ago, but other work has hinted that older ($\gtrsim 1$ Gyr) red giant branch (RGB) stars may exist in this galaxy. Our deeper data indeed reveal evidence for an RGB. Underlying old ($\gtrsim 1$ Gyr) populations are therefore present in even the most metal-poor systems, implying that star formation started at $z \gtrsim 0.1$. The RGB tip (TRGB) magnitude and the properties of Cepheid variables identified from our program indicate that I Zw 18 is farther away ($D = 19.0 \pm 1.8$ Mpc) than previously believed.'
---
Introduction
============
Within the framework of hierarchical formation, dwarf ($M$ $\lesssim$ 10$^9$ M$_{\odot}$) galaxies are the first systems to collapse and start forming stars, supplying the building blocks for the formation of more massive galaxies through merging and accretion. As remnants of this process, present-day dwarfs may have been sites of the earliest star-formation (SF) activity in the universe. However, the most metal-poor ($12 + \log (O/H) \lesssim 7.6$, corresponding to $Z$ $\lesssim$ 1/20 Z$_{\odot}$) dwarf irregular (dIrr) and blue compact dwarf (BCD) galaxies have been repeatedly pointed out as good candidate “primeval” galaxies in the nearby universe, with possible ages less than 100-500 Myr (e.g., Izotov & Thuan 1999). If some of these objects turn out to be young galaxies, their existence would support the view that SF in low-mass systems has been inhibited until the present epoch (e.g., Babul & Rees 1992). The discovery of a nearby primordial galaxy would thus have revolutionary cosmological implications. On the other hand, the lack of such systems would provide strong constraints on the chemical and physical evolution of low-metallicity dIrr/BCDs.
The only direct way to unambiguously infer the evolutionary status of a metal-poor dIrr/BCD galaxy is to resolve it into stars with deep HST observations, and study stellar features in the color-magnitude diagram (CMD). The brightest of these features that contains stars of significant age is the Red Giant Branch (RGB), formed by low-mass stars with ages $\sim 1$-13 Gyr that are burning H in a shell around a He core. In the last 15 years all metal-poor dIrr galaxies in the Local Group and BCDs with $D \lesssim 15$ Mpc have been imaged with HST. Interestingly, an RGB has been detected in all those galaxies for which photometric data exist that go deep enough to reach the RGB tip (TRGB, brightest phase of the RGB) in the CMD (e.g., SBS 1415+437, Aloisi et al. 2005, and references therein).=-2
The only possible exception so far is the BCD I Zw 18. With an oxygen abundance $12 + \log (O/H) = 7.2$, corresponding to $Z = 1/50$ Z$_{\odot}$, I Zw 18 is the “prototype” member of its class, and the most metal-poor galaxy in the nearby universe. Many HST studies have focused on the evolutionary status of I Zw 18. After other groups had already resolved the brightest individual stars in the galaxy, our group was the first to go deep enough to detect asymptotic giant branch (AGB) stars in HST/WFPC2 images with ages of at least several hundreds Myr and possibly up to a few Gyrs (Aloisi et al. 1999). These results were confirmed by Östlin (2000) through deep HST/NICMOS imaging. More recently, Izotov & Thuan (2004) presented new deep HST/ACS observations. Their $I$ vs. $V-I$ CMD shows no sign of an RGB, from which they concluded that the most evolved (AGB) stars are not older than 500 Myr and that I Zw 18 is a bona fide young galaxy. This result was subsequently challenged by Momany et al. (2005) and our group (Tosi et al. 2006) based on a better photometric analysis of the same data. This showed that many red sources do exist at the expected position of an RGB, and that their density in the CMD drops exactly where a TRGB would be expected. However, the small number statistics, the large photometric errors, and the incompleteness at the TRGB, did not allow a conclusive statement about the possible existence of an RGB in I Zw 18. =-2
New HST/ACS Observations
========================
We were awarded 24 additional orbits with ACS over a three-month period starting in October 2005 (GO program 10586, PI Aloisi) to better understand the evolutionary status of I Zw 18. The observations were obtained in 12 different epochs in F606W and F814W to: (1) build a deeper CMD to search for RGB stars; (2) detect and characterize Cepheids at the lowest metallicity available in the local universe; and (3) use both the Cepheids and a possible TRGB detection to determine an accurate distance to I Zw 18. In this paper we will present the preliminary results from the new deep photometry. Preliminary results from the study of the variable stars in this metal-poor galaxy are discussed in Fiorentino et al. (2007, in this volume). Detailed descriptions of the observational program, the implications for the evolutionary state and distance of I Zw 18, and the results on Cepheid variability will be presented in forthcoming papers in the refereed literature.=-2
PSF-fitting photometry was performed on deep images that were obtained by combining the exposures in each filter with MultiDrizzle. After application of CTE and aperture corrections, the count rates were transformed to Johnson-Cousins $V$ and $I$ magnitudes. Values shown and discussed hereafter are corrected for $E(B-V) = 0.032$ mag of Galactic foreground extinction, but not for any extinction intrinsic to I Zw 18. The archival ACS data in F555W and F814W (GO program 9400, PI Thuan) were also re-processed in a similar manner. The two ACS datasets were then combined using several different approaches and rejection schemes. The results discussed here were obtained by demanding that stars should be detected in all the four deep images ($V$ and $I$ for both datasets). At the expense of some depth, this approach has the advantage of minimizing the number of false detections and therefore providing relatively “clean” CMDs.
{width="0.93\hsize"}
{width="0.80\hsize"}
Results and Interpretation
==========================
Fig. 1a shows the resulting $I$ vs. $V-I$ CMD of I Zw 18. The CMD shows faint red stars exactly at the position where an RGB would be expected (see the Padua isochrones overplotted in the figure). Figure 2 shows the luminosity function (LF) of the red stars. It shows a sharp drop towards brighter magnitudes, exactly as would be expected from a TRGB. The magnitude of the discontinuity, $I
= 27.32 \pm 0.10$, implies a distance modulus $m-M = 31.39 \pm 0.20$ (e.g., Bellazzini et al. 2001), i.e., $D = 19.0 \pm 1.8$ Mpc. This assumes that the evolved RGB stars have negligible intrinsic extinction. The TRGB distance is consistent with the preliminary analysis of the first few Cepheid variables identified by our program. Comparison of their Wesenheit relation to the Wesenheit relation inferred from new theoretical Cepheid pulsation models at the metallicity of I Zw 18 yields $m-M =
31.28 \pm 0.26$ (Fiorentino et al. 2007, in this volume). This agreement further supports our interpretation of the LF drop in Figure 2 as a TRGB feature.=-2
The evidence for an RGB in I Zw 18 is further strengthened by comparison to another BCD, SBS 1415+437, observed by us with a similar HST/ACS set-up (Aloisi et al. 2005). This galaxy is not quite as metal poor as I Zw 18 ($12 + \log (O/H) = 7.6$) and is somewhat nearer at $D
\approx 13.6$ Mpc. But taking into account the differences in distance and completeness, the CMDs of these galaxies look very similar. Since SBS 1415+437 has an unmistakable RGB sequence, this suggests that such an RGB sequence exists in I Zw 18 as well. It has previously been suggested that I Zw 18 has no stars with ages $\gtrsim 500$ Myr. It is therefore interesting to ask the question whether there exist Star Formation Histories (SFHs) which do not have stars older than $\sim 1$ Gyr (and by extension, do not have an RGB), yet which can still fit: (1) the observed number of faint red stars in I Zw 18 ($V-I \gtrsim
1.0$, $I \gtrsim 27.3$); and (2) the observed LF drop-off at $I
\approx 27.3$. Preliminary results of SFH modeling that we have performed indicate that such SFHs do not exist.
Conclusions
===========
We have obtained new deep HST/ACS observations of I Zw 18 that provide improved insight into the evolutionary status of this benchmark metal-poor BCD. Our results indicate that this galaxy contains RGB stars, in agreement with findings for other local metal-poor BCDs studied with HST. Underlying old ($\gtrsim 1$ Gyr) populations are therefore present in even the most metal-poor systems, so they must have started forming stars at $z \gtrsim 0.1$. Deeper studies (well below the TRGB) will be needed to pinpoint the exact onset of star formation in these galaxies. We find that at $D = 19.0 \pm 1.8$ Mpc, I Zw 18 is more distant than the values $\sim 15$ Mpc that have often been assumed in previous work. This may explain why it has remained difficult for so long to unambiguously detect or rule out the presence of old resolved (RGB) stars in this object. The data that we are compiling on Cepheid stars in I Zw 18 will be unique for probing the properties of variable stars at metallicities that have never before been probed.
Support for proposals \#9361 and \#10586 was provided by NASA through a grant from STScI, which is operated by AURA, Inc., under NASA contract NAS 5-26555.
1999, *AJ* 118, 302
2005, *ApJ* 631, L45
1992, *MNRAS* 255, 346
2001, ApJ, 556, 635
2000, A&A, 359, 601
1999, *ApJ* 511, 639
2004, *ApJ* 616, 768
2005, *A&A* 439, 111
2000, *ApJ* 535, L99
2006, *Proc. IAU Symp. 235*, eds. F. Combes & J. Palous, in press
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.