id
string | text
string | image
image |
|---|---|---|
230405623/5
|
Figure4: Patient distribution of genetic subtype (a), lesion type (b), SBR grade (c) and pCR status (d) in the dataset.
| |
230405623/2
|
Figure 1: Example breast CDI s images from the Cancer-Net BCa open-source benchmark dataset for each SBR grade.
a part of a global open-source initiative dedicated to accelerating advancement in machine learning to aid clinicians in the fight against cancer.
## Methodology
To construct the Cancer-Net BCa benchmark dataset, we produced CDI s acquisitions for a pre- treatment (T0) patient cohort of 253 patient cases across 10 institutions via the American College of Radiology Imaging Network (ACRIN) 6698/I-SPY2 study [3 – 6]. More specifically, acquisitions were conducted with a four b-value imaging protocol (0 s/mm $^{2}$, 100 s/mm $^{2}$, 600 s/mm $^{2}$, 800 s/mm $^{2}$, 3-direction) on a 1.5 or 3.0 Tesla scanner using a dedicated breast radiofrequency coil. The pixel spacing for the acquisitions ranged from 0.83 mm to 2.08 mm with a median of 1.29 mm, with both slice thickness and spacing between slices ranged from 4.0 to 5.0 mm with a median of 4.0. The native and synthetic signals produced via a signal synthesizer were mixed together to obtain a final CDI s signal [1]. Each patient case is also associated with one of three possible SBR grades: I (Low), II (Intermediate), and III (High). Example images from each SBR type is shown in Fig. 1. The pCR state after neoadjuvant chemotherapy (No pCR/pCR) is also provided for each patient, with an example of each pCR state shown in Fig. 2.
## Results and Discussion
| |
230405623/1
|
# A Multi-Institutional Open-Source Benchmark Dataset for Breast Cancer Clinical Decision Support using Synthetic Correlated Diffusion Imaging Data
Chi-en Amy Tai Hayden Gunraj Alexander Wong Vision and Image Processing Lab, University of Waterloo {amy.tai, hayden.gunraj, alexander.wong} @uwaterloo.ca
Recently, a new form of magnetic resonance imaging (MRI) called synthetic correlated diffusion (CDI s) imaging was introduced and showed considerable promise for clinical decision support for cancers such as prostate cancer when compared to current gold-standard MRI techniques. However, the efficacy for CDI s for other forms of cancers such as breast cancer has not been as well- explored nor have CDI s data been previously made publicly available. Motivated to advance efforts in the development of computer-aided clinical decision support for breast cancer using CDI s, we introduce Cancer-Net BCa, a multi-institutional open-source benchmark dataset of volumetric CDI s imaging data of breast cancer patients. Cancer-Net BCa contains CDI s volumetric images from a pre-treatment cohort of 253 patients across ten institutions, along with detailed annotation meta- data (the lesion type, genetic subtype, longest diameter on the MRI (MRLD), the Scarff-Bloom-Richardson (SBR) grade, and the post-treatment breast cancer pathologic complete response (pCR) to neoadjuvant chemotherapy). We further examine the demographic and tumour diversity of the Cancer-Net BCa dataset to gain deeper insights into potential biases. Cancer-Net BCa is publicly available as a part of a global open-source initiative dedicated to accelerating advancement in machine learning to aid clinicians in the fight against cancer.
## Introduction
A new form of magnetic resonance imaging (MRI) called synthetic correlated diffusion (CDI s) imaging was recently introduced and showed considered promise for clinical decision support for cancers such as prostate cancer when compared to current gold-standard MRI techniques such as T2-weighted (T2w) imaging, diffusion-weighted imaging (DWI), and dynamic contrast-enhanced (DCE) imaging [1]. However, the efficacy for CDI s for other forms of cancer such as breast cancer has not been as well-explored nor have CDI s data been previously made publicly available. The development of computer-aided clinical decision support for breast cancer using CDI s has begun to be analyzed and shown to have superior results compared to other gold-standard imaging for the prediction of breast cancer patient response from neoadjuvant chemotherapy prior to treatment [2]. Motivated to advance efforts in the development of computer-aided clinical decision support for breast cancer using CDI s for diagnosis, prognosis/grading, treatment planning and more, we in- troduce Cancer-Net BCa, a multi-institutional open-source benchmark dataset of volumetric CDI s imaging data of breast cancer patients with detailed annotation metadata for each patient. We further examine the demographic and grade diversity of the Cancer-Net BCa dataset to gain deeper insights into potential biases. The Cancer-Net BCa benchmark dataset has been made publicly available 1 as
1 https://www.kaggle.com/datasets/amytai/cancernet-bca
| |
170404598/16
|
SIMONA NISTOR
where
$$
\stackrel{\mathclap{\tiny\mbox{$N$}}}{\Riem}\left(X_1,X_2\right) =R^N\left(X_1,X_2,X_1,X_2\right).
$$
It is clear that $\langle H,\eta_\alpha\rangle=0$, and then $\trace A_{\eta_\alpha}=2\langle H,\eta_\alpha\rangle=0$, for any $\alpha\in \left\{1,2,\cdots n-3\right\}$. As $A_{\eta_\alpha}$ is symmetric, we note that $\det A_{\eta_\alpha}\neq 0$ for any $\alpha$ and $\sum_{\alpha=1}^{n-3}\det A_{\eta_\alpha}\leq 0$. Then, we get
$$
K-\stackrel{\mathclap{\tiny\mbox{$N$}}}{\Riem}\left(X_1,X_2\right)\leq \det A_\eta.
$$
Let us consider $\mu_1$ and $\mu_2$ the principal curvatures of $A_\eta$. Then
\begin{align*} \det A_\eta =& \mu_1\mu_2=\frac{\left(\mu_1+\mu_2\right)^2-\left(\mu_1^2+\mu_2^2\right)}{2} \\
= & \frac{\left(\trace A_\eta\right)^2-\left|A_\eta\right|^2}{2} \\
= & \frac{4|H|^2-\left|A_\eta\right|^2}{2}. \end{align*}
From $\eqref{ineq-1}$ one obtains
$$
\left|A_\eta\right|^2\leq 4|H|^2-2K+2\stackrel{\mathclap{\tiny\mbox{$N$}}}{\Riem}\left(X_1,X_2\right).
$$
Since $K\geq 0$ and $\stackrel{\mathclap{\tiny\mbox{$N$}}}{\Riem}\ \leq k_0$, it follows that
$$
\left|A_\eta\right|^2\leq 4|H|^2+2k_0.
$$
Therefore, $\left|A_\eta\right|^2$ is bounded from above by the constant $4|H|^2+2k_0$ and then $\left|A_H\right|^2$ is bounded from above.
It is well known that a complete surface with $K\geq 0$ is parabolic (see [11]), i.e., any subharmonic function bounded from above is constant. Thus, as $\left|S_2\right|^2$ is bounded from above and subharmonic, it follows that $\left|S_2\right|^2$ is a constant. Using $\left|\nabla S_2\right|^2=16\left|\nabla A_H\right|^2$ and $\eqref{Laplac-S2}$ one obtains that $\nabla A_H=0$, and therefore $M$ is flat or pseudoumbilical.
5.1.
## Exemples of submanifolds with $\nabla A_H=0$
. As we have seen, a $PMC$ surface in a space form $N^n(c)$, $n\geq 4$, is trivially biconservative. But, if the surface is only $CMC$ then it is not necessarily biconservative. In [17] it was proved that if a surface is biconservative and $CMC$ in $N^4(c)$, with $c\neq0$, then the surface has to be $PMC$, i.e., the trivial case for our problem. We just recall here that, if $c=1$, then a $PMC$ surface in $\mathbb{S}^4$ is either a minimal surface of a small hypersphere of radius $a$, $a\in (0,1)$, in $\mathbb{S}^4$, or a $CMC$ surface in a small or great hypersphere in $\mathbb{S}^4$ (see [24,25]). Of course, if we consider a $CMC$ biconservative surface $M^2$ of genus $0$ in $\mathbb{R}^4$, it is pseudoumbilical and therefore it is $PMC$, i.e., $M^2$ is a $2$ -sphere (see [10]). In $\mathbb{R}^4$, there were obtained all $CMC$ biconservative surfaces which are not $PMC$. They are given by the isometric immersion $\varphi:\mathbb{R}^2\rightarrow\mathbb{R}^4$ defined by
$$
\varphi(u,v)=\overline{\gamma}(u)+(v+a)\overline{e}_4,
$$
where $\overline{\gamma}:\mathbb{R}\rightarrow\mathbb{R}^3$ is a smooth curve parametrized by arc length with positive constant curvature $k$ and free torsion $\tau$. By direct computation we obtain that the second fundamental form of the surface is given by
$$
B\left(\partial_u,\partial_u\right)=k(u)N(u), \qquad B\left(\partial_u,\partial_v\right)=0, \qquad B\left(\partial_v,\partial_v\right)=0,
$$
| |
170404598/4
|
SIMONA NISTOR
- (1) the stress-bienergy tensor of $\varphi$ is determined by
$$
S_2= - \frac{m^2}{2}|H|^2 I + 2m A_H;
$$
$\trace S_2 = m^2|H|^2\left(2-\frac{m}{2}\right)$;
- (3) the relation between the divergence of $S_2$ and the divergence of $A_H$ is given by
$$
\Div S_2=-\frac{m^2}{2}\grad \left(|H|^2\right)+2m\Div A_H;
$$
$\left|S_2\right|^2=m^4|H|^4\left(\frac{m}{4}-2\right)+4m^2\left|A_H\right|^2$.
Remark 2.8. From equation $\eqref{divergente0}$, we see that if $M$ is biconservative it does not follow that $\Div A_H$ automatically vanishes. In fact, only when $|H|$ is constant the biconservativity is equivalent to $\Div A_H=0$.
## 3. Other characterizations of biconservative submanifolds
In this section we will characterize the biconservative submanifolds which satisfy some additional geometric hypotheses.
We begin with a study on the basic properties of submanifolds with $A_H$ parallel, as they are the “ simplest ” biconservative surfaces. First, we define the principal curvatures of a submanifold $M^m$ of $N^n$ as being the eigenvalue functions of $A_H$.
Proposition Let $\varphi:M^m\rightarrow N^n$ be a submanifold and $\lambda_1\geq \cdots \geq \lambda_m$ the principal curvatures of $M$. If $\nabla A_H=0$, then:
$M$ is
- biconservative; $\lambda_i$ are constant functions on $M$, in particular $M$ is $CMC$; $A_{\nabla^\perp_X H}(Y)-A_{\nabla^\perp_Y H}(X)=\left(R^N(X,Y)H\right)^T$, for any $X,Y\in C(TM)$; $\trace A_{\nabla^\perp_\cdot H}(\cdot)=-\trace\left(R^N(\cdot,H)\cdot\right)^T$.
Proof. For the sake of completeness, we give only the proof of the second item. More precisely, we show that $\lambda_i$ are constant functions on $M$. Let us consider an arbitrary point $p\in M$. Since $A_H(p)$ is symmetric, then $A_H(p)$ is diagonalizable. We denote by $\lambda_{1,p}\geq\cdots\geq\lambda_{m,p}$ the eigenvalues of $A_H(p)$ and then we define the continuous functions $\lambda_i:M\rightarrow \mathbb{R}$, $\lambda_i(p)=\lambda_{i,p}$, for any $p\in M$ and any $i=\overline{1,m}$.
Further, we consider $\left\{e_i\right\}_{i=\overline{1,m}}$ an orthonormal basis in $T_p M$ which diagonalize $A_H(p)$, i.e., $\left(A_H(p)\right)\left(e_i\right)=\lambda_i(p)e_i$, for any $i=\overline{1,m}$.
Consider $q\in M$, $q\neq p$, and $\gamma:[a,b]\rightarrow M$ a smooth curve such that $\gamma(a)=p$ and $\gamma(b)=q$. We define the vector fields $E_i=E_i(t)$, along $\gamma$, such that $DE_i(t)/dt=0$, for any $t$ and $E_i(a)=e_i$. It is easy to see that $W(t)=\left(A_H(\gamma(t))\right)\left(E_i(t)\right)$ is also a vector field along $\gamma$ and
\begin{align*} \frac{DW}{dt}(t)= & \left(\nabla_{\gamma'(t)}A_H\right)(E_i(t))+A_H\left(\frac{D E_i}{dt}(t)\right)\\
= & \ 0. \end{align*}
Now, since $D\left(\lambda_i(p)E_i\right)(t)/dt=0$, we get that $W(t)$ and $\lambda_i(p)E_i$ are parallel vector fields along $\gamma$. Since for $t=a$ they are equal, it follows that they coincide for any $t$, and in particular, for $t=b$. Therefore, $\lambda_i(p)$, $i=\overline{1,m}$, are eigenvalues of $A_H(q)$. As $q$ was chosen in an arbitrary way, we get that $\lambda_i$ are constant functions on $M$, for any $i=\overline{1,m}$.
| |
170404598/1
|
# ON BICONSERVATIVE SURFACES
SIMONA NISTOR
Abstract. We study in a uniform manner the properties of biconservative sur- faces in arbitrary Riemannian manifolds. Biconservative surfaces being charac- terized by the vanishing of the divergence of a symmetric tensor field $S_2$ of type $(1,1)$, their properties will follow from general properties of a symmetric tensor field of type $(1,1)$ with free divergence. We find the link between the biconser- vativity, the property of the shape operator $A_H$ to be a Codazzi tensor field, the holomorphicity of a generalized Hopf function and the quality of the surface to have constant mean curvature. Then we determine the Simons type formula for biconservative surfaces and use it to study their geometry.
## 1. Introduction
In the last decade the theory of biconservative submanifolds proved to be a very interesting research topic (see, for example, [1,4–7,16,19,22,23]). This theory arose from the theory of biharmonic submanifolds, but the class of biconservative subman- ifolds is richer than the later one.
Let $\left(M^m,g\right)$ and $\left(N^n,h\right)$ be two Riemannian manifolds. A biharmonic map is a critical point of the bienergy functional
$$
E_2:C^{\infty}(M,N)\rightarrow\mathbb{R},\quad E_{2}(\varphi)=\frac{1}{2}\int_{M}|\tau(\varphi)|^{2}\ v_g,
$$
where $\tau(\varphi)$ is the tension field of a smooth map $\varphi:M\rightarrow N$, and it is characterized by the vanishing of the bitension field $\tau_2(\varphi)$ (see). If $\varphi:\left(M^m,g\right)\rightarrow \left(N^n,h\right)$ is a biharmonic map and a Riemannian immersion, then $M$ is called a biharmonic submanifold of $N$.
According to D. Hilbert (see [9]), to a functional $E$ we can associate a symmetric tensor field $S$ of type $(1,1)$, called the stress-energy tensor, which is conservative, i.e., $\Div S=0$, at the critical points of $E$. In the particular case of the bienergy functional $E_2$, G. Y. Jiang (see [13]) defined the stress-bienergy tensor $S_2$ by
\begin{align*} \langle S_2(X),Y\rangle=&\frac{1}{2}|\tau(\varphi)|^2\langle X,Y\rangle+\langle d\varphi,\nabla\tau(\varphi)\rangle\langle X,Y\rangle\
&-\langle d\varphi(X),\nabla_Y\tau(\varphi)\rangle-\langle d\varphi(Y),\nabla_X\tau(\varphi)\rangle, \end{align*}
and proved that
$$
\Div S_2=-\langle\tau_2(\varphi),d\varphi\rangle.
$$
Therefore, if $\varphi$ is biharmonic, then $\Div S_2=0$ (see [13,14]).
One can see that if $\varphi:\left(M^m,g\right)\rightarrow\left(N^n,h\right)$ is a Riemannian immersion then $\Div S_2=0$ if and only if the tangent part of the bitension field vanishes. A subman- ifold $M$ is called biconservative if $\Div S_2=0$.
Key words and phrases. Biconservative surfaces, biharmonic submanifolds, mean curvature vec- tor field, Codazzi tensor fields.
The author was supported by a grant of the Romanian National Authority for Scientific Research and Innovation, CNCS - UEFISCDI, project number PN-II-RU-TE-2014-4-0004.
| |
170404598/3
|
where $B\in C(\odot^2 T^*M\otimes NM)$ is the second fundamental form of $M$ in $N$ and the $\trace$ is considered with respect to the metric on $M$.
## 2. Preliminaries
First we recall some notions, formulas and general results about tensor fields and submanifolds that we will use later.
It is well-known that a symmetric tensor field $T$ of type $(1,1)$ on a Riemannian manifold $\left(M^m,g\right)$ can be identified with a symmetric tensor field $\tilde{T}$ of type $(0,2)$
$$
\langle T(X),Y\rangle=\tilde{T}(X,Y), \qquad X,Y\in C(TM),
$$
and, we will use the same notation $T$ instead of $\tilde{T}$.
Proposition 2.1. Let $\left(M^m,g\right)$ be a Riemannian manifold and consider $T$ and $S$ two symmetric tensor fields of type $(1,1)$. Then
$$
\langle\Delta^R T,S\rangle = \langle \nabla T,\nabla S \rangle - \Div Z,
$$
with $\Delta^R T = -\trace \nabla^2 T$, $Z\in C(TM)$, $Z=\langle \nabla_{X_i} T,S\rangle X_i$, where $\left\{X_i\right\}_{i=\overline{1,m}}$ is an orthonormal local frame field.
Proposition 2.2. Let $\left(M^m,g\right)$ be a Riemannian manifold and consider $T$ a sym- metric tensor field of type $(1,1)$ and $\alpha$ is a smooth function on $M$. Then
$$
\Div \left(T\left(\grad \alpha\right)\right) = \langle \Div T,\grad \alpha \rangle + \langle T, \Hess \alpha\rangle,
$$
Definition A submanifold $\varphi:M^m\rightarrow N^n$ is called pseudoumbilical if $A_H=|H|^2 I$, where $I$ is the identity tensor field of type $(1,1)$.
Using the Codazzi equation, we easily find the next result.
Proposition Let $\varphi:M^m\rightarrow N^n$ be a submanifold. Then
$$
\trace \nabla A_H = \frac{m}{2}\grad \left(|H|^2 \right) + \trace A_{\nabla_{\cdot}^\perp H}(\cdot)+ \trace\left(R^N(\cdot,H)\cdot\right)^T.
$$
Corollary Let $\varphi:M^m\rightarrow N^n(c)$ be a submanifold, $c\in\mathbb{R}$. Then
$$
\trace \nabla A_H = \frac{m}{2}\grad \left(|H|^2 \right) + \trace A_{\nabla_{\cdot}^\perp H}(\cdot).
$$
Let $\varphi:M^m\rightarrow N^n$ be a submanifold. Computing $\tau_2(\varphi)$ by splitting it in the tangent and in the normal part and using (2.3) we get the following characterizations for biconservative submanifolds (various expressions for $\tau_2(\varphi)$ were obtained in [2, 14,20,21]).
Proposition 2.6. Let $\varphi:M^m\rightarrow N^n$ be a submanifold. Then the following condi- tions are equivalent:
$M$ is
- biconservative; $\trace A_{\nabla^\perp_{\cdot} H}(\cdot)+\trace \nabla A_H +\trace \left(R^N(\cdot,H)\cdot\right)^T=0$; $\frac{m}{2}\grad\left(|H|^2\right)+2\trace A_{\nabla^\perp_{\cdot} H}(\cdot) + 2\trace \left(R^N(\cdot,H)\cdot\right)^T=0$; $2\trace \nabla A_H-\frac{m}{2}\grad\left(|H|^2\right)=0$.
We end this section with the following result.
Proposition 2.7. Let $\varphi:M^m\rightarrow N^n$ be a submanifold. Then we have:
| |
201001150/5
|
Figure3: Correlation between different similarity mea- sures and diversity measure and the improvement ($\Delta$) due to domain-specific BERT models.
## Analysis
After we empirically show the importance of se- lecting in-domain source data, the next question is: can we find a cost-effective way to nominate in-domain source data?
## Measuring Similarity
We use three measures of the similarity between source and target data. We then observe whether these similarity values correlate with the usefulness of pretrained models in $\S$ 5.2.
Language model perplexity (PPL) has been used to provide a proxy to estimate corpus sim- ilarity (Baldwinetal., 2013). We construct Kneser- Ney smoothed 3-gram models (Heafield, 2011) on source data and use the perplexity of target data relative to these language models as the similarity between source and target data.
Jensen-Shannon divergence (JSD), based on term distributions, has been successfully used for domain adaptation (Ruder and Plank, 2017). We first measure the probability of each term (up to 3-gram) in source and target data, separately. Then, we use the Jensen-Shannon divergence between these two probability distributions as the similarity between source and target data.
Targetvocabularycovered(TVC) measures the percentage of the target vocabulary present in the source data, where only content words (nouns, verbs, adjectives) are counted. Dai et al. (2019) show that it is very informative in predicting the effectiveness of pretrained word vectors.
In addition, RuderandPlank (2017) show that the diversity of source data is as important as do- main similarity for domain adaptation. Inspired by this, we also explore a very simple diversity mea- sure: type token ratio (TTR, $\frac{\# \, \text{unique} \, \text{tokens}}{\# \, \text{tokens}}$), that measures the lexical diversity of the source data.
To mitigate the impact of source data size on these measurements, for each source data, we sam- ple five sub-corpora, each of which contains 10M tokens. Then we measure the similarity of source and target data and the diversity of source data as the average values of these sub-corpora.
## Correlation Analysis
To analyze how the effectiveness of domain- specific BERT models correlate to the similarity be- tween source and target data, we employ the Pear- son correlation analysis to find out the relationships between improvements due to domain-specific BERT models and similarity between source and target data. For example, considering the BTC task, we use the performance of the original BERT as baseline, and measure the improvement due to Twitter BERT as $1.0$, whereas the corresponding value using BioBERT is $-2.9$. Note that we repeat all the experiments five times; therefore, we collect 300 source-target data points in total.
The correlation results are visualized in Figure 3. JSD has the strongest correlation (0.519) with the improvement due to domain-specific models, while the other two measures also have modest correla- tion ($0.481$ for PPL and $0.436$ for TVC). Recall that the calculation of JSD takes uni-grams, bi- grams and tri-grams into consideration, whereas PPL considers tri-grams only and the TVC consid- ers uni-grams only. Correlations between different measures indicate that these measures are able to reach agreement on whether source and target are similar. We find no correlation between the TTR of source data and the improvement.
## Summary
We conduct a case study of pretraining BERT on social media text. Through extensive experiments, we show the importance of selecting in-domain source data. Based on empirical analysis, we rec- ommend measures to help select pretraining data for best performance on new applications.
## Acknowledgments
We would like to thank anonymous reviewers for their helpful comments. XD also thanks Shubin Du and Ying Zhou for early investigation of this work. XD is supported by Sydney University’s En- gineering and Information Technologies Research Scholarship and a CSIRO Data61 top up scholar- ship.
| |
201001150/4
|
\begin{tabular}{c|l|c|c|c|c|c|c}
\rightarrowprule
%&&\multicolumn{6}{c}{\bf BERT}\bf Target Text type& \bf Corpus & \bf BERT & \bf Bio & \bf Clinical & \bf Sci & \bf Twitter & \bf Forum & & (3.3B) & (18B) & (0.5B) & (3.1B) & (0.9B) & (0.6B) \midrule
\multirow{4}{*}{Tweets} & Airline ($C$) & 80.5\scriptsize{$\pm$ 0.3} & 79.0\scriptsize{$\pm$ 0.5} & 78.8\scriptsize{$\pm$ 0.8} & 78.8\scriptsize{$\pm$ 0.9} & 80.8\scriptsize{$\pm$ 0.6} & \underline{\bf 81.6\scriptsize{$\pm$ 0.5}} % Dec-4
& BTC ($N$) & 78.0\scriptsize{$\pm$ 0.5} & 75.2\scriptsize{$\pm$ 0.3} & 76.9\scriptsize{$\pm$ 0.5} & 77.4\scriptsize{$\pm$ 0.4} & \bf 79.0\scriptsize{$\pm$ 0.5} & 77.0\scriptsize{$\pm$ 0.4} % Dec-4
& SMM4H-18 task3 ($C$) & 76.5\scriptsize{$\pm$ 0.9} & 75.4\scriptsize{$\pm$ 1.1} & 75.6\scriptsize{$\pm$ 0.7} & 75.4\scriptsize{$\pm$ 1.0} & 77.0\scriptsize{$\pm$ 1.0} & \bf 77.2\scriptsize{$\pm$ 1.3} % Task3 % Dec-4
& SMM4H-18 task4 ($C$) & 89.4\scriptsize{$\pm$ 0.5} & 87.7\scriptsize{$\pm$ 0.4} & 88.1\scriptsize{$\pm$ 0.8} & 88.7\scriptsize{$\pm$ 0.8} & 90.3\scriptsize{$\pm$ 0.3} & \underline{\bf 91.1\scriptsize{$\pm$ 0.6}} % Task4 % Dec-4
%& WNUT-16 ($N$) & 49.7 & 38.5 & 43.9 & 46.2 & \bf 49.9 & 46.7 % Nov-30
\hline
\multirow{4}{*}{Forum} & CADEC ($N$) & 71.9\scriptsize{$\pm$ 0.6} & 72.1\scriptsize{$\pm$ 0.6} & 72.1\scriptsize{$\pm$ 0.8} & \underline{\bf 73.2\scriptsize{$\pm$ 0.4}} & 72.1\scriptsize{$\pm$ 1.0} & 72.9\scriptsize{$\pm$ 0.6} % Nov-30
% & CADEC ($C$) & 91.6 & \bf 92.2 & 91.4 & 92.1 & 90.9 & 91.9 % & SMM4H-18 T3 ($C$) & 76.3 & 76.4 & 76.4 & 74.9 & \bf 77.7 & 77.5 & SemEval-14 laptop ($N$) & 81.1\scriptsize{$\pm$ 0.8} & 79.3\scriptsize{$\pm$ 0.3} & 78.5\scriptsize{$\pm$ 0.4} & \bf 81.6\scriptsize{$\pm$ 1.1} & 81.3\scriptsize{$\pm$ 0.6} & 81.4\scriptsize{$\pm$ 1.1} % Nov-30
& SemEval-14 restaurant ($N$) & 87.5\scriptsize{$\pm$ 0.6} & 84.9\scriptsize{$\pm$ 0.3} & 85.5\scriptsize{$\pm$ 0.7} & 86.7\scriptsize{$\pm$ 0.5} & 87.4\scriptsize{$\pm$ 0.7} & \underline{\bf 89.3\scriptsize{$\pm$ 0.5}} % Dec-4
& SST-2 ($C$) & 92.4\scriptsize{$\pm$ 0.2} & 91.1\scriptsize{$\pm$ 0.5} & 90.4\scriptsize{$\pm$ 0.3} & 91.4\scriptsize{$\pm$ 0.4} & 92.3\scriptsize{$\pm$ 0.4} & \underline{\bf 93.4\scriptsize{$\pm$ 0.4}} % Dec-2
%& & Amazon review & 61.5 & 60.5 & 60.4 & 60.2 & 61.3 & \bf 62.6 \hline
\multirow{3}{*}{Non-social media} & EBM ($N$) & 41.5\scriptsize{$\pm$ 0.5} & 42.1\scriptsize{$\pm$ 0.2} & 41.1\scriptsize{$\pm$ 0.5} & \bf 42.4\scriptsize{$\pm$ 0.7} & 40.5\scriptsize{$\pm$ 0.5} & 41.5\scriptsize{$\pm$ 0.5} % Nov-30
& i2b2-10 ($N$) & 85.8\scriptsize{$\pm$ 0.1} & \underline{\bf 87.4\scriptsize{$\pm$ 0.2}} & \underline{\bf 87.4\scriptsize{$\pm$ 0.1}} & 87.3\scriptsize{$\pm$ 0.2} & 84.8\scriptsize{$\pm$ 0.2} & 85.2\scriptsize{$\pm$ 0.1} % Nov-30
& JNLPBA ($N$) & 72.5\scriptsize{$\pm$ 0.3} & \underline{\bf 74.2\scriptsize{$\pm$ 0.2}} & 71.9\scriptsize{$\pm$ 0.1} & 73.6\scriptsize{$\pm$ 0.3} & 72.2\scriptsize{$\pm$ 0.2} & 72.5\scriptsize{$\pm$ 0.2} % Nov-30
& Paper Field ($C$) & 74.5\scriptsize{$\pm$ 0.1} & 74.3\scriptsize{$\pm$ 0.1} & 73.3\scriptsize{$\pm$ 0.1} & \underline{\bf 75.1\scriptsize{$\pm$ 0.1}} & 74.1\scriptsize{$\pm$ 0.1} & 73.3\scriptsize{$\pm$ 0.2} % & & CoNLL03 & \bf 91.0 & 89.6 & 90.5 & 87.5 & 90.1 & 88.4 \bottomrule
\end{tabular}
Table2: Effectiveness of different BERT models, evaluated on downstream tasks. # tokens in each pretraining data are listed in brackets. $C$: Classification task, for which we report macro-F1; $N$: NER task, for which we report span-level micro-F1. We repeat all experiments five times with different random seeds. Mean values are reported. underline: the best result is significantly better than the second best result (paired student’s t-test, p: $0.05$).
(a) False positives.
(b) False negatives.
Figure2: Error predictions on CADEC. Dotted line cir- cle: errors by the BERT model. Dashed line: YelpBERT. Solid line: SciBERT.
representations on short tweets.
We also observe that, when domain-specific models are applied on a target task with out-of- domain data, they achieve much lower results than the original BERT. For example, BioBERT achieves lower results than the original BERT on 7 out of 8 target social media tasks. It only achieves a better result on CADEC, which is about medica- tions. Recall that all these domain-specific BERT models use the pretrained weights of the original BERT as initialization. On one hand, we argue that this observation may challenge the conven- tional wisdom that the larger the pretraining data is, the better the pretrained model is. Training on out-of-domain source data may cause negative im- pact, at least for the two-stage pretraining approach we consider. On the other hand, this observation reinforces recent work showing the importance of task-adaptive pretraining (Gururanganetal., 2020).
ErroranalysisonCADEC We conduct an error analysis on CADEC, because it is at the intersec- tion between social media tenor (online posts) and medication field (adverse drug events), and thus could be similar to multiple sources. We compare the error predictions by the two best performing BERT models – ForumBERT and SciBERT, as well as the baseline BERT model. In Figure 2a, we observe that both domain-specific BERT models can reduce greatly the number of false positives made by the baseline BERT. Specifically, 159 false positives made by the baseline BERT are fixed by the domain-specific BERT models. However, domain-specific BERT models do not reduce a lot the number of false negatives – gold mentions not recognized. There are 258 gold mentions recog- nized by none of three models, and only 41 false negatives by the baseline BERT are fixed by the domain-specific BERT models (Figure 2b).
| |
201001150/1
|
# Cost-effective Selection of Pretraining Data: A Case Study of Pretraining BERT on Social Media
\begin{tabular}{cccc}
Xiang Dai$^{1,2}$ & Sarvnaz Karimi$^{1}$ & \textbf{Ben Hachey$^{3}$} & \textbf{Cecile Paris$^{1}$}
\end{tabular}
\begin{tabular}{cccc}
\multicolumn{4}{c}{$^{1}$CSIRO Data61, Sydney, Australia}\multicolumn{4}{c}{$^{2}$University of Sydney, Sydney, Australia}\multicolumn{4}{c}{$^{3}$Harrison ai, Sydney, Australia}\multicolumn{4}{c}{\tt \{dai.dai,sarvnaz.karimi,cecile.paris\}@csiro.au}\multicolumn{4}{c}{\tt ben.hachey@gmail.com} \end{tabular}
Recent studies on domain-specific BERT mod- els show that effectiveness on downstream tasks can be improved when models are pre- trained on in-domain data. Often, the pretrain- ing data used in these models are selected based on their subject matter, e.g., biology or com- puter science. Given the range of applications using social media text, and its unique language variety, we pretrain two models on tweets and forum text respectively, and empirically demon- strate the effectiveness of these two resources. In addition, we investigate how similarity mea- sures can be used to nominate in-domain pre- training data. We publicly release our pre- trained models at https://bit.ly/35RpTf0.
## Introduction
Sequence transfer learning (Ruder, 2019), that pre- trains language representations on unlabeled text (source) and then adapts these representations to a supervised task (target), has demonstrated its effectiveness on a range of NLP tasks (Radford et al., 2018; Devlin et al., 2019; Liu et al., 2019). Approaches vary in model, pretraining objective, pretraining data and adaptation strategy. We con- sider a widely used method, BERT (Devlinetal., 2019). It pretrains a transformer-based model us- ing a masked language model objective and then fine-tunes the model on the target task. We inves- tigate the impact of the domain (i.e., the similar- ity between the underlying distribution of source and target data) of pretraining data on the effec- tiveness of pretrained models. We also propose a cost-effective way to select pretraining data.
Recent studies on domain-specific BERT mod- els, which are pretrained on specialty source data, empirically show that, when in-domain data is used for pretraining, target task performance can be im- proved (Lee et al., 2019; Alsentzer et al., 2019; Huang et al., 2019; Beltagy et al., 2019). These publicly available domain-specific BERT models are valuable to the NLP community. However, the selection of in-domain data usually resorts to intu- ition, which varies across NLP practitioners (Dai et al., 2019). According to Halliday and Hasan (1989), the context specific usage of language is affected by three factors: field (the subject matter being discussed), tenor (the relationship between the participants in the discourse and their purpose) and mode (communication medium, e.g., ‘spoken’ or ‘written’). 1 Generally, the selection of pretrain- ing data in existing domain-specific BERT models is based on the field rather than the tenor. For example, BioBERT (Lee et al., 2019) and SciB- ERT (Beltagyetal., 2019) are both pretrained on scholar articles, but on different fields (biology and computer science).
We conduct a case study of pretraining BERT on social media text which has very different tenor from existing domain-specific BERT models. Our contributions are two-fold: (1) We release two pre- trained BERT models trained on tweets and forum text, and we demonstrate the effectiveness of these two resources on a range of NLP data sets using social media text; and, (2) we investigate the corre- lation of source-target similarity and task accuracy using different domain-specific BERT models. We find that simple similarity measures can be used to nominate in-domain pretraining data (Figure 1).
## Related Work
SelectingdatatopretrainBERT There are two known strategies: (1) collecting very large generic data, such as web crawl and news (Radfordetal., 2019; Liuetal., 2019; Baevskietal., 2019); and, (2) selecting in-domain data, which we refer to as
1 We do not explicitly consider mode in this study, because all data used are written text.
| |
201001150/3
|
with sequence-pair training (Liuetal., 2019).
Twitter We use English tweets ranging from Sep 1 to Oct 30, 2018 4 to pretrain our Twitter BERT. There are in total 60 million English tweets, con- sisting of $0.9$ B tokens. Although we aim to avoid tailored pre-processing strategies to make a fair comparison with other domain-specific BERT mod- els, we find 44% of these tweets contain url and 78% contain other user names (@, if a tweet replies another tweet, @ is added automatically). We thus employ minimal processing by: (1) replacing to- kens starting with ‘@’, referring to a Twitter user’s account name, with a special token [TwitterUser]; and, (2) replacing urls as a special token [URL]. We hypothesize that the surface form of these tokens do not contain useful information.
Forum We use local businesses reviews released by Yelp 5 to pretrain our Forum BERT. There are in total five million reviews, consisting of $0.6$ B tokens. No preprocessing is conducted on the text.
We used four Nvidia P100 GPUs for the pretrain- ing. Training of each model took seven days.
## Effectiveness of Pretrained BERT Models
To evaluate the effectiveness of our pretrained BERT models, we experiment on a range of clas- sification and Named Entity Recognition (NER) data sets. Both text classification and NER are fundamental NLP tasks that can employ generic architectures on top of BERT. For the classifica- tion task, the representation of the first token (i.e., [CLS]) is fed into the output layer for the final prediction. For the NER task, the representations of the first sub-token within each token are taken as input to a token-level classifier to predict the token’s tag. We did not explore more complex ar- chitectures, such as adding LSTM or CRF on top of BERT (Beltagyetal., 2019; Baevskietal., 2019), because our aim is to demonstrate the efficacy of domain-specific BERT models and to observe the impact of pretraining data, rather than to achieve state-of-the-art performance on these data sets.
Our BERT results follow the standard two- stage approach of finetuning the pretrained model. Domain-specific BERTs add a stage in the mid- dle: finetuning BERT on domain-specific unlabeled data (cf. Figure 1).
## Target Tasks
We use eight target tasks with their text sampled from Twitter and forums, to examine whether our BERT models can lead to improvements, com- pared to the original BERT. These tasks are Air- line 6: classifying sentiment on tweets about ma- jor U.S. airlines; BTC: identifying location, per- son, and organization on tweets (Derczynskietal., 2016); SMM4H-18: classifying whether the user reports an adverse drug events (task3) (Weis- senbacheretal., 2018), or intends to receive a sea- sonal influenza vaccine (task4) on tweets about health (Joshi et al., 2018); CADEC: identifying adverse drug events etc. on reviews about medica- tions (Karimietal., 2015); SemEval-14: identify- ing product or service attributes on reviews about laptops and restaurants (Pontikietal., 2014); SST: classifying sentiment on movie reviews (Socher etal., 2013).
In addition, we use four tasks that do not use so- cial media text to investigate how our BERT mod- els perform on out-of-domain target tasks: Paper Field: classifying the research topic based on the title of scholar articles about various fields (Belt- agy et al., 2019); EBM: identifying intervention, outcome etc. on scholar articles about clinical tri- als (Nye et al., 2018); i2b2-10: identifying treat- ment, test and problem on clinical notes about health (Uzuneretal., 2011); JNLPBA: identifying RNA, DNA etc. on scholar articles about biol- ogy (Kimetal., 2004).
## Results
We observe that our BERT models achieve the high- est F1 score on 6 out of 8 target tasks that use social media text (Table 2). On CADEC (medica- tions) and SemEval-14 laptop, SciBERT achieves the highest score due to the overlapping fields (i.e., medication and computer hardware, respectively). We note, however, that our Forum BERT achieves very close results. This demonstrates the effec- tiveness of our pretrained models on target tasks using social media text. To our surprise, on target tasks using tweets, forum BERT achieves better results than Twitter BERT on 3 classification tasks. On one hand, this may be explained by Baldwin et al. (2013) ’s observation that forum text is the ‘median’ data, which is similar to all other types of social media text. On the other hand, it also reveals the challenge of pretraining contextual language
4 Internetarchive, Accessed 1 June 2020.
5 YelpChallenge, Accessed 1 June 2020.
| |
230600409/2
|
common visual prompting. This strategy has been recently attempted in some recent advances, such as FewVLM (Jin et al. 2021) and BLIP-2 (Li et al. 2023). In this paper, we also conduct some toy experiments on VL benchmarks, of which results show that this strategy can help BERT (De- vlinetal.2018) obtain decent capabilities on VL tasks, e.g. 68.0% on VQA2.0 (Goyaletal.2017) and 76.3% on SNLI- VE (Xie et al. 2019). Although the performance still lags behind large-scale vision-language pre-trained (VLP) mod- els, such as CLIP-ViL (Shenetal.2021) and METER (Dou etal.2022), it yields an affordable way for quick VL adap- tion.
Nevertheless, visual prompting greatly increases the com- putation complexity of PLMs. For instance, directly adding all the visual patches of ViT (Dosovitskiy et al. 2020) to BERT (Devlin et al. 2018) will increase the already high computation by up to 384%. Moreover, the use of complete image features is often redundant to the VL models, as re- vealed in the previous VL research (Yangetal.2016;Zhou etal.2019,2021). During experiments, we also observe that the deployment of visual prompts greatly affects the final performance, i.e. at which layer we insert the prompts. For instance, the performance of BERT ranges from 64.97% to 69.54% on VQA2.0 with different placements.
To address these shortcomings, we propose a novel multi- modal transfer learning approach for PLMs called Dynamic Visual Prompting (DVP) in this paper. Instead of using the entire image features, DVP first applies a cross-attention module to collect task-related and text-relevant visual infor- mation, which is further linearly projected onto the semantic space of PLMs as the prompt token. Meanwhile, to auto- matically set the optimal placement of these tokens, we also equip DVP with a novel search algorithm based on $k$ -armed bandit theory, which regards the token insertion as a policy action and effectively estimates the search weights via nu- merous single-shot trials (Zhou et al. 2020). In addition, to achieve parameter-efficient adaption of PLMs, we also com- bine DVP with the recently popular adapter approach (Sung, Cho, and Bansal 2022) for downstream VL tasks. Such a combination can not only greatly save the parameter expen- diture, but also enable PLMs to make a quick shift between NLP and VL tasks, as shown in Fig. 1.
To validate DVP, we apply it to two representative PLMs, which are the encoder-based BERT (Devlinetal.2018) and the encoder-decoder based T5 (Raffeletal.2020). Extensive experiments are performed on a set of VL reasoning bench- marks including VQA2.0 (Goyal et al. 2017), GQA (Hud- sonandManning2019) and SNLI-VE (Xieetal.2019). To validate the generalization, we also apply DVP to a recent LLM called LLaMA (Touvron et al. 2023) on ScienceQA (Lu et al. 2022). The experimental results show that com- pared with the common visual prompting, DVP can save up to 80% computation while improving the performance sig- nificantly, e.g. +1.21% on VQA2.0 for BERT and +2.19% on GQA for T5. When combined with Adapter (Sung,Cho, and Bansal 2022), DVP can well maintain the overall per- formance of PLMs on these benchmarks, and it only needs to update about 5.0% -6.0% parameters of the model for VL adaptions. In addition to pre-training costs, DVP also ex- hibits much better efficiency than VLP models, e.g. 4.0G FLOPs of BERT-DVP v.s. 956.4G FLOPs of OSCAR (Li etal.2020). Meanwhile, when combined with Adapter, the updated parameters are only about 5.1% of OSCAR. These results well confirm our motivation about DVP.
Overall, the contributions of this paper are three-fold:
- We propose to directly adapt PLMs as a stand-alone model to VL tasks via inserting visual prompt tokens, which can avoid the building of heavy fusion networks and make use of the context reasoning ability of PLMs.
- We propose a novel transfer learning method called D ynamic V isual P rompting (DVP) for efficient PLM adaption, which includes a cross-attention module to ob- tain compact visual tokens and a $k$ -armed bandit based search algorithm for automatic prompt placement.
- DVP can reduce the computation of common visual prompting methods by up to 80% while achieving bet- ter performance on multiple benchmarks. Compared with VLP models, DVP can also help PLMs obtain competi- tive performance on VL tasks while saving the parame- ters and training overhead substantially.
## Related work
With the great success of pre-training and fine-tuning paradigm in NLP, large-scale vision-language (VL) pre- training has also become a standard step in VL research (Chenetal.2023). In the early stage, the pre-trained vision encoders are introduced to provide the region knowledge. VisualBERT (Li et al. 2019), ViLBERT (Lu et al. 2019), LXMERT (TanandBansal2019), Uniter (Chenetal.2020), Oscar (Lietal.2020) use the pre-trained Faster-RCNN (Ren etal.2015) to extract region features and combine them with text features for deep fusion. To break through the limita- tion of object detectors in inference efficiency, PixelBERT (Huang et al. 2020), Grid-VLP (Yan et al. 2021), SOHO (Huangetal.2021) uses ResNet (Heetal.2016) to extract the grid features of the image. These methods not only real- ize end-to-end training of VLP models, but also achieve bet- ter performance on VL tasks. Although VLP models achieve excellent performance on downstream VL tasks, the PLMs are only used as a language encoder and still require another large fusion branch and expensive VL pretraining.
Recently, Prompt tuning (Guetal.2021;Liuetal.2021; Wei et al. 2021; Han et al. 2022) aims to insert a set of prompt tokens into the input sequence of PLMs, thereby alleviating the difference between the data distributions of pre-training and downstream tasks (Liu et al. 2023b). To avoid laborious manual tuning, recent advances resort to learnable tokens for downstream tasks, which is also termed soft prompting (Lester, Al-Rfou, and Constant 2021). The great success of prompt tuning in NLP also sparks its ap- plication to computer vision (CV) and vision-language (VL) studies. In terms of VL studies, CoOp (Zhou et al. 2022b) keeps the parameters of CLIP unchanged, and puts learn- able parameters as soft prompting for the input text, which fully uses the zero-shot retrieval ability of CLIP. To solve the problem that CoOp is easy to overfit on the basic classes, CoCoOp (Zhou et al. 2022a) implements instance-adaptive
| |
230600409/6
|
Table 3: Comparison between KAB-APP and the Manual search. “ Com-Prompt. ” refers to the common visual prompt- ing mentioned above. “ At 1 $st$ -layer ” denote that deploying DVP at the input layer of PLMs.
\begin{tabular}{@{}cccccc@{}}
\rightarrowprule
\textbf{Model} & \textbf{Setting} & \textbf{FLOPs} & \textbf{\begin{tabular}[c]{@{}c@{}}VQAv2\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}GQA\end{tabular}} & \textbf{\begin{tabular}[c]{@{}c@{}}SNLI-VE\end{tabular}} \midrule
\multirow{4}{*}{BERT} & Com-Prompt. & 167.6G & 68.00 & 51.34 & \textbf{76.26} & At 1$st$-layer & 34.6G & 64.97 & 50.01 & 75.78 & Manual & 34.1G & \textbf{69.54} & 52.55 & 76.00 & KAB-APP & 34.0G & 69.21 & \textbf{52.55} & 74.76 \midrule
\multirow{4}{*}{T5} & Com-Prompt. & 192.4G & 69.24 & 50.85 & \textbf{76.19} & At 1$st$-layer \tablefootnote{The first layer of the decoder of T5.} & 37.2G & 69.42 & 52.27 & 75.60 & Manual & 36.8G & 69.82 & 53.04 & 75.60 & KAB-APP & 36.8G & \textbf{69.82} & \textbf{53.04} & 75.60 \bottomrule
\end{tabular}
all image features is a reliable way of adapting PLMs to VL tasks, of which performance is consistent across tasks and models. However, compared with other solutions, its compu- tation is much more expensive, i.e. +4-+5 times. For [CLS] prompting, using a static visual prompt is inferior for most VL tasks, especially the ones that require fine-grained rea- soning, i.e. VQA2.0 and GQA. For the DVP methods, the dynamic prompt tokens can obtain better performance than the static [CLS] prompting in most cases, confirming our as- sumption about dynamic prompting. However, the increase of dynamic tokens does not always lead to better perfor- mance under all settings. Meanwhile, the effectivenesses of DVP on BERT and T5 are also different. Compared to T5, we directly place DVP at the input layer of BERT, which has yet to learn sufficient text semantics. This finding im- plies that the placement of prompt tokens is critical to DVP. Overall, these results confirm the feasibility of DVP, but it still needs great improvements in token placement.
The impact of prompt placement. We further examine the impact of prompt placement via manual setting in Tab. 2. From this table, we can first confirm that the prompt place- ment is vital to performance. In terms of BERT, the per- formance of different DVP placements varies vastly, which changes are 4.57%, 2.54%, and 2.06% in VQA2.0, GQA, and SNLI-VE, respectively. In terms of T5, the performance change still exists but is less obvious, which are 0.6%, 1.18%, and 0.79%, respectively. To explain, we only try the placement of T5 from the decoding layers, where the textual semantic is well-built via its encoder branch. In this case, the quality of DVP is better ensured, and the similar cases can be also found in Tab. 1. Furthermore, we can see that the effect is also different on VL tasks. For instance, VQA performs better when the prompts are inserted to the higher layers and SNLI-VE does the opposite. For DVP, we specu- late that VQA firstly focuses on understanding the question itself, and then use its semantics to extract key information in the image to generate an effective prompt. In contrast, SNLI-VE may need more understanding of image seman- tics, and deploy it at a lower layer can better facilitate VL comprehension. Thus, we can conclude that the best inter- action of DVP varies for different tasks, which substantially confirms the necessity of automatic prompt placement.
Effectiveness of KAB-APP. To validate KAB-APP, we compare it with the manual search in Tab. 3, which is ob- tained by training all possible insertions. It can be seen that the search results of KAB-APP is very close to that of man- ual search, which can be regarded as the upper-bound of DVP. On BERT, KAB-APP can achieve competitive perfor- mance against the manual search, although the best insertion layers are slightly different on VQA2.0, i.e. 11- $th$ and 10- $th$ layers, respectively. In terms of T5, of which text semantics are well learned, the best placements of KAB-APP and man- ual search are consistent on all tasks, greatly showing the effectiveness of our search algorithm. From Tab. 3, we can also see that via finding the optimal placement, the perfor- mance of DVP can be greatly improved. For instance, KAB- APP is better than using all visual prompts (Com-Prompt.) on VQA2.0 and GQA, while reducing the computation by 80.7%. Meanwhile, compared to the static placement, i.e. “ at 1- $st$ layer ”, the performance gains become more obvi- ous. These results well confirm our assumption and validate the effectiveness of KAB-APP.
The combination with Adapter. To further improve ef- ficiency, we introduce Adapter (Sung, Cho, and Bansal 2022) to parameter-efficiently adapt PLMs into the down- stream tasks. As shown in the bottom two rows of Tab. 4, the adapter-based methods achieve similar performance to fine-tuning. Specifically, the performance on BERT is slightly improved compared to the fine-tuning mode, e.g. +1.07%, +0.92% and +0.94% on VQA2.0, GQA and SNLI- VE, respectively. In terms of T5, the DVP with adapters can achieve 99.3%, 99.8% and 98.8% performance of fine- tuning manner while only updating 5% parameters. From the above experiments, we can conclude that the proposed DVP method can easily cooperate with other PETL methods for adapting PLMs in a more efficient way.
Comparison with visual-language pre-trained models. We also compare DVP with a bunch of VLP models on VL benchmarks in Tab. 4, which are often built with an- other deep fusion networks. Here, DVP is deployed accord- ing to the results of KAB-APP. Notably, compared to VLP models, DVP directly transfers PLMs into VL tasks with- out additional requirement of large-scale VL pre-training. When adapting pre-trained model into downstream tasks, DVP takes fewer computing resources especially when be- ing combined with the adapters, i.e. updates only 5.9% and 5.0% parameters of BERT and T5, respectively. When training similar parameters, the proposed DVP on BERT achieves competitive performance against ViLT on VQA2.0, GQA and SNLI-VE, respectively, while only taking 60.8% FLOPs. Compare to the SOTA method METER (Douetal. 2022), the proposed BERT-DVP $_{adp}$ method takes 2.2% trainable parameters and 14.1% FLOPs to achieve 70.28% performance on VQA2.0. Overall, these results confirm the feasibility of DVP in adapting PLMs to VL tasks.
The generalization ability of DVP. In Tab. 5, we exam- ine the generalization of DVP to the recently proposed LLM called LLaMA (Touvronetal.2023) on ScienceQA. When applying DVP to LLaMA-7B, another VL pre-training is also not required, and updated parameters are only 9M, which is about 0.13% of LLaMA-7B. Compared with BLIP- 2 (Li et al. 2023) using Vicuna-7B as the base language
| |
230600409/4
|
Figure2: Illustration of the proposed method. The right sub-figure illustrates the process of dynamic visual prompting (DVP), which can produce compact yet highly related visual tokens for prompting. The left sub-figure depicts the process of the proposed k-armed bandit based automatic prompt placement algorithm (KAB-APP). With a short search period, KAB-APP can find out the optimal deployment of DVP on different PLMs for different VL tasks.
## Automatic Prompt Placement
In our experimental trials, we find that the model perfor- mance is disparate with the different placements of DVP. Meanwhile, for VL tasks with various definitions, a static setting is suboptimal to PLM adaption. To this end, we pro- pose a $k$ - A rmed B andit based A utomatic P rompt P lacement algorithm (KAB-APP) as shown in the left half of Fig. 2.
Problem Definition. The target of KAB-APP is to auto- matically find out the insertion layer that can lead to the best performance of DVP. Thus, we consider the prompt place- ment as a policy action and model the search process as a $k$ -armed bandit problem.
Concretely, we equally initialize the action preference for each layer of PLMs as $h \in \mathbb{R}$, that is, the probability of each layer being selected. During each search step, the policy for the $K$ - $th$ layer $\pi_t(K)$ is obtained by
$$
\begin{aligned}
\pi_t(K) &= \frac{e^{\mathbf{H}_t(K)}}{\sum_{m=1}^{M}e^{\mathbf{H}_t(l)}},
\end{aligned}
$$
where $\mathbf{H}_t=[h_t^1,...,h_t^K,...,h_t^M]$ denotes the preference of each layer at $t$ -th setp. Based on Eq. 9, the action weight is then updated by
$$
\begin{aligned}
\Delta R &= R_t-R_b, \\
\mathbf{H}_{t+1}(K) &= \mathbf{H}_{t}(K)+\alpha\Delta R\pi_t(K)(1-\pi_t(K)),
\end{aligned}
$$
where $\alpha$ is the learning rate for updating $\mathbf{H}$, $R_t$ is the reward of inserting DVP to $K$ - $th$ layer at the $t$ - $th$ step, $R_b$ is the baseline reward obtained by $n$ sampling times. Here we use the average value of $n$ samplings’ rewards at the $t$ - $th$ step.
Search Algorithm. The procedure of KAB-APP is de- scribed in Algorithm 1. Firstly, we establish an action space consisted of $M$ insertion layers for PLMs. Meanwhile, we equip all the possible insertion layers with independent cross-modal attention modules.
At each training step, an inserted layer is randomly sam- pled from $M$, and this layer $K$ and the corresponding cross- attention module $\phi_K$ are activated for training. In a short training interval, we sample $n$ candidate insertion layers for validation by the RL policy $\pi_t$, and use the validation accu- racy or the power of the negative loss of $e$, as the rewards. Then we update preferences $\mathbf{H}$ by using the gradient update rules in Eq. 10.
KAB-APP aims to search the best insertion layer for DVP on the language model. With the above search algorithm, the action weights can be fairly estimated via numerous single- shot validations (Zhou et al. 2020). Since the search space is limited, the whole process can be accomplished within a few periods. In practice, we control this search process to the limited steps, which only requires an additional 30-40% of the time of fine-tuning a PLM.
## Combination with Adapter
We further combine DVP with an advanced parameter- efficient transfer learning approach called Adapter (Sung, Cho,andBansal2022;Luoetal.2023) to reduce the scale of training parameters. It can also help PLMs to make a quick shift between NLP and VL tasks. In practice, given a PLM, we can search the optimal DVP insertion layer $K$ with KAB- APP, and then use cross-modal attention layer $\phi$ to generate lightweight DVP at $K$ -th layer. Furthermore, we equipped language model with Adapters. The process above not only satisfies high parameter efficiency, but also attains optimal performance.
## Experiments
## Datasets
VQA2.0 (Goyal et al. 2017) is a benchmark dataset that builds on VQA1.0 (Antol et al. 2015) and contains open- ended questions about images from MSCOCO (Ren,Kiros, andZemel2015). Following previous works (Yuetal.2019), we consider VQA2.0 as a classification task with 3129 an- swer categories. Following ViLT (Kim,Son,andKim2021), we use VQA2.0 train set and val set for fine-tuning while re- taining 1000 examples of the val set for validation.
| |
230600409/1
|
# Adapting Pre-trained Language Models to Vision-Language Tasks via Dynamic Visual Prompting
Shubin Huang 1, Qiong Wu 1, Yiyi Zhou 12 *, Weijie Chen 3, Rongsheng Zhang 3, Xiaoshuai Sun 12,
Pre-trained language models (PLMs) have played an increas- ing role in vision-language (VL) learning, but they usually re- quire a deep multi-modal branch for VL reasoning, resulting in excessive computation and memory overhead. Recently, visual prompting is a feasible way to adapt PLMs to VL tasks, but we notice that the use of all visual tokens will greatly ex- ecrate the already high computation, and the token placement is also vital to performance. Based on these observations, we propose a novel transfer learning approach for PLMs in this paper, termed D ynamic V isual P rompting (DVP). Concretely, DVP first deploys a cross-attention module to obtain text- related and compact visual prompt tokens, thereby greatly reducing the input length of PLMs. To obtain the optimal placement, we also equip DVP with a reinforcement-learning based search algorithm, which can automatically merge DVP with PLMs for different VL tasks via a very short search pro- cess. In addition, we also combine DVP with the recently popular adapter approach to keep the most parameters of PLMs intact during adaption, which also help PLMs achieve a quick shift between single- and multi-modal tasks. We ap- ply DVP to two representative PLMs, namely BERT and T5, and a recent large language model called LLaMA. Extensive experiments are conducted on a set of VL reasoning bench- marks including VQA2.0, GQA, SNLI-VE and ScienceQA. The experimental results not only show the merits of DVP in performance and efficiency, e.g. +2.28% accuracy and -80% FLOPs on VQA2.0, but also confirm its superiority in adapt- ing pre-trained language models to VL tasks. Our code is anonymously released at https://github.com/hsb1357173526/ Dynamic Visual Prompting.
## Introduction
Recent years have witnessed the rapid development of pre- trained language models (PLMs) (Devlinetal.2018;Lewis 2020), which have become the de facto standard in natu- ral language processing (NLP). Recently, the emergence of ChatGPT (OpenAI2022) and LLaMA (Touvronetal.2023) further confirms the significance of PLMs in exploring gen- eral artificial intelligence. The advent of PLMs also leads
Figure 1: The application of the proposed D ynamic V isual P rompting (DVP) to pre-trained language models (PLMs). With low parameter and training expenditure, DVP can adapt PLMs to vision-language tasks and help them make a quick shift between single- and multi-modal tasks.
to the prevalence of large-scale pre-training in the vision- language (VL) field (Lu et al. 2019; Tan and Bansal 2019; Lietal.2020;Shenetal.2021;Douetal.2022).
However, directly adapting PLMs to VL tasks is pro- hibitively expensive. Above all, PLMs often act as a build- ing block for advanced VL models. Similar to vision mod- els (Ren et al. 2015; He et al. 2016; Dosovitskiy et al. 2020), PLMs are only used as a modal-specific encoder. For VL reasoning tasks like Visual Question Answering (VQA) (Goyal et al. 2017), the VL models still need to employ a deep fusion branch upon the multi-modal encoders, making the model extremely cumbersome. In addition to parameter redundancy, this deployment also undermines the ability of PLM in context reasoning since it only serves to embed the text words. To this end, the VL model often requires another large-scale pre-training on massive VL data (Luetal.2019; Son,andKim2021;Douetal.2022).
In this case, we aim to directly adapt PLMs to vision- language tasks with low parameter and training overhead. A natural solution is to directly feed all the extracted visual features into PLMs as the prompt tokens. Here, we term it
| |
230600409/3
|
by adding the vision feature of each image to learnable prompts, which makes prompt tuning more robust. The above methods mainly use text words or learnable tokens as the prompts for NLP and VL models. There are also some very recent works applying image information for VL task prompting. With an encoder-decoder architecture, SimVLM (Wang et al. 2021) uses the patch features extracted from images by ViT as prefix. Frozen (Tsimpoukellietal.2021) based on GPT only trains image encoder, and uses extracted visual features as visual prompt tokens for language model, which achieves excellent performance. Overall, existing vi- sual prompting methods often use the entire image features for VL adapting, while the critical issues of redundant visual information and excessive computation are still overlooked.
## Method
## Preliminary
Before introducing our approach, we first recap the basic vi- sual prompting for PLMs. Concretely, given a PLM as $G(\cdot)$, and the image-text example as $(I,T)$, the target of adapting PLMs to VL tasks is to minimize the loss defined by
$$
\setlength{\abovedisplayskip}{3pt}
\setlength{\belowdisplayskip}{3pt}
\begin{aligned}
\operatorname*{argmin}_{\theta} \mathcal{L}\big(G(I,T|\theta)\big),
\end{aligned}
$$
where $\theta$ is the parameters of $G$. $\mathcal{L}$ is the objective function of the downstream task. For visual prompting, a natural so- lution is to use the all visual features as the prompt tokens, and then project them onto the semantic space of PLMs.
However, in this solution, the length of input sequence $L$ is largely increased by the number of visual tokens $N$, which leads to the additional computations by $O(N^2+2NL)$. Be- sides, using all visual information for VL reasoning is often redundant to the model (Rao et al. 2021; Ryoo et al. 2021; Bolyaetal.2022).
To alleviate the computational burdens of PLMs and re- duce the redundancy in visual prompts, we propose to use a lightweight network $\phi$ to generate compact visual prompts, thereby reducing the length of input sequence. The optimiza- tion objective can be further defined as
$$
\setlength{\abovedisplayskip}{3pt}
\setlength{\belowdisplayskip}{3pt}
\begin{aligned}
\operatorname*{argmin}_{\theta,\phi} \mathcal{L}\big(G(\phi(I,T),T|\theta)\big),
\end{aligned}
$$
which can obtain substantial improvements in both model inference and computation overhead.
## Dynamic Visual Prompting
In this paper, we propose a Dynamic Visual Prompting (DVP) approach towards efficient VL adaption. In principle, DVP adopts cross-modal attention to dynamically generate visual prompts relevant to the input text and image, as shown in the right half of Fig. 2.
Given an image $I$, we first use a frozen visual encoder to extract its features, denoted as $\mathbf{F}_v \in \mathbb{R}^{N \times d}$, where $N$ is the number of visual features and $d$ is the feature dimension. To obtain the compact visual tokens related to the input text, we use the text features in PLMs as the query vectors, $\mathbf{F}_t \in \mathbb{R}^{L \times d}$, where $L$ refers to the length of text sentences. With the projection weight matrices, $\mathbf{W}_Q \in \mathbb{R}^{d \times d}$, $\mathbf{W}_K \in \mathbb{R}^{d \times d}$, and $\mathbf{W}_V \in \mathbb{R}^{d \times d}$, for $Q, K, V$ transformations in our cross- modal attention module $\phi$. DVP is obtained by
$$
\setlength\abovedisplayskip{6pt}
\begin{aligned}
\mathbf{F}_{DVP} = Softmax(\frac{\mathbf{F}_t\mathbf{W}_Q(\mathbf{F}_v\mathbf{W}_K)^T}{\sqrt{d}})\mathbf{F}_v\mathbf{W}_V.
\end{aligned}
$$
Here, we also employ multi-head attention mechanism, so the above formula can be further rewritten as
$$
\begin{aligned}
\mathbf{F}_{DVP} &= Concat(head_1, ..., head_h)\mathbf{W}_O, \\
head_i &= Softmax(\frac{\mathbf{F}_t\mathbf{W}^i_Q(\mathbf{F}_v\mathbf{W}^i_K)^T}{\sqrt{\frac{d}{n}}})\mathbf{F}_v\mathbf{W}^i_V,
\end{aligned}
$$
where $n$ is the number of attention heads, $\mathbf{W}_O \in \mathbb{R}^{d \times d}$ and $\mathbf{W}^i_Q, \mathbf{W}^i_K, \mathbf{W}^i_V \in \mathbb{R}^{d \times \frac{d}{n}}$ are the weight matrices. To this end, the length of generated $\mathbf{F}_{DVP}$ becomes $L$, and the computational complexity is reduced from $O\big((N+L)^2\big)$ to $O\big((L+L)^2\big)$ in calculating attention matrix, where $N$ is often larger than $L$ in VL tasks (Kim, Son, and Kim 2021; Shenetal.2021;Douetal.2022).
However, the computation overhead is still expensive and we further reduce the length of prompt tokens to improve the efficiency by using the global features of PLMs as query vectors, such as [CLS] token of BERT, pooling feature of T5’s encoder or [EOS] token of LLaMA.
For encoder-based language models like BERT, we define its layers as $\Theta=[\Theta_1,...,\Theta_{K},...,\Theta_{M}]$, where $K$ is the in- dex of insertion layer and $M$ refers to the number of layers. Here, $\mathbf{F}_t$ denotes the hidden features obtained before the $K$ - $th$ layer we want to insert, and we use its [CLS] token fea- tures $\mathbf{F}_t^{[CLS]}$ as query vector. Afterwards, DVP of BERT at insertion layer $K$ is obtained by
$$
\setlength{\abovedisplayskip}{1pt}
\setlength{\belowdisplayskip}{1pt}
\begin{aligned}
\mathbf{F}_t &= \Theta_{K-1}\big(\Theta_{K-2}...\Theta_1\big(E(T)\big)\big), \\
\mathbf{F}_{DVP} &= \phi(\emph{query}=\mathbf{F}_t^{[CLS]}, \emph{key}=\mathbf{F}_v, \emph{value}=\mathbf{F}_v),
\end{aligned}
$$
where $E$ denotes the text embedding. Next, $\mathbf{F}_{DVP}$ is con- catenated with $\mathbf{F}_t$ and sent to subsequent layers
$$
[\mathbf{F}_{DVP};\mathbf{F}_t] = \Theta_M\big(\Theta_{M-1}...\Theta_{K}\big([\mathbf{F}_{DVP};\mathbf{F}_t]\big)\big) ,
$$
Finally, we use $\mathbf{F}_t^{[CLS]}$ to connect the specific VL task clas- sification head for prediction.
For encoder-decoder based PLMs like T5, we define its encoder as $\Psi$, decoder layers as $\Theta=[\Theta_1,...,\Theta_{K},...,\Theta_{M}]$ and decoder input vector as $\mathbf{f} \in \mathbb{R}^{d}$. Here, $\mathbf{F}_t$ means the output of $\Psi$ and we utilize the pooling features of $\rho(\mathbf{F}_t)$ as text query vector, where $\rho$ means the operation of average mean pooling. Then $\mathbf{F_{DVP}}$ of T5 at insertion decoder layer $K$ is obtained by
$$
\begin{aligned}
\mathbf{F}_t &= \Psi\big(E(T)\big), \\
\mathbf{F}_{DVP} &= \phi(\emph{query}=\rho(\mathbf{F}_t), \emph{key}=\mathbf{F}_v, \emph{value}=\mathbf{F}_v).
\end{aligned}
$$
Afterwards, $\Theta_K$ ’s output is spliced with $\mathbf{F}_{DVP}$ and sent to the subsequent decoder layers:
$$
\setlength{\abovedisplayskip}{1pt}
\setlength{\belowdisplayskip}{1pt}
\begin{aligned}
\mathbf{f} &= \Theta_{K-1}\big(\Theta_{K-2}...\Theta_1\big(E(f)\big)\big), \\
[\mathbf{F}_{DVP};\mathbf{f}] &= \Theta_M\big(\Theta_{M-1}...\Theta_{K}\big([\mathbf{F}_{DVP};\mathbf{f}]\big)\big),
\end{aligned}
$$
Lastly, we use $\mathbf{f}$ attached to the classification head for pre- diction.
| |
221016850/2
|
Third, based on the MIMIC-III dataset 3, we demonstrate the RAC model’s effectiveness in the most challenging full codes prediction testing set from inpatient clinical notes. The RAC model wins over all the previously reported SOTA results considerably. RAC establishes a new SOTA, considerably outperforming the current best Macro-F1 by 18.7%, and reaches past the human-level coding baseline. For further information on prediction results, see Table 1 and 2 of [1], as well as the Papers with Code’s leaderboard in the medical code prediction [5].
xRAC Framework and Human-Grounded Explainability Evaluation: Next, we present a general e x plainable R ead, A ttend, and C ode (xRAC) framework that generates evidentiary text snippets for a predicted code, which are oriented towards the needs of a deployment scenario. The first attention score-based approach, xRAC-ATTN, utilizes the label-wise attention mechanism first introduced in [4] to select key sentences for prediction decisions. In the second model-agnostic knowledge- distillation-based method, xRAC-KD, a large"teacher" RAC model is distilled into a collection of"student" linear models in post-hoc manner without sacrificing much accuracy of the teacher model while retaining many advantages of linear models including explainability and smaller model size which is beneficial for deployment. More methodological details are available in Section 2 of [2].
Then, a simplified but thorough human-grounded evaluation with two groups of annotators, one group (Group A) with and one group without (Group B) medical coding expertise 4, is conducted to evaluate the explainability of our xRAC framework, xRAC-ATTN and xRAC-KD. Both groups are given the same question sheet, which has explanation text snippets extracted to support the appearance of the predicted codes. Each annotator must select one of the three choices that are highly informative, informative, and irrelevant. More information on the evaluation task design and execution can be found in Section 3.2 of [2].
We find that the proposed xRAC framework can be easily explainable and the supporting evidence text highlighted by xRAC-ATTN is of higher quality than xRAC-KD even though xRAC-KD has potential advantages in production deployment scenarios. One can see from Table 2 in [2] that there is much larger gap in xRAC-KD between Group A and Group B than between xRAC-ATTN. This implies that xRAC-ATTN is a more viable choice than xRAC-KD to extract a text snippet from clinical notes to support code prediction. However, Table 3 in [2] shows that the consistency score measured by Jaccard Similarity between two groups is lower than 40% even with xRAC-ATTN.
This suggests that the automated extraction system must still rely on professional coders’ feedback, and there is room remaining to improve for a lay person without expertise to correctly code. We believe that this is a very meaningful step toward the goal of developing accurate, explainable and automated ML systems for medical code prediction from clinical notes. To our knowledge, most previous works are still in early stages in terms of providing textual references and explanations of the predicted codes and no studies to date have thoroughly analyzed in this regard, especially for complex transformer-based models such as the RAC model.
## Conclusion
In this study, we present for the first time a human-coding baseline for medical code prediction on the subsampled MIMIC-III-full-label inpatient clinical notes testing set task. We have developed an attention-based RAC model that sets the new SOTA records, and the resulting RAC model outperforms the human-coding baseline to a great extent on the same task. The performance improvements can be attributed to effectively learning the common embedding space between the clinical note and medical codes by utilizing attention mechanisms that efficiently address the severe long-tail sparsity issues. Additionally, we present a xRAC framework to obtain supporting evidence text from clinical notes that justify predicted medical codes from medical code prediction systems. We have demonstrated that the proposed xRAC framework may help even complex transformer-based models to attain high accuracy with a decent level of explainability (which is of high value for deployment scenarios) through qualitative human-grounded evaluations. We also show for the first time that, given the current state of explainability methodologies, using the proposed explainable yet accurate medical codes prediction system still requires professional coders’ expertise and competencies.
3 The MIMIC-III Dataset (MIMIC v1.4) is a freely accessible medical database containing de-identified medical data of over 40,000 patients staying in the BIDMC between 2001 and 2012.
4 Group A had two annotators without medical coding experience and Group B had six certified professional coders.
| |
221016850/1
|
# Medical Codes Prediction from Clinical Notes: From Human Coders to Machines
Byung-Hak Kim AKASA, Inc. South San Francisco, CA hak.kim@AKASA.com
## Introduction
Prediction of medical codes from clinical notes is a practical and essential need for every healthcare delivery organization within current medical systems 1. Automating annotation will save significant time and excessive effort that human coders spend today. However, the biggest challenge is directly identifying appropriate medical codes from several thousands of high-dimensional codes from unstructured free-text clinical notes. This complex medical codes prediction problem from clinical notes has received substantial interest in the NLP community, and several recent studies [1,6,3,8] have shown the state-of-the-art (SOTA) code prediction results of full-fledged deep learning-based methods. This progress raises the fundamental question of how far automated machine learning (ML) systems are from human coders’ working performance, as well as the important question of how well current explainability methods apply to advanced neural network models such as transformers. This is to predict correct codes and present references in clinical notes that support code prediction, as this level of explainability and accuracy of the prediction outcomes is critical to gaining trust from professional medical coders.
## Explainable and Accurate Medical Codes Prediction
RAC Model and Human-Level Coding Baselines: We first present our R ead, A ttend, and C ode (RAC) model to learn the assignment mappings of medical codes that can process unstructured clinical notes and attend to text areas annotating medical codes. The main building blocks of RAC are built upon by connecting convolved embeddings with Transformer-encoder based self-attention and code-title guided attention modules. The architecture details can be found in Section 4 of [1]. In principle, we look at the medical codes prediction problem as a set-to-set assignment learning problem from the set of input sentence vectors to the set of code labels and employ the problem’s unique permutation equivariant property 2 in the design.
Second, the human coding baseline is estimated via a primitive web interface built using the Label Studio [7]. Given the reference diagnosis and procedure codes that Beth Israel Deaconess Medical Center (BIDMC) assigned, we have two professional CPC certified coders independently code the different set of notes and evaluate where their total annotations have differed from the references. The intention is to provide the same possible conditions as for the RAC system. We found that estimated inter-coder agreement rates (i.e., human coding baseline) were not as high as we initially thought and are far exceeded by the RAC system’s 3.9 times higher rate in Micro-Jaccard similarity. Section 3 of [1] conatins the evaluation design details and baseline results.
1 A human coder or health care provider scans the medical documentation in electronic health records, identifying essential information and annotating codes for that particular treatment or service. With a wide range of medical services and providers (primary care clinics, specialty clinics, emergency departments, mother-baby units, outpatient and inpatient units, etc.), the complexity of human coders’ tasks increases as the medical industry advances, while productivity standards decrease as charts take more time to review.
2 Consider permuting the notes’ sentence vectors; medical code outputs will be the same but permuted.
| |
190800995/5
|
In Fig. 2, we show the numerically obtained value of $\overline{C}_\infty M/4$ for various $k$ and $q$. The alternating pattern ex- pected from Eq. (13) is clearly visible [panels (a) to (c)]. While the numerical value of the nonzero plateau differs from expectation when $q \ll k$ (and $k-q \ll k$, not shown), Eq. (13) gives an accurate prediction when $\binom{k}{q}/\binom{k}{\lfloor k/2 \rfloor}$ is close to one (with $\lfloor \ldots \rfloor$ the floor function). To investi- gate this further, in panels (d) and (e), we show $C_\infty M/4$ versus $k$. The growth with $k$ is slower for $q=5$ than for $q=3$, which is in turn slower than the almost linearly growing $q=1$ case [46].
To summarize, we have shown that supersymmetry is (almost) always present in the SYK model with generic four-body interactions. It is only absent in those classes without particle-hole symmetry, i.e., in classes AI and AII. The type of SUSY, in particular the number $\mathcal{N}_\text{loc}$ of local supercharges and their symmetry properties fol- low a pattern that finds a natural interpretation when $\Gamma_{j\le \mathcal{N}_\mathrm{loc}}$ are viewed as emergent Majorana fermions in a one-dimensional topological phase with SYK model boundary physics. These SUSY features all link directly to features in time-dependent correlation functions of fermion-parity-odd observables. For $q$ -body retarded cor- relation functions, this includes the shape of the ramp in the short-time regime, due to a link between $\mathcal{N}_\text{loc}$ and the Dyson index $\beta$; and the value of the long-time plateau due to the imprint of how $\Gamma_j$ transforms under $T_\pm$ on its mi- croscopic Majorana structure. These $q$ -body correlation functions, even with large $q$, can be measured in digital quantum simulation of the SYK model [53]. The single- particle Green’s function ($q=1$) is accessible through scanning tunneling spectroscopy [20, 21]. We stress that the features in the correlation functions are dynamical consequences of SUSY, which are less frequently consid- ered than ground-state consequences [54, 55].
We thank David Tong for helpful discussions. This work was supported by the ERC Starting Grant No. 678795 TopInSy.
## Appendix A: Eightfold symmetry classification of the SYK model
In the main text, we make extensive use of the eight- fold symmetry classification of the SYK model with four-body interactions. This classification, introduced in Ref. 46, reveals an Altland-Zirnbauer structure behind the period-eight pattern of Wigner-Dyson classes high- lighted in Ref. 22. The considerations of both Refs. 22 and 46 are based on the work of Fidkowski and Kitaev on the topological classification of interacting fermions in one dimension [45], which Ref. 22 noted to apply to the SYK model upon viewing it as existing at the end of a one-dimensional fermionic topological phase. Here we summarize the key ideas in Ref. 46 and highlight certain aspects relevant for the main body of the paper.
The eightfold symmetry classification is based on the presence of antiunitary symmetries and their relation to fermion parity. First note that the SYK Hamilto- nian [Eq. (2) in the main text] conserves fermion parity, $[H,P]=0$ with
$$
P =
\begin{cases}
i^{k/2} \gamma_1 \gamma_2 \cdots \gamma_k & \text{even $k$} \\
i^{(k+1)/2} \gamma_1 \gamma_2 \cdots \gamma_k \gamma_\infty & \text{odd $k$}
\end{cases}
$$
where the Majorana $\gamma_\infty$ ensures a well-defined fermion parity for odd $k$, but does not contribute to local opera- tors, including the Hamiltonian. All Majorana operators including $\gamma_\infty$ are odd in fermion parity and thus satisfy $\{P,\gamma_q \} =0$. When $k$ is odd, the Hamiltonian also com- mutes with the operator
$$
Z = - i^{(k-1)/2} \gamma_1 \gamma_2 \cdots \gamma_k
$$
that in turn commutes with all local Majorana opera- tors $[Z,\gamma_{q \neq \infty}]=0$, but anticommutes with the additional Majorana at infinity $\{Z,\gamma_{\infty} \}=0$ and accordingly with fermion parity $\{Z, P\} = 0$.
Depending on the number of Majorana operators, it may be possible to find local unitary operators $C_\pm$ that satisfy [56]
$$
C_\pm \gamma_{q \neq \infty}^* C_\pm^\dagger = \pm \gamma_{q \neq \infty} .
$$
In other words, the antiunitary operator $T= C_+ \mathcal{K}$ commutes with all local Majorana operators, whereas $A = C_- \mathcal{K}$ anticommutes with them.
For even $k$, it is always possible to find both operators $C_+$ and $C_-$ [56]. Using Eq. (A1) we conclude
$$
T P T^{-1} = (-1)^{k/2} P,
$$
i.e., $T$ commutes with fermion parity when $k\mod 4 = 0$ and anticommutes with fermion parity when $k\mod 4 = 2$. Thus, when both operators commute, $T$ is an antiu- nitary operator that maps each parity subblock of the Hamiltonian to itself, otherwise, the parity subblocks are exchanged by $T$. We refer to the first case as time- reversal symmetry: The level spacing statistics of each subblock are determined by the presence of $T$ and by the sign $T^2 = \pm 1$. We refer to the second case as particle-hole symmetry, as the parity sectors are exchanged. Particle- hole symmetry sets correlations across different parity sectors, a feature we exploit in the main text. In the fol- lowing, we distinguish the two cases by the notation $T_+$ for time-reversal ($[T_+,P]=0$) and $T_-$ for particle-hole symmetry ($\{T_-,P \} =0$), the same notation we use in the main text.
For odd $k$, only one of two operators $C_\pm$ is local. In particular, when $k =4n+1$, only $C_+$ is local, and when $k = 4n+3$, only $C_-$ is local (with integer $n$). When only $C_-$ is local, we can use fermion parity to define $T= P C_- \mathcal{K}$, an antiunitary operator that commutes with
| |
190800995/7
|
Ensemble-averaged $q$ -body correlation function in symmetry classes AI [$k \mod 8 =0$, panels (a) and (b)] and AII [$k \mod 8 =4$, panels (c) and (d)]. The color code de- notes the different values of $k$ and $q$; cf. the inset in panel (a). Here, $q=5$ serves is a representative example for $q=4n+1$, and $q=3$ for $q=4n+3$. The correlation function decays to zero in all cases shown here. Only when $k=12$ [panels (c) and (d)] do the two parity sectors show correlations, which is a small-size effect. All results are averaged over a large ensemble, ranging from $768$ (for $k=20$, $q=5$) to $2^{17}$ (for $k=8$) realizations of the couplings $J_{qrst}$. Error bars are ei- ther smaller than the line width (for small $k$) or smaller than the disorder-induced fluctuations of the lines (for large $k$).
For completeness, we give these operators in terms of the eigenstates. To give explicit relations, we need to fix certain phase relations between the eigenstates. Defining time-reversal symmetry as
$$
T_+ \ket{\psi_{2\mu+1}^+} = \ket{\psi_{2\mu}^+}
$$
fixes the phase relation between Kramers doublets that we denote by $\ket{\psi_{2\mu}^p}$ and $\ket{\psi_{2\mu+1}^p}$. This choice sets $\varepsilon_{2\mu} = \varepsilon_{2\mu+1}$. We fix the phase between opposite parities by choosing
$$
Z \ket{\psi_{\mu}^p} = \ket{\psi_{\mu}^{-p}} .
$$
In this basis, the supercharges defined above read
$$
\begin{aligned}
\Gamma_1 &= \sum_{p\mu} \left(\ket{\psi_{2\mu}^p} \bra{\psi_{2\mu+1}^{-p}} + \ket{\psi_{2\mu+1}^p} \bra{\psi_{2\mu}^{-p}} \right) \\
\Gamma_2 &= -i \sum_{p\mu} \left(\ket{\psi_{2\mu}^p} \bra{\psi_{2\mu+1}^{-p}} - \ket{\psi_{2\mu+1}^p} \bra{\psi_{2\mu}^{-p}} \right) \\
\Gamma_3 &= \sum_{p\mu} \left(\ket{\psi_{2\mu}^p} \bra{\psi_{2\mu}^{-p}} - \ket{\psi_{2\mu+1}^p} \bra{\psi_{2\mu+1}^{-p}} \right) .
\end{aligned}
$$
Their product is accordingly
$$
\begin{aligned}
\Gamma_4
&= -i \Gamma_1 \Gamma_2 \Gamma_3
= \sum_{p\mu} \ket{\psi_{\mu}^p} \bra{\psi_{\mu}^{-p}} = Z
\end{aligned}
$$
Ensemble-averaged $q$ -body correlation function in symmetry classes BDI ($k\mod 8=1$) and CII ($k\mod 8 = 5$). In panels (a) and (b), we show different system sizes in class BDI, where the different colors denote $k$ and $q$; cf. inset in panel (d). As in Fig. 3, $q=5$ serves is a representative example for $q=4n+1$, and $q=3$ for $q=4n+3$. In panels (c) and (d), we show different systems sizes in class CII. The dotted lines show the expectation for $\overline{C}_{q,\infty}$ based on random matrix theory; cf. Eq. (13) in the main text. All results are averaged over a large ensemble, ranging from $192$ (for $k=21$, $q=5$) to $2^{14}$ (for $k=9$) realizations of the couplings $J_{qrst}$. Similarly to Fig. 3, error bars are either smaller than the line width (for small $k$) or smaller than the disorder-induced fluctuations of the lines (for large $k$).
and the nonlocal supercharge
$$
\begin{aligned}
\Gamma_\infty &= -i \sum_\mu \left(\ket{\psi_{\mu}^+} \bra{\psi_{\mu}^{-}} - \ket{\psi_{\mu}^-} \bra{\psi_{\mu}^{+}} \right) .
\end{aligned}
$$
Writing the fermion parity $P$ in terms of the eigenstates
$$
\begin{aligned}
P &= \sum_\mu \left(\ket{\psi_{\mu}^+} \bra{\psi_{\mu}^{+}} - \ket{\psi_{\mu}^-} \bra{\psi_{\mu}^{-}} \right)
\end{aligned}
$$
and using $P=-i Z \gamma_\infty$ confirms that $\Gamma_\infty \equiv \gamma_\infty$.
## Appendix C: Correlation functions
a. Symmetry classes AI and AII. Classes AI ($k \mod 8 = 0$) and AII ($k \mod 8 =4$) are the two classes without SUSY. In these classes, parity sectors are un- correlated, hence, upon evaluating Eq. (9) of the main text, we expect to see neither a ramp structure nor a plateau at long times $t \gg 1/\Delta_\infty$. To support this expec- tation with numerical evidence, we show the ensemble- averaged $q$ -body correlation function at infinite tempera- ture [Eq. (10) in the main text] in Fig. 3. In Fig. 3 (a), we show the $(q=5)$ -body correlation function for $k=8,16$
| |
190800995/2
|
\begin{tabular}{c|cccccccc}
\rightarrowprule
$k \mod 8$ & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 \colrule
Label & AI & BDI & D & DIII & AII & CII & C & CI \colrule
$T_+^2$ & $+1$ & $+1$ & $0$ & $-1$ & $-1$ & $-1$ & $0$ & $+1$ $T_-^2$ & $0$ & $+1$ & $+1$ & $+1$ & $0$ & $-1$ & $-1$ & $-1$ \botrule
\end{tabular}
TABLEI. Time-reversal symmetry $T_+$ and particle-hole sym- metry $T_-$ in the SYK model. The symmetries may be absent (denoted by $0$), or present and square to $-1$ or $+1$.
Wigner-Dyson way highlighted in Ref. 22. In fact, as we now briefly review, it gives rise to the eight real Altland-Zirnbauer classes. For even $k$, either time- reversal symmetry $T_+$ or particle-hole symmetry $T_-$ is present. For odd $k$, both $T_+$ and $T_-$ are present; in this case $T_+ \gamma_{\infty} T_+^{-1} = (-1)^{(k+1)/2}\gamma_{\infty}$. Their product, the unitary operator $Z= T_+T_-$, equals the product of all local Majorana operators up to a complex phase and corresponds to a chiral symmetry [45, 46]. A key feature of $Z$, which we will use repeatedly for diagnosing local- ity, is $[Z,\gamma_{q\neq\infty}]=0$ and $\{Z,\gamma_\infty\}=0$. The squares $T_\pm^2$ vary with $k$ and label the eight real Altland-Zirnbauer classes [48]. While the Altland-Zirnbauer and Wigner- Dyson picture give the same level spacing statistics, the former also takes cross-parity correlations into account. We summarize the symmetry classification in Table I and review it in detail in Appendix A.
SUSY is known to imply a degeneracy between the par- ity sectors [36]: the supercharges $Q_a$ exchange bosonic states with parity eigenvalues $p=+1$ and fermionic states with parity eigenvalue $p=-1$ [36]. Thus, the su- percharges anticommute with fermion parity, $\{P, Q_a \} =0$. The presence of particle-hole symmetry also guar- antees degeneracy between parity sectors, which as we now note, also implies SUSY. Parity degeneracy directly follows from particle-hole symmetry because $\ket{\psi^p_\mu}$ and $T_- \ket{\psi^p_\mu}$ have the same energy $\varepsilon_\mu=\varepsilon_\mu^p= \varepsilon_\mu^{-p}$ (since $[T_-,H]=0$), but opposite parity ($\lbrace T_-,P \rbrace = 0$). Therefore, $\ket{\psi_\mu^p} \bra{\psi_\mu^{-p}}$ is an odd-parity zero mode, i.e., an operator that commutes with the Hamiltonian, but anticommutes with fermion parity [46]. This in turn implies SUSY: The operator $\tilde{Q}_\mu = \sqrt{\varepsilon_\mu} \ket{\psi_\mu^+} \bra{\psi_\mu^{-}}$ sat- isfies $\tilde{Q}_\mu \tilde{Q}_\mu^\dagger = \varepsilon_\mu \ket{\psi_\mu^+} \bra{\psi_\mu^+}$ and $\tilde{Q}_\mu^\dagger \tilde{Q}_\mu = \varepsilon_\mu \ket{\psi_\mu^{-}} \bra{\psi_\mu^{-}}$, and hence the linear combinations $Q_{1,\mu} = \tilde{Q}_\mu + \tilde{Q}_\mu^\dagger$ and $Q_{2,\mu} = i (\tilde{Q}_\mu - \tilde{Q}_\mu^\dagger)$ are Hermitian, anticommute, and square to $\varepsilon_\mu$ times the projector on the two parity- degenerate states. Consequently, the two supercharges
$$
\begin{aligned}
Q_1 = \sum_\mu (\tilde{Q}_\mu + \tilde{Q}_\mu^\dagger) , & & Q_2 = -i \sum_\mu (\tilde{Q}_\mu - \tilde{Q}_\mu^\dagger)
\end{aligned}
$$
satisfy Eq. (1) and anticommute with $P$. Particle-hole symmetry is present unless $k=4n$. Thus, all but two of the symmetry classes are supersymmetric.
Given the presence of six supersymmetric classes, there are a number of questions regarding the interplay of SUSY and the symmetry classification. How does $\mathcal{N}$ depend on the symmetry class? How do $Q_j$ transform under $T_\pm$ and how does this translate to the structure of the supercharges? We next turn to these questions.
We start with counting $\mathcal{N}$. A direct approach is based on counting level degeneracies. This follows from the ob- servation that the “ spectrally flattened ” Hermitian su- percharges $\Gamma_{j}=Q_{j}/\sqrt{H}$ satisfy
$$
\begin{aligned}
\{\Gamma_{j},\Gamma_{k}\} = 2\delta_{jk}, & & [H,\Gamma_{k}] = 0 , & & \{P,\Gamma_{k}\}=0.
\end{aligned}
$$
They are thus many-body zero mode forms of Majorana fermions. An even $\mathcal{N}$ of such zero modes give rise to a $2^{\mathcal{N}/2}$ -dimensional fermionic degeneracy space for each of the $\varepsilon_\mu$ with one of the $|\psi_\mu^p\rangle$ chosen as “ vacuum ”. (With a suitable choice, the $\Gamma_j$ -fermion parity of an eigenstate matches the state’s physical fermion parity.) This pro- cedure is similar in spirit to the standard construction of supermultiplets [49]. For the six supersymmetric SYK classes, a twofold degeneracy is guaranteed by $T_-$ and a further twofold (Kramers) degeneracy is present when- ever $T_+^2=-1$, resulting in an overall fourfold degeneracy. This suggests $\mathcal{N}=2$, except for DIII and CII where this count gives $\mathcal{N}=4$. What this counting does not address is how many $\Gamma_j$ (and hence $Q_j$) are local to the SYK model. Next we investigate this to obtain the decompo- sition $\mathcal{N}=\mathcal{N}_\text{loc}+\mathcal{N}_\infty$ with $\mathcal{N}_\text{loc}$ counting the number of supercharges involving only $\gamma_{q\neq\infty}$. We first discuss the symmetry classes D and C before demonstrating the im- plications of locality in classes BDI and CI. For brevity, we derive the supercharges in classes DIII and CII with $T_-^2 = -1$ in Appendix B and only summarize the results here.
We begin with classes D and C. Here $k$ is even hence all $\gamma_q$ are local. Therefore, our argument above applies directly: we find $\mathcal{N}=\mathcal{N}_\text{loc}=2$. All the other supersym- metric classes have odd $k$, thus potentially $\mathcal{N}\neq\mathcal{N}_\text{loc}$ due to $\gamma_\infty$. As we shall see, in all of these classes $\mathcal{N}=\mathcal{N}_\text{loc}+1$ with $\mathcal{N}$ following its degeneracy-based value above. This is intuitive because $\gamma_\infty\equiv \Gamma_\infty$ auto- matically satisfies Eq. (5) (in particular, it anticommutes with any local parity-odd operator), thus $\mathcal{N}_\text{loc}$ is at most $\mathcal{N}-1$. To formally establish $\mathcal{N}_\text{loc}$, and the transforma- tion of $\Gamma_{j}$ under $T_\pm$, we work in the energy eigenbasis, $H = \diag (\lbrace \varepsilon_\mu \rbrace) \otimes \openone_{2^{\mathcal{N}/2}}$, with $P = \openone_{M/2}\otimes\tau_3$. (Here and below, $\tau_j$ and $\sigma_j$ are Pauli matrices; $\tau_j$ act in parity grading and $\sigma_j$ in the space of Kramers doublets, where applicable. We will often omit trivial tensor factors.) In this basis, class D (C) has particle-hole symmetry [up to a phase $\diag (\lbrace \exp(i\varphi_\mu) \rbrace)$ omitted here and be- low] $T_- = \tau_{1(2)} \mathcal{K}$ (with $\mathcal{K}$ for complex conjugation); this follows from $T_-^2=\pm1$ and parity being the only degener- acy, $\openone_{2^{\mathcal{N}/2}}=\tau_0$. We have $\Gamma_{1,2}=\tau_{1,2}$, which correspond to the two supercharges introduced in Eq. (4) [50].
To study classes BDI and CI, we focus on a degeneracy space with energy $\varepsilon_\mu$ and first establish the form of $T_\pm$ and thus $Z$ in this space. $T_+^2=+1$ implies that parity is again the only degeneracy, so $T_- = \tau_{1(2)} \mathcal{K}$ in class BDI (CI). $T_+\ket{\psi_\mu^p} \propto \ket{\psi_\mu^p}$ implies that the most general form is $T_+=\exp(i\varphi_{\mu}\tau_{3})\mathcal{K}$. With a suitable choice of
| |
190800995/6
|
all fermions; when $C_+$ is local, we can use $T=C_+ \mathcal{K}$. This implies that $T$ and $P$ commute
$$
\begin{aligned}
T P T^{-1}
&= (-1)^{(k+1)/2} \gamma_1 \gamma_2 \cdots \gamma_k T \gamma_\infty T^{-1} \\
&= P,
\end{aligned}
$$
where we chose $T \gamma_\infty T^{-1} = (-1)^{(k+1)/2} \gamma_\infty$ for conve- nience. We thus identify $T = T_+$. Due to the presence of $Z$, we can define a second antiunitary operator $T_-$ via $T_- = T_+^{-1} Z$ such that $Z = T_+ T_-$ as quoted in the main text. Since $Z$ anticommutes with fermion parity, $T_-$ also anticommutes with it, $\{T_- ,P \} =0$.
The squares of $T_+$ and $T_-$ can be determined from their explicit construction in terms of Majorana operators [45]. Without loss of generality, we choose the Majorana oper- ators such that $\gamma_{2n+1}^* = \gamma_{2n+1}$ are real while $\gamma_{2n}^* = -\gamma_{2n}$ are purely imaginary.
When $k$ is even, the basis choice introduced above re- sults in
$$
C_+ =
\begin{cases}
\gamma_2 \gamma_4 \cdots \gamma_{k} & ~ k = 4 n \\
\gamma_1 \gamma_3 \cdots \gamma_{k-1} & ~ k = 4 n +2 ,
\end{cases}
$$
thus, $C_+$ is always the product of $k/2$ Majorana oper- ators. In both cases, $C_+ = C_+^*$ is real ($C_+$ is either the product of only real Majorana operators or the product of an even number of purely imaginary operators). When $k=4n$, we have
$$
\begin{aligned}
T_+^2 = C_+^ 2 = \gamma_2\gamma_4 \cdots \gamma_k \gamma_2 \gamma_4 \cdots \gamma_k ,
\end{aligned}
$$
giving $T_+^2 = (-1)^{k/4}$, and when $k=4n+2$,
$$
\begin{aligned}
T_+^2 = C_+^ 2 = \gamma_1\gamma_3 \cdots \gamma_{k-1} \gamma_1 \gamma_3 \cdots \gamma_{k-1} ,
\end{aligned}
$$
which gives $T_+^2 = (-1)^{(k-2)/4}$.
When $k$ is odd, we need to distinguish the two cases $k=4n+1$ and $k=4n+3$. For $k=4n+1$ we find
$$
C_+ = \gamma_2 \gamma_4 \cdots \gamma_{k-1} .
$$
Since $k$ is odd, the nonlocal Majorana $\gamma_\infty^* = -\gamma_\infty$ is purely imaginary, giving
$$
C_+ \gamma_\infty^* C_+^\dagger = -(-1)^{(k-1)/2} \gamma_\infty = (-1)^{(k+1)/2} \gamma_\infty
$$
where we used that $C_+$ contains $(k-1)/2$ Majorana operators $\gamma_{q \neq \infty}$. Accordingly, we identify $T_+ = C_+ \mathcal{K}$ and conclude $T_+^2 = (-1)^{(k-1)/4}$. We further identify $T_- = T_+^{-1} Z$ with $T_-^2 = T_+^2 T_+ Z T_+^{-1} Z$. Using $T_+ Z T^{-1} = (-1)^{(k-1)/2} Z = Z$ and $Z^2 = 1$, we realize that $T_-^2 = T_+^2$.
For $k = 4n+3$ we find
$$
C_- = \gamma_1 \gamma_3 \cdots \gamma_{k} ,
$$
which is a product of $(k+1)/2$ Majorana operators. Since $C_- \mathcal{K}$ and $P$ both anticommute with all local Ma- jorana operators, the antiunitary operator $P C_- \mathcal{K}$ com- mutes with them. Again, $\gamma_\infty^* = - \gamma_\infty$, which gives
$$
P C_- \mathcal{K} \gamma_\infty (P C_- \mathcal{K})^{-1} = (-1)^{(k+1)/2} \gamma_\infty,
$$
where we used that $C_-$ is the product of $(k+1)/2$ local Majorana operators and $\{P,\gamma_\infty \} = 0$. Thus, we iden- tify $T_+ = P C_- \mathcal{K}$. The square $T_+^2 = P C_- P^* C_- = (-1)^{k+1} C_-^2$ and $C_-^2 = (-1)^{(k+1)/4}$, such that $T_+^2 = (-1)^{(k+1)/4}$. Accordingly, $T_-^2 = T_+^2 T_+ Z T_+^{-1} Z = (-1)^{(k-1)/2} T_+^2 = - T_+^2$.
We summarize these results in Table I in the main text.
## Appendix B: Supercharges in classes DIII and CII
In this section, we construct the supercharges in classes DIII and CII along the lines of the same strategy that we followed in the main text for the other classes. We show and discuss how the supercharge structure summarized in the main text arises in these classes.
In classes DIII and CII, time reversal symmetry with $T_+^2 = -1$ implies Kramers degeneracy: The states $\ket{\psi_\mu^p}$ and $T_+ \ket{\psi_\mu^p}$ are orthogonal. Together with parity degen- eracy this gives a fourfold degeneracy of energy eigenval- ues and $\mathcal{N}=4$. In its diagonalized form, the Hamiltonian reads $H=\diag (\{\varepsilon_\mu \}) \otimes \openone_{2^{\mathcal{N}/2}}$ with $\openone_{2^{\mathcal{N}/2}}=\tau_0\sigma_0$, where the matrices $\tau_\mu$ act in parity grading space and $\sigma_\nu$ in the space of Kramers doublets (hence $\tau_\mu \sigma_\nu$ is a shorthand for the Kronecker product $\tau_\mu \otimes \sigma_\nu$; additional tensor prod- ucts with the identity are implied).
Without loss of generality, we can choose $Z=\tau_1$. The four $\Gamma_j$ satisfying Eq. (5) in the main text are now the familiar Dirac matrices, chosen as $\Gamma_j=\tau_1\sigma_j$ ($j=1,2,3$) and $\Gamma_\infty=\tau_2$. Checking the (anti)commutation with $Z$ we indeed find that only $\Gamma_{1,2,3}$ are local. Conversely, identifying $\Gamma_\infty\equiv \gamma_\infty$ and requiring that it transforms under $T_+$ as $T_+ \gamma_\infty T_+^{-1} = (-1)^{(k+1)/2} \gamma_\infty$ sets
$$
T_+= \begin{cases}
i\tau_3\sigma_2\mathcal{K} & \text{class DIII} \\
i\sigma_2\mathcal{K} & \text{CII}
\end{cases}
$$
in this basis. The form of particle-hole symmetry $T_-$ follows from $T_+ T_-= Z$.
The product $\Gamma_4 = -i \Gamma_1 \Gamma_2 \Gamma_3$ is also Hermitian, par- ity odd, local, and linearly independent of $\Gamma_{1,2,3}$. Fur- thermore, $Q_4= \diag (\lbrace \sqrt{\varepsilon_\mu} \rbrace) \otimes \Gamma_4$ squares to the Hamil- tonian. It is, however, not a supercharge because $\Gamma_4$ does not anticommute with $\Gamma_{1,2,3}$. (Nevertheless, $\Gamma_4$ con- tributes to correlation functions, as discussed in the main text.) We thus find $\mathcal{N}_\text{loc}=3$.
Using the explicit form of $T_+$ obtained above and $T_- = T_+^{-1} Z$, we find
$$
T_\pm \Gamma_{j \le \mathcal{N}_\mathrm{loc}} T_\pm^{-1} =
\begin{cases}
+\Gamma_{j \le \mathcal{N}_\mathrm{loc}} & \text{class DIII} \\
-\Gamma_{j \le \mathcal{N}_\mathrm{loc}} & \text{class CII.}
\end{cases}
$$
This implies that
$$
T_\pm \Gamma_4 T_\pm^{-1} = - \Gamma_4
= \begin{cases}
- \Gamma_4 & \text{class DIII} \\
+ \Gamma_4 & \text{class CII.}
\end{cases}
$$
| |
190800995/4
|
$q$ -body
time-dependent correlation function at infinite temperature, averaged over an ensemble of up to $2^{16}$ Gaussian distributed $J_{qrst}$, for classes (a) D, (b) C, (c) DIII, and (d) CI. The different colors denote different $k$ and $q$, cf. inset in (a), the dashed lines represent $q=3$ and the solid lines $q=5$. The ramp shape follows the Dyson index $\beta$ and hence links to the number of supercharges. The long-time plateau $\overline{C}_{q,\infty}$ is studied in Fig. 2. Error bars are either smaller than the line width (for small $k$) or smaller than the disorder-induced fluctuations of the lines (for large $k$).
time-dependent $q$ -body correlation function
$$
C_{q}^+ (t) = - i \Theta (t) \frac{1}{\binom{k}{q}} \sum_{a,n_a =q} \langle \lbrace \Upsilon_a (t) , \Upsilon_a (0) \rbrace \rangle ,
$$
where $\langle \ldots \rangle$ denotes thermal average. Although the sig- natures we reveal are present at any temperature, we find an especially transparent relationship at infinite temper- ature, where the correlation function reads
$$
\begin{aligned}
C_{q}^+ (t) =& -i \Theta (t) \frac{1}{\binom{k}{q}} \sum_{a,n_a =q} \frac{1}{M} \sum_{p\mu\nu} \left| \bra{\psi_\mu^p} \Upsilon_a \ket{\psi_\nu^{-p}} \right|^2 \\
& \times 2 \cos \left(t \left(\varepsilon_\mu^p - \varepsilon_\nu^{-p} \right) \right).
\end{aligned}
$$
When $t \gg 1/\Delta_\infty$, terms with $\varepsilon_\mu^p \neq \varepsilon_\nu^{-p}$ give a quickly os- cillating contribution $\delta C_q^+ (t)$ that averages to zero. Only states with $\varepsilon_\mu^p = \varepsilon_\nu^{-p}$ give a time-independent contribu- tion $C_{q,\infty}$. Thus, $C_q^+ (t) = -i \Theta(t) [C_{q,\infty} + \delta C_q^+ (t)]$ with
$$
\!\!\!\! C_{q,\infty}=\frac{1}{\binom{k}{q}}\sum_{a,n_{a}=q}\frac{2}{M}\sum_{\mu} \tr \Upsilon_{a\mu}^{2},\quad\Upsilon_{a\mu}=P_{\mu}\Upsilon_{a}P_{\mu},
$$
where in converting the equal-energy sum to a trace, we introduced the projection $P_{\mu}$ to the eigenspace with energy $\varepsilon_{\mu}$ and used that $\Upsilon_{a}$ is Hermitian and parity odd.
Next we convert Eq. (11) into a sum over $\Gamma_{j<\infty}$. We start by expanding $\Upsilon_{a\mu}=\sum_{j<\infty}y_{j}P_{\mu}\Gamma_{j}$ (with real $y_j$)
Normalized plateau $\overline{C}_{q,\infty} M/4$ of the $q$ -body correla- tion function, averaged over an ensemble of up to $2^{14}$ Gaussian distributed $J_{qrst}$. The color encodes the number $k$ of Majo- ranas, cf. panels (d) and (e). In all classes, $\overline{C}_\infty M/4$ alternates with $q$ approximately as predicted in Eq. (13); the agreement is excellent when $\binom{k}{q}/\binom{k}{\lfloor k/2 \rfloor}$ is close to one. In panel (d), we show that $\overline{C}_\infty M/ (4c)$ [with $c$ the random matrix expectation based on Eq. (13)] increases as a function of $k$, but with a rate that decreases upon increasing $q$ [panel (e)]. Statistical error bars are smaller than the marker size.
which holds as within an eigenspace, the (projected) operators $P_{\mu}\Gamma_{j<\infty}$ form a basis for local, Hermitian, parity-odd operators. If $\Upsilon_{a}$ transforms the same (op- posite) way to $\Gamma_{j}$ under $T_{\pm}$ then generically $y_{j}\neq0$ ($y_{j}=0$). Now using the trace-orthogonality of the $\Gamma_{j<\infty}$ and $\tr \Gamma_{j}^{2} = 2^{\mathcal{N}/2}=\mathcal{N}$ (for $\mathcal{N}=2,4$), we find
$$
C_{q,\infty}=\frac{1}{\binom{k}{q}} \sum_{a,n_{a}=q} \frac{2}{M\mathcal{N}} \sum_{\mu} \sum_{j<\infty} \left[\tr(P_{\mu}\Upsilon_{a}P_{\mu}\Gamma_{j}) \right]^{2}.
$$
A simple estimate for $C_{q,\infty}$ can be given assuming that expanding $P_\mu\Gamma_{j<\infty}=\sum_{a} v_{\mu j,a}\Upsilon_{a}$ results in random co- efficients $v_{\mu j,a}$ subject only to normalization and symme- try constraints. Denoting such a random vector average by $\overline{(\ldots)}$, we find
$$
\frac{\overline{C}_{q,\infty} M}{4}=\begin{cases}
\frac{\mathcal{N}}{\beta}\mathcal{N}_{\text{loc}} & q:\,\Upsilon_{a}\triangleq\Gamma_{j\leq\mathcal{N}_{\text{loc}}},\\
\frac{\mathcal{N}}{\beta}\delta_{\beta,4} & \text{otherwise},
\end{cases}
$$
where $\Upsilon_{a} \triangleq \Gamma_j$ here means that $\Upsilon_{a}$ transforms the same way as $\Gamma_j$ under $T_\pm$. Thus, each $\Gamma_{j<\infty}$ give the same con- tribution to $\overline{C}_{q,\infty}$ when they contain $q$ -Majorana terms and zero otherwise. The nonzero value for $\beta=4$ when $\Upsilon_{a}\not\triangleq \Gamma_{j\leq \mathcal{N}_\text{loc}}$ arises due to $\Gamma_4$ since $\Gamma_4\not\triangleq\Gamma_{j\leq \mathcal{N}_\text{loc}}$. Con- sidering the Majorana structure of $\Gamma_j$ in Table II, Eq. (13) translates to an alternating pattern of $\overline{C}_{q,\infty}$ as $q$ is var- ied in a given symmetry class, with complementary $\overline{C}_{q,\infty}$ values for classes with opposite $s\mathcal{N}_\text{loc}$.
| |
190800995/1
|
# Supersymmetry in the nonsupersymmetric Sachdev-Ye-Kitaev model
Jan Behrends 1 and
Benjamin B ´e ri 1,2
Supersymmetry is a powerful concept in quantum many-body physics. It helps to illuminate ground state properties of complex quantum systems and gives relations between correlation func- tions. In this work, we show that the Sachdev-Ye-Kitaev model, in its simplest form of Majorana fermions with random four-body interactions, is supersymmetric. In contrast to existing explicitly supersymmetric extensions of the model, the supersymmetry we find requires no relations between couplings. The type of supersymmetry and the structure of the supercharges are entirely set by the number of interacting Majorana modes, and are thus fundamentally linked to the model’s Altland- Zirnbauer classification. The supersymmetry we uncover has a natural interpretation in terms of a one-dimensional topological phase supporting Sachdev-Ye-Kitaev boundary physics, and has conse- quences away from the ground state, including in $q$ -body dynamical correlation functions.
The Sachdev-Ye-Kitaev (SYK) model [1, 2] is a toy model that provides insight into diverse physical phe- nomena, ranging from the holographic principle [3 – 5] to quantum chaos [6 – 11], and non-Fermi liquid behavior of strongly correlated electron systems [12 – 18]. Similar to black holes, it is believed to scramble quantum informa- tion with maximal efficiency [17, 19].
The simplest variant of the SYK model describes $k$ Majorana fermions that interact through a random four- body term [2]. Its proposed physical realizations include mesoscopic systems based on Majorana fermions in vor- tices or quantum dots [20, 21], or the ends of a one- dimensional topological phase [22].
Various generalizations of the SYK model have been considered, including models with $n$ -body inter- actions [23, 24] and supersymmetric extensions [25 – 29]. Typically, exact supersymmetry (SUSY) requires fine- tuning of the parameters [30 – 33]. In the supersymmetric SYK extensions, this fine-tuning corresponds to requiring certain relations between different couplings [25].
In this work, we show that already the simplest four- body SYK model, without any fine-tuning, is supersym- metric for all but two values of $k \mod 8$. The type of SUSY depends only on $k$. The supercharges will be shown to relate to ramps and long-time plateaus in time- dependent correlation functions [34], which thus provide signatures of SUSY far from equilibrium. In particu- lar, we find that the number of supercharges is linked to the presence and nature of time-reversal symmetry and is reflected in the ramp shape [35]. We also show that the number and structure of supercharges set the plateau features in $q$ -body time-dependent correlation functions.
Throughout this work, we focus on SUSY in the sense of supersymmetric quantum mechanics [30, 31, 36 – 42]. SUSY is characterized by $\mathcal{N}$, the number of mutually anticommuting Hermitian fermionic supercharges that square to the Hamiltonian [36]
$$
\begin{aligned}
\lbrace Q_a , Q_b \rbrace = 2 H \delta_{ab} , & & [H,Q_a] = 0 .
\end{aligned}
$$
The Hamiltonian we consider describes four-body in- teractions between $k$ Majorana modes [2]
$$
H = \sum_{t =0}^{k-1} \sum_{s=0}^{t-1} \sum_{r=0}^{s-1} \sum_{q=0}^{r-1} J_{qrst} \gamma_q \gamma_r \gamma_s \gamma_t + E_0
$$
with real (as required by Hermiticity) but otherwise structureless couplings $J_{qrst}$, and the constant $E_0$ that ensures positive energies. The Hermitian Majorana op- erators $\gamma_q = \gamma_q^\dagger$ satisfy the anticommutation relation $\lbrace \gamma_q , \gamma_r \rbrace = 2 \delta_{qr}$, and span an $M$ -dimensional Hilbert space with $M=2^{\lceil k/2 \rceil}$. Since each term in the Hamiltonian (2) contains an even number of Majoranas, it conserves fermion parity $P$, given by
$$
P = \begin{cases}
i^{k/2} \gamma_1 \gamma_2 \ldots \gamma_k & \text{even $k$} \\
i^{(k+1)/2} \gamma_1 \gamma_2 \ldots \gamma_k \gamma_\infty & \text{odd $k$.}
\end{cases}
$$
To work in a Hilbert space with well-defined fermion parity, the additional Majorana $\gamma_\infty$ “ at infinity” must be included when $k$ is odd [45]. The operator $\gamma_\infty$ is not local to the SYK model; considering, e.g., a realization in a superconducting vortex [20], it represents a degree of freedom with support far away from the vortex where the SYK Majoranas $\gamma_{j\neq \infty}$ reside. Like the local Ma- joranas, $\gamma_\infty$ is Hermitian and satisfies $\{\gamma_q ,\gamma_r \} = 2 \delta_{qr}$. Since $[H,P]=0$, all eigenstates of $H$ can by labeled by their parity eigenvalue $p=\pm 1$, giving $H \ket{\psi_\mu^p} = \varepsilon_\mu^p \ket{\psi_\mu^p}$ and $P \ket{\psi_\mu^p} = p \ket{\psi_\mu^p}$.
The number of interacting Majorana modes, specif- ically $k\mod 8$, sets the model’s antiunitary symme- tries [22, 45 – 47]. These come in two variants $T_\pm$, antiu- nitary operators satisfying $T_\pm \gamma_{q\neq \infty} T_\pm^{-1} = \gamma_{q\neq \infty}$. They further satisfy $T_\pm P T_\pm^{-1}= \pm P$. We refer to $T_+$ as time- reversal symmetry because it commutes with fermion parity and hence sets correlations within a parity sector. Conversely, we call $T_-$ particle-hole symmetry. Crucially for this work, since $T_-$ flips fermion parity, its presence implies correlations between parity sectors.
The consideration of both $T_+$ and $T_-$ implies [46] a classification with more structure than the threefold
| |
190800995/3
|
the relative phases between the two parity sectors we can thus use $Z=T_+ T_-=\tau_1$; in this basis $T_+=\mathcal{K}$ ($T_+=\tau_3\mathcal{K}$) for class BDI (CI). The two $\Gamma_j$ satisfying Eq. can again be chosen as $\Gamma_{1,2}=\tau_{1,2}$. However, checking the (anti)commutation with $Z$ we find that only $\Gamma_1$ is local. Conversely, we can identify $\Gamma_2\equiv\Gamma_\infty\equiv \gamma_\infty$; this is consistent both with $\gamma_\infty$ itself satisfying Eq. (5) and its transformation under $T_+$. We thus find $\mathcal{N}_\text{loc}=1$.
In classes DIII and CII, we find $\mathcal{N}_\mathrm{loc} =3$ local super- charges as we show in detail in Appendix B. The spec- trally flattened supercharges can be written as Kronecker products $\Gamma_j =\tau_1 \sigma_j$. Their product $\Gamma_4 = -i \Gamma_1 \Gamma_2 \Gamma_3 = \tau_1$ is also local, but does not anticommute with $\Gamma_{j \le \mathcal{N}_\mathrm{loc}}$; it does, however, contribute to correlation function, as we discuss in the following. As in classes BDI and CI, the nonlocal supercharge is $\Gamma_\infty = \tau_2$.
The values $\mathcal{N}_\text{loc}$, together with the sign $s$ in $T_\pm \Gamma_{j\leq\mathcal{N}_\text{loc}}T_\pm^{-1} = s\Gamma_{j\leq\mathcal{N}_\text{loc}}$ have a natural interpretation if one views the SYK model as arising at the end of a one-dimensional topological phase in class BDI [22, 45]. These systems admit a $\mathbb{Z}_8$ classification: At one of their ends, they have $k_s$ Majoranas satisfying $T_\pm \gamma_q T_\pm^{-1}=s\gamma_q$; the topological index is $\nu= (k_+-k_-) \mod 8$. Thus, the eight topological classes can be labeled by $\nu = 0,1,2,3,4,-3,-2,-1$ with the integers counting the number and sign of unpaired Majoranas. In the SUSY classes, we find the same pattern for $s\mathcal{N}_\text{loc}$ against $k \mod 8$ ($T_\pm \gamma_{q\neq \infty} T_\pm^{-1}=\gamma_{q \neq \infty}$ implies $k_+=k$, $k_-=0$), see Table II. The $\mathcal{N}_\text{loc}$ supercharges $\Gamma_{j\leq\mathcal{N}_\text{loc}}$ can thus be viewed as the many-body emergence of the minimal num- ber and type of unpaired Majoranas consistent with $k$.
Next we turn to the structure of the supercharges in terms of the Majorana fermions $\gamma_q$. For this, we employ another operator basis of the Hilbert space, the products of $n_a$ Majorana operators $\gamma_{q\neq\infty}$
$$
\Upsilon_a = i^{n_a (n_a -1)/2} \gamma_{i_1 (a)} \gamma_{i_2 (a)} \cdots \gamma_{i_{n_a} (a)}
$$
with $i_j (a) \neq i_{j'} (a)$. $\Upsilon_a$ are Hermitian, unitary, and or- thonormal with respect to the trace, $\tr \left[\Upsilon_a \Upsilon_b \right] /M = \delta_{ab}$. In total, there are $2^{k}$ local operators $\Upsilon_a$ [51]. As we aim to expand $\Gamma_{j\neq \infty}$, i.e., Hermitian odd-parity operators in terms of $\Upsilon_a$, we use only those $\Upsilon_a$ with odd $n_a$, and use only real expansion coefficients.
Both time-reversal and particle-hole symmetry have the same (anti-) commutation properties when acting on $\Upsilon_a$. Since $T_\pm \gamma_q T_\pm^{-1} = \gamma_q$, only the phase of $\Upsilon_a$ [cf. Eq. (6)] may change when applying $T_\pm$, giving
$$
T_\pm \Upsilon_a T_\pm^{-1} = (-1)^{n_a (n_a -1)/2} \Upsilon_a.
$$
That is, $T_\pm$ and $\Upsilon_a$ commute when $n_a = 4n + 1$, and anti- commute when $n_a = 4n +3$. This, together with $v_{j,a}\in \mathbb{R}$ below, implies that, when expanding the supercharges,
$$
\Gamma_j = \sum_a v_{j,a} \Upsilon_a ,\quad \sum_a v_{j,a}^2=1,
$$
only terms with $n_a = 4 n+1$ contribute to $\Gamma_j$ when $[T_\pm ,\Gamma_j] = 0$, and only terms with $n_a = 4n+3$ contribute
\begin{tabular}{c|cccccc}
\rightarrowprule
$k \mod 8$ & 1 & 2 & 3 & 5 & 6 & 7 \colrule
Label & BDI & D & DIII & CII & C & CI $\beta$ & $1$ & $2$ & $4$ & $4$ & $2$ & $1$ \colrule
$s\mathcal{N}_\text{loc}$ & $1$ & $2$ & $3$ & $-3$ & $-2$ & $-1$ $\Gamma_{j\leq\mathcal{N}_\text{loc}}$ & $4n+1$ & $4n+1$ & $4n+1$ & $4n+3$ & $4n+3$ & $4n+3$ \colrule
$\Gamma_{4}$ & & & $4n+3$ & $4n+1$ & & \botrule
\end{tabular}
TABLE II. The Dyson index $\beta$, number $\mathcal{N}_\text{loc}$, signature $T_\pm \Gamma_{j\leq\mathcal{N}_\text{loc}}T_\pm^{-1}=s\Gamma_{j\leq\mathcal{N}_\text{loc}}$, and the Majorana fermion struc- ture of $\Gamma_{j\neq \infty}$. (The supercharges $Q_j$ have the same proper- ties, since T ± HT − 1 = H.) In the Majorana expansion of $\Gamma_j$, only those $\Upsilon_a$ with $n_a = 4n+1$ or $n_a = 4n+3$ contribute; the two options are shown in the last two rows of the table. The horizontal line visually distinguishes $\Gamma_4$ from the three supercharges because it does not anticommute with them. A blank entry indicates that $\Gamma_4$ does not exist in these classes.
when $\lbrace T_\pm,\Gamma_j \rbrace =0$. In classes DIII and CII we also consider $\Gamma_4=-i\Gamma_1\Gamma_2\Gamma_3$ whose transformation proper- ties follow from those of $\Gamma_{1,2,3}$. The resulting expansion structure is summarized in Table. II.
Having discussed the interplay of SUSY and the sym- metry classification, we now identify signatures of SUSY, $\mathcal{N}_\text{loc}$, and the supercharge structure in various observ- ables. A simple link between $\mathcal{N}_\text{loc}$ and observables exists due to the fact that the number of different $\Gamma_{j\leq \mathcal{N}_\text{loc}}$ and their linearly independent odd-parity products, i.e., in- cluding $\Gamma_4$ in classes CII and DIII, equals the degrees of freedom $\beta$ (i.e., the Dyson index linked to $T_+$ [52]) of the Hamiltonian’s off-diagonal matrix elements. In fact, the most general Hermitian linear combinations of these $\Gamma_j$ have the same type of offdiagonals, up to an imaginary unit, as the Hamiltonian: real for $\beta=1$ (classes BDI and CI), complex for $\beta=2$ (classes D and C) and real quaternion for $\beta=4$ (classes DIII and CII).
In the SUSY classes, the value of $\beta$ sets the en- ergy level correlations, including the long-range spec- tral rigidity, across opposite parity sectors (these are uncorrelated without SUSY) which lead to “ ramps ” in time-dependent correlation functions of parity- odd ob- servables. (For single-Majorana examples see Refs. 34 and 46.) These ramps occur at time scales below $2\pi$ times the inverse mean level spacing $1/\Delta_\infty$, and have $\beta$ - dependent shape [35]. In particular, the ramp connects to a long-time plateau smoothly when $\beta=1$, sharply when $\beta=2$, and with a kink when $\beta=4$. In Fig. 1, we show ensemble-averaged $q$ -body correlation functions [Eq. (10) below] in classes D, C, DIII, and CI. For completeness, we show the correlation function in the remaining sym- metry classes, including those that do not support SUSY, in Appendix C.
Besides this direct correspondence between the super- charges and ramp structure, we additionally find more subtle consequences of SUSY: The long-time ($t \gg 1/\Delta_\infty$) plateau in $q$ -body correlation functions is also related to the number and structure of the supercharges, cf. Fig. 2. To quantify this relationship, we consider the retarded
| |
221207153/5
|
Gilles Chabrier etal.:ImpactofanewH / Heequationofstateontheevolutionofmassivebrowndwarfs
\begin{tabular}{lcccccc}
\hline\hline
Object & & $M$ [${\mathrm{M_{Jup}}}$] & age [Gyr] & $\mathrm{T_{eff}}$ [K] & $\log(L/L_\odot)$ & $R/R_\odot$ \hline HD4113C & Observations$^{1,8}$ & 66$\pm 5$ & 5$^{+1.3}_{-1.7}$ & 500-600 & $-6.30\pm 0.22$ & % & B2003 & 63 & 5 & 1120 & -5.06 & 0.079 % & & 66 & 5 & 1520 & -4.50 & 0.081 & P2020 & 63 & 5 & 1108 & -5.06 & 0.080 & & 66 & & 1149 & -4.99 & 0.080 & {\bf present} & 63 & 5 & 1081 & -5.11 & 0.079 & & 66 & & 1144 & -5.02 & 0.079 % Eps Indi Ba & Observations$^2$ & 75$\pm 0.82$ & {\rm old} & 1320 & -4.70 & 0.083 Eps Indi Ba & Observations$^2$ & 66.92$\pm 0.36$ & 3.5$^{+0.8}_{-1.0}$ & & & & Observations$^5$ & 68$\pm 0.9$ & & 1312$\pm9$ & -4.70$\pm 0.02$ & & P2020 & 66-67 & 3.0 & 1324-1395 & -4.72 - -4.63 & 0.083 \
& & & 3.5 & 1262-1325 & -4.81 - -4.73 & 0.082 \
& & & 3.8 & 1233-1293 & -4.85 - -4.78 & 0.081 \
& {\bf present} & 66-67 & 2.8 & 1321-1354 & -4.74 - -4.70 & 0.081 \
& & & 3.0 & 1291-1321 & -4.78 - -4.74 & 0.081 \
& & & 3.5 & 1241-1265 & -4.86 - -4.83 & 0.080 \
Eps Indi Bb (or C) & Observations$^2$ & 53.25$\pm 0.29$ & 3.5$^{+0.8}_{-1.0}$ & & & & Observations$^{5}$ & 53$\pm 0.3$ & & 975$\pm 11$ & -5.23$\pm 0.02$ & & P2020 & 53 & 3.0 & 1071 & -5.08 & 0.084 \
& & & 3.5 & 1023 & -5.17 & 0.083 \
& & & 3.8 & 996 & -5.21 & 0.083 \
& {\bf present} & 53 & 2.8 & 1060 & -5.10 & 0.083 & & & 3.0 & 1042 & -5.13 & 0.083 & & & 3.5 & 998 & -5.22 & 0.083 \
& & & 3.8 & 973 & -5.26 & 0.082 GL 758 B & Observations$^3$ & 42$^{+19}_{-7}$ $(>30)$ & 1-6\,({\rm older?}) & 650 & -6.07$\pm 0.03$ & & Observations$^4$ & $38.1^{+1.7}_{-1.5}$ & $\gtrsim 6$ & & & & P2020 & 37 & 5 & 680 & -5.84 & 0.087 & & & 8 & 600 & -6.07 & 0.086 & {\bf present} & 37 & 5 & 673 & -5.86 & 0.087 & & & 8 & 593 & -6.09 & 0.085 WISE J0720-0846B & Observations$^6$ & 66$\pm 4$ & $>$ a few & 1250$\pm$40 & -4.82$\pm 0.07$ & % & B2003 {\bf 2x-CHECK} & 67 & 5 & 1270 & -4.83 & 0.078 & P2020 & 66 & 3 & 1324 & -4.72 & 0.082 & & & 4 & 1216 & -4.88 & 0.082 & {\bf present} & 66 & 3 & 1291 & -4.78 & 0.081 & & & 4 & 1201 & -4.92 & 0.080 2M0805+48 & Observations$^7$ & 66$^{+5}_{-14}$ & $\ge 4$ & & & 2M1059-21 & Observations$^7$ & $67^{+4}_{-5}$ & & & & & P2020 & 66 & 5 & 1149 & -4.99 & 0.080 & & & 10 & 964 & -5.33 & 0.077 & {\bf present} & 66 & 5 & 1142 & -5.02 & 0.079 & & & 10 & 975 & -5.33 & 0.076 Gliese 229 B & Observations$^8$ & 71.4$\pm 0.6$ & $<$10 & & -5.208$\pm 0.007$ & & P2020 & 70.5 & 8 & 1171 & -4.99 & 0.078 & & & 9 & 1140 & -5.04 & 0.077 \
& & & 10 & 1109 & -5.09 & 0.077 \
& {\bf present} & 70.5 & 8 & 1133 & -5.06 & 0.077 & & & 9 & 1103 & -5.11 & 0.076 \
& & & 10 & 1075 & -5.16 & 0.076 \
& & 71.5 & 10 & 1110 & -5.10 & 0.076 \hline \hline
\end{tabular}
Table2. E ff ective temperature, luminosity and radius of the various BDs examined in the text obtained with the P2020 and present models for the observed masses and ages. Observations: $^1$ Cheetham et al. (2018), $^2$ Chen et al. (2022), $^3$ Bowler et al. (2018), $^4$ Brandt et al. (2019), $^5$ Cardoso (2012), $^6$ Dupuy et al. (2019), $^7$ Sahlmann et al. (2020), $^8$ Brandt et al. (2021). Models are Phillips et al. (2020, P2020) and present calculations. We have taken 1 ${\mathrm{M_{Jup}}}$ = $9.5\times 10^{-4}\,{\mathrm{M}_\odot}$ in the evolutionary models.
Using high-contrast imaging with the SPHERE instrument at the VLT, Cheetham et al. (2018) obtained the first images of the cold BD HD4113C (Sp = T9), which is part of a dynami- cal system with a M-dwarf companion. The dynamical mass is ${\mathrm{M_{dyn}}}=66\pm5\,{\mathrm{M_{Jup}}}$, while comparison of the observed spectrum of HD4113C with atmospheric models (Morley et al. 2012, 2014 and Tremblin et al. 2015) yields $\mathrm{T_{eff}} \sim500$ -600 K, $\log\, g$ = 4.5-5.0 and a radius $R\sim 1.4$ -1.5 ${\mathrm{R_{Jup}}}$, much larger than predicted for old, high-mass BDs. Using stellar evolution models (Mowlavi et al. 2012), the derived age of the parent star is found to be $\tau=5^{+1.3}_{-1.7}$
| |
221207153/7
|
Gilles Chabrier etal.:ImpactofanewH / Heequationofstateontheevolutionofmassivebrowndwarfs
Fig.4. Left: cooling curves for a $67\,{\mathrm{M_{Jup}}}$ (top) and a $53\,{\mathrm{M_{Jup}}}$ (bottom) brown dwarf, respectively, representative of the $\epsilon$ Ind Bab system. Solid line: present models, based on the CD21 EOS; dashed line: Phillips et al. (2020) models, based on the CMS19 EOS. Right: brown dwarf isochrones typical of the inferred age of the $\epsilon$ Ind Bab system, calculated with the present (red) and Phillips et al. (2020) (black) models.
Fig.5. Left: e ff ective temperature as a function of mass for massive BDs for 3 isochrones calculated with the present models, based on the CD21 EOS and ATMO atmosphere models, the Phillips et al. (2020) models, based on the CMS19 EOS and ATMO atmosphere models and the Bara ff e et al. (2003) models, based on the SCvH EOS and the COND atmosphere m odels. Right: luminosity as a function of mass for 3 isochrones for BDs with dynamical mass measurements. References for the data are given in Table 2. All models have an equivalent helium mass fraction $Y_{eq} = 0.292$.
al. (2019). The BD companion has a mass ${\mathrm{M_{dyn}}}=66\pm4\,{\mathrm{M_{Jup}}}$, an e ff ective temperature $\mathrm{T_{eff}}= 1250\pm40$ K and a luminosity $\log (L/L_\odot)=-4.8\pm0.15$. This suggests an age older than a few Gyr, consistent with the age estimates for the primary star. As shown in Table 2, both the P2020 and present models yield a nearly perfect agreement with the observational determinations for an age $\sim$ 3-4 Gyr, with the new ones predicting a slightly younger age (more rapid cooling). For the primary star, WISE J0720-0846A, as noted in Dupuy et al. (2019), models (BHAC 2015) overestimate the luminosity for its mass (${\mathrm{M_{dyn}}}=99\pm6\,{\mathrm{M_{Jup}}}$), or conversely underestimate its mass for its luminosity, at about 2 $\sigma$ (see their Table 5 and Fig. 7). As mentioned earlier, the new CD21 EOS does not significantly change this analysis for such high (stellar) masses (see Table 1).
It is worth noting that the mass and age determinations of WISE J0720-0846B are very similar to the ones of the BD HD4747B ($66\pm 3\,{\mathrm{M_{Jup}}}$, $2.9^{+0.5}_{-0.4}$ Gyr, $\log (L/L_\odot)=-4.55\pm0.08$; Brandt et al. 2021, Table 10), for which the present models give a very good agreement (see Table 2).
## 3.5. 2M1059 and 2M0805
Sahlmann et al. (2020) have measured the complete astromet- ric orbits for the systems 2M0805 + 48 and 2M1059-21. They
| |
221207153/2
|
Gilles Chabrier etal.:ImpactofanewH / Heequationofstateontheevolutionofmassivebrowndwarfs
Fig.1. Interior $T$ - $\rho$ profiles in the hydrogen phase diagram for di ff erent astrophysical bodies at about 5 Gyr, as labelled in the figure (red lines) and for a 0.075 ${\mathrm{M}_\odot}$ object at 10 $^8$, $8\times$ 10 $^8$ and $8\times$ 10 $^{9}$ yr (blue dotted, short-dashed and long-dashed lines, re- spectively). The SCvH, CP and QMD labels stand for Saumon, Chabrier & vanHorn (1995) EOS, Chabrier & Potekhin (1998) EOS and QMD simulations (see Chabrier et al. (2019) and Chabrier & Debras (2021) for further details on the line sym- bols).
eral perspective, it should be mentioned that the aforementioned recent analysis of Brandt et al. (2021) suggests another, di ff er- ent issue for young ($<$ 1 Gyr) and low-mass ($\lesssim 40\,{\mathrm{M_{Jup}}}$) BDs. In that case, the trend is the opposite: models underpredict lu- minosities for given mass and age, or equivalently overestimate (resp. underestimate) the mass (resp. the age) for the correct age (resp. mass). For objects in-between, mass $\sim$ 40-70 ${\mathrm{M_{Jup}}}$, age $\sim$ 1- 5 Gyr, models agree well ($<1\sigma$) with observations (see Table 10 of Brandt et al. (2021)). This points to two di ff erent issues in BD cooling theory in the extreme mass and age limits. In the present paper, we will focus on the first issue, for which we show that part of the solution lies in the EOS. In contrast, the second problem stems more likely from remaining uncertainties in BD atmosphere models.
## 2. Impact of the new equations of state on the internal structure, cooling history and hydrogen-burning limit
Figure 1 displays temperature-density profiles of various astro- physical bodies, from 1 ${\mathrm{M}_\odot}$ to 1 ${\mathrm{M_{Jup}}}$ for an age of about 5 Gyr, in the phase diagram of hydrogen, for the $T$ - $\rho$ range covered by the CMS19 and Chabrier & Debras (2021, CD21) EOSs. This diagram is similar to the ones portrayed in these papers for H and He and we refer the reader to these papers for more details. The key issue here is that, as illustrated by the central part of the diagram labelled QMD (for Quantum Molecular Dynamics), all BD cooling tracks enter the domain where interactions be- tween H and He species can no longer be ignored, as assumed in the so-called ideal (or ’additive’) volume law approximation, used in CMS19 and Saumon et al. (1995, SCvH) (see Fig. 1 of CD21 for further details). Recently, Chabrier & Debras (2021) took into account the impact of these interactions on the thermo- dynamic properties of the H / He mixture by incorporating in the EOS the QMD calculations performed by Militzer & Hubbard (2013). The figure also displays the cooling sequence of a 0.075 ${\mathrm{M}_\odot}$ object at 10 $^8$, $\sim$ 10 $^9$ and $\sim$ 10 $^{10}$ yr, respectively. As seen in the figure (see also the Conclusion of CMS19), essentially all objects below $0.1\,{\mathrm{M}_\odot}$ older than about $\sim 0.1$ Gyr will enter the domain where H-He interactions cannot be ignored and will af- fect to some degree their mechanical and thermal properties, thus their structure and evolution. It is the aim of the present paper to examine in detail this impact.
The left panel of Fig. 2 displays the evolution of the cen- tral density, $\rho_c$, temperature, $T_c$, and degeneracy parameter, $\psi_c=(T_c/T_F)\approx 3.314\times10^{-6}T_c(\mu_e/\rho)^{2/3}$,where $T_F$ and $\mu_e$ denote the electron Fermi temperature and mean molecular weight, respec- tively, of a 0.075 ${\mathrm{M}_\odot}$ object (see Chabrier & Bara ff e 2000). The evolution is calculated with our recent so-called ATMO atmosphere models and updated solar abundances (Ca ff au et al. 2011) for a global helium and heavy element abundance $Y=0.275, Z_\odot=0.017$, respectively (Phillips et al. 2020), but with 3 di ff erent H / He EOS, namely Saumon et al. (1995, SCvH), Chabrier et al. (2019, CMS19) and Chabrier & Debras (2021, CD21). The di ff erences between these di ff erent EOS, and the improvements in the treatment of the H-He interactions, are described in detail in these papers. As already mentioned in Phillips et al. (2020) and Chabrier & Debras (2021) and clearly seen in the figure, the new EOSs yield denser and cooler struc- tures for a given object, thus more correlated and degenerate (lower $\psi$) interiors. This in turn increases the cooling rate of the object. This is illustrated in the right panel of Fig. 2, which por- trays the late evolution of the e ff ective temperature, luminosity and radius of a 0.073 (black), 0.074 (blue) and 0.075 (red) ${\mathrm{M}_\odot}$ object for the same 3 EOSs. While these 3 quantities, notably $\mathrm{T_{eff}}$ and $L$, become constant after $\sim 5\times 10^8$ yr at $0.073\,{\mathrm{M}_\odot}$ with the SCvH EOS, indicating the stellar-substellar boundary, they keep decreasing with the two other EOSs, this limit occuring at $0.074\,{\mathrm{M}_\odot}$ with the CMS19 EOS and at $0.075\,{\mathrm{M}_\odot}$ with the CD21 EOS. Table 1 displays the characteristic of the steady H-burning limit (HBL) obtained with the 3 EOSs. We define the HBL as the limit below which nuclear equilibrium ($L_{\rm nuc}=L_\star$) will never be reached, thus below which cooling and gravitational contraction will last for ever. This corresponds to the physical limit of the stellar main sequence, i.e. the stellar-substellar boundary. Any object below this limit, i.e. cooler and fainter than the corre- sponding e ff ective temperature and luminosity (see Table 1) will be a brown dwarf, whatever its age. Note, however, that the op- posite is not true: objects hotter and brighter than these limits can be either stars or BDs, depending on their age. The H-burning minimum mass (HBMM), i.e. the minimum mass to sustain hy- drogen fusion, that does not depend on the age, is found to be 0.075 ${\mathrm{M}_\odot}$ ($\sim78.5\,{\mathrm{M_{Jup}}}$) with the most recent CD21 EOS. For sake of comparison, we also give in this table the characteris- tics of the HBL obtained in Bara ff e et al. (2003). These models use the SCvH EOS, the same helium and heavy element abun- dances as mentioned above but the so-called COND model at- mospheres. These comparisons enable us to disentangle the im- pact of the atmosphere and EOS models, respectively, upon the cooling properties of an object at the H-burning limit. The atmo-
| |
221207153/6
|
Gilles Chabrier etal.:ImpactofanewH / Heequationofstateontheevolutionofmassivebrowndwarfs
Gyr. The COND models (Bara ff e et al. 2003) for such temper- atures predict a mass $M=36\pm5\,{\mathrm{M_{Jup}}}$ for the BD, in strong conflict with the dynamical mass. Conversely, for $M=66\,{\mathrm{M_{Jup}}}$, these models predict $\mathrm{T_{eff}}\sim 1200\pm170$ K at the age of the sys- tem, significantly higher than the afore estimates from the best- fit spectral models. Table 2 displays the temperature, luminos- ity and radius obtained with the Phillips et al. (2020) models, which use the CMS19 models, and the present ones, with the CD21 EOS. Both models use the same abundances for helium, $Y=0.275$, and metals, $Z_\odot=0.017$, yielding an ’equivalent’ helium abundance $Y_{eq}=0.292$. This comparison highlights the impact of the H / He interactions in the most recent EOS mod- els. As seen in the table, even though, for the relevant mass and age range, the models including the new EOS yield temperatures cooler and luminosities fainter than the Bara ff e et al. (2003) and Phillips et al. (2020) models, relieving part of the tension, it is clear that they are still far from resolving the disagreement with the observational determinations.
As noted by Cheetham et al. (2018), this discrepancy may be caused by the object being an unresolved binary brown dwarf system or by the presence of an additional object in the system, which could have biased the RV data and caused an overestimate of the dynamical mass. An equal mass binary of 500-600 K ob- jects with $R\sim 1\,{\mathrm{R_{Jup}}}$, notably, would provide a good match to the observed data while being in good agreement with the model predictions, namely $\mathrm{T_{eff}}=600\pm40$ K for a 33 ${\mathrm{M_{Jup}}}$ object at $5\pm1$ Gyr (e.g. Phillips et al. 2020 or present models).
## 3.2. Eps Indi Bab
Recently, Chen et al. (2022) reported dynamical masses for the binary BD system $\epsilon$ Indi Ba (Sp = T1-1.5) and $\epsilon$ Indi Bb (Sp = T6, also called $\epsilon$ Indi C), with individual masses ${\mathrm{M_{dyn}}}=66.92\pm0.36\,{\mathrm{M_{Jup}}}$ and ${\mathrm{M_{dyn}}}=53.25\pm0.29\,{\mathrm{M_{Jup}}}$, respectively, with a $\approx 5\%$ precision. With an age of 3.5 $^{+0.8}_{-1.0}$ Gyr from $\epsilon$ Indi A’s activity, this system provides a stringent constraint for BD cooling models, notably for old, massive BDs. Field-aged ob- jects of spectral types T1 and T6 have e ff ective temperatures in the range $\mathrm{T_{eff}}\simeq 1300$ -900 K (e.g. Filipazzo et al. 2015). This is consistent with the inferred temperatures for Indi B and C, namely $\mathrm{T_{eff}}\simeq 1300$ -1340 K and $\mathrm{T_{eff}}\simeq 880$ -940 K, respectively (King et al. 2010). Correct theoretical evolutionary models must thus allow these objects to reach these spectral types and tem- peratures within the aforementioned timescale for the observed metallicity of the host star $\epsilon$ Indi A ($[Fe/H]=0.13$). nan of the widely used Chabrier et al. (2000), Burrows et al. (2001) and Saumon & Marley (2008) ’cloudy’ models can fullfill this con- straint and predict too high temperatures and luminosities. The ’COND’ Bara ff e et al. (2003) predict faster cooling rates, due to the less opaque atmospheres. However, as noted in § 2, the lower opacity in the COND atmosphere models stems, notably, from missing line opacities. Therefore, this analysis suggests that the aforementioned models underpredict the cooling rates for $\epsilon$ Indi Ba and Bb.
Figure 4 displays the $L$ - $t$ and $L$ - $M$ relationships obtained for this system with the Phillips et al. (2020) cooling models, based on the CMS19 H / He EOS, and the present ones, based on the CD21 EOS, using the same ATMO atmosphere models and ef- fective helium + heavy element abundance $Y_{eff}=0.292$. As seen in the Figure, while the Phillips et al. (2020) models yield an age of between $\sim 3.0$ and 3.8 Gyr for the system for the observed lu- minosities, the present model predict slightly younger ages, of 2.8 Gyr for Ba and 3.5 Gyr for Bb. Both models are in good agreement with the observational determination. Assuming co- evality, this corresponds to a $\sim 0.7$ Gyr uncertainty on the age of these BDs, for this age and mass range. As seen in Fig. 13 of Chen et al. (2022), the Saumon & Marley (2008) hybrid mod- els predict a significantly older age (5 Gyr) for the system, at odds with the observational determination, suggesting a too slow cooling rate for these models. The corresponding e ff ective tem- peratures and surface gravities between 2.8 and 3.5 Gyr with the present models are $\mathrm{T_{eff}}=$ 1354-1265 K, $\log \,g=$ 5.45 for $\epsilon$ Indi Ba, for a mass 67 ${\mathrm{M_{Jup}}}$, and $\mathrm{T_{eff}}=$ 1060-998 K, $\log \,g=$ 5.34 for $\epsilon$ Indi Bb, for a mass 53 ${\mathrm{M_{Jup}}}$, respectively (see Table 2). As illustrated in Figure 4 and discussed earlier, the faster cooling rate for the most massive BDs with the CD21 EOS than with the CMS19 one stems the sharp increase of degeneracy in this mass range (see e.g. Fig. 1 of Chabrier & Bara ff e 2000). The cooling rate then appears to be a bit too slow as a function of the mass, with $\Delta \log L/\Delta \log M\simeq 3.95$ instead of 5.37 between $\epsilon$ Indi Ba and Bb, i.e. from early T to late T, even though the discrepancy remains within 2 $\sigma$. As mentioned earlier, this stems more likely from remaining issues with the atmosphere models than with the EOS.
## 3.3. Gl 758 B
Combining radial velocity and astrometry, Bowler et al. (2018) determined a dynamical mass of ${\mathrm{M_{dyn}}}=42^{+19}_{-7}\,{\mathrm{M_{Jup}}}$ for the T7- T8 BD Gl 758 B, with a robust lower limit of 30.5 ${\mathrm{M_{Jup}}}$ at the 4 $\sigma$ level, for nominal ages 1-6 Gyr adopted for the host star (Vigan et al. 2016). More recently, using Hipparcos and Gaia data, Brandt et al. (2021) derived the most precise mass mea- surement to date for this system, with ${\mathrm{M_{dyn}}}=38.1^{+1.7}_{-1.5}\,{\mathrm{M_{Jup}}}$, while their analysis of activity and rotation of Gl 758A favors an age $\gtrsim$ 6 Gyr. As for the previous objects, substellar evolu- tionary models generally underestimate the mass of Gl 758 B. As noted by the authors above, this discrepancy can be rec- onciled if the system is older, which is consistent with activ- ity indicators and recent isochrone fitting of the host star, or alternatively if the models are systematically overluminous by $\approx$ 0.1-0.2 dex. Atmospheric model fitting yields for this object a bolometric luminosity $\log (L/L_\odot)=-6.07\pm0.03$, $\mathrm{T_{eff}}=650$ K and $\log g=5.0$. All current models essentially underpredict the mass for an age 1-6 Gyr or, alternatively, are overluminous by $>0.1$ dex at this age. This is in the opposite sense from results by Dupuy et al. (2009, 2014), who found that substellar cool- ing models underpredict the luminosities of brown dwarfs with dynamical masses by $\approx$ 0.2-0.4 dex. Altogether, the most likely explanation for the disagreement in mass probably resides in the age of Gl 758. Older ages of 6-9 Gyr would readily put the pre- dicted and dynamical distributions in excellent agreement and are indeed suggested from the low activity level, lack of X-ray emission and slow projected rotational velocity (see Bowler et al. (2018) and references therein). Indeed, more recent isochrone fitting are converging on an older value that agrees better with activity indicators, with an average of $5.3$ - $7.5$ Gyr (Brewer et al. 2016, Luck 2017). This is supported by the results displayed in Table 2, with the new models yielding a nearly perfect agree- ment with the observations for a mass $\sim 30$ - $40\,{\mathrm{M_{Jup}}}$ and an age $\sim$ 5-8 Gyr.
## 3.4. WISE system J072003.20
Individual dynamical masses for the nearby M9.5 + T5.5 binary system WISE J0720-0846AB have been determined by Dupuy et
| |
221207153/4
|
Gilles Chabrier etal.:ImpactofanewH / Heequationofstateontheevolutionofmassivebrowndwarfs
Fig.3. E ff ective temperature, luminosity and radius for $M\le0.075\,{\mathrm{M}_\odot}$ for 0.1, 1, 2, 5 and 10 Gyr (from top to bottom), respectively, for the same atmosphere models and helium and heavy element compositions but 2 di ff erent EOSs, namely Chabrier & Debras (2021) (present) and Chabrier et al. (2019) (Phillips et al. 2020).
& Bara ff e 2000, Chabrier et al. 2009), at ${\rm T_{{eff}}}_{\rm HBL} \simeq 2075$ K, $\log(L/L_\odot)_{\rm HBL}\simeq -3.9$ and $R/R_\odot \simeq 0.086$. The inversion of the radius trend occurs near the location of the L2.5 dwarf 2MASS J0523-1403 ($M=67.54\pm12.79\,{\mathrm{M_{Jup}}}$, Sp = L2.5 (Filippazzo et al. 2015)). Although at first glance this seems to point to a dis- agreement with the models in the determination of the HBL, it must be kept in mind that the age of the objects in the Dieterich et al. (2014) sample, although supposedly larger than 1 Gyr, is unknown. This inevitably translates into some uncertainty in the comparison between observations and evolutionary calcula- tions. As seen in the right panel of Fig. 2, however, the HBL is only reached at much older ages, about 10 Gyr with the new EOS. Therefore, most of the objects identified in Dieterich et al. (2014), which lie at less than 25 pc and belong dominantly to the young disk, are too young to yield a proper determination of the HBL. On the other hand, as seen in figure 3, which por- trays the $\mathrm{T_{eff}}$ - $M$, $L$ - $M$ and $R$ - $M$ relations below the HBMM for 5 isochrones, an average age of about 2 Gyr for the observed sample near the $R$ - $\mathrm{T_{eff}}$ and $R$ - $L$ minimum location yields an ex- cellent agreement between the observational determinations and the new models. As highlighted in the rightmost panel of this figure, the minimum in the radius-mass relation does not occur exactly at the HBMM but at about 65 ${\mathrm{M_{Jup}}}$ (see, e.g., Fig. 1 of Chabrier et al. 2009), consistent with the aforementioned BD 2MASS J0523-1403 determination, and is more and more pro- nounced as objects get older. We see from the 3 panels that the observational identifications of the minimum radius in the sam- ple of Dieterich et al. (2014) are consistent with objects of mass $\sim 0.065$ - $0.075\,{\mathrm{M}_\odot}$ and age $\sim1$ -2 Gyr. We also see that, whereas the newest EOS a ff ects only modestly the e ff ective temperature and the luminosity, for a given mass and age, compared with the CMS19 EOS used in the models of Phillips et al. (2020), it a ff ects more severely the radius, notably for the most massive BDs. This reflects the di ff erent rate at which degeneracy pro- gresses throughout the body, starting at about $\sim 0.1$ Gyr, as seen in the left panel of Fig. 2. The Phillips et al. (2020) models have been updated accordingly and are available online (see at the end of the paper).
It is important to mention at this stage that the procedure used by Dieterich et al. (2014) remains model atmosphere de- pendent, notably in a temperature domain characterised by cloud condensation and sedimentation in the models. Even though the BT-Settl atmosphere models, used by Dieterich et al. (2014), have been shown to provide a remarkable agreement with ob- served M and L spectra, alternative ’cloudless’ models have been proposed, which are very successfull in reproducing BD atmo- sphere spectra (Tremblin et al. 2015, 2016, 2017). ’Cloudless’ in these models does not mean that clouds do not form, but rather that they are not responsible for shaping the spectral evolution of brown dwarfs. Instead, the (massive) BD spectral evolution is due to a thermo-compositional instability triggered by the chem- ical conversion $CO\rightarrow CH_4$ in their atmospheres (Tremblin et al. 2019).
Concerning the domain of very-low mass stars (VLM, $M\le 0.4\,{\mathrm{M}_\odot}$) on the main sequence (age $\gtrsim 1$ Gyr), the impact of the new EOS (CD21) compared with the SCvH one used in the Bara ff e et al. (2015, BHAC) models is quite modest and thus will not be displayed here.
## 3. Comparison with the data
As mentioned in the previous section, near-infrared spectro- scopic and imaging surveys have uncovered a population of short-period, spectral binaries composed of low-mass stars and brown dwarfs, allowing a precise dynamical determination of the mass of these latter objects. Activity and kinematic constraints for the age of the primary of these massive T-dwarfs lead to a range of 2-8 Gyr. In this section, we compare the e ff ective tem- peratures obtained with the COND models (Bara ff e et al. 2003), the Phillips et al. (2020) models, based on the Chabrier et al. (2019) EOS, and a subset of models based on the most recent Chabrier & Debras (2021) EOS, with the various observations of BD dynamical systems.
## 3.1. HD4113C
| |
221207153/1
|
# Impact of a new H/He equation of state on the evolution of massive brown dwarfs. New determination of the hydrogen burning limit
Gilles Chabrier 1, 2, Isabelle Bara ff e 2, 1, Mark Phillips 3, 2,and Florian Debras 4
We have explored the impact of the latest equation of state (EOS) for dense hydrogen-helium mixtures (Chabrier & Debras 2021), which takes into account the interactions between hydrogen and helium species, upon the evolution of very low mass stars and brown dwarfs (BD). These interactions modify the thermodynamic properties of the H / He mixture, notably the entropy, a quantity of prime importance for these fully convective bodies, but also the onset and the development of degeneracy throughout the body. This translates into a faster cooling rate, i.e. cooler isentropes for a given mass and age, and thus larger brown dwarf masses and smaller radii for given e ff ective temperature and luminosity than the models based on previous EOSs. This means that objects of a given mass and age, in the range $M\lesssim 0.1\,{\mathrm{M}_\odot}$, $\tau\gtrsim 10^8$ yr, will have cooler e ff ective temperatures and fainter luminosities. Confronting these new models with several observationally determined BD dynamical masses, we show that this improves the agreement between evolutionary models and observations and resolves at least part of the observed discrepancies between the properties of dynamical mass determinations and evolutionary models. A noticeable consequence of this improvement of the dense H / He EOS is that it yields a larger H-burning minimum mass, now found to be $0.075\,{\mathrm{M}_\odot}$ ($78.5\,{\mathrm{M_{Jup}}}$) with the ATMO atmosphere models for solar metallicity. These updated brown dwarf models are made publicly available.
## 1. Introduction
Tremendous progress has been accomplished along the years in the theoretical description of brown dwarfs (BD), enabling us to better understand their fundamental properties and cooling his- tories. The most recent progresses include a more complete de- scription of their atmosphere, thus of their spectral energy distri- bution (e.g. Morley et al. 2012, 2014, Tremblin et al. 2015, 2016, 2017), and of their interior, notably concerning the equation of state (EOS) of dense hydrogen and helium (Chabrier et al. 2019, CMS19). A new generation of BD evolutionary models has been derived recently, that incorporates both new (so-called ’ATMO’) atmosphere models and the new CMS19 H / He EOS (Phillips et al. 2020). As shown in these models, a noticeable impact of this new EOS is to yield denser and cooler, thus more degener- ate objects at the same mass than the ones computed with the Saumon-Chabrier-vanHorn (SCvH) EOS (Saumon et al. 1995). This yields slightly faster cooling rates, thus cooler temperatures and lower brightness at a given age (Phillips et al. 2020).
Yet, a puzzling issue has emerged within the recent years, thanks to the determination of the dynamical masses of several ’massive’ T dwarfs (Cheetham et al. 2018, Dieterich et al. 2018, Dupuy et al. 2019, Bowler et al. 2018, Sahlmann et al. 2020, Brandt et al. 2019). These observations reveal dynamical masses significantly larger than the theoretical predictions for the deter- mined e ff ective temperature and age. A common feature of all these objects is their relatively high mass ($\sim60$ -75 ${\mathrm{M_{Jup}}}$), near the hydrogen burning minimum mass (HBMM), similar spec- tral type (late T) and cool e ff ective temperatures ($\mathrm{T_{eff}}\lesssim 1200$ K), which suggests ages in the range $\sim 5$ -10 Gyr, at the very least $>1$ Gyr. A noticeable exemple of this puzzling issue is the recent dynamical mass determination of Gliese 229 B, with a larger mass than estimated previously (Brandt et al. 2020). For all these late, massive T dwarfs, all existing BD evolutionary models generally underpredict the mass for the nominal ages and temperatures, suggesting too small cooling rates for these objects. This trend is confirmed by the recent thorough analysis of BD companion dynamical masses using Hipparcos and Gaia EDR3 data and age determinations based on activity-rotation- age calibrations by Brandt et al. (2021). For old ($\gtrsim$ 5 Gyr) and high-mass ($\gtrsim 60\,{\mathrm{M_{Jup}}}$) BDs, the models overpredict luminosities for the measured mass and age, or equivalently underestimate (resp. overestimate) the mass (resp. the age) for the correct age (resp. mass). As noticed by Dieterich et al. (2018), concerning the Eps Indi B-C system, models cannot make such massive ob- jects reach such a cool $\mathrm{T_{eff}}$ within the age of the Galaxy. Due to the degeneracy between mass and age in the substellar domain, there are no reliable age indicators for isolated BDs, and thus no robust constraint can be derived regarding substellar cooling rates and evolutionary models. In contrast, the aforementioned BD companions to higher mass stars with masses close to the stellar-substellar limit allow us to test a boundary value of the theory of substellar structure and evolution.
A related diagnostic of these observations is the suggestion that the HBMM, identified as a minimum in the radius-e ff ective temperature and radius-luminosity relation, is larger and lies at a $\sim 400$ K larger e ff ective temperature than predicted by all cur- rent models (Dieterich et al. 2014).
In this paper, we show that the most recent improvements in EOS calculations of dense hydrogen / helium mixtures contribute to resolve these issues in BD cooling theory. From a more gen-
| |
221207153/3
|
Gilles Chabrier etal.:ImpactofanewH / Heequationofstateontheevolutionofmassivebrowndwarfs
Fig.2. Left: Evolution of the central density, temperature and degeneracy parameter for a 0.075 ${\mathrm{M}_\odot}$ object calculated with 3 dif- ferent EOS, namely CD2021 (solid line), CMS2019 (long-dashed line) and SCvH (dotted line). Right: Evolution of the e ff ective temperature and luminosity for 0.075 (red), 0.074 (blue) and 0.073 (black) ${\mathrm{M}_\odot}$ and radius for 0.075 ${\mathrm{M}_\odot}$, with the same 3 EOS.
\begin{tabular}{lcccc}
\hline\hline
EOS & ${\mathrm{M}_{\rm HBMM}}/M_\odot$ & ${\rm T_{{eff}}}_{\rm HBL}$ [K] & $\log(L/L_\odot)_{\rm HBL}$ & $R_{\rm HBL}/R_\odot$ \hline
SCvH+COND (Baraffe et al. 2003) & 0.072 & 1560 & -4.47 & 0.081 SCvH+ATMO & 0.073 & 1807 & -4.14 & 0.087 CMS'19+ATMO (Phillips et al. 2020) & 0.074 & 1800 & -4.16 & 0.085 CD'21+ATMO (present) & 0.075 & 1800 & -4.19 & 0.083\hline
\end{tabular}
Table1. Mass, e ff ective temperature, luminosity and radius characteristic of the H-burning limit ($t$ = 10 Gyr) with the SCvH (1995), Chabrier et al. (2019) and Chabrier & Debras (2021) EOSs, all calculated with the ATMO atmosphere models. For the sake of comparison, the results with the COND atmosphere models are also presented.
sphere models have been greatly improved between the COND and ATMO models. Line opacities, in particular, were missing in the former ones, yielding a much faster cooling rate and thus explaining the cooler and fainter limits at the HBL. But the EOS also bears some impact on the HBL. As seen in Table 1, the larger cooling rate with the new EOS translates into a larger HBMM with a smaller radius than in the previous calculations. On the other hand, the luminosities at the HBL remain barely a ff ected ($\lesssim 10\%$) and the e ff ective temperature remains essen- tially the same. This is explained by the fact that the threshold for hydrogen fusion at the center of the star occurs at a fixed temperature and thus, for the same interior-atmosphere (T $_c$ - $\mathrm{T_{eff}}$) boundary condition, at the same e ff ective temperature. In con- trast, the nuclear energy rate, thus the luminosity depends on the mass and then slightly di ff ers between the 3 di ff erent HBMM’s.
On the other hand, Dieterich et al. (2014) have determined the e ff ective temperature and bolometric fluxes of 63 objects ranging in spectral type from M6V to L4, bracketing the stel- lar / substellar boundary, by comparing observed optical and in- frared photometric colors on 9 bands with synthetic ones, de- rived from the so-called BT-Settl model atmospheres (Allard et al. 2012, 2013), used in the Bara ff e et al. (2015) evolutionary calculations. The optimisation procedure for determining e ff ec- tive temperatures also indicates which model spectrum in the BT-Settl grid provides the overall best fit to the observed pho- tometry. Once the e ff ective temperatures and bolometric fluxes of the objects were determined by this procedure, the radii of the objects with known trigonometric parallax were determined from the Stefan-Boltzmann law (see Dieterich et al. (2014) for details). As the di ff erences between stellar and substellar objects become more pronounced at ages $>1$ Gyr, objects with known youth signatures were rejected from the sample. Based on this analysis, Dieterich et al. (2014) find evidence for a local min- imum in the radius-temperature and radius-luminosity trends, that indicates the vicinity of the stellar main sequence - brown dwarf sequence boundary (see e.g. Burrows et al. 1997, Chabrier
| |
221207153/8
|
Gilles Chabrier etal.:ImpactofanewH / Heequationofstateontheevolutionofmassivebrowndwarfs
find masses ${\mathrm{M_{dyn}}}=66^{+5}_{-14}\,{\mathrm{M_{Jup}}}$ and spectral type T5.5 for 2M0805 + 48 B and ${\mathrm{M_{dyn}}}=67^{+4}_{-5}\,{\mathrm{M_{Jup}}}$, T3.5 for 2M1059-21 B.
The striking feature of their analysis is that the mass for the T3.5 2M1059-21B object is significantly higher than its two spectral type equivalents DENIS J2252-1730B (T3.5), 41 $\pm 4\, {\mathrm{M_{Jup}}}$, and 2MASS J1534-2952A (T4.5), 51 $\pm 5\, {\mathrm{M_{Jup}}}$. As noted by the authors, the mass of 2M0805 + 48B is almost equal to the one of WI0720-08B estimated by Dupuy et al. (2019), which has the same spectral type, yet a mass higher than the other three T5 dwarfs with measured dynamical masses, which all have masses lower than $<61\,{\mathrm{M_{Jup}}}$ as an upper limit (see Table 6 of Dupuy et al. 2019). As shown by the authors (see their Fig. 13), the masses derived for each member of our two pairs are compatible at the 1 $\sigma$ level with the 5 to 12 Gyr isochrones of Bara ff e et al. (2015), while $>$ 1 Gyr isochrones show reasonable agreement with the observational data.
This, as the other previous analysis, highlights the fact that the higher observational masses than other T-dwarfs of similar spectral type and than predicted by the models, suggesting that models underpredict the mass for a given temperature or lumi- nosity, concerns essentially the most massive ($\gtrsim 60\,{\mathrm{M_{Jup}}}$) BDs, old enough to have reach the T spectral type domain, i.e. older than $\gtrsim 1$ Gyr for such masses. This strengthens our suggestion that this property stems essentially from the higher degeneracy, thus the faster cooling of these objects, which culminates at the highest central density of the stellar-substellar domain (see Fig. 1 of Chabrier & Bara ff e 2000). The new models predict cooler temperatures and fainter luminosities for a given mass and age compared with the Bara ff e et al. (2003) models, by up to about 100-150 K and $\sim 0.2$ - $0.3$ dex for the present ones, resolving at least partly the discrepancies mentioned above (see Table 2).
Here again, we note that the mass and age of this system are similar to the ones of the BD HD19467B listed in Brandt et al. (2021), with an excellent agreement for the models.
## 3.6. Gliese 229B
Combining Keck / HIRES radial velocities, imaging with HiCIAO / Subaru and the HST, and absolute astrometry from Hipparcos and GAIA, Brandt et al. (2020) measured the dynam- ical mass ${\mathrm{M_{dyn}}}=70\pm 5\,{\mathrm{M_{Jup}}}$ for the T7 BD Gliese 229B. Not only this value is higher than the $\simeq 64.0^{+2}_{-1}\,{\mathrm{M_{Jup}}}$ predicted by the Bara ff e et al. (2003) or Saumon & Marley (2008) models for such a low luminosity, $\log(L/L_\odot)=-5.208$, but to be com- patible with the observational determinations, the models would predict an age of 7-10 Gyr. Such an age seems to be excluded by kinematic and activity indicators that rather suggest a range 2-6 Gyr. Gliese 229B thus joins the club of ultracool BDs near the HBMM too massive, for their age, for the model predictions. For a $70\,{\mathrm{M_{Jup}}}$ at an age of 10 Gyr, the Phillips et al. (2020) mod- els (see their Fig. 8), calculated with the Chabrier et al. (2019) EOS, are $\sim 0.1$ dex less luminous than the Bara ff e et al. (2003, B03) and $\sim 0.4.$ dex less luminous than the hybrid cloud tracks of SM08, helping to relieve some of the tension. As seen in Table 2, the present models, based on the new CD21 EOS, predict $\gtrsim$ 40 K cooler and $\sim -0.1$ dex fainter models than the Phillips et al. (2020) ones for a mass and luminosity consistent with the lat- est observationally inferred values, for an age of about 9-10 Gyr, improving the model-observations agreement.
Best-fit atmosphere models to the observed spectrum of Gliese 229B by Nakajima et al. (2015) yield acceptable solutions in the range 750 K $\le\mathrm{T_{eff}}\le$ 900 K, $4.75\le \log g\le 5.0$. As seen in the Table, these values are quite small compared to the ones from the models. It must be stressed, however, that this fitting proce- dure is based on one single source of cloudy atmosphere models, namely Tsuji (2002, 2005). A more robust determination of the e ff ective temperature and gravity requires further detailed com- parisons with more recent atmosphere models, a point we have already stressed in § 2.
As a last remark, it is worth noting that a possibility to resolve the discrepancy between models and observations for Gl229 B would be that this latter is itself an unresolved tight binary (see e.g. Brandt et al. 2021).
A case close to Gliese 229B is HR 7672 B, with a dy- namical mass ${\mathrm{M_{dyn}}}=72.7\pm{0.8}\,{\mathrm{M_{Jup}}}$ and a weighted aver- age luminosity $\log(L/L_\odot)=-4.25\pm 0.05$ (Brandt et al. 2021), while the activity analysis suggests a rather young age, centered around 2 Gyr (Brandt et al. 2019, 2021). For this age and a mass $M=70\,{\mathrm{M}_\odot}$ ($=72.7\,{\mathrm{M_{Jup}}}$), the present models predict $\log(L/L_\odot)=-4.25$, $\mathrm{T_{eff}}=$ 1730 K, while the Phillips et al. (2020) ones yield $\log(L/L_\odot)=-4.16$, $\mathrm{T_{eff}}=$ 1785 K.
In order to illustrate the results listed in Table 2, Fig. 5 dis- plays the $\mathrm{T_{eff}}$ -mass and $L$ -mass data comparisons (as in Fig. 8 of Phillips et al. 2020), including the most recent determinations, for massive BDs for 4 isochrones, namely 1, 2, 5 and 8 Gyr, based on the CMS19 (Phillips et al. 2020) and CD21 (present) EOS models.
To briefly summarise this section, we note that the models based on the most recent EOSs, notably the latest CD21 one, reach cooler temperatures and fainter luminosities than the pre- vious generation for a given mass and age, illustrating the faster cooling rates for massive, old ($\gtrsim 1$ Gyr) BDs, yielding in gen- eral a much better agreement with the observational determina- tion. As seen in Table 2, some tension remains potentially for Gl229B but most importantly for HD4113C, which appears to be substantially cooler and fainter than predicted by the mod- els. Given the general agreement for all the other objects, this suggests that these two BDs could be unresolved systems.
## 4. Conclusion
In this paper, we have explored the impact of the latest EOS for dense hydrogen-helium mixtures upon the structure and evo- lution of very low mass stars and brown dwarfs. Whereas the previously used Saumon et al. (1995) and Chabrier et al. (2019) EOSs are based on the so-called additive volume law, which im- plies that the interactions between hydrogen and helium species are not taken into account, these interactions are included in the Chabrier & Debras (2021) EOS, based on the QMD simulations of Militzer & Hubbard (2013) (see CD21 for details). These interactions modify the thermodynamic properties of the H / He mixture in two ways. They yield cooler and denser, thus more compact structures, i.e. smaller radii and thus fainter luminosity for given mass and age. They also yield cooler entropy profiles, a quantity of prime importance for these fully convective bodies, which a ff ects the onset and the development of electron degen- eracy throughout the body (see Fig. 9 of CD21). This translates into a faster cooling rate and thus larger masses for given e ff ec- tive temperatures and luminosities at a given age in the brown dwarf regime (see Table 2). Confronting these new models with several observationally determined BD dynamical masses, we show that, indeed, this improves the agreement between evolu- tionary models and observations and resolves at least part of the observed discrepancies for massive, rather old brown dwarfs. In
| |
200211922/5
|
Table2. Best-fitting parameters for Model 1: crabcor*TBabs (diskbb + relxill)
\begin{tabular}{llccccccc}
\rightarrowprule
Model & Parameter &Spec. A & Spec. B & Spec. C & Spec. D & Spec. E & Spec. F & Spec. G \midrule
{\tt crabcor}&$C$ &\multicolumn{7}{c}{1.097 (f)}\specialrule{0em}{1pt}{1.1pt}
&$\Delta\Gamma$ &\multicolumn{7}{c}{0.01 (f)}\specialrule{0em}{1pt}{1.1pt}
{\tt TBabs}&$N_\mathrm{H}$ (cm$^{-2}$) &\multicolumn{7}{c}{$4.0\times10^{21}$ (f)}\specialrule{0em}{1pt}{1.1pt}
{\tt diskbb} &$T_\mathrm{col}$ (keV)&$0.858_{-0.014}^{+0.011}$&$0.825\pm0.012$&$0.835_{-0.018}^{+0.015}$&$0.787_{-0.010}^{+0.008}$&$0.787_{-0.009}^{+0.007}$&$0.771\pm0.007$&$0.772_{-0.009}^{+0.010}$\specialrule{0em}{1pt}{1.1pt}
&$N_\mathrm{DISC}$&$3889_{-219}^{+366}$&$4151_{-273}^{+237}$&$3561_{-297}^{+312}$&$4862_{-247}^{+329}$&$4923_{-234}^{+398}$&$5025_{-212}^{+230}$&$4891_{-311}^{+328}$\specialrule{0em}{1pt}{1.1pt}
{\tt relxill}
&$q$&$3.69_{-0.56}^{+1.03}$&$3.78_{-0.64}^{+1.23}$&$3.29_{-0.52}^{+1.04}$&$3.98_{-0.70}^{+1.41}$&$3.82_{-0.57}^{+1.09}$&$3.67_{-0.55}^{+1.20}$&$3.71_{-0.65}^{+1.13}$ \specialrule{0em}{1pt}{1.1pt}
&$a_*$&\multicolumn{7}{c}{$0.67_{-0.08}^{+0.15}$}\specialrule{0em}{1pt}{1.1pt}
&$i$ (deg)&\multicolumn{7}{c}{$36.3_{-3.4}^{+5.3}$}\specialrule{0em}{1pt}{1.1pt}
&$A_\mathrm{Fe}$&\multicolumn{7}{c}{$5.05_{-0.26}^{+1.21}$}\specialrule{0em}{1pt}{1.1pt}
&$\Gamma_\mathrm{r}$&$2.79\pm0.06$&$2.59_{-0.04}^{+0.06}$&$2.65_{-0.07}^{+0.08}$&$2.71_{-0.09}^{+0.06}$&$2.66_{-0.08}^{+0.07}$&$2.69_{-0.08}^{+0.06}$&$2.80_{-0.07}^{+0.12}$ \specialrule{0em}{1pt}{1.1pt}
&$\mathrm{log}\xi$&$3.72_{-0.16}^{+0.13}$&$3.54_{-0.13}^{+0.22}$&$3.53_{-0.15}^{+0.46}$&$3.70_{-0.21}^{+0.13}$&$3.72_{-0.19}^{+0.29}$&$3.73_{-0.12}^{+0.15}$&$3.71_{-0.18}^{+0.05}$\specialrule{0em}{1pt}{1.1pt}
&$R_\mathrm{f}$&$0.35_{-0.10}^{+0.07}$&$0.29_{-0.03}^{+0.06}$&$0.27_{-0.04}^{+0.07}$&$0.44_{-0.11}^{+0.05}$&$0.41_{-0.08}^{+0.06}$&$0.44_{-0.07}^{+0.16}$&$0.61_{-0.14}^{+0.12}$ \specialrule{0em}{1pt}{1.1pt}
&$N_\mathrm{r}$&$0.184_{-0.031}^{+0.038}$&$0.091_{-0.011}^{+0.016}$&$0.099_{-0.022}^{+0.024}$&$0.065_{-0.016}^{+0.013}$&$0.051_{-0.012}^{+0.011}$&$0.04_{-0.009}^{+0.008}$&$0.057_{-0.012}^{+0.026}$ \midrule
& $\chi^2$ &\multicolumn{7}{c}{390.40} & $\nu$ & \multicolumn{7}{c}{452} & $\chi^2_{\nu}$ & \multicolumn{7}{c}{0.86} \bottomrule
\end{tabular}
Notes. Columns 3-9 show successively the results of Spec. A-G. The parameters with “ f ” in parenthesis indicate they were fixed at values given. All errors for one parameter of interest were calculated with 90% confidence level. 1. Parameters including the spin $a_*$, the inclination angle $i$ and the iron abundance $A_\mathrm{Fe}$ of relxill were linked together among different spectra. 2. Parameters including the temperature $T_\mathrm{col}$, normalization constant $N_\mathrm{DISC}$, emissivity index $q$, photon index $\Gamma _\mathrm{r}$, ionization state $\mathrm{log}\xi$, reflection fraction $R_\mathrm{f}$ and the normalization $N_\mathrm{r}$ were independent for each spectrum.
upper limit, which indicates that the density in the disc is larger than the maximal value in the model.
We found that the spin parameter pegged at -0.998 when the iron abundance was decreased to $\sim{3}$. To better understand this surprising finding, we did another trial. We fixed the iron abundance at values between 3.5-6.5 with a stepsize of 1.0, and found four best fits with $\chi^2_{\nu}$ < 1 (d.o.f = 453). We, then, explored the $\chi^2$ for spins from 0 to 1.0 with stepsize of 0.01 for them (Figure 7). These four models all obtain moderate spin black holes at 90% statistical confidence level, but when the iron abundance is higher than $\sim{5.5}$, at more than 90% statistical confidence level, we note the reduced sensitivity of models to large values of spin. The increase of iron abundance induces more photoelectric absorption making the Fe K-edge near 8 keV deeper. At the same time, the strength of the Fe k emission in the band of 6-8 keV increases (Garcíaetal.2013). More fluorescent iron photons near the black hole would be scattered down below $\sim{6}$ keV to make a stronger red wing, as expected.
Then, we further explored the dependence of the spin parameter $a_*$ on the inclination angle $i$. We fit spectra for 60 evenly-spaced values of the $a_*$ in the range of 0.4-1.0 and 30 evenly-spaced values of the $i$ with the range of 20 $^{\circ}$ -50 $^{\circ}$ (Figure 8). When $i$ is larger than $\sim{36}^{\circ}$, a positive relationship is shown. Moreover, the model lost the ability to give an upper limit on the spin parameter at 99% confidence level.
The inclination angle ($36.3_{-3.4}^{+5.3}$ degrees) in this paper is consis- tent with the value ($32_{-4}^{+3}$ degrees) in, which may indicate that the inclination angle of the inner disc is misaligned with the orbital inclination angle ($\sim{21}^{\circ}$). However, for a transient system, the timescale for accretion to torque the black hole into alignment is approximated to be $10^{6}-10^{8}$ years (Martin etal.2008). Therefore, the alignment is expected to occur early in the typical lifetime of transients, which are characteristically Gyrs old (White&Ghosh1998; Fragosetal.2013). In our estimation, the most likely resolution to this apparent tension lies in the reflection modeling. The inclination angle estimation via X-ray reflection fit- ting method is principally determined by the blue wing of the broad Fe line. The high density model leads to increasing soft X-ray flux. Recent reflection analyses of Cyg X-1 by Tomsicketal. (2018), and GX 339-4 by Garcíaetal. (2015) and Jiangetal. (2019) suggest that reflection models which underestimate the density of disc introduce systematic changes of order 10 $\degr$ in the inclination angle.
Shafeeetal. (2006) first reported the spin of the black hole in 4U 1543 via the continuum-fitting method. They estimated its spin to be $0.8\pm0.1$. Then, and Morningstar&Miller (reported two spin measurements, $0.3\pm0.1$ and $0.43^{+0.22}_{-0.31}$, respectively, both constrained by combining the continuum-fitting and X-ray reflection fitting methods. These three works utilized the dynamical parameters which were reported in Park et al. (2004), but found conflicting values of spin. For measuring the spin of a black hole via the continuum-fitting method to succeed, one con- strains the size of the emitting region via the efficient blackbody-like property of the optically thick disc. To relate the emitting area to a dimensionless ISCO and thereby spin, it is critical to have accurate measurements of the distance to the source, the mass of the black hole, and the inclination angle of the accretion disc (McClintock etal.2011; McClintocketal.2006; Gouetal.2011).
The spin and the inclination angle measured by modeling the reflection emission with relxill in this paper is consistent with those reported by Morningstar&Miller (2014). However, the spin measurement is in conflict with the one reported by Miller et al. (2009). Milleretal. (2009) assumed the inclination angle of accre- tion disc is equal to the orbital inclination angle.
Accordingly, we test the implication of the lower inclination angle on the spin measurement. When the inclination angle param- eter in REXILL model was fixed at the orbital inclination angle of 21.0 $^{\circ}$, which is considered as Model 2, we find that the fit be-
| |
200211922/7
|
Table3. Best-fitting parameters for Model 2: crabcor*TBabs *(diskbb + relxill), in which $i$ = 21 deg
\begin{tabular}{llccccccc}
\rightarrowprule
Model & Parameter &Spec. A & Spec. B & Spec. C & Spec. D & Spec. E & Spec. F & Spec. G \midrule
{\tt crabcor}&$C$ &\multicolumn{7}{c}{1.097 (f)}\specialrule{0em}{1pt}{1.1pt}
&$\Delta\Gamma$ &\multicolumn{7}{c}{0.01 (f)}\specialrule{0em}{1pt}{1.1pt}
{\tt TBabs}&$N_\mathrm{H}$ ($cm^{-2}$) &\multicolumn{7}{c}{$4.0\times10^{21}$ (f)}\specialrule{0em}{1pt}{1.1pt}
{\tt diskbb} &$T_\mathrm{col}$(keV)&$0.842_{-0.013}^{+0.015}$&$0.814_{-0.015}^{+0.016}$&$0.829\pm0.015$&$0.784\pm0.013$&$0.783\pm0.007$&$0.770_{-0.007}^{+0.008}$&$0.770_{-0.011}^{+0.010}$ \specialrule{0em}{1pt}{1.1pt}
&$N_\mathrm{DISC}$&$4349_{-291}^{+220}$&$4471_{-374}^{+376}$&$3778_{-301}^{+229}$&$5021_{-383}^{+388}$&$5124_{-216}^{+221}$&$5110_{-272}^{+251}$&$5052_{-306}^{+355}$ \specialrule{0em}{1pt}{1.1pt}
{\tt relxill}
&$q$&$2.72_{-0.13}^{+0.10}$&$2.69\pm0.13$&$2.44\pm0.17$&$2.78_{-0.14}^{+0.08}$&$2.67\pm0.09$&$2.59_{-0.06}^{+0.11}$&$2.76_{-0.15}^{+0.10}$\specialrule{0em}{1pt}{1.1pt}
&$a_*$&\multicolumn{7}{c}{$> 0.83$}\specialrule{0em}{1pt}{1.1pt}
&$i$ (deg)&\multicolumn{7}{c}{$21.0$ (f)}\specialrule{0em}{1pt}{1.1pt}
&$A_\mathrm{Fe}$&\multicolumn{7}{c}{$7.72_{-1.62}^{+1.23}$}\specialrule{0em}{1pt}{1.1pt}
&$\Gamma_\mathrm{r}$&$2.64_{-0.03}^{+0.04}$&$2.48_{-0.03}^{+0.04}$&$2.51_{-0.03}^{+0.04}$&$2.56_{-0.04}^{+0.09}$&$2.52\pm0.03$&$2.53_{-0.02}^{+0.04}$&$2.59_{-0.05}^{+0.14}$ \specialrule{0em}{1pt}{1.1pt}
&$\mathrm{log}\xi$&$4.52_{-0.15}^{+4.52}$&$4.19_{-0.35}^{+0.25}$&$4.26_{-0.34}^{+0.15}$&$4.07_{-0.34}^{+0.25}$&$4.29_{-0.20}^{+0.09}$&$4.19_{-0.23}^{+0.16}$&$4.01_{-0.23}^{+0.35}$\specialrule{0em}{1pt}{1.1pt}
&$R_\mathrm{r}$&$0.21_{-0.09}^{+0.06}$&$0.23_{-0.01}^{+0.06}$&$0.18_{-0.03}^{+0.04}$&$0.3_{-0.04}^{+0.06}$&$0.32\pm0.04$&$0.32_{-0.04}^{+0.05}$&$0.36_{-0.06}^{+0.13}$ \specialrule{0em}{1pt}{1.1pt}
&$N_\mathrm{r}$&$0.111_{-0.011}^{+0.015}$&$0.063_{-0.008}^{+0.010}$&$0.064_{-0.007}^{+0.010}$&$0.04_{-0.003}^{+0.015}$&$0.032_{-0.003}^{+0.004}$&$0.023_{-0.003}^{+0.004}$&$0.031_{-0.005}^{+0.016}$ \midrule
& $\chi^2$ &\multicolumn{7}{c}{443.71} & $\nu$ & \multicolumn{7}{c}{453} & $\chi^2_{\nu}$ & \multicolumn{7}{c}{0.98} \bottomrule
\end{tabular}
Notes. Columns 3-9 show successively the results of Spec. A-G. The parameters with “ f ” in parenthesis indicate they were fixed at values given. All errors for one parameter of interest were calculated with 90% confidence level. 1. Parameters including the spin $a_*$ and the iron abundance $A_\mathrm{Fe}$ of relxill were linked together among different spectra. 2. Parameters including the temperature $T_\mathrm{col}$, normalization constant $N_\mathrm{DISC}$, emissivity index $q$, photon index $\Gamma _\mathrm{r}$, ionization state $\mathrm{log}\xi$, reflection fraction $R_\mathrm{r}$ and the normalization $N_\mathrm{r}$ were independent for each spectrum.
Table4. Best-fitting parameters for Model 3: crabcor*TBabs *(diskbb + relxill), in which $i$ = 21 deg and $a_{*}$ = 0.30
\begin{tabular}{llccccccc}
\rightarrowprule
Model & Parameter &Spec. A & Spec. B & Spec. C & Spec. D & Spec. E & Spec. F & Spec. G \midrule
{\tt crabcor}&$C$ &\multicolumn{7}{c}{1.097 (f)}\specialrule{0em}{1pt}{1.1pt}
&$\Delta\Gamma$ &\multicolumn{7}{c}{0.01 (f)}\specialrule{0em}{1pt}{1.1pt}
{\tt TBabs}&$N_\mathrm{H}$ ($cm^{-2}$) &\multicolumn{7}{c}{$4.0\times10^{21}$(f)}\specialrule{0em}{1pt}{1.1pt}
{\tt diskbb} &$T_\mathrm{col}$(keV)&$0.853\pm0.013$&$0.83\pm0.015$&$0.837_{-0.016}^{+0.013}$&$0.798\pm0.009$&$0.791_{-0.008}^{+0.007}$&$0.778_{-0.007}^{+0.006}$&$0.783_{-0.010}^{+0.008}$\specialrule{0em}{1pt}{1.1pt}
&$N_\mathrm{DISC}$&$4127_{-172}^{+262}$&$4103_{-341}^{+346}$&$3586_{-242}^{+331}$&$4595_{-259}^{+324}$&$4868_{-212}^{+250}$&$4862_{-192}^{+239}$&$4603_{-241}^{+360}$\specialrule{0em}{1pt}{1.1pt}
{\tt relxill}
&$q$&$2.75\pm0.15$&$2.75_{-0.15}^{+0.07}$&$2.49_{-0.20}^{+0.18}$&$2.76_{-0.19}^{+0.15}$&$2.72_{-0.11}^{+0.10}$&$2.61_{-0.13}^{+0.12}$&$2.65\pm0.23$\specialrule{0em}{1pt}{1.1pt}
&$a_*$&\multicolumn{7}{c}{$0.30$ (f)}\specialrule{0em}{1pt}{1.1pt}
&$i$ (deg)&\multicolumn{7}{c}{$21.0$ (f)}\specialrule{0em}{1pt}{1.1pt}
&$A_\mathrm{Fe}$&\multicolumn{7}{c}{$4.5_{-0.21}^{+0.46}$ }\specialrule{0em}{1pt}{1.1pt}
&$\Gamma_\mathrm{r}$&$2.65\pm0.03$&$2.5_{-0.03}^{+0.05}$&$2.54\pm0.03$&$2.62_{-0.06}^{+0.08}$&$2.54\pm0.02$&$2.56\pm0.03$&$2.73_{-0.12}^{+0.06}$ \specialrule{0em}{1pt}{1.1pt}
&$\mathrm{log}\xi$&$4.31_{-0.63}^{+0.22}$&$3.89_{-0.36}^{+0.29}$&$4.03_{-0.27}^{+0.30}$&$3.73_{-0.21}^{+0.31}$&$4.06_{-0.18}^{+0.24}$&$4.0_{-0.20}^{+0.17}$&$3.63_{-0.16}^{+0.27}$\specialrule{0em}{1pt}{1.1pt}
&$R_\mathrm{r}$&$0.18\pm0.03$&$0.19_{-0.02}^{+0.03}$&$0.17\pm0.02$&$0.29\pm0.06$&$0.27_{-0.02}^{+0.04}$&$0.29\pm0.03$&$0.39_{-0.06}^{+0.09}$ \specialrule{0em}{1pt}{1.1pt}
&$N_\mathrm{r}$&$0.115_{-0.012}^{+0.025}$&$0.069_{-0.008}^{+0.012}$&$0.07_{-0.008}^{+0.009}$&$0.049_{-0.009}^{+0.013}$&$0.034\pm0.003$&$0.026\pm0.003$&$0.046_{-0.014}^{+0.009}$ \midrule
& $\chi^2$ &\multicolumn{7}{c}{456.69} & $\nu$ & \multicolumn{7}{c}{454} & $\chi^2_{\nu}$ & \multicolumn{7}{c}{1.01} \bottomrule
\end{tabular}
Notes. Columns 3-9 show successively the results of Spec. A-G. The parameters with “ f ” in parenthesis indicate they were fixed at values given. All errors for one parameter of interest were calculated with 90% confidence level. 1. The iron abundance $A_\mathrm{Fe}$ of relxill were linked together among different spectra. 2. Parameters including the temperature $T_\mathrm{col}$, normalization constant $N_\mathrm{DISC}$, emissivity index $q$, photon index $\Gamma _\mathrm{r}$, ionization state $\mathrm{log}\xi$, reflection fraction $R_\mathrm{r}$ and the normalization $N_\mathrm{r}$ were independent for each spectrum.
## ACKNOWLEDGEMENTS
We thank the useful discussion with Prof. J.Orosz, Prof. Youjun Lu, Dr. Erlin Qiao, Dr. Weiwei Xu, and Dr. Zhu Liu. We would also like to thank the reviewer for his/her valuable input. Lijun Gou are supported by the National Program on Key Research and Development Project through grant No. 2016YFA0400804, and by the National Natural Science Foundation of China with grant No. U1838114, and by the Strategic Priority Research Program of the Chinese Academy of Sciences through grant No. XDB23040100. We also thank RXTE/PCA public data and facilities. This work is made under the help with tools available on Astrophysics Sci- ence Archive Research Centre (HEASARC), belonging to NASA’s Coddard Space Flight Centre (GSFC).
| |
200211922/2
|
holes in Milky Way galaxy (Brown&Bethe1994; Timmesetal. 1996), however, only roughly two dozen dynamically-confirmed black hole X-ray binaries have been confirmed (Remillard&Mc- Clintock 2006), and 4U 1543-47 (4U 1543) is one of them. This transient source was first discovered by Uhuru satellite in 1971 (Matilsky et al. 1972). Then, it went into outburst again in 1983, 1992 and 2002, respectively (Kitamotoetal.1984; Harmonetal. 1992; Parketal.2004). It was a very long time after the first dis- covery that the compact primary was confirmed to be a black hole (Rhoades&Ruffini1974; Oroszetal.1998).
As to its spin parameter, Shafeeetal. (2006) first reported its spin with the continuum-fitting method. They estimated its spin to be $0.8\pm0.1$. Then, and Morningstar&Miller (reported two spin measurements, $0.3\pm0.1$ and $0.43^{+0.22}_{-0.31}$, respectively, both constrained by combining the continuum-fitting and the X-ray reflection fitting methods. These three works utilized the previous mass of $9.4 \pm{1.0} M_{\odot}$ and distance of $7.5 \pm{1.0}$ kpc, which were reported in Parketal. (2004). Except that Milleretal. (2009) used the inclination angle of $32_{-4}^{+3}$ degrees constrained by their own fits to the iron line, the other two works used the inclination angle of $20.7 \pm{1.5}$ degrees which is equal to the binary orbital inclination angle. Recently, an updated set of dynamical parameters have been identified which have significant differences compared to earlier (J.Orosz, private communication).
In this paper, we revisited 7 Rossi X-ray Timing Explorer (RXTE, Zhangetal.1993) data-sets of 4U 1543 to check the spin parameter via the X-ray reflection fitting method, i.e., carefully ex- ploring the reflection component. We made a joint-fit for all spectra in order to achieve a better signal-to-noise ratio. we have adopted the updated reflection-emission model, relxill (Garcíaetal.2014a; Dauser et al. 2014). The whole paper is organized as follows. In Section 2, we provide details of the data reduction and selection. In Section 3, we describe the analysis of spectra and the spin re- sult. In Section 4 and Section 5, we present our discussions and conclusions, respectively.
## DATA REDUCTION AND SELECTION
We revisited data sets for 4U 1543 which were observed by RXTE during its 2002 outburst (Parketal.2004). There are 130 contin- uous pointed observations in total (the long exposures were split), collected by the Proportional Counter Array (PCA, Jahoda et al. 1996). We only focused our analysis on the best-calibrated propor- tional counter unit, namely PCU2, as in previous work (Parketal. 2004; Jahodaetal.2006; Shaposhnikovetal.2012). All layers of the PCU2 were combined. We neglected 49 observations whose count rate is smaller than 10 counts s $^{-1}$. The remaining observations are presented in the hardness-intensity diagram (HID, Figure 1).
RXTE /PCA data of bright X-ray binaries are fundamentally limited not by counting statistics but by the systematic measure of calibration certainty in the detector. We apply a calibration correc- tion, pcacorr (Garcíaetal.2014b), which improves the instrumen- tal response to a quality of 0.1% precision. We include this 0.1% as a systematic error. A second correction, crabcorr (Steineretal. 2010), standardizes the PCA absolute flux calibration to the Toor & Seward (1974) values for the Crab. This latter tool improves not on the precision of the detector, but on the accuracy of our measurement.
We firstly subtracted background and made deadtime correc- tion for the RXTE data. Next, the calibration tool pcacorr was applied. Then, a 0.1% systematic error was added to the spectra.
Figure1. The evolutionary tracks in hardness-intensity diagram for obser- vations except those with count rate smaller than 10 counts s $^{-1}$. The vertical axis presents the count rate in energy band 3-45 keV. The horizontal axis presents the hardness ratio (HR) defined as the ratio of count rate between 5-8.6 keV and 8.6-18 keV. The 7 red open circles represent the data (two overlapping) we used to determine the spin of the black hole.
Finally, we performed RXTE data analysis over the energy range between 2.8-45.0 keV using XSPEC 12.9.0g software package (Ar- naud1996). The quoted errors were given with a 90% confidence level ($\Delta \chi ^2=2.71$) if not specified.
We selected 7 observations (MJD 52459-MJD 52463, defined as Spec. A-G) which show strong reflection components. In Table 1, we give the detailed information for these observations. In order to show the reflection features more clearly, we analysed the Spec. A-G between 2.8-45.0 keV, omitting 4.5-8.0 keV and 15.0-35.0 keV with the model crabcor*TBabs*(diskbb+powerlaw) in XSPEC.
For the model crabcor, the normalization coefficient of $C=1.097$ and the slope difference of $\Delta\Gamma = 0.01$, are applied. For the model TBabs, which is used to account for the galactic absorption by the interstellar medium (ISM) along the line of sight, the Wilms et al. (2000) set of solar abundances and the Verner et al. (1996) photoelectric cross sections were specified accordingly. Since the effective low energy of RXTE is limited at 2.8 keV, the data cannot constrain the column density ($N_\mathrm{H}$) well. The column density 1 was fixed at $4.0\times10^{21}$ cm $^{-2}$ as in and Morningstar &Miller (2014).
The fits to all 7 spectra are statistically unacceptable with $\chi^2_{\nu}$ = 31.32 (2129.78/68), 32.59 (2216.33/68), 23.26 (1581.93/68), 36.65 (2492.28/68), 63.75 (4335.57/68), 49.58 (3371.47/68) and 39.70 (2699.92/68), respectively. Data-to-model ratios are plotted in Fig- ure 2. The positive features in residuals are the broadened iron line and Compton hump characteristic of reflection emission.
## ANALYSIS AND RESULTS
XMM-Newton
1 Milleretal. (2003) measured the column density for 4U 1543-47 to be (3.8 $\pm$ 0.2) $\times10^{21}$ cm $^{-2}$ by analysing the /EPIC-pn spectrum (0.3-10.0 keV). This value is consistent with the one we currently use in the fit. The new value of the column density has a negligible effect on our fitting results.
| |
200211922/6
|
Figure3. Making joint-fit to Spec. A-G using model: crabcor*TBabs*(diskbb+relxill). Data-to-model ratios and contributions to the total $\chi^2$ are presented in left and right panels, respectively. Different colors represent different spectra. The model achieved a satisfactory fit with $\chi^2_{\nu}=390.40/452$.
comes worse than when the inclination angle is free ($\Delta \chi^2 = 53.51$ for increasing 1 d.o.f). The best-fitting values are listed in Table 3. The temperature and the normalization of the thermal emission do not change significantly. The photon index parameters become smaller than the Newtonian value ($q=3$), which indicates the coro- nal model changes from a compact geometry ($q>3$ in Model 1) to extended. As for reflected emission, the emissivity index decreased. The spin pegged at 0.998 in this condition, possibly owning to the higher iron abundance of $7.72_{-1.62}^{+1.23}$, a more nonphysical value. The ionization state becomes much higher, and the reflection fraction becomes smaller.
As an extension of Model 2, we define a new Model 3 in which we keep the inclination fixed at 21.0 $^{\circ}$ and also fix the spin parameter to the value found by Milleretal. (2009): $a_{*}$ =0.3. The best-fitting values are listed in Table 4. Compared to Model 2, its $\chi^2$ increased 12.99 for 1 d.o.f. Except the iron abundance was constrained at $4.50_{-0.21}^{+0.46}$, which is lower than that in Model 2, other parameters were not appreciably affected. We also plot the contribution to $\chi^2$ for Spec. A resulting from Model 1-3 in Figure 9. The most pronounced changes in comparing the model differences are residuals around the iron line region ($\sim$ 5.0-8.0 keV).
## CONCLUSIONS
We have measured the spin of 4U 1543 via modeling its reflected components in 7 SPL state observations carefully. The spectra con- sist of 4 different components: the galactic absorption, thermal emis- sion from the accretion disc, power-law emission and reflected emis- sion. We did joint-fit to all the spectra to improve the signal-to-noise ratio of X-ray reflected component. We used the reflection model, relxill, to fit the data. We find a super-solar iron abundance for the disc. At the same time, the disc is highly ionized.
The model with free inclination angle and spin (Model 1) de- scribes the spectra best. The inclination angle of the inner accretion disc is constrained to be $36.3_{-3.4}^{+5.3}$ degrees at 90 statistical confi- dence. When the inclination angle is fixed at the orbital inclination value of 21.0 $^{\circ}$ in Model 2 or Model 3, the statistic becomes sig- nificantly worse, and the spin is larger than 0.83. The best-fitting inclination differs from that of the orbital plane by more than 10 degrees. This may be owed to the systematic limitations of current models which underestimate the density of disc.
Our results indicate a moderate rotation rate for the black hole in 4U 1543. The spin parameter is established to be $0.67_{-0.08}^{+0.15}$ at 90% statistical confidence. At the 99% statistical confidence level, we exclude spins below $a_{*}$ < 0.5 (which also excludes any retrograde geometries).
| |
200211922/4
|
tended to peg at the maximal negative value of -0.998. Therefore, we let the iron abundance $A_\mathrm{Fe}$ free.
The model achieved a statistically good fit with $\chi^2_{\nu}$ = 0.86 (390.4/452) for 7 observations (Table 2). All parameters including the spin and the inclination angle are well constrained. The spin parameter $a_*$ is obtained to be $0.67_{-0.08}^{+0.15}$. The inclination angle $i$ is obtained to be $36.3_{-3.4}^{+5.3}$ degrees. The iron abundance $A_\mathrm{Fe}$ is obtained to be $5.05_{-0.26}^{+1.21}$. Figure shows the data-to-model ratios and the contributions to the total $\chi^2$ of the best-fitting. No distant reflection from the outer disc, the wind or the surface of companion (Wang-Ji etal.2018; Xuetal.2018), was necessary, which is attributed to that the RXTE is not sensitive to the narrow line. When the distant reflection component is added using xillver, the statistic is not improved with $\chi^2_{\nu}$ = 0.87 (385.82/445). The spin is $0.78_{-0.13}^{+0.98}$, and the inclination angle is $41.35_{-6.69}^{+10.86}$ degrees, which is still consistent with the model without xillver.
In order to investigate the effect of different values of column density on our model, especially the main parameters, we tried to let the parameter $N_\mathrm{H}$ free. However, the model was unable to provide any meaningful constraint on $N_\mathrm{H}$ (a detection of $N_\mathrm{H}$ only 1 $\sigma$). This is not surprising given the band of sensitivity for the PCA. Most importantly, this has negligible impact on the fit parameters for the model.
We investigated the model dependence on the high energy cut- off. We made $E_\mathrm{cut}$ vary among 7 observations. Comparing with the best-fitting result we describe above, the fit was only improved with $\Delta \chi^2 = 9.33$ for reducing 7 degrees of freedom (d.o.f), which is not a significant improvement. Meanwhile, the fit did not constrain $E_\mathrm{cut}$ well. This changed setting for $E_\mathrm{cut}$ did not affect profiles of thermal, power-law and reflection emission largely. It still requires a higher super-solar iron abundance of 6.61 $_{-1.17}^{+2.06}$. Moreover, the free $E_\mathrm{cut}$ did not change the inclination angle and the spin of the black hole largely. The spin parameter $a_*$ is obtained to be $0.73_{-0.14}^{+0.10}$. The inclination angle $i$ is obtained to be $34.2_{-3.7}^{+6.8}$ degrees.
For the broken emissivity profile, i.e. the inner index is free, the outer index is fixed at 3, and the break radius is fixed at 15 $R_{\rm{g}}$, is assumed, the fit statistics are not improved with $\chi^2_{\nu} = 0.86$ (388.72/452). It is also found that the best-fitting parameters are not changed significantly, such as the spin parameter $a_*$ is $0.71_{-0.08}^{+0.11}$, the inclination angle $i$ is $35.8_{-2.7}^{+3.6}$ degrees, and the inner indexes are $\sim3-4$. For the single emissivity profile, we explored the implication of freezing the emissivity indexes at typical value ($q$ = 3) instead of being free. The best-fitting model only gave a lower limit 0.65 of spin at 90% confidence level with $\chi^2_{\nu} = 0.88$ (403.92/459), which is slightly worse than the case with free $q$ ($\Delta\chi^2=13.52$ for increasing 7 d.o.f). The inclination angle $i$ is also greater than the orbital inclination angle by approximately 10 $^{\circ}$.
We explored the $\chi^2$ parameter space using the command “ step- par ” for the spin and the inclination angle. During the searching process, at each step the parameters of interest were fixed at incre- mentally stepped values while all other parameters were allowed to fit. For the spin parameter, the stepsize of 0.01 was used from 0.0 to 1.0 (Figure 4). For the inclination angle parameter, the stepsize of 0.1 $^{\circ}$ was explored from 20 $^{\circ}$ to 50 $^{\circ}$ (Figure 5). Three levels of confidence (68%, 90% and 99%) are also marked in both figures. The spin is constrained well between $\sim{0.58-0.82}$ at 90% sta- tistical confidence which is consistent with a moderate spin black hole. Negative and low spins (< 0.5) at more than 99% statistical confidence are ruled out. The inclination angle is constrained to be $\sim{32^{\circ}-42^{\circ}}$ at 90 statistical confidence.
## DISCUSSIONS
In this paper, we have carefully explored the constraint on the spin of the black hole in 4U 1543 on the basis of its reflection emission. We selected 7 SPL state spectra which show strong reflection compo- nent. These spectra were selected from the 2002 outburst observed by RXTE.
According to the phenomenological model of crab- cor*TBabs*smedge(diskbb+powerlaw+ Gauss) in Section 3, the central energy of the Gaussian profile is less than 6.4 keV, suggesting the presence of strong gravitational redshift around the black hole, and the reflection region is con- centrated quite close to the black hole. To improve sensitivity to faint reflection features, we have fitted the 7 spectra simultaneously (Garcíaetal.2015). Milleretal. (2009) found that diskbb and phys- ically rigorous relativistic disc models such as kerrbb performed similarly in their ability to characterize the thermal continuum for the purposes of isolating the reflection signal.
From the best-fitting parameters in Table 2, the color temper- ature $T_\mathrm{col}$ of accretion disc drops down with decreasing luminosity. The relative consistency of $N_\mathrm{DISC}$ indicates that the disc radius and coronal covering cannot change appreciably over the 5-day span in which these observations were accrued. The photon index is rel- atively constant and constrained between 2.6-2.8. The emissivity index is also relatively constant but larger than the canonical value ($q=3$). The high values of $\mathrm{log}\xi$ (> 3.5) indicate that the accretion disc is highly ionized. The reflection fraction $R_\mathrm{f}$, which is defined as the ratio that the power-law emission hit on the disc to that es- caped to the infinite, is in the range of $\sim{0.25-0.65}$. It goes up modestly with decreasing luminosity. Because the hot gas layer at the surface of the disc dilutes the reflection signal by scattering and blurring reflection features (Nayakshin&Kallman2001). The hotter the surface of disc, the more significant the dilution is.
We also report a super-solar iron abundance for the binary sys- tem 4U 1543. Based on the full reflection Model l, we investigated the degeneracy between the iron abundance and the spin by fixing the spin value and fitting for the iron abundance. We divided the spin between the range of 0.4-1.0 into 60 evenly-spaced values, and it is found that 50 of 60 fitted iron abundance $A_{Fe}$ is in the range of 3.5-8.0 in units of solar abundance (Figure 6). The relation indicates that the iron abundance has a weakly positive correlation with the spin. The iron abundance is greater than 4 given the 99% confidence range level.
The very large iron abundance is not unique to 4U 1543. Similar results have been reported in other stellar-mass black hole binaries such as GX 339-4 ($A_\mathrm{Fe}=5\pm1$ solar in García et al. 2015 and $A_\mathrm{Fe}=6.6\pm0.5$ solar in), V404 Cyg ($A_\mathrm{Fe}\sim{5}$ solar in), and Cyg X-1 ($A_\mathrm{Fe}=4.7\pm0.1$ solar in Parkeretal.2015 and $A_\mathrm{Fe}=$ 4.0-4.3 solar in Waltonetal.2016). At present, there is no satisfactory physical explanation for the oc- currence of high iron abundance in these systems. The most likely explanation is the atomic data shortcomings in current reflection models. Tomsicketal. (2018) explored the super-solar iron abun- dance of Cyg X-1 using observation of the intermediate state. They found that the higher electron density ($n_{e} \approx 4 \times 10^{20}$ cm $^{-3}$) model was compatible with solar iron abundance using the high-density model reflionx_hd (a new version of reflionx). However, the range of the photon index in that model is 1.4-2.3, which can not be applied to the observations in our paper. While the maximum den- sity in high-density version of relxill (relxillD) is only 10 $^{19}$ cm $^{-3}$. When it is used to fit the data, the density is pegged to its
| |
200211922/1
|
# The spin measurement of the black hole in 4U 1543-47 constrained with the X-ray reflected emission
Yanting Dong, $^{1,2}$ Javier A. Garc a, $^{3,4}$ James F. Steiner, $^{5,6}$ Lijun Gou $^{1,2}$
$^{1}$ The
National Astronomical Observatories, Chinese Academy of Sciences, Beijing, 100101, China $^{2}$ University of Chinese Academy of Sciences, No.19(A) Yuquan Road, Shijingshan District, Beijing, 100049, China $^{3}$ Cahill Centre for Astronomy and Astrophysics, California Institute of Technology, Pasadena, CA 91125, USA $^{4}$ Dr. Karl Remeis-Observatory and Erlangen Centre for Astroparticle Physics, Sternwartstr. 7, 96049 Bamberg, Germany $^{5}$ CfA, 60 Garden St. Cambridge, MA 02138, USA $^{6}$ MIT Kavli Institute for Astrophysics and Space Research, MIT, 70 Vassar Street, Cambridge, MA 02139
4U 1543-47 is a low mass X-ray binary which harbours a stellar-mass black hole located in our Milky Way galaxy. In this paper, we revisit 7 data sets which were in the Steep Power Law state of the 2002 outburst. The spectra were observed by the Rossi X-ray Timing Explorer. We have carefully modelled the X-ray reflection spectra, and made a joint-fit to these spectra with relxill, for the reflected emission. We found a moderate black hole spin, which is $0.67_{-0.08}^{+0.15}$ at 90% statistical confidence. Negative and low spins (< 0.5) at more than 99% statistical confidence are ruled out. In addition, our results indicate that the model requires a super-solar iron abundance: $5.05_{-0.26}^{+1.21}$, and the inclination angle of the inner disc is $36.3_{-3.4}^{+5.3}$ degrees. This inclination angle is appreciably larger than the binary orbital inclination angle ($\sim$ 21 degrees); this difference is possibly a systematic artefact of the artificially low-density employed in the reflection model for this X-ray binary system.
accretion, accretion discs – X-rays: binaries – stars: individual: IL Lupi, 4U 1543-47
## INTRODUCTION
An astronomical black hole can be readily characterized with two parameters, mass ($M$) and spin ($a_*$), hence it could be described by Kerr metric (Kerr1963). Once we know these two parameters, we can make a complete description to the system. Compared to the mass, the spin is relatively harder to be constrained mainly because it only manifests in the most proximate, strong gravity region. The spin is commonly defined in terms of the dimensionless parameter $a_* = Jc/GM^2$ ($-1 \le a_* \le 1$, where $J$ is the angular momentum of the black hole, $c$ is the speed of light, and $G$ is the gravitational constant).
As to the spin measurement, currently there are two leading approaches: the continuum-fitting method (Zhang et al. 1997; Li etal.2005) and the X-ray reflection fitting method (Iwasawaetal. 1997; Milleretal.2002). Both approaches are based on the funda- mental assumption that the inner edge of the accretion disc extends down to the innermost stable circular orbit ($R_\mathrm{in} = R_\mathrm{ISCO}$), which is a monotonic function of spin parameter $a_*$. $a_{*}=$ -1, 0 and 1 correspond to $R_\mathrm{ISCO}=$ 9, 6 and 1 $R_\mathrm{g}$, respectively, where $R_\mathrm{g}$ is the gravitational radius and is defined to be $R_\mathrm{g} = GM/c^2$. Given the inner radius of the accretion disc, one can readily obtain the spin.
The continuum-fitting method can measure the inner radius by modeling the thermal continuum of the accretion disc using ker- rbb2. kerrbb2 is a combination model of kerrbb and bhspec. kerrbb has three fit parameters, $a_*$, the hardening factor $f$ and the mass accretion rate $\dot{M}$, only two of which can be determined at one time. A look-up table between $f$ and the scaled luminosity using bhspec is generated. Then kerrbb and the table allow one to directly fit for a* and $\dot{M}$ (refer sec. 4.2 of McClintocketal.2006). This method relies on accurate measurement to the system param- eters of mass, distance, and inclination angle (often assumed to be identical to the orbital inclination angle) for the source (Gouetal. 2009; Steineretal.2011; Chenetal.2016). The X-ray reflection fitting method mainly models the relativistic reflection spectrum, which is a combination of fluorescent lines, absorption edges and recombination continua (Wangetal.2017; Waltonetal.2019; Gar- cíaetal.2018). One of the advantages of this technique is that it does not require information on the binary parameters, furthermore, it can make an independent constraint on the inclination angle of the inner disc. There have been consistent check for these two meth- ods on several sources, and they generally showed consistent results (Reynolds2019).
It is expected that there exists billions of stellar-mass black
★ E-mail: ytdong@nao.cas.cn † E-mail: lgou@nao.cas.cn
| |
200211922/3
|
Figure 2. Spec. A-G were analysed using a simple mixture of an absorbed power-law together with multicolor disc blackbody model, respectively. The continuum models were fit over the energy band of 2.8-45.0 keV, ignoring 4.5-8.0 keV and 15-35keV region. Data-to-model ratios are plotted, with those regions added back in after the fit. The dotted green line represents the energy of 6.4 keV, which corresponds to the neutral iron line. The broad iron line and Compton hump, i.e. reflection component, are quite prominent.
Table1. Properties of Spec. A-G
\begin{tabular}{ccccccc}
\rightarrowprule
Spec.&Date&MJD&\tabincell{c}{Count Rate(cts/s/PCU)} & \tabincell{c}{Exp.(s)}&\tabincell{c}{HR(8.6-18/5-8.6 keV)}&\tabincell{c}{$\chi^2_{\nu}$ $^a$(63 d.o.f)}\midrule
A & July 04 & 52459 & 2348 & 1072 &0.35&67.70B & July 05 & 52460 & 1991 & 1120&0.39&63.76C & July 06 & 52461 & 1836 & 800 &0.38&77.15D & July 07 & 52462 & 1432 & 1328 &0.32&57.47E & July 07 & 52462 & 1406 & 3376&0.31&63.49F & July 08 & 52463 & 1179 &3056&0.29&74.34G& July 08 & 52463 & 1184 &1520&0.29&83.88\bottomrule
\end{tabular}
𝑎 The spectrum is fitted with crabcor*TBabs*smedge(diskbb+powerlaw+Gauss) in XSPEC.
We fitted Spec. A-G with a phenomenological model, crab- cor*TBabs*smedge(diskbb+powerlaw +Gauss), in which Gauss and smedge (Ebisawa et al. 1994) are used to model the reflection features. The central energy of the iron line was constrained between 6.0 and 6.97 keV. The width and the normalization were allowed to be free. The width of smedge is fixed at 7.0 keV, the smeared edge could change from 7.0 to 9.0 keV, and the optical depth floated freely. We only focused on the best-fitting models. The detailed information on the quality of the fit for each spectrum is shown in Table 1. The temperatures of thermal emission are $\sim$ 0.75-0.85 keV. The photon indexes of power-law emission are $\sim$ 2.45-2.75, which indicate that the source is in the Steep Power Law (SPL) state during these observations. Line peaks are less than 6.4 keV, which suggest the presence of strong gravitational redshift around the black hole.
We then fit the full ionized reflection spectrum with an so- phisticated model. The reflection information, however, was weak relative to the disc and corona continua, we, therefore, make a joint- fit to Spec. A-G. We used a relativistic reflection model relxill (García et al. 2014a; Dauser et al. 2014) which is a combination of the reflection model xillver (García&Kallman2010; García etal.2011, 2013) and the relativistic convolution kernel relconv (Dauseretal.2010, 2012, 2013). This model is designed to fit the re- flection and the power-law components simultaneously. It has been widely used in recent years for the reflection exploration in stellar- mass black hole binaries and AGNs, sometimes also in neutron star binaries. The returned parameter list contains inner index ($q_\mathrm{in}$), outer index ($q_\mathrm{out}$), and break radius ($R_\mathrm{br}$) which describe the radial dependence of the emissivity of reflection emission; spin parameter ($a_*$), inclination angle ($i$), inner radius ($R_\mathrm{in}$), outer radius ($R_\mathrm{out}$), redshift ($z$) to the source (set to 0 for Galactic systems), photon in- dex ($\Gamma_\mathrm{r}$), ionization state ($\mathrm{log}\xi$), iron abundance ($A_\mathrm{Fe}$), high energy cut-off ($E_\mathrm{cut}$), reflection fraction ($R_\mathrm{f}$), and normalization ($N_\mathrm{r}$).
The overall self-consistent model we adopt here is crab- cor*TBabs(diskbb+relxill). For relxill model, we assumed a single emissivity profile ($q_\mathrm{in}=q_\mathrm{out}=q$) and the inner radius of accretion disc extended down to the ISCO radius ($R_\mathrm{in} = R_\mathrm{ISCO}$). Some parameters were independent for each spectrum: the tempera- ture $T_\mathrm{col}$ and normalization constant $N_\mathrm{DISC}$ of thermal emission; the emissivity index $q$, photon index $\Gamma _\mathrm{r}$, ionization state $\mathrm{log}\xi$, reflection fraction $R_\mathrm{f}$ and normalization $N_\mathrm{r}$ of reflection component. The other parameters were linked together among 7 spectra. The spin parame- ter $a_*$ and the inclination angle $i$ were free. The outer radius was set to default value: $R_\mathrm{out}=400$
$R_\mathrm{g}$.
Because the power-law is extremely steep, we can’t detect the high energy cut-off in these observations. We, then, fixed $E_\mathrm{cut}$ at 300 keV which is a physically reasonable and sufficiently large value for our purposes. Meanwhile, it is beneficial for reducing the complexity of the model. When the iron abundance was fixed at unity (i.e. solar abundance), the model returned an ac- ceptable but not a good fit with $\chi^2_{\nu}$ = 1.56 (707.96/453), and the spin
| |
210406888/2
|
Radial distribution functions $g_\mathrm{OO}(r)$ for LD – SCW (red line with circles) and HD – SCW (green line with squares) compared to experimental data at $\rho=0.23$ g/cm $^{3}$ (Ref. 52, dashed line) and $0.58$ g/cm $^{3}$ (Ref. 17, dotted line).
In this Letter we investigate — with a clear focus on changes of the THz spectral response — the properties of water at two supercritical state points that represent gas-like and liquid-like H-bond dynamics in the spirit of Widom crossover. We uncover to what extent the dra- matic changes in the dynamical H-bond network fluctu- ations associated with crossing the percolation line af- fect cross-correlations in particle motion in a spatially- resolved frequency-domain picture, and how the inter- molecular dipolar correlations of supercritical fluids qual- itatively change their THz responses compared to ambi- ent water.
For this study we selected two supercritical state points (at $T=660$ K) located deep in the two regimes sepa- rated by the Widom line [6] (and by the percolation line of SCW [17]), namely at $\rho=0.2$ g/cm $^{3}$ (low-density, LD – SCW) and at $0.6$ g/cm $^{3}$ (high-density, HD – SCW). For reference, we also performed ambient condition sim- ulations at $300$ K and $0.997$ g/cm $^{3}$ (room-temperature water, RTW). It is noted in passing that we must resort to force field MD (see Supplemental Material, SM [40]) in order to provide the statistical data base that allows us to spatially decompose the long-ranged intermolecular dependencies at THz frequencies. As a first test of the adequacy of these simulations, we investigate the global structure of SCW as revealed by $g_\mathrm{OO}(r)$ in Fig. 1. These radial distribution functions (RDFs) compare most favor- ably with the available experimental data at matching densities (0.23 g/cm $^{3}$ [52] and 0.58 g/cm $^{3}$ [17]) as do the corresponding O – H and H – H correlations [40].
The corresponding microscopic structure is reflected in the H-bonding properties of the fluids. In contrast to RTW, where essentially all molecules form an inter- connected three-dimensional H-bond network, SCW is
Probability that a water molecule belongs to a cluster of size $n_\mathrm{C}$ for LD – SCW (red, where $P(n_\mathrm{C} = 1)=34$ and $P(2)=17$%) and HD – SCW (green). The inset shows the distribution of the water molecules according to the number of H-bonds formed including RTW (black).
characterized by the appearance of isolated patches (“ clusters ”) that fluctuate rapidly on a (sub-) picosecond time scale [53]. The lifetime of these clusters is of course somewhat dependent on the applied H-bond definition [54] and we follow here a variant of the standard geomet- ric criterion [40].
In Fig. 2 we analyze the clustering properties of SCW at the two state points compared to RTW.
While for LD – SCW the relative abundance of the clusters decays almost exponentially with the cluster size, having high- est values for water monomers followed by dimers, the distribution for HD – SCW is clearly bimodal with one maximum at non – H-bonded individual molecules and an- other one close to the total available number of water molecules; these $P(n_\mathrm{C})$ distributions are in good agree- ment with previous simulations [53]. In the inset of Fig. 2 the propensity of the water molecules to form H-bonds is illustrated. For LD – SCW, non – H-bonded and singly H- bonded molecules predominate, while this shifts to singly and doubly H-bonded ones at the higher density. Impor- tantly, the H-bond statistics in both cases is markedly different from RTW, where four-coordinated molecules form a dense network (i.e. a single “ cluster ” so that $P(n_\mathrm{C} = 128) \approx 1$) and $\langle n_\mathrm{HB} \rangle = 3.68 \pm 0.07$. For compari- son: $\langle n_\mathrm{HB} \rangle = 1.84 \pm 0.11$ and $1.00 \pm 0.12$ for HD and LD – SCW, respectively, in agreement with other simulations [54, 55]. We checked on larger systems that these con- clusions are stable when increasing the system size [40]. An independent experimental confirmation of the dra- matic reduction of H-bonding in SCW comes from the Compton profiles obtained from IXS [19] that provide indirect access to $\langle n_\mathrm{HB} \rangle$ via the number of electrons in- volved in hydrogen bonding. The $\langle n_\mathrm{HB} \rangle$ value is then
| |
210406888/4
|
similar manner.
The radially-resolved IR spectra of water reveal in the THz spectral range a characteristic loss of spatial correlations with decreasing density, see panels a – c of Fig. 4.
While for RTW, dipolar correlations up to the second hydration sphere are clearly detectable in quali- tative agreement with previous AIMD analyses [36], for HD – SCW only the cross-correlations within the first hy- dration sphere at roughly 3 Å are important. Even these short-range correlations fade away upon further expan- sion of supercritical water to LD conditions. For LD – SCW, the major correlation features at this distance be- come less intense and steadily red-shift relative to RTW by as much as 250 cm $^{-1}$. Such red-shifts of the posi- tive and negative cross-correlation peaks reflect the in- creased librational freedom of the water molecules in the supercritical state due to the lower degree of H-bonding upon decreasing the density. This general trend is in line with what has been observed in the subcritical regime [28] upon raising the temperature.
In the supercritical regime, our findings therefore ex- tend the Widom line crossover concept from thermody- namics and dynamics [3, 4, 6] now into correlated dipo- lar response and THz vibrational spectra. The presence of weak correlations in the radially-resolved spectra of SCW suggests that some nearest-neighbor dipolar cou- plings are still effective even in supercritical water, while the long-range correlations are strongly suppressed in the LD – SCW case due to the completely disrupted tetrahe- dral H-bond network of water. This directly relates the THz response of gas-like and liquid-like SCW to the per- colation line [17].
Contrasting next the radially-resolved vibrational cor- relations (see panels d – f in Fig. 4) to the dipolar THz re- sponse (panels a – c) it is striking to find out that the distinctly modulated pattern characteristic for RTW be- comes greatly simplified at supercritical conditions. With reference to RTW [36], this implies for SCW the pro- gressively diminishing spatial range and magnitude of cross-correlations of the H-bond umbrella motion at ca. 80 cm $^{-1}$ (2.4 THz), H-bond network stretching at 200 cm $^{-1}$ (6 THz), librations at roughly 500 cm $^{-1}$ (15 THz), as well as the absence of any noteworthy cor- relations in the intramolecular vibrations range at still higher (mid – IR) frequencies. This reflects the radical changes in the nature of the underlying atomic motion at supercritical compared to ambient conditions. The steeply rising positive correlations close to zero frequency can be attributed to increasingly “ ballistic ” movement of those water molecules which are no longer firmly en- gaged in an extended H-bond network in SCW. This pseudo-hard sphere character of water molecules in SCW and the long-range nature of the observed velocity cross- correlations are reminiscent of the similarly far-reaching correlations in the sedimentation of colloidal particles due to instantaneous density fluctuations [58]. Next, the weak and broad feature at 200 – 300 cm $^{-1}$ does not show much correlations even with nearest-neighbor molecules (expected at $r \approx 3$ Å), but instead it is due to single- molecule motion (visible at $r \approx 0$ Å) mainly of H atoms (see Fig. S4 in the SM [40]). This librational single- particle motion is distinctly different from the “ network mode ” of RTW at 200 cm $^{-1}$ (6 THz) that is caused by intermolecular stretching motion of tetrahedrally H- bonded water molecules [36] as evidenced by the promi- nent cross-correlations at $r \approx 3$ Å (panel f in Fig. 4). Our peak assignment of the spectra in Figs. 4 and 3 is strongly supported by QENS experiments [18] which pro- vide a DOS for protons (see Fig. 9 therein) that features a clearly structured response with two pronounced max- ima at roughly 60 and 520 cm $^{-1}$ (7 and 65 meV) for RTW, while only one broad unimodal resonance is ob- served for HD – SCW (at $\rho=0.59$ g/cm $^{3}$ and 653 K) close to 240 cm $^{-1}$ ($\approx 30$ meV). This nicely agrees with our partial VDOS for H atoms as depicted in Fig. S4 [40] with two peaks at about 45 and 500 cm $^{-1}$ for RTW and a single broadband feature at 275 cm $^{-1}$ for HD – SCW.
In conclusion, dramatic changes of the THz response of supercritical water are predicted in this spectral re- gion characteristic of the tetrahedral H-bond network of liquid water, which are traced back to significantly de- creasing long-range intermolecular correlations. In par- ticular, the celebrated “ network mode ” of ambient water at around 200 cm $^{-1}$ (6 THz) vanishes at supercritical conditions, while the libration band red-shifts by as much as 250 cm $^{-1}$. Thus, instead of being bimodal, the super- critical THz spectra show a single but unusually broad, mainly librational feature around 200 – 300 cm $^{-1}$. Un- derstanding the molecular mechanism that leads to the observed THz resonances will be key to interpret future experiments of supercritical water and solutions. In par- ticular, the far-reaching couplings characteristic to the THz response of ambient liquid water are found to be suppressed when crossing the percolation line by mov- ing from high to low densities and thus from liquid-like to gas-like supercritical water, respectively. The associ- ated changes of the H-bond network motion across the Widom line are encoded in changes of the correlated dipo- lar response as probed by THz vibrational spectroscopy, which therefore complements thermodynamic and dy- namic techniques to investigate this crossover concept in supercritical water and beyond. Finally, THz spec- troscopy with its ability to detect changes in the H-bond network between high – and low-density water might be an ideal technique to also shed light on the Widom line in the framework of the controversially discussed (second) liquid-liquid critical point of supercooled water.
We are particularly grateful to Hajime Torii for clar- ifying the details of his IR corrections. We acknowl- edge partial financial support from DFG (MA 1547/11 and Cluster of Excellence “ RESOLV ” EXC 1069). Com- putations were performed at LRZ M ü nchen, HPC –
| |
210406888/1
|
# Correlated Particle Motion and THz Spectral Response of Supercritical Water
Maciej Ś miechowski
Molecular dynamics simulations of supercritical water reveal distinctly different distance- dependent modulations of dipolar response and correlations in particle motion compared to ambient conditions. The strongly perturbed H-bond network of water at supercritical conditions allows for considerable translational and rotational freedom of individual molecules. These changes give rise to substantially different infrared spectra and vibrational density of states at THz frequencies for den- sities above and below the Widom line that separates percolating liquid-like and clustered gas-like supercritical water.
Supercritical matter is an ideal testing ground to sys- tematically study fundamental aspects of fluids as a function of density [1–8]. Although formally being sin- gle phase systems, supercritical fluids nevertheless show two distinct regions that are separated by the so-called Widom line, distinguishing gas-like and liquid-like dy- namical regimes that are reminiscent of the respective subcritical domains [3, 4, 6–8]. Hence, supercritical flu- ids are not only far more complex in nature than believed for a long time, but the distinct regimes also encompass vastly different properties. Transcending fundamentals, the study of liquids above their critical point (CP) is also motivated by their increasing importance for large- scale industrial processes, serving e.g. as environmentally friendly “ green solvents ” [9, 10].
Most notably supercritical water (SCW; CP: $T_{\rm c}=647$ K, $p_{\rm c}=22.1$ MPa, $\rho_{\rm c}=0.32$ g/cm $^{3}$) becomes an increasingly important processing medium for key trans- formations in benign aqueous environments [11]. The latter owes much to the dramatically reduced dielec- tric constant of SCW (from $\sim 80$ at ambient conditions to $\sim 6$ just above the CP) [12]. On the other hand, SCW becomes fairly conducting up to the point of met- alizing above 7000 K in the GPa pressure range [13]. All this makes SCW a “ tunable fluid environment ” with amazing properties [14]. The macroscopic phenomena that underlie these peculiarities at the microscopic level seem to be driven chiefly by significant changes in the hydrogen bond (H-bond) network of water at extreme conditions. Such structure of SCW is well-studied us- ing neutron/x-ray diffraction (ND/XRD) [15–17], quasi- elastic neutron scattering (QENS) [18] with energy trans- fers up to 100 meV ($\approx 800$ cm $^{-1}$) and deep inelastic neu- tron (Compton)/x-ray (Raman) scattering (DINS/IXS) [19–21] in the eV range, thus spanning many orders of magnitude in time but also in length scales depending on momentum transfer. For instance, the dynamic struc- ture factor obtained from incoherent QENS of SCW [18] yields, in the limit $Q<1$ Å $^{-1}$, the density of states of H atoms up to ca. 800 cm $^{-1}$ ($\approx 24$ THz). Complement- ing structure, dynamic relaxation [22] of SCW has been probed as well by NMR.
Experiments on SCW can be interpreted at the molec- ular level by analyzing molecular dynamics (MD) simu- lations that can be directly compared with experimental data [23]. The electronic structure of SCW is accessible via ab initio MD (AIMD) [24] that allow one to com- pute dipole moments and vibrational spectra [25] as well as IXS spectra [21]. Moreover, AIMD simulations also provide computational access to reactions in SCW [26].
When it comes to elucidating H-bond dynamics, in- frared (IR) spectroscopy remains the prime technique to study water [27] including SCW [28,29]. IR spectroscopy is sensitive primarily to atomic and/or molecular motion that gives rise to dipole moment changes along the as- sociated oscillatory displacements [30]. Upon lowering the frequency, i.e. the excitation energy, one can succes- sively probe fast intramolecular vibrations in the 4000 – 500 cm $^{-1}$ (120 – 15 THz) mid-IR range, intermolecular dynamics such as H-bond motion or hindered molecu- lar rotations and translations in the 500 – 50 cm $^{-1}$ (15 – 1.5 THz) far-IR regime [30], and even slower phenomena such as dielectric relaxation at GHz frequencies [31], thus allowing one to probe specific processes on their intrin- sic timescales. Vibrational spectroscopy of water in the mid-IR regime is particularly sensitive to cluster size, H- bond abundance and H-bond strength as the intramolec- ular stretching band of H 2 O is the traditional probe of these [28, 29, 32]. In recent years, however, the far- IR region of the vibrational spectrum, now commonly referred to as the “ terahertz ” (THz) regime (note that 100 cm ${}^{-1} \approx 3$ THz $\approx$ 12 meV), has been appreciated to offer direct insights into the picosecond dynamics — and thus into intermolecular motion such as H-bond net- work dynamics being characteristic to liquid water. Ad- vances in THz laser spectroscopy closed the “ THz gap ” in IR spectroscopy and now permit unprecedented in- sights for instance into the modified H-bond dynamics in the vicinity of molecules [33–35]. For ambient water, it has been shown theoretically that THz spectroscopy in- deed probes specific intermolecular modes of the H-bond network [36]. For aqueous solutions of simple ions, the THz response due to these perturbations has been com- puted [37, 38] and THz peaks could even be assigned to distinct intermolecular modes in the case of molecules [39].
| |
210406888/3
|
found to decrease from 3.40 for RTW to 0.73 for SCW (at $\rho=0.4$ g/cm $^{3}$ and 670 K) [19] in agreement with our findings.
The significant changes in the fluid structure of SCW are expected to alter significantly its IR spectrum [28,29]. Particularly in the THz range, being utmost sensitive to H-bond dynamics, temperature increase is known to lead to major changes as observed experimentally already at subcritical conditions. Below the CP, the two main features of the RTW spectrum, that is the librational mode at $\approx 680$ cm $^{-1}$ ($\approx 20$ THz) and the intermolecu- lar H-bond stretching mode at $\approx 200$ cm $^{-1}$ (6 THz), are known to merge gradually into a single feature centered at $\approx 450$ cm $^{-1}$ ($\approx 13.5$ THz) upon increasing the tem- perature to 510 K along the gas – liquid coexistence curve [28]; note that SCW spectra in the THz range have not yet been measured.
Our THz spectra computed [40] from the Fourier trans- form of the dipole derivative time correlation function (TCF) are shown in Fig. 3. It is well known that flexible force fields without polarizability and/or charge transfer cannot reproduce the THz spectra of RTW, in particular not the eminent H-bond network mode at $\approx 200$ cm $^{-1}$ (6 THz). Several schemes to correct for this deficiency have been proposed and both so-called “ Torii corrections ” [56,57] reach a considerable agreement with experimental and AIMD IR spectra of RTW [36]. In Torii’s schemes, which we generalized in the SM [40] to flexible water models, polarization effects due to H-bonding are in- corporated. In the more advanced version [57], this is achieved by adding both intermolecular charge fluxes to existing H-bonds and local electric field effects to each molecule when computing the total dipole moment ac- cording to Eqs. (S25) – (S28).
We scrutinized [40] the ap- plicability of these corrections to SCW and conclude that they can be neglected in this regime, which is backed up by comparison to available spectra (see Sec. IV.E and Fig. S7 [40]), in stark contrast to RTW. The THz spec- tra obtained with the most recent Torii correction [57] as well as the uncorrected original spectra clearly illustrate this point, see Fig. 3.
In the supercritical regime the THz lineshape function changes dramatically compared to ambient conditions: it becomes unimodal in the first place and centered at roughly 300 cm $^{-1}$ (9 THz) at high density, while for LD – SCW it shifts down to about 200 cm $^{-1}$ (6 THz) accord- ing to Fig. 3. Even more pronounced are the changes in the vibrational density of states (VDOS), see the in- set or Fig. S4 [40]. The librational band at $\approx 500$ cm $^{-1}$ (15 THz) disappears for SCW in both the gas-like and liquid-like dynamical regimes, while the principal peak due to the oxygen atoms movement at $<100$ cm $^{-1}$ ($<3$ THz) shifts to zero frequency and dramatically in- creases in intensity, reflecting the only slightly hindered translational (“ ballistic ”) movement of water molecules in the supercritical fluid. The latter is also evident from
Computed total THz absorption spectra of water at the LD – SCW (red), HD – SCW (green), and RTW (black) state points. The Torii – corrected [57] spectra (solid lines) are compared to the uncorrected spectra (dashed lines), see text. The inset shows the corresponding vibrational density of states with separate contributions due to O (dashed lines) and H (dotted lines) atoms. The zero-frequency peak of LD – and HD – SCW is off scale and the data are magnified in Fig. S4 [40].
the value of the self-diffusion coefficient estimated via the Green-Kubo relation from the center-of-mass veloc- ity TCF. It increases from 0.23 for RTW to 4.92 for HD – SCW to 14.17 Å $^{2}$ /ps for LD – SCW in accordance with the available QENS data [18] (7.65 and 13.70 Å $^{2}$ /ps, respec- tively). Considering separately the O/H atoms VDOS (as defined in Sec. III of the SM [40]), the strikingly chang- ing nature of the $\approx 200$ cm $^{-1}$ (6 THz) band is revealed. While in RTW it is mainly due to oxygen motion, in SCW it is the hydrogen motion that underlies this vibrational activity due to much more rotational freedom of the in- dividual water molecules. This already gives a glimpse into the fundamentally different nature of this resonance at supercritical versus ambient conditions, vide infra.
The intra- and intermolecular correlations underlying both the IR spectra and VDOS can be resolved when applying a spatial decomposition scheme [36]. To this end, either molecular dipole velocities or mass-weighted atomic velocities are projected on a grid to obtain spa- tially resolved charge current or mass-weighted velocity density vector fields which are then auto-correlated [40]. For isotropic liquids, such as RTW and SCW, angular av- eraging leaves only a radial dependence of the spectrum, at each frequency, on the distance from the reference point. The resulting radially-resolved spectral response might be understood most easily in analogy to RDFs: while the latter describe radial correlations in particle density with increasing separation $r$, the former captures vibrational correlations (in IR or particle dynamics) in a
| |
220808195/5
|
where $\boldsymbol{0}$ is a distinguished symbol, which is not in $\Sigma$, called the zero string. The zero string is an anhilitator, i.e., it has the property that $\boldsymbol{0} \circ \sigma = \sigma \circ \boldsymbol{0} = \boldsymbol{0}$ for any $\sigma \in \Sigma$.
Importantly, because $\mathcal{A}$ is complete and deter- ministic, there is exactly one non- $\boldsymbol{0}$ entry in every row of $\boldsymbol{M}^{\sigma}$. Now, we define an operation $\otimes$ be- tween two such matrices. Since the automaton is complete and deterministic, we know there exists a unique $j'$ such that $M^\sigma_{ij'} \neq \boldsymbol{0}$ and, by the same argu- ment, there exists a unique $k'$ such that $M^\sigma_{j'k'} \neq \boldsymbol{0}$. Then, in terms of $j'$ and $k'$, we have
$$
(\boldsymbol{M}^{\sigma} \otimes \boldsymbol{M}^{\sigma'})_{ik} =
\begin{cases}
M^{\sigma}_{ij'}\circ M^{\sigma}_{j' k'}, & \textbf{if} k = k'\\
\boldsymbol{0}, & \textbf{otherwise}
\end{cases}
$$
Clearly, $\boldsymbol{M}^{\sigma} \otimes \boldsymbol{M}^{\sigma'}$, for any $\sigma, \sigma' \in \Sigma$, en- forces that the resulting product still has the prop- erty that there is exactly one element of every row that is not equal to $\boldsymbol{0}$. =-1 5 With $\boldsymbol{M}^{\boldsymbol{\sigma}}$, where $\boldsymbol{\sigma} = \sigma_1 \cdots \sigma_K \in \Sigma^*$ is a string of length $K$, we denote the product of matrices $\boldsymbol{M}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}^{\sigma_K}$.
Proposition Let $\mathcal{A} = \left\langle \Sigma, Q, q_0, F, \delta \right\rangle$ be a complete deterministic finite-state automaton. Let $\mathcal{M} \defeq \Big\{\boldsymbol{M}^{\sigma} \mid \sigma \in \Sigma\Big\}$ be the set of $\mathcal{A}$ ’s symbol transition matrices. Then $(\mathcal{M}^*, \otimes)$ with $\otimes$ defined in Eq. (2) is a transition monoid with $\boldsymbol{E}$ as a dis- tinguished identity element. 6 =-1
Proof. See App. A.
Now we are in a position to discuss the precise sense in which subsequential finite-state transduc- ers are homomorphisms. In the case of a finite-state transducer $\mathcal{T} = \left\langle \Sigma, \Gamma, Q, q_0, F, \delta \right\rangle$, there are two symbol transition matrices. The input symbol tran- sition matrix is defined analogously to that of a finite-state acceptor. However, the output symbol transition matrix is defined slightly differently:
$$
M^\sigma_{ik} \defeq
\begin{cases}
\boldsymbol{\gamma},& \textbf{if} q_i \xrightarrow[]{\sigma / \boldsymbol{\gamma}} q_k \in \delta\\
\boldsymbol{0}, & \textbf{otherwise}
\end{cases}
$$
Indeed, for any product of $K$ output symbol tran- sition matrices $\boldsymbol{M}^{\boldsymbol{\sigma}} = \boldsymbol{M}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}^{\sigma_K}$, the entries of $\boldsymbol{M}^{\boldsymbol{\sigma}}$ have a clear interpretation. Let $\boldsymbol{\sigma} = \sigma_1 \cdots \sigma_K$ and suppose $\boldsymbol{M}^{\boldsymbol{\sigma}}_{ik} = \boldsymbol{\gamma} \neq \boldsymbol{0}$. Then, we know that if we start in state $q_i$ and read in input string $\boldsymbol{\sigma}$ we end up in state $q_k$ and output string $\boldsymbol{\gamma}$. There are exactly $|Q|$ non-zero entries in $\boldsymbol{M}^{\boldsymbol{\sigma}}$. Thus, for any given finite-state transducer there is both an input and output transition monoid, which we denote as $(\mathcal{M}^*_\Sigma, \otimes)$ and $(\mathcal{M}^*_\Gamma, \otimes)$, respectively. Then, the intuition is that SFSTs constitute a homo- morphism over the closure of the symbol transition matrices. We state the more formal result below.
Proposition 3. Let $\mathcal{T} = \left\langle \Sigma, \Gamma, Q, q_0, F, \delta \right\rangle$ be a complete subsequential finite-state transducer. Let $(\mathcal{M}^*_\Sigma, \otimes)$ be the transition monoid associated with the input alphabet $\Sigma$ and let $(\mathcal{M}^*_\Gamma, \otimes)$ be the transition monoid associated with the output alphabet $\Gamma$. Then there exists a homomorphism $f : \mathcal{M}^*_\Sigma \rightarrow \mathcal{M}^*_\Gamma$.
Proof. See App. A.
At a higher level, Proposition 3 simply fixes the bug present in Example 4 by incorporating the state into the values. Moreover, Proposition 3 is a clear generalization of Proposition 1 in that if we have a single-state SFST, we have $1\times 1$ matrices that may be viewed as single symbols and the operation $\otimes$ defined in Eq.(2) reduces to string concatenation.
BrieflybacktoMontague. To put the above re- sults in context, we showed by Proposition 3 that arbitrary complete SFSTs are compositional in the sense of Montague. Specifically, SFSTs encode a homomorphism between the free monoid $\left(\Sigma^*, \cdot\right)$, which is isomorphic to $\left(\mathcal{M}_{\Sigma}^*, \otimes \right)$, and the transition monoid $\left(\mathcal{M}_{\Gamma}^*, \otimes \right)$. This instantiates compositional- ity by directly linking it to the robust class-based characteristics of subsequential functions. There are of course other ways, but subsequentiality gives us a theoretical underpinning to use learning SFSTs from finite data as a benchmark for compositional- ity, and motivates our experiments and analysis in the coming sections. =-1
## Experimental Methods
Next we introduce the methodology for generating our datasets as well as the neural and symbolic models employed in the empirical portion of our paper. =-1
## Generating Random SFSTs
We generate random SFSTs using the following stochastic process. We first generate random un- labeled directed graphs that correspond to the symbol-specific transition matrices. Given a set
5 This property is reminiscent of the “ tails ” of a subsequen- tial function (Oncinaetal., 1993)
6 The element $\boldsymbol{E}$ can be thought of as an $N \times N$ matrix where every element is $\varepsilon$.
| |
220808195/7
|
Figure 2: As the number of states in SFST increases, accuracy drops to nearly zero.
give a practical speed-up, we limit the number of samples we provide to OSTIA to 1000, which is only 5% to 10% of what the neural transducers train on. This keeps OSTIA’s run time roughly equivalent to its neural counterparts, but unlevels the playing field, as it were. =-1
## Results and Discussion
We train a neural network, as described in § 4.3, on the datasets taken from our sampled SFSTs. We discuss and analyze the results below. =-1
MinimumTransitionCoverage. We define tran- sition coverage of a given transition as the percent- age of samples in the training dataset that cross that transition. The results reported in Fig. 1 reveal a threshold on the number of samples per transition required to comfortably learn the transducer: $400$ samples. This may seem unsurprising given neural networks’ “ notorious thirst ” for data (LakeandBa- roni, 2018). In the vast majority of the cases when we train on a dataset that does not meet our transi- tion coverage threshold, the neural network does not generalize to held-out data. And, indeed, in the few cases where they manage to have non-zero accuracy, we observe that early transitions in such SFSTs have attained sufficient coverage, and are responsible for the above zero performance. These findings give credence to the idea that there is a relatively simple complexity metric on the SFSTs, i.e., transition coverage, that determines whether or not the neural model will generalize.
BiggerModelsGeneralizeBetter. Additionally, we find that our discovered transition coverage threshold is not constant across all network sizes. For instance, when we increase the size of the hid- den layers in the encoder and decoder from $200$ to
Figure3: As SFST state size increases, average accuracy (blue line) decreases. Increasing the neural model’s size improves accuracy (purple line).
$300$ dimensions,
the sequence-to-sequence models are able to generalize on datasets where the transi- tion coverage is lower; see Fig. 3, where the purple line is the average accuracy of the larger model. With the exception of the higher transition cover- age threshold, these models follow the same trend as their lower dimensional counterparts. =-1
OSTIAisSlow. In terms of wallclock time, we find that an open-source implementation of OS- TIA 8 does not scale to dataset sizes above $1000$. This makes it impossible to perform an apples- to-apples comparison of our neural sequence-to- sequence models against OSTIA. On the one hand, reducing the size of the training dataset disadvan- tages OSTIA. On the other hand, providing OSTIA with the full $20,000$ samples did not terminate af- ter $3$ days on a single dataset. OSTIA provably halts after a finite number of steps, but given the above, a proper comparison with neural models is not possible. =-1
## What about SCAN?
We now turn to the SCAN dataset and examine it in light of our findings above. First, we encode SCAN as an SFST. In so doing, we find it to be an outlier in terms of the number of states it requires, which far exceeds the $100$ states of our largest SFST. In fact, we calculate that the full SFST encoding SCAN has $7,728$ states; see a small example in Fig. 4. With a finite dataset size of $20,000$ input – output pairs, it should not be possible to learn SCAN with high accuracy. However, unlike other datasets of a similar size, SCAN turns out to be nearly perfectly learnable in our (random-split) experiments. This
8 The code is available at github.com/alenaks/OSTIA.
| |
220808195/12
|
## Proofs
Proposition1. Let $\Sigma$ and $\Gamma$ be two alphabets. The function $f$ is a string homomorphism between $(\Sigma^*, \circ)$ and $(\Gamma^*, \circ)$ iff it is a minimal non-empty complete subsequential finite-state transducer with one state. =-1
$\Rightarrow$ Assume
a homomorphism $f$ and an alphabet $\Sigma$ are given. Construct a finite-state transducer with $Q = F = \{q_0\}$. Let $\Gamma = \{f(\sigma) \mid \sigma \in \Sigma\}$. Create a transition $q_0 \xrightarrow[]{\sigma / f(\sigma)} q_0$ for every $\sigma \in \Sigma$. This finite-state transducer is minimal (i.e. non-empty and already has one state), therefore it is the minimal representation of $f$; see Choffrut (2003) for more details. =-1
$\Leftarrow$ Since
$\mathcal{T} = \left\langle \Sigma, \Gamma, Q, q_0, F, \delta \right\rangle$ is non-empty, we know that its only state $q_0$ is also a final state. Thus, $\mathcal{T}$ maps $\varepsilon \mapsto \varepsilon$, i.e., the identity element to the identity element. Let $\boldsymbol{\sigma} \in \Sigma^*$ and suppose that under $\mathcal{T}$ we have $\boldsymbol{\sigma} \mapsto \boldsymbol{\gamma}$. Suppose $\boldsymbol{\sigma} = \boldsymbol{\sigma}' \boldsymbol{\sigma}''$. Since $\mathcal{T}$ is subsequential, there is a unique path in $\mathcal{T}$ that transduces $\boldsymbol{\sigma} \mapsto \boldsymbol{\gamma}$ with exactly $|\boldsymbol{\sigma}|$ transitions. Let $\boldsymbol{\gamma}'$ be the output yield of the first $|\boldsymbol{\sigma}'|$ transitions and let $\boldsymbol{\gamma}''$ be the output yield of the next $|\boldsymbol{\sigma}''|$ transitions. Thus, $\boldsymbol{\sigma}' \mapsto \boldsymbol{\gamma}'$ and $\boldsymbol{\sigma}'' \mapsto \boldsymbol{\gamma}''$ which shows that $\mathcal{T}$ is a homomorphism since $\boldsymbol{\gamma} = \boldsymbol{\gamma}' \boldsymbol{\gamma}''$ and $\mathcal{T}$ a single-state transducer and we remain in that state. Finally, we assumed completeness so that $f$ is a total function over $\Sigma^*$. =-1
Proposition2. Let $\mathcal{A} = \left\langle \Sigma, Q, q_0, F, \delta \right\rangle$ be a complete deterministic finite-state automaton. Let $\mathcal{M} \defeq \Big\{\boldsymbol{M}^{\sigma} \mid \sigma \in \Sigma\Big\}$ be the set of $\mathcal{A}$ ’s symbol transition matrices. Then $(\mathcal{M}^*, \otimes)$ with $\otimes$ defined in Eq. (2) is a transition monoid with $\boldsymbol{E}$ as a distinguished identity element. 9 =-1
Proof. We define $\boldsymbol{E}$ to be the identity element, i.e., for any $\boldsymbol{A} \in \mathcal{M}^*$, we define $\boldsymbol{E} \otimes \boldsymbol{A} \defeq \boldsymbol{A} \defeq \boldsymbol{A} \otimes \boldsymbol{E} = \boldsymbol{A}$. Closure follows from the fact for $\boldsymbol{A}, \boldsymbol{B} \in \mathcal{M}^+$, we have that $\boldsymbol{A} \otimes \boldsymbol{B}$ has by construction exactly one non-zero entry in every row, as elaborated upon in the main text. To check associativity, consider
$$
\left(\boldsymbol{M}^{\boldsymbol{\sigma}} \otimes \boldsymbol{M}^{\boldsymbol{\sigma}'} \otimes \boldsymbol{M}^{\boldsymbol{\sigma}''}\right)_{ik} = M^{\boldsymbol{\sigma}} _{ij} M^{\boldsymbol{\sigma}'}_{jj'} M^{\boldsymbol{\sigma}''}_{j'k}
$$
for some $i, j, j', k$. By the associativity of string concatenation (including when it is augmented to include the zero string), we have that
$$
(M^{\boldsymbol{\sigma}} _{ij} M^{\boldsymbol{\sigma}'}_{jj'}) M^{\boldsymbol{\sigma}''}_{j'k} = M^{\boldsymbol{\sigma}} _{ij} (M^{\boldsymbol{\sigma}'}_{jj'} M^{\boldsymbol{\sigma}''}_{j'k})
$$
which in turn implies that
$$
\left((\boldsymbol{M}^{\boldsymbol{\sigma}} \otimes \boldsymbol{M}^{\boldsymbol{\sigma}'}) \otimes \boldsymbol{M}^{\boldsymbol{\sigma}''}\right)_{ik} = \left(\boldsymbol{M}^{\boldsymbol{\sigma}} \otimes (\boldsymbol{M}^{\boldsymbol{\sigma}'} \otimes \boldsymbol{M}^{\boldsymbol{\sigma}''})\right)_{ik}
$$
Thus, we have that $(\mathcal{M}^*, \otimes)$ is a monoid.
Proposition Let $\mathcal{T} = \left\langle \Sigma, \Gamma, Q, q_0, F, \delta \right\rangle$ be a complete subsequential finite-state transducer. Let $(\mathcal{M}^*_\Sigma, \otimes)$ be the transition monoid associated with the input alphabet $\Sigma$ and let $(\mathcal{M}^*_\Gamma, \otimes)$ be the transition monoid associated with the output alphabet $\Gamma$. Then there exists a homomorphism $f : \mathcal{M}^*_\Sigma \rightarrow \mathcal{M}^*_\Gamma$.
Proof. We define $f$ as follows. First, we define $f(\boldsymbol{E}) \defeq \boldsymbol{E}$. Next, consider an arbitrary element $\mathcal{M}^*_\Sigma$. Because $\mathcal{M}^*_\Sigma$ is a Kleene closure, we may write this arbitrary element as $\boldsymbol{M}_{\Sigma}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}_{\Sigma}^{\sigma_K}$. Now we define $f(\boldsymbol{M}_{\Sigma}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}_{\Sigma}^{\sigma_K}) \defeq \boldsymbol{M}_{\Gamma}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}_{\Gamma}^{\sigma_K}$. Thus,
\begin{align*} &f(\boldsymbol{M}_{\Sigma}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}_{\Sigma}^{\sigma_K}) \\
&= f\left((\boldsymbol{M}_{\Sigma}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}_{\Sigma}^{\sigma_k}) \otimes (\boldsymbol{M}_{\Sigma}^{\sigma_{k+1}} \otimes \cdots \otimes \boldsymbol{M}_{\Sigma}^{\sigma_K})\right) \
&= \left(\boldsymbol{M}_{\Gamma}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}_{\Gamma}^{\sigma_{k}}\right) \otimes \left(\boldsymbol{M}_{\Gamma}^{\sigma_{k+1}} \otimes \cdots \otimes \boldsymbol{M}_{\Gamma}^{\sigma_K}\right)\\
&= \boldsymbol{M}_{\Gamma}^{\sigma_1} \otimes \cdots \otimes \boldsymbol{M}_{\Gamma}^{\sigma_K} \end{align*}
This proves $f$ is a homomorphism.
9 The element $\boldsymbol{E}$ can be thought of as an $N \times N$ matrix where every element is $\varepsilon$.
| |
220808195/2
|
trary SFSTs to generate many different string-to- string datasets and evaluate the behavior of neu- ral sequence-to-sequence models when learning them. By controlling the formal properties of the SFSTs we sample from, we are able to make pre- cise statements about the learnability of systematic phenomena. =-1
Empirically, we find that neural sequence-to- sequence models are, in many cases, capable of per- fectly learning SFSTs from finite data. Moreover, we observe an interesting tendency for neural mod- els to either generalize correctly or to fail outright — with little middle ground. Our analysis reveals a possible explanation for this — generalization seems to be possible when the training data has sufficient coverage, i.e., when every transition in a given transducer is crossed in a minimum number of training examples ($\approx 400$ in our experiments).
We then turn to analyze a popular hand-crafted dataset, SCAN, through the lens of an SFST that encodes it. We find that SCAN is peculiar in that it seems to serve as a counterexample to our transi- tion coverage finding. This suggests that there is a more nuanced story to tell: We predict the learnabil- ity of a language based on a notion of complexity native to subregular languages, but it may be that a more consistently predictive complexity metric would come from a higher point on the Chomsky hierarchy. Future work might seek to limit the number of outlier languages like SCAN in order to identify a notion of complexity that is native to the architecture of the model itself. Such a notion of complexity would identify the level of abstraction that best reflects the representations learned by the model. A fruitful avenue towards this goal might lie in exploring more complex formalisms. =-1
## Finite-State Transducers
This section provides a short technical overview of finite-state transducers and motivates our choice to learn this class of relations.
## Why Learn Finite-State Transductions?
Our study focuses on learning a particular kind of transduction. Specifically, we focus on restricted classes of regular relations, which are those rela- tions computable by finite-state transducers. We believe this is a natural starting point since this class of formal languages is mathematically well- studied, has provable learning guarantees, and has a long use history in linguistics and NLP (Mohri, 1997). =-1
Finite-state transducers also encompass most previous work on compositionality: many datasets, e.g., SCAN (LakeandBaroni, 2018) and gSCAN (Ruis et al., 2020), describe finite string relations and are, therefore, finite-state by definition. These handcrafted datasets have many advantages, like easy interpretability and domain specificity since they directly encode particular relevant relation- ships like movement over a grid or specific lin- guistic phenomena. However, this realism pays the price of diminished robustness of any findings over such datasets (RogersandPullum, 2011), see WhiteandCotterell (2021) for more discussion of this point. By removing the ability to simply adjust properties of the underlying function class, and the transducers which compute it, one loses the possi- bility to experiment more robustly over a function type, rather than just one token instantiation of it. -1
Rather than manually designing individual datasets ourselves, we generate unboundedly many new datasets via randomly sampled SFSTs. This offers a principled view of the problem of learn- ing artificial languages by simply varying proper- ties of the class of transducers that generate them. Furthermore, as we will see in § 6, one may view existing compositionality tasks as learning a spe- cific SFST. We contend this view enables a deeper understanding of modeling results. Both specific artificial languages, such as the compositionality datasets mentioned above, and those randomly sam- pled from a particular function class such as the work presented in this paper, are worth studying. However, our approach has been missing from the compositionality discourse.
## Basic Theory
Now we will overview the basic elements of finite- state theory that will be necessary for the rest of the paper; we start with some definitions. =-1
Definition1. A finite-state automaton (FSA) is a 5-tuple $\mathcal{A}=\left\langle \Sigma, Q, q_0, F, \delta \right\rangle$ where
$\Sigma$ is
- an input alphabet whose elements are denoted $\sigma$; $Q$ is a finite set of states whose elements are denoted $q$; $q_0 \in Q$ is the unique start state; $F \subseteq Q$ is the set of final states;
| |
220808195/6
|
of states $Q$ (let $N = |Q|$) and an input alphabet $\Sigma$, we sample a matrix $\boldsymbol{B}^{\sigma} \in \mathbb{B}^{N \times N}$ for every $\sigma \in \Sigma$ where $\mathbb{B} = \{0, 1\}$. During sampling, we enforce the constraint that there be at most one non- zero entry in every row vector $\mathbf{b}^{\sigma}_{i}$. This constraint ensures that the resulting SFST is subsequential by construction. To achieve this, we sample from $\{1, \ldots, N\}$ uniformly at random for each of the $N$ rows to determine the location of the non-zero entry in each of the $N$ rows. In terms of the SFST, if the entry $b^{\sigma}_{ik} = 1$, then our generated SFST has a transition from state $q_i \overset{\sigma}{\longrightarrow} q_k$ with input symbol $\sigma$. Then, for every transition $q_i \overset{\sigma}{\longrightarrow} q_k$ in our gen- erated SFST, we sample its output symbol $\gamma$ from a uniform distribution over the output alphabet $\Gamma$. This results in a transition $q_i \overset{\sigma / \boldsymbol{\gamma}}{\longrightarrow} q_k$. Finally, to get a canonical representation for particular SFSTs, we perform finite-state minimization (Choffrut, 2003). Minimization also ensures that our sampled SF- STs are canonicalized SFSTs which allows us to check duplicates and ensure the same SFST is not sampled more than once. =-1
## Generating the Datasets
Next we discuss the creation of the datasets we use to investigate the learnability of our sampled SFSTs. We start by sampling $1000$ unique SFSTs using the process described in § 4.1. All SFSTs have an input alphabet $\Sigma$ of 10 symbols and output alphabet of 30 symbols. We consider SFSTs with states numbering from $10$ to $100$ in increments of $10$, and we sample $100$ unique SFSTs for each num- ber. Additionally, we sample extra SFSTs with $21$ to $39$ states because it was at this point that we em- pirically observed performance drop-off during a preliminary investigation. In total, our experiments use $2800$ unique SFSTs.
SamplingInput–outputPairs. To generate the input – output pairs to train and evaluate a neural model, we perform a random walk through the SFST where we select a transition (including the option to halt if we are in a final state) uniformly at random. Following LakeandBaroni (2018), the maximum length of an input string is capped at $50$, i.e., we reject walks with more than $50$ steps during sampling. Using this process, we collect $20,000$ unique input – output pairs from each SFST. We considered larger dataset sizes ($40,000$) for the SFSTs with $29$ to $39$ states since we observe an accuracy drop-off in this region. All datasets are randomly split $80$ – $20$ into training and test sets.
Figure1: Accuracy versus frequency of transition cov- erage, with an inflection point at $400$.
## Neural Sequence-to-Sequence Models
Our experiments make use of?
’s (2019) open- source neural transduction library. 7 Our experi- ments consider an LSTM encoder – decoder with attention in the style of Bahdanau et al. (2015). We use the following hyperparameters: $200$ - dimensional hidden states in the encoder and de- coder, each of which have $2$ layers, maximum gradi- ent clipping normalization of $5$, dropout set to $0.5$, and a batch size of $64$. Additionally, our alphabet tokens are embedded as $100$ -dimensional vectors. We train the model for $100,000$ epochs using the Adam optimizer (KingmaandBa, 2015) with the default learning rate of $0.001$. To determine the effect of model capacity, we also consider a neural sequence-to-sequence model with $300$ -dimensional hidden states and all other hyperparameters kept the same. =-1
## OSTIA
The onward subsequential transducer inference al- gorithm (OSTIA; Oncina et al., 1993) learns the class of subsequential functions from positive pre- sentations of input – output strings. OSTIA works by first building a prefix-tree transducer of the train- ing data, which is then transformed through a series of state-merging operations into the SFST encoding the function the data is drawn from. If a charac- teristic sample is contained in the learning data, OSTIA finds a correct transducer in polynomial (cubic) time. Since OSTIA is designed specifically to learn subsequential relations, it provides a useful baseline. Unfortunately, due to its cubic runtime, OSTIA is too slow to use on larger datasets. To
7 The code is available at shijie-wu/neural-transducer.
| |
220808195/4
|
It is widely held that compositionality is a cor- nerstone of human language. However, language researchers use the term compositionality to refer to a variety of different concepts (Hupkes et al., 2020b). Here, we discuss the specific definition we employ throughout the paper and give a theoretical justification for the use of finite-state transducers as an instantiation of that definition. =-1
## Montague’s Compositionality 3
Montague famously gave a mathematical definition of what it means for a language to be compositional (Montague, 1970)— specifically, he proposed that a relation, e.g., the mapping from syntax to seman- tics, be called compositional if and only if it is a homomorphism, i.e., if it is a structure-preserving map between an input and an output algebra (An- dreas, 2019). To make this notion more formal, we have to be specific about what structure we hope our function preserves. In this paper, we will exclu- sively focus on monoidal structure. We emphasize, however, that the notion of a homomorphism is not restricted to monoids.
Definition3. A monoid is a set $A$ equipped with a binary operation $\bullet$ such that =-1
- For $f, g\in A$, $f \bullet g \in A$ (closure);
- For $f,g,h \in A$, $f \bullet (g \bullet h) = (f \bullet g) \bullet h$ (associativity);
- There exists a unique element $e$ such that for every $g \in A$, we have that $g \bullet e = e \bullet g = g$ (identity).
Definition 4. A free monoid over strings is the structure $(\Sigma^*, \circ)$ where $\Sigma^*$ is the Kleene closure of the alphabet $\Sigma$ and $\circ$ is string concatenation. 4 The empty string $\varepsilon$ is the identity element as $\varepsilon \circ \boldsymbol{\sigma} = \boldsymbol{\sigma} \circ \varepsilon = \boldsymbol{\sigma}$ for any $\boldsymbol{\sigma} \in \Sigma^*$.
Definition Let $(A, \bullet_A)$ and $(B, \bullet_B)$ be two monoids with unique identity elements $e_A$ and $e_B$. We call a function $f : A \rightarrow B$ a homomorphism if it obeys the following two properties:
- 0em
3 A clarifying note natural language not
$f(e_A) = e_B$;
- $f(x \bullet_A y) = f(x) \bullet_B f(y), \quad \forall x, y \in A$
We call a homomorphism between two free monoids a string homomorphism. As it turns out, there is a precise connection between string homomorphisms and finite-state theory.
Proposition 1. Let $\Sigma$ and $\Gamma$ be two alphabets. The function $f$ is a string homomorphism between $(\Sigma^*, \circ)$ and $(\Gamma^*, \circ)$ iff it is a minimal non-empty complete subsequential finite-state transducer with one state. =-1
Proof. See App. A.
Proposition 1 starts to shed light on the connec- tion between Montague’s notion of composition- ality and finite-state transducers. However, this connection is quite weak because multi-state trans- ducers are not covered. We remedy this disparity in the subsequent section.
## Transducers as Homomorphisms
Now we offer a more formal treatment of the exact sense in which SFSTs may be considered homo- morphism and thus fall under Montague’s defini- tion of compositionality. As shown by Proposi- tion 1, in general, SFSTs do not encode string ho- momorphisms. Indeed, it is straightforward to find a counterexample that hammers this point home.
Example 4. Below we exhibit a two-state subse- quential finite-state transducer that is not a string homomorphism. =-1
In the above example, we have $ab \mapsto ab$, but also $a \mapsto a$ and $b \mapsto a$. Thus, it is not a homomorphism.
Example 4 is dissatisfying; it contradicts the intuition that an SFST encodes some notion of Montague-esque compositionality. Luckily, as it turns out, we can find a precise sense in which an SFST is indeed a homomorphism. The idea is to lift the free monoid into a matrix. Given a complete SFSA $\mathcal{A} = \left\langle \Sigma, Q, q_0, F, \delta \right\rangle$ over $|Q| = N$ states, for every symbol $\sigma \in \Sigma$ define an $N
\times N$symbol transition matrix$\boldsymbol{M}^{\sigma}$ whose entries are =-1
$$
M^\sigma_{ik} \defeq
\begin{cases}
\sigma,& \textbf{if} q_i \xrightarrow[]{\sigma} q_k \in \delta\\
\boldsymbol{0}, & \textbf{otherwise}
\end{cases}
$$
: Some readers, post-publication, re- ported that they interpreted our paper as making a general claim about the syntax – semantics interface. Specifically, that it can be well modeled as a regular string- to-string transduction. This was our intent. Instead, our goal is to exhibit a simple benchmark based on SFSTs that is formally compatible with Montague’s definition of composi- tionality and allows for extreme experimentation. =-1
4 When clear from context, we write $\sigma\circ\sigma'$ as $\sigma\sigma'$. =-1
| |
220808195/1
|
# Benchmarking Compositionality with Formal Languages
Josef Valvoda Naomi Saphra Jonathan Rawski Adina Williams Ryan Cotterell University of Cambridge New York University San Jos ´ e State University FAIR ETH Z ¨ u rich
Recombining known primitive concepts into larger novel combinations is a quintessentially human cognitive capability. Whether large neu- ral models in NLP can acquire this ability while learning from data is an open question. In this paper, we investigate this problem from the perspective of formal languages. We use deter- ministic finite-state transducers to make an un- bounded number of datasets with controllable properties governing compositionality. By ran- domly sampling over many transducers, we explore which of their properties contribute to learnability of a compositional relation by a neural network. We find that the models either learn the relations completely or not at all. The key is transition coverage, setting a soft learn- ability limit at $400$ examples per transition. -1
## Introduction
Compositionality is a hallmark of human language (Montague, 1970; Partee, 1995; Fodor, 1998). It is arguably a requirement for any model to count as a model of language, or to achieve human-like natural language understanding. Compositional- ity seems to be such a deep property of language that speakers draw conclusions about the overall meaning of sentences even when the meanings of individual words are not known. For instance, En- glish speakers reading the Jabberwocky (Carroll, 1871) comprehend that the noun phrase slithy toves is built from the composition of the adjective slithy with the plural noun toves, despite lacking a clear understanding of what slithy or toves — let alone their composition — could mean. In cognitive sci- ence, whether neural networks can learn to com- bine a limited number of primitives (in the case of language, word or morphemes) to describe a complex environment has been debated for over 30 years (FodorandPylyshyn, 1988; Marcus, 1998). =-1
In recent work, researchers have explored the in- herent limitations of neural models to exhibit com- positionality by analyzing sequence-to-sequence model performance on small, controlled datasets Dankers et al., 2022; White and Cotterell, 2022). However, the conclusions of these studies are of- ten murky. For instance, LakeandBaroni (2018) cast doubt on neural models’ ability to do com- positional generalization using their toy SCAN dataset, but shortly thereafter, Bastingsetal. (2018) demonstrated that an out-of-the-box sequence-to- sequence model could indeed fully master the task. -1
Instead of hand-crafting small challenge datasets, we propose to test for compositionality by ran- domly sampling from a whole class of string-to- string functions. In doing so, we draw on two linguistic traditions. On the one hand, we follow Montague’s assertion that no important theoretical difference exists between natural and artificial lan- guages (Montague, 1970). Following this logic, the question of whether neural networks composi- tionally process human language is fundamentally equivalent to asking whether they compositionally process artificial languages. On the other hand, we draw lessons from the field of grammatical in- ference (de la Higuera, 2010; Rawski and Heinz, 2019), and evaluate neural sequence-to-sequence models on many automatically generated artificial languages sampled from particular classes of func- tions — as is standard practice at grammatical infer- ence competitions (Balleetal., 2017). =-1
In this paper, we study the class of string func- tions encoded by subsequential finite-state trans- ducers (SFSTs), a restricted class of general finite- state transducers (Mohri, 1997). We sample arbi-
| |
220808195/3
|
$\delta: Q \times \Sigma \cup \{\varepsilon \} \rightarrow Q$ is
- the transition relation and $\varepsilon$ is an empty string.
We denote transitions, i.e., when $q' \in \delta(q, \sigma)$, with the more suggestive notation $q \xrightarrow[]{\sigma} q'$. We say that the automaton $\mathcal{A}$ accepts a string $\boldsymbol{\sigma} \in \Sigma^*$ iff there exists a path 1 through the automaton states $q_0 \xrightarrow[]{\sigma_1} q_1 \xrightarrow[]{\sigma_2} q_2 \cdots \xrightarrow[]{\sigma_N} q_N$ where $q_0$ is the initial state and $q_N \in F$ and $\sigma_1 \cdots \sigma_N = \boldsymbol{\sigma}$. In our notation $\sigma_n$ can be the empty string $\varepsilon$. We call $\boldsymbol{\sigma}$ the yield of the path $q_0 \xrightarrow[]{\sigma_1} q_1 \xrightarrow[]{\sigma_2} q_2 \cdots \xrightarrow[]{\sigma_N} q_N$.
Furthermore, we call an automaton complete when one may transition from every state to every symbol, i.e., $\delta(q, \sigma)$ is defined for all $q \in Q$ and $\sigma \in \Sigma$. And, we say an automaton is deterministic, if, given a state $q$ and an alphabet symbol $\sigma \in \Sigma$, there is at most one transition for $\sigma$ from $q$, i.e., $|\delta(q, \sigma)| \leq 1$. We have that $|\delta(q, \sigma)| = 1$ for all $q \in Q$ and $\sigma \in \Sigma$ if the automaton is both complete and deterministic. In our examples below we denote the final state as a circle with a double border. -1
Example1. Below we exhibit a complete determin- istic finite-state automaton.
The above finite-state automaton accepts the lan- guage $\{(b^{i}a^{j})^n \mid i,j,n \in \mathbb{Z}_{+}\}$.
Definition2. A finite-state transducer (FST) is a 6-tuple $\mathcal{T}=\left\langle \Sigma, \Gamma, Q, q_0, F, \delta \right\rangle$:
$\Sigma$ is
- an input alphabet whose elements are denoted $\sigma$; $\Gamma$ is an output alphabet whose elements are denoted $\gamma$; $Q$ is a finite set of states whose elements are denoted $q$; $q_0 \in Q$ is the unique start state; $F \subseteq Q$ is the set of final states; $\delta: Q \times \Sigma \cup \{\varepsilon \} \rightarrow Q \times \Gamma^*$ is the transition relation.
We denote transitions, i.e., when $(q',\gamma) \in \delta(q, \sigma)$, with the more suggestive notation $q \xrightarrow[]{\sigma / \boldsymbol{\gamma}} q'$. We say that the transducer $\mathcal{T}$ transduces a string $\boldsymbol{\sigma} \in \Sigma^*$ to a string $\boldsymbol{\gamma} \in \Gamma^*$ iff there exists a path $q_0 \xrightarrow[]{\sigma_1 / \boldsymbol{\gamma}_1} q_1 \xrightarrow[]{\sigma_2 / \boldsymbol{\gamma}_2} q_2 \cdots \xrightarrow[]{\sigma_N / \boldsymbol{\gamma}_N} q_N$ where $q_0$ is the initial state and $q_N \in F$, $\sigma_1 \cdots \sigma_N = \boldsymbol{\sigma}$ and $\boldsymbol{\gamma}_1 \cdots \boldsymbol{\gamma}_N = \boldsymbol{\gamma}$. In our notation either $\sigma_n$ or $\boldsymbol{\gamma}_n$ can be the empty string $\varepsilon$, from which it follows that the length $|\boldsymbol{\sigma}|$ must not equal $|\boldsymbol{\gamma}|$. We call $\boldsymbol{\sigma}$ the input yield and $\boldsymbol{\gamma}$ the output yield, respectively, of the path $q_0 \xrightarrow[]{\sigma_1 / \boldsymbol{\gamma}_1} q_1 \xrightarrow[]{\sigma_2 / \boldsymbol{\gamma}_2} q_2 \cdots \xrightarrow[]{\sigma_N / \boldsymbol{\gamma}_N} q_N$.
As in the case of an FSA, we say an FST is complete if $\delta$ is defined for all states and all sym- bols, i.e., $|\delta(q, \sigma)| > 0$ for $q \in Q$ and $\sigma \in \Sigma$. Furthermore, we say an FST is subsequential if it is deterministic with respect to the input, i.e., if $|\delta(q, \sigma)| \leq 1$ for all $q \in Q$ and $\sigma \in \Sigma$ and the FST does not have transitions of the form $q \xrightarrow[]{\varepsilon / \boldsymbol{\gamma}} q'$. By construction, subsequential transducers (SFSTs) are functional, i.e., the string-to-string relations they encode are functions rather than relations. In- deed, it is this functionality that makes them a useful tool for the analysis of neural sequence-to- sequence models. The class of subsequential func- tions are those describable with SFSTs. 2 They are a subclass of the regular relations, but a superclass of the finite relations. =-1
Example 2. Below is an example of a non- deterministic finite-state transducer:
The above transducer only has two paths $q_0 \xrightarrow[]{a / b} q_1 \xrightarrow[]{b / \varepsilon} q_2$ and $q_0 \xrightarrow[]{\varepsilon / b} q_2$. The first transduces $ab \mapsto b$ and the second $\varepsilon \mapsto b$. =-1
Example 3. Below is an example of a complete subsequential transducer over the input alphabet $\Sigma = \{a, b\}$ and output alphabet $\Gamma = \{a, b\}$.
Note the absence of $\varepsilon$ on the input side; however, we do have $\varepsilon$ on the output side.
## Compositionality Formalized
1 A path is a sequence of transitions.
2 Note there are other algebraic and logical characterisa- tions (Oncinaetal., 1993; Bhaskaretal., 2020).
| |
220808195/9
|
dataset removes artefacts in SCAN which are not central to the compositional generalization. They find that models perform worse on gSCAN than on SCAN. More recently, Boginetal. (2022) iden- tify that unobserved “ local structures ” in composi- tionality datasets are harder to learn if no similar structures are observed during training. -1
## Grammatical Inference
Grammatical inference studies the ability to learn- ing classes of formal language from data. Our work focuses on the learning of a restricted class of functions generated by a correspondingly restricted class of finite-state transducers. This allows us to synthesize our study of compositionality in neural models as rule-based inference by neural models, which we can restrict in principled ways. Finite- state machines generalise many techniques in NLP: probabilistic finite-state automata, hidden Markov models, Markov chains, $n$ -grams, probabilistic suf- fix trees, deterministic stochastic probabilistic au- tomata, weighted automata, and other syntactic objects which generate distributions over sets of possible infinite cardinality of strings, sequences, words, trees, and graphs (Vidaletal., 2005). Many grammatical inference studies of neural networks test them on samples drawn from some determinis- tic finite-state acceptors (Cleeremansetal., 1989). See Jacobsson (2005) for a review. =-1
Others experiment with neural networks to see if they can learn languages higher up the Chom- sky hierarchy. LSTMs (Hochreiter and Schmid- huber, 1997) can perform dynamic counting and variably learn simple counter languages such as some $k$ -Dyck languages and $a^{n}b^{n}$ patterns (Weiss 2020; Hewitt et al., 2020), which are generated by a finite-state machine with a counter on top ¨ utzenberger, 1962). In contrast, Avcu et al. (2017) show that LSTM and other RNN architec- tures often fail to learn long-distance dependencies drawn from simpler subregular language classes, even on large benchmarks (MahalunkarandKelle- her, 2019). Nelsonetal. (2020) study the inference of sequence-to-sequence networks, showing that RNN, LSTM, and GRU (Cho et al., 2014) sys- tematically fail to learn a wide range of regular string copying functions generated from a family of two-way transducers, which characterize reg- ular string-to-string functions. When augmented with attention, they reliably learn every function, and the attention history mirrors the derivations of the corresponding two-way transducers. These independently productive strands of work in com- positionality and inference suggest that our work is a reasonable starting point for future interactions.
## Conclusion
We study whether neural sequence-to-sequence models are capable of learning string-to-string transductions with Montague-style composition- ality, i.e., where compositional behavior is defined to be homomorphic. To execute our study, we first provide a theoretical justification of why SFSTs meet Montague’s definition. In the empirical por- tion of the paper, we randomly sample $2800$ SFSTs using the process described in § 4.1, and, then, sam- ple input – output pairs from each SFST to create our unique string-to-string transduction datasets. We find that neural networks tend to either gen- eralize completely or fail miserably — with little middle ground. Moreover, we identify a simple complexity metric, transition coverage, that seems to reliably allow us to predict when an SFST in our randomly sampled dataset is learnable from the dataset sampled from it. =-1
Finally, our paper discusses how analyzing SCAN as an SFST provides a counterexample to our contention that transtion coverages reliably pre- dicts the learnability of an SFST from a given dataset. It seems that while transition coverage is a good metric for subregular languages, there are datasets generated by synchronous context-free grammars that are learnable despite requiring a large number of transitions when encoded as an SFST. For example, SCAN’s learnability is likely due to the fact that the synchronous context-free grammar used to generate it is relatively small and, thus, under a metric such as production coverage it would be considered simple. In conclusion, to get a more complete view of the factors underly- ing learnability, it may be fruitful to consider not only SFSTs, but other formalisms that describe more complex formal relations. We hypothesize that there might yet be a class of automata whose native complexity will more consistently predict the learnability of a language task. =-1
## Acknowledgements
We would like to thank the following people for their valuable insights that made this work possi- ble: Tim Vieira, Hagen Blix, Dieuwke Hupkes,
| |
220808195/8
|
Figure4: A part of the minimal SFST encoding SCAN; in its entirety, it has $> 7000$ states.
result stands in contrast to our randomly generated SFSTs, which exhibit a consistent relationship be- tween the complexity of a formal language and its learnability. =-1
We offer a possible theoretical behavior for SCAN’s surprising learnability. The class of sub- regular relations has a native complexity metric built into the formalism: The size of the SFST itself, as measured by the number of transitions. However, the results on the SCAN dataset indicate that the number of transitions in an SFST is not the native complexity metric of neural networks under consid- eration. This should not come as a surprise because neural networks can learn context-free transduc- tions, which are not even representable by an SFST. While it is interesting that we empirically identify a scaling law that consistently applies to our ran- domly sampled SFSTs, it is not the whole story. =-1
Indeed, SCAN is not a randomly sampled dataset from the class of SFSTs: It is generated by a hand- crafted synchronous context-free grammar (Lake and Baroni, 2018, Figure 6). While SCAN is a complex SFST, requiring thousands of transitions, it can be encoded by a very small synchronous context-free grammar. Casting SCAN as an SFST therefore misrepresents its native complexity. Thus, it is left for future work to identify a class of au- tomata whose native complexity metric can consis- tently predict the learnability of arbitrary language tasks; such a class will surely be higher than SF- STs on the Chomsky hierarchy given our reported results. =-1
## Related Work
Our paper builds on two common strains of re- search: The construction of datasets to benchmark compositional behavior in neural networks and re- search in grammatical inference. =-1
## Compositionality Datasets
There is a growing number of artificial language datasets focused on compositionality. Lake and Baroni (2018) introduced a SCAN dataset, made up of simple navigational text commands. The task is to translate the command in the simple natural language into sequences of actions. One succes- sor to SCAN is the NACS dataset (Bastingsetal., 2018), which is comparable to SCAN, but instead of mapping multiple input signals to a single du- plicated output symbol (e.g., walk twice $\rightarrow$ WALK WALK), NACS does the opposite (WALK WALK $\rightarrow$ walk twice). Since SCAN is a finite language, its inverse NACS is also a finite language, and it can, thus, also be encoded as an SFST. However, this does not hold true for general SFSTs. Inverting an SFST often results in non-subsequential transducer because the output tape of SFSTs is, in general, not deterministic. Another successor is gSCAN (Ruisetal., 2020) focuses on grounding SCAN-like commands in states of a grid world. This makes gSCAN closer to?
’s (2016) grid world grounding for their agents. In contrast to SCAN, gSCAN re- quires the agent to learn differences between sizes and colours of different geometric shapes and in- teract with them, by moving them around the grid world. Executing a gSCAN command is therefore much more difficult than to execute its SCAN coun- terpart. As Ruis et al. (2020) assert, the gSCAN
| |
180603635/5
|
MULTIPLICITY UPON RESTRICTION TO THE DERIVED SUBGROUP
- where $R$ is the regular representation of $\pi_0(Z(\varphi_2))$ and $\chi_R$ its character, thus
$$
\chi_R(g) =
\begin{cases}
0 & \text{if $g$ is not the identity}, \\
|\pi_0(Z(\varphi_2))| & \text{if $g$ is the identity}.
\end{cases}
$$
- Therefore,
$$
\dim\hom_{\pi_0(Z(\varphi_2))}[R, \lambda_\star \chi_1]
=
\chi_1(1).
$$
- By $(*)$ it follows that the pullback of the distribution $\sum _{\chi_2} \chi_2(1) \Theta(\pi_{(\varphi_2, \chi_2)})$ on ${\mathrm G}_2(k)$ to ${\mathrm G}_1(k)$ is equal to $\sum _{\chi_1} \chi_1(1) \Theta(\pi_{(\varphi_1, \chi_1)})$, where the sum is taken over those $\chi_1$ with a given restriction to $Z(\widehat{{\mathrm G}}_1)^{W_k}$. Thus the pullback of the distribution $\sum _{\chi_2} \chi_2(1) \Theta(\pi_{(\varphi_2, \chi_2)})$ on ${\mathrm G}_2(k)$ to ${\mathrm G}_1(k)$ is a stable distribution on ${\mathrm G}_1(k)$ which is what we set out to prove.
Remark2. A weaker version of our conjecture says that the pullback to ${\mathrm G}_1(k)$ of the stable character $\sum_\chi \chi(1)\Theta_\chi$ on ${\mathrm G}_2(k)$ is $\sum_\mu \mu(1)\Theta_\mu$ on ${\mathrm G}_1(k)$, where both of the sums are over the characters of component groups defining fixed pure inner forms that are ${\mathrm G}_2$ and ${\mathrm G}_1$, respectively.
## 3. S ome remarks on the multiplicity formula
Conjecture relating $\mm(\pi_2,\pi_1)$ with $\dim \hom_{\pi_0(Z(\varphi_2))}[\lambda_\star \chi_1, \chi_2]$ can be con- sidered as a set of assertions keeping $\pi_2$ fixed and varying $\pi_1$, or keeping $\pi_1$ fixed and varying $\pi_2$, say inside an $L$ -packet for ${\mathrm G}_2(k)$. It is easy to see that for ${\mathrm G}_1$ and ${\mathrm G}_2$ two reductive groups over a local field $k$, and $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$, a $k$ -homomorphism that is a central isogeny when restricted to their derived subgroups, the image of ${\mathrm G}_1(k)$ inside ${\mathrm G}_2(k)$ is a normal subgroup, and therefore every irreducible rep- resentation of ${\mathrm G}_1(k)$ that appears inside a given irreducible representation $\pi_2$ of ${\mathrm G}_2(k)$ does so with the same multiplicity (depending of course on $\pi_2$). This section aims to prove this as a consequence of our Conjecture 1.
This section is meant to prove that $\dim \hom_{\pi_0(Z(\varphi_2))}[\lambda_\star \chi_1, \chi_2]$ remains con- stant when $\chi_2$ is a fixed character of $\pi_0(Z(\varphi_2))$ but $\chi_1$ varies among characters of $\pi_0(Z(\varphi_1))$. This is achieved by combining Corollary 4 with Lemma 5. We begin with the following lemma whose straightforward proof will be omitted.
Lemma 3. Let $N$ be a normal subgroup of a finite group $G$ with $A=G/N$ an abelian group. Let $\pi$ be an irreducible representation of $N$. Then any two irreducible repre- sentations $\pi_1$ and $\pi_2$ of $G$ containing $\pi$ on restriction to $N$ are twists of each other by characters of $G/N$, i.e.,
$$
\pi_2 \cong \pi_1 \otimes \chi,
$$
for $\chi \colon G/N \longrightarrow \mathbb C^\times$.
Corollary 4. If $N$ is a normal subgroup of a group $G$ with $A=G/N$ a finite abelian group, and $\pi$ an irreducible representation of $N$, then all irreducible $G$ -submodules of $\Ind_N^G(\pi)$ appear in it with the same multiplicity.
Lemma 5. Let ${\mathrm G}_1$ and ${\mathrm G}_2$ be two connected reductive groups over a local field $k$, and let $\lambda \colon {\mathrm G}_1 \longrightarrow {\mathrm G}_2$ be a $k$ -homomorphism that is a central isogeny when restricted to their derived subgroups, and giving rise to a homomorphism ${}^{L}\lambda \colon {}^{L} {\mathrm G}_2 \longrightarrow {}^{L}{\mathrm G}_1$ of
| |
180603635/7
|
MULTIPLICITY UPON RESTRICTION TO THE DERIVED SUBGROUP
denote by $\lambda^\star$. Since $\lambda$ gives an isomorphism ${\mathrm G}_1(k)/P_1(k) \longrightarrow {\mathrm G}_2(k)/P_2(k)$, we have the following commutative diagram:
$$
\begin{xy}
\xymatrix{
\mathcal{R}_{\mathrm{fin}}({\mathrm G}_2(k)) \ar[r]^{\lambda^\star} & \mathcal{R}_{\mathrm{fin}}({\mathrm G}_1(k)) \\
\mathcal{R}_{\mathrm{fin}}(M_2(k)) \ar[r]^{\lambda^\star}
\ar[u]^{\Ind_{P_2(k)}^{{\mathrm G}_2(k)}}
&
\mathcal{R}_{\mathrm{fin}}(M_1(k))
\ar[u]^{\Ind_{P_1(k)}^{{\mathrm G}_1(k)}}
\\
}
\end{xy}
$$
Lemma 7. Let $\sigma_2$ be an irreducible, essentially tempered representation of $M_2(k)$ with strictly positive exponents along the center ${\mathrm Z}(M_2)(k)$ of $M_2(k)$. Write
$$
\lambda^\star \sigma_2
=
\sum_\alpha m_\alpha \sigma_{1,\alpha},
$$
a sum of irreducible, essentially tempered representations of $M_1(k)$ with (finite) multi- plicities $m_\alpha$. Let $\pi_2$ be the Langlands quotient of the standard module $\Ind_{P_2(k)}^{{\mathrm G}_2(k)} \sigma_2$, and $\pi_{1,\alpha}$ the Langlands quotients of $\Ind_{P_1(k)}^{{\mathrm G}_1(k)} \sigma_{1,\alpha}$. Then
$$
\lambda^\star \pi_2 = \sum_\alpha m_\alpha \pi_{1,\alpha}.
$$
Proof. Clearly,
$$
\lambda^\star
\Ind_{P_2(k)}^{{\mathrm G}_2(k)} \sigma_2
=
\Ind_{P_1(k)}^{{\mathrm G}_1(k)} \lambda^\star \sigma_2
=
\sum_\alpha m_\alpha \Ind_{P_1(k)}^{{\mathrm G}_1(k)} \sigma_{1,\alpha}.
$$
Since “ taking maximal semi-simple quotient ” commutes with direct sum, our result follows from Lemma 6.
Corollary 8. If Conjecture 1 is true for tempered representations, then it is true in general.
Proof. Every representation $\pi_2$ of ${\mathrm G}_2(k)$ can be realized as a Langlands quotient of a standard module $\Ind_{P_2(k)}^{{\mathrm G}_2(k)}\sigma_2$ for an essentially tempered representation $\sigma_2$ of $M_2(k)$. The Langlands parameter $\varphi_2 \colon W_F' \longrightarrow {}^{L} {\mathrm G}_2$ for $\pi_2$ is the same as the Langlands parameter $\varphi_2$ for $\sigma_2$ considered as a map $W_F'
\stackrel{\varphi_2}{\longrightarrow} {}^{L} M_2
\longrightarrow {}^{L} {\mathrm G}_2$. The component groups of these parameters, and thus the representations of these component groups, correspond as discussed in [10] * §5. Therefore, our result is a consequence of Lemma 7.
## 5. C onsequences of the conjecture
If the group of connected components $\pi_0(Z_{\widehat{\mathrm G}_1}(\varphi_1))$ is known to be abelian, as is the case when ${\mathrm G}_1$ is any of the groups $\SL_n$, $\U_n$, $\SO_n$, and $\Sp_n$, then our conjec- ture predicts that for any homomorphism $\lambda \colon {\mathrm G}_1 \longrightarrow {\mathrm G}_2$ of connected reductive algebraic groups that is an isomorphism up to center (i.e., $\bar\lambda \colon {\mathrm G}_1/{\mathrm Z}_1 \longrightarrow {\mathrm G}_2/{\mathrm Z}_2$ is an isomorphism of algebraic groups, where ${\mathrm Z}_i$ is the center of ${\mathrm G}_i$), any irreducible representation of ${\mathrm G}_2(k)$ when restricted via $\lambda$ to ${\mathrm G}_1(k)$ decomposes as a sum of irreducible representations of ${\mathrm G}_1(k)$ with multiplicity $\leq 1$.
We note that by our earlier work [1], we know that multiplicity is $\leq 1$ when- ever the pair $({\mathrm G}_1,{\mathrm G}_2)$ is $(\SL_n,\GL_n)$, or (when the characteristic of $k$ is not two)
| |
180603635/12
|
JEFFREY D. ADLER AND DIPENDRA PRASAD
proving it for tempered representations under the assumption that the Langlands correspondence has many of the expected properties.
For their hospitality, we thank the CIRM (Luminy), and the first-named au- thor also thanks the Tata Institute of Fundamental Research (Mumbai). We thank Hengfei Lu for his comments on this work. The final writing of this work was supported by a grant of the Government of the Russian Federation for the state support of scientific research carried out under the supervision of leading scien- tists, agreement 14.W03.31.0030 dated 15.02.2018.
| |
180603635/2
|
JEFFREY D. ADLER AND DIPENDRA PRASAD
We list a few such situations in §5, and prove multiplicity one for restriction from $\GU(n)$ to $\U(n)$ (§6). Along the way, we prove multiplicity one in some other cases where it follows from elementary considerations. In §7, we present an example of a depth-zero supercuspidal representation of quasi-split $\GU(2d,2d)$ that decomposes with multiplicity two upon restriction to $\SU(2d,2d)$. Finally (§8), we give a general procedure for constructing higher multiplicities.
## 2. T he C onjecture on multiplicities
Let ${\mathrm G}^{{\textup{qs}}}_1$ and ${\mathrm G}^{{\textup{qs}}}_2$ be two connected quasi-split reductive groups over a local field $k$ and let $\lambda: {\mathrm G}^{{\textup{qs}}}_1\rightarrow {\mathrm G}^{{\textup{qs}}}_2$ be a $k$ -homomorphism that is a central isogeny when restricted to their derived subgroups. In what follows we will be twisting ${\mathrm G}^{{\textup{qs}}}_1$ by a cohomology class in $H^1(\Gal(\bar{k}/k),{\mathrm G}^{{\textup{qs}}}_1(\bar{k}))$ to construct a pure inner form ${\mathrm G}_1$ of ${\mathrm G}^{{\textup{qs}}}_1$; simultaneously, by twisting ${\mathrm G}^{{\textup{qs}}}_2$ by the image of this class under the map $H^1(\Gal(\bar{k}/k),{\mathrm G}^{{\textup{qs}}}_1(\bar{k})) \rightarrow
H^1(\Gal(\bar{k}/k),{\mathrm G}^{{\textup{qs}}}_2(\bar{k}))$, we will have a pure inner form${\mathrm G}_2$of${\mathrm G}^{{\textup{qs}}}_2$, together with map of algebraic groups that we will still call$\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$which will appear in considerations below, all coming from an element of$H^1(\Gal(\bar{k}/k),{\mathrm G}^{{\textup{qs}}}_1(\bar{k}))$.
The map $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$ gives rise to a “ restriction ” map from representations of ${\mathrm G}_2(k)$ to those of ${\mathrm G}_1(k)$, and from Silberger [11] one knows that the restric- tion of an irreducible representation of ${\mathrm G}_2(k)$ is a finite direct sum of irreducible representations of ${\mathrm G}_1(k)$. In particular, we obtain a functor $\lambda^\star : \mathcal{R}_{\mathrm{fin}}({\mathrm G}_2(k)) \rightarrow \mathcal{R}_{\mathrm{fin}}({\mathrm G}_1(k))$, where $\mathcal{R}_{\mathrm{fin}}(H)$ denotes the category of smooth, finite-length repre- sentations of a group $H$.
Let ${}^L {\mathrm G}_1 = {\widehat {{\mathrm G}}}_1\rtimes W'_k$ and $ {}^L{\mathrm G}_2
= {\widehat {{\mathrm G}}}_2\rtimes W'_k$be the$L$-groups associated to the quasi-split reductive groups${\mathrm G}^{{\textup{qs}}}_1$and${\mathrm G}^{{\textup{qs}}}_2$respectively. The map$\lambda: {\mathrm G}^{{\textup{qs}}}_1\rightarrow {\mathrm G}^{{\textup{qs}}}_2$also gives rise to a homomorphism of$L$ -groups:
$$
{}^L\lambda: {}^L {\mathrm G}_2 \longrightarrow {}^L{\mathrm G}_1,
$$
as well as a homomorphism of their centers:
$$
{}^L \lambda
:
Z({\widehat{{\mathrm G}}_2})^{W_k}\longrightarrow Z({{\widehat {{\mathrm G}}}_1})^{W_k}.
$$
In particular, a character $\chi_1$ of $\pi_0(Z({{\widehat {{\mathrm G}}}_1})^{W_k})$ gives rise to a character $\chi_2$ of $\pi_0(Z({{\widehat {{\mathrm G}}}_2})^{W_k})$ which by the Kottwitz isomorphism (assuming $k$ to be nonarchimedean at this point):
$$
H^1(\Gal(\bar{k}/k),{\mathrm G}^{{\textup{qs}}}_i(\bar{k}))
\cong
\hom (\pi_0(Z({{\widehat{\mathrm G}}_i})^{W_k}), \mathbf{Q}/{\mathrm Z}),
$$
constructs pure inner forms ${\mathrm G}_1$ of ${\mathrm G}_1^{{\textup{qs}}}$ and ${\mathrm G}_2$ of ${\mathrm G}^{{\textup{qs}}}_2$, together with a map $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$ as before.
Let $\varphi_2: W'_k \rightarrow {}^L{\mathrm G}_2$, and $\varphi_1= {}^L\lambda \circ \varphi_2 : W'_k \rightarrow {}^L{\mathrm G}_1$ be associated Langlands parameters, where $W'_k = W_k\times \SL_2(\mathbf{C})$, with $W_k$ the Weil group of $k$. Then ${}^L\lambda$ gives rise to a homomorphism of centralizers of the images of the parameters $\varphi_1$ with values in ${}^L{\mathrm G}_1$ and $\varphi_2$ with values in ${}^L{\mathrm G}_2$, and also a homomorphism of the groups of connected components of their centralizers:
$$
\pi_0({}^L \lambda)
:
\pi_0(Z_{\widehat{{\mathrm G}}_2}(\varphi_2))
\longrightarrow
\pi_0(Z_{{\widehat {{\mathrm G}}}_1}(\varphi_1)).
$$
| |
180603635/11
|
MULTIPLICITY UPON RESTRICTION TO THE DERIVED SUBGROUP
Let $G$ be a group, and $N$ a normal subgroup of $G$ such that
$$
G/N \cong \mathbf{Z}/2 \oplus \mathbf{Z}/2.
$$
A good example to keep in mind is $G=Q_8=\{\pm 1, \pm i, \pm j, \pm k\}$, the quaternion group of order $8$, and $N=\{\pm 1\}$. Let $\omega_1$ and $\omega_2$ be two distinct, nontrivial characters of $G$ that are trivial on $N$.
Suppose $\pi$ is an irreducible representation of $G$ such that
$$
\pi \cong \pi\otimes \omega_1 \cong \pi\otimes\omega_2.
$$
By [4] * §2, $\pi|_N$ must be one of the following:
- (1) a sum of four inequivalent, irreducible representations, or (2) a sum of two copies of an irreducible representation.
Deciding which of these two options we have is a subtle question, and this is what we wish to do here.
Let $N_1 = \ker\{\omega_1 \colon G \longrightarrow \mathbf{Z}/2\}$, so that $G\supset N_1\supset N$. Since $\pi \cong \pi\otimes \omega_1$, $\pi|_{N_1} = \pi_1\oplus\pi_2$, a sum of inequivalent, irreducible representations. Further, since $\pi\cong\pi\otimes\omega_2$, we have
$$
(\pi_1\oplus\pi_2) \cong (\pi_1\oplus\pi_2)\otimes \omega_{21},
$$
where $\omega_{21} = \omega_2|_{N_1}$, a nontrivial character of $N_1$ of order 2. Therefore, we have the following two possibilities
$\pi_1 \cong \pi_1\otimes \omega_{21}$,
- $\pi_2 \cong \pi_1\otimes \omega_{21}$,
In case (i), $\pi_1$, which is an irreducible representation of $N_1$, decomposes when restricted to $N$ into two inequivalent irreducible representations, and therefore $\pi$ has at least two inequivalent irreducible subrepresentations when restricted to $N$, hence in case (i),
$$
\pi|_N = \text{a sum of $4$ inequivalent, irreducible representations}.
$$
In case (ii), clearly $\pi|_N$ is twice an irreducible representation.
How does one then construct an example of an irreducible representation $\pi$ of $G$ for which $\pi|_{N}$ is twice an irreducible representation? We start with an irreducible representation $\pi_1$ of $N_1$ such that the following equivalent conditions hold:
$\pi_1$ does
- not extend to a representation of $G$; $\pi_1^g \not\cong \pi_1$ for some $g\in G$.
Given such a representation $\pi_1$ of $N_1$, next we must ensure that
$$
\pi_1^g \cong \pi_1\otimes \omega_{21} \text{for} g \in G \setminus N.
$$
If we understand $N_1$, together with the action of $G$ on the representations of $N_1$, then the condition
$$
\pi_1^g \cong \pi_1\otimes \omega_{21} \not\cong \pi_1
$$
is checkable, constructing an irreducible representation $\pi = \Ind_{N_1}^G \pi_1$ of $G$ such that
$$
\pi|_N= 2 \pi_1|_N.
$$
Acknowledgements. As mentioned in the introduction, most of this work is several years old. In the mean time, there is the recent work of K. Choiy [2] formulating the conjectural multiplicity formula quite independently of us, and
| |
180603635/6
|
JEFFREY D. ADLER AND DIPENDRA PRASAD
the $L$ -groups. Let $\varphi_2: W'_k \rightarrow {}^L{\mathrm G}_2$, and $\varphi_1= {}^L\lambda \circ \varphi_2 : W'_k \rightarrow {}^L{\mathrm G}_1$ be asso- ciated Langlands parameters. Then for the associated homomorphism of finite groups $\lambda^\star \colon %
\longrightarrow \pi_0(Z_{\widehat{{\mathrm G}}_2}(\varphi_2)){\pi_0(Z_{{\widehat {{\mathrm G}}}_1}(\varphi_1))}$, the image is normal with abelian cokernel.
Proof. It suffices to prove the lemma separately in the two cases:
$\lambda \colon {\mathrm G}_1 \longrightarrow {\mathrm G}_2$ is
- injective as a homomorphism of algebraic groups. $\lambda \colon {\mathrm G}_1 \longrightarrow {\mathrm G}_2$ is surjective as a homomorphism of algebraic groups.
We will do only the first case, the other being very similar.
Assume then that $\lambda \colon {\mathrm G}_1 \longrightarrow {\mathrm G}_2$ is injective, and thus $\widehat{\lambda} \colon \widehat{{\mathrm G}}_2 \longrightarrow \widehat{{\mathrm G}}_1$ is surjective with kernel say $\widehat{Z}$. Use $\varphi_2 \colon W'_k \longrightarrow {}^{L}{\mathrm G}_2$ and $\varphi_1= {}^{L}\lambda \circ \varphi_2:W'_k \longrightarrow {}^{L}{\mathrm G}_1$ to give $\widehat{{\mathrm G}}_2$ and $\widehat{{\mathrm G}}_1$ a $W'_k$ -group structure such that we have an exact sequence of $W'_k$ -groups:
$$
1
\longrightarrow \widehat{Z}
\longrightarrow \widehat{{\mathrm G}}_2
\longrightarrow \widehat{{\mathrm G}}_1
\longrightarrow 1.
$$
This gives rise to a long exact sequence of $W'_k$ -cohomology sets:
$$
1
\longrightarrow \widehat{Z}^{W'_k}
\longrightarrow \widehat{{\mathrm G}}^{W'_k}_2
\longrightarrow \widehat{{\mathrm G}}^{W'_k}_1
\longrightarrow H^1(W'_k, \widehat{Z})
\longrightarrow \cdots.
$$
Equivalently, we have the exact sequence of groups:
$$
1
\longrightarrow Z_{\widehat{{\mathrm G}}_2}(\varphi_2) /\widehat{Z}^{W'_k}
\longrightarrow Z_{\widehat{{\mathrm G}}_1}(\varphi_1)
\longrightarrow A
\longrightarrow 1,
$$
where $A \subset H^1(W'_k, \widehat{Z})$, a locally compact abelian group. Taking $\pi_0$ of the terms in the above exact sequence which all fit together in a long exact sequence of $\pi_i$ ’s (higher homotopy groups), the assertion in the lemma follows on noting that if $E_1 \rightarrow E_2$ is a surjective map of locally compact and locally connected topological groups, then the induced map $\pi_0(E_1) \rightarrow \pi_0(E_2)$ is also surjective.
## 4. R eduction of the conjecture to the case of tempered representations
As before, let ${\mathrm G}_1$ and ${\mathrm G}_2$ be two reductive groups over a local field $k$, and let $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$ be a $k$ -homomorphism that is a central isogeny when restricted to their derived subgroups, giving rise to the restriction functor $\lambda^\star : \mathcal{R}_{\mathrm{fin}}({\mathrm G}_2(k)) \rightarrow \mathcal{R}_{\mathrm{fin}}({\mathrm G}_1(k))$.
Lemma 6. Let $V$ be a finite-length representation of ${\mathrm G}_2(k)$ with maximal semi-simple quotient $Q$. Then $\lambda^\star Q$ is the maximal semi-simple quotient of $\lambda^\star V$, a finite-length repre- sentation of $G_1(k)$.
Proof. It suffices to observe that a finite-length representation of ${\mathrm G}_2(k)$ is semisim- ple if and only if its image under $\lambda^\star$ is a finite-length, semi-simple representation of ${\mathrm G}_1(k)$. If ${\mathrm Z}({\mathrm G}_1)(k)\cdot{\mathrm G}_1(k)$ is of finite index in ${\mathrm G}_2(k)$, such as when $k$ is of char- acteristic zero, then this is easy to see. By a theorem of Silberger [11], irreducible representations of ${\mathrm G}_2(k)$ remain finite-length semi-simple representations when restricted to ${\mathrm G}_1(k)$, and the lemma follows in general.
To set up the next result, let $P_2=M_2N_2$ be a Levi factorization of a parabolic subgroup in ${\mathrm G}_2$. If we let $P_1=\lambda^{-1}(P_2)$, $M_1=\lambda^{-1}(M_2)$, and $N_1=\lambda^{-1}(N_2)$, then $P_1=M_1N_1$ is a Levi factorization of a parabolic subgroup in ${\mathrm G}_1$. Then $\lambda \colon M_1 \longrightarrow M_2$ gives us a restriction functor $\mathcal{R}_{\mathrm{fin}}(M_2(k)) \rightarrow \mathcal{R}_{\mathrm{fin}}(M_1(k))$ that we will also
| |
180603635/4
|
JEFFREY D. ADLER AND DIPENDRA PRASAD
- (3) The restriction to ${\mathrm G}_1(k)$ of an irreducible representation $\pi_2$ of ${\mathrm G}_2(k)$ is a finite-length (completely reducible) representation of ${\mathrm G}_1(k)$, whose irre- ducible components are all in the same $L$ -packet. This $L$ -packet for ${\mathrm G}_1(k)$ depends only on the $L$ -packet for ${\mathrm G}_2(k)$ containing $\pi_2$. If the Langlands parameter of our $L$ -packet for ${\mathrm G}_2(k)$ is $\varphi_2: W'_k \rightarrow {}^L{\mathrm G}_2$, then the Langlands parameter of our $L$ -packet for ${\mathrm G}_1(k)$ is $\varphi_1:= {}^L\lambda \circ \varphi_2 : W'_k \rightarrow {}^L{\mathrm G}_1$. (This is part (a) of the conjecture.)
- (4) If Conjecture 1 is true, then the pullback from ${\mathrm G}_2(k)$ to ${\mathrm G}_1(k)$ of the dis- tribution
$$
\sum _{\chi_2} \chi_2(1) \Theta(\pi_{(\varphi_2, \chi_2)})
$$
- where the sum is taken over those characters $\chi_2$ of the component group which have a particular restriction to $Z(\widehat{G}_2)^{W_k}$ defining the group ${\mathrm G}_2(k)$ assumed to be a pure inner form of a fixed quasi-split group ${\mathrm G}_2^{{\textup{qs}}}(k)$, is a stable distribution on ${\mathrm G}_1(k)$ as we check now.
- By Conjecture, the pullback of the distribution $\Theta_{\pi_2} = \Theta(\pi_{(\varphi_2, \chi_2)})$ on ${\mathrm G}_2(k)$ to ${\mathrm G}_1(k)$ is
$$
\sum_{\pi_1} m(\pi_2,\pi_1) \Theta(\pi_1)
=
\sum_{\chi_1}\Theta(\pi_{(\varphi_1, \chi_1)})
\dim \hom_{\pi_0(Z(\varphi_2))}[\chi_2, \lambda_\star \chi_1].
$$
- Therefore, the pullback to ${\mathrm G}_1(k)$ of the distribution $\sum _{\chi_2} \chi_2(1) \Theta(\pi_{(\varphi_2, \chi_2)})$ on ${\mathrm G}_2(k)$ is (assuming Conjecture 1)
$$
\sum _{\chi_1, \chi_2}
\chi_2(1) \Theta(\pi_{(\varphi_1, \chi_1)})
\dim \hom_{\pi_0(Z(\varphi_2))}[\chi_2, \lambda_\star \chi_1],
$$
- which is the same as
$$
\sum _{\chi_1,\chi_2}
\Theta(\pi_{(\varphi_1, \chi_1)})
\dim \hom_{\pi_0(Z(\varphi_2))}[\chi_2(1) \chi_2, \lambda_\star \chi_1],
$$
- where the sum is taken over all pairs of characters $\chi_1,\chi_2$ with particular restrictions to $Z(\widehat{{\mathrm G}}_1)^{W_k}$ and $Z(\widehat{{\mathrm G}}_2)^{W_k}$. Observe that those characters $\chi_2$ whose restrictions to $Z(\widehat{{\mathrm G}}_2)^{W_k}$ are not compatible with the restriction of $\chi_1$ to $Z(\widehat{{\mathrm G}}_1)^{W_k}$ contribute $0$ to the sum. Therefore, we can take the sum over all $\chi_2$. The sum then is the same as
$$
\tag{$*$}
\sum _{\chi_1}
\Theta(\pi_{(\varphi_1, \chi_1)})
\dim \hom_{\pi_0(Z(\varphi_2))}[R, \lambda_\star \chi_1],
$$
- where $R = \sum\chi_2(1) \chi_2$ is the regular representation of $\pi_0(Z(\varphi_2))$.
- By Schur orthogonality,
$$
\dim\hom_{\pi_0(Z(\varphi_2))} [\chi_2, \lambda_\star \chi_1]
=
\frac{1}{|\pi_0(Z(\varphi_2))|}
\sum_{g \in \pi_0(Z(\varphi_2))}\chi_1(\lambda^\star g) \bar{\chi}_2(g),
$$
- where $\lambda^\star$ denotes the map $\pi_0({}^{L} \lambda) \colon \pi_0(Z(\varphi_2)) \longrightarrow \pi_0(Z(\varphi_1))$. So
\begin{align*} \dim\hom_{\pi_0(Z(\varphi_2))} [R, \lambda_\star \chi_1] & = \frac{1}{|\pi_0(Z(\varphi_2))|} \sum_{g\in \pi_0(Z(\varphi_2))}\chi_1(\lambda^\star g) \chi_R(g), \end{align*}
| |
180603635/10
|
JEFFREY D. ADLER AND DIPENDRA PRASAD
$\pi = \cInd_{\U(\mathcal{L})}^{\U(V)} \sigma$.
Then $\pi\oplus \pi^{g_0}$ extends to an irreducible representation $\widetilde\pi$ of $\GU(V)$ whose restriction to $\SU(V)$ decomposes with multiplicity two.
Proof. From Moy-Prasad [8] * Proposition 6.6, $\pi$ is an irreducible, supercuspidal representation of $\U(V)$. Let $\pi$ also denote one of its extensions to $Z(\GU(V)) \U(V)$. From the last sentence of [7] * Theorem 5.2, $\pi^{g_0} \not\cong\pi$, so the sum $\pi\oplus \pi^{g_0}$ ex- tends to an irreducible (also supercuspidal) representation $\widetilde\pi$ of $\GU(V)$. By the induction-restriction formula (observe that by the explicit description of $\U(\mathcal{L})$, $\det: \U(\mathcal{L})
\rightarrow E_1$is surjective, and hence$\U(\mathcal{L})\SU(V)=\U(V)$),
\begin{align*} \pi|_{\SU(V)} & = \cInd_{\SU(\mathcal{L})}^{\SU(V)} (\sigma|_{\SU(\mathcal{L})}), \\
\pi^{g_0}|_{\SU(V)} & = \cInd_{\SU(\mathcal{L})}^{\SU(V)} (\sigma^{g_0}|_{\SU(\mathcal{L})}) . \end{align*}
Since $\rho\otimes\rho\chi \cong \rho\chi\otimes \rho$ as representations of ${\textup{S}}(\U(2d)\times\U(2d))(\mathfrak{f})$, we have that $\sigma\cong\sigma^{g_0}$ as representations of $\SU(\mathcal{L})$, so
$$
\widetilde\pi|_{\SU(V)}
= (\pi\oplus \pi^{g_0})|_{\SU(V)}
= 2 \cdot \cInd_{\SU(\mathcal{L})}^{\SU(V)}
(\sigma|_{\SU(\mathcal{L})}).
\qedhere
$$
In order to have an example of multiplicity at least two, it is thus sufficient to find a representation $\rho$ of $\U(2d)(\mathfrak{f})$ such that $\rho\not\cong\rho\chi$, as in the theorem. In fact, most irreducible Deligne-Lusztig cuspidal representations of $\U(2d)(\mathfrak{f})$ will have this property, as they restrict irreducibly to $\SU(2d)(\mathfrak{f})$.
Remark11. In a future work, we will expand upon the example given in the The- orem, whose essence is the following. Given a supercuspidal representation of ${\mathrm G}_2(k)$ whose restriction to ${\mathrm G}_1(k)$ has regular components (in the sense of Kaletha [5]), then the components occur with multiplicity one. (Nevins [9] already verified this for many cases.) If the components are not regular, then higher multiplicities can occur.
Our example begins with $\rho$, an irreducible cuspidal representation of $\U(2d)(\mathfrak{f})$ that arises via Deligne-Lusztig induction from a character $\theta$ of the group of $\mathfrak{f}$ - points of an anisotropic torus $\mathsf{T} \subset\U(2d)$. Suppose also that the restriction of $\theta$ to $\mathsf{T}(\mathfrak{f})\cap \SU(2d)(\mathfrak{f})$ remains regular so that the restriction of $\rho$ to $\SU(2d)(\mathfrak{f})$ remains irreducible. The torus $\mathsf{T}\times \mathsf{T} \subset \U(2d)\times\U(2d)$ lifts to give an unramified torus $T\subset \GU(V)$, and the character $\theta\otimes\theta\chi$ can be inflated and extended to give a character $\Theta$ of $T$. The representation $\widetilde\pi$ of $\GU(V)$ that we have constructed in the theorem is a regular supercuspidal representation in the sense of Kaletha [5], but the irreducible components of its restriction to $\SU(V)$ are not since our character $\Theta$ of $T$, when restricted to $T \cap \SU(V)$, is not regular because of the presence of the element $g_0\in\GU(V)$.
For depth-zero supercuspidal representations of quasi-split unitary groups, the parahoric that we have used is the only one that can lead to higher multiplicities.
## 8. G eneralities on constructing higher multiplicities
In this section, we discuss some generalities underlying the example of the previous section, which will be useful for constructing higher multiplicities in general.
| |
180603635/1
|
# MULTIPLICITY UPON RESTRICTION TO THE DERIVED SUBGROUP
JEFFREY D. ADLER AND DIPENDRA PRASAD
A bstract. We present a conjecture on multiplicity of irreducible representations of a subgroup $H$ contained in the irreducible representations of a group $G$, with $G$ and $H$ having the same derived groups. We point out some consequences of the conjecture, and verification of some of the consequences. We give an explicit example of multiplicity $2$ upon restriction, as well as certain theorems in the context of classical groups where the multiplicity is $1$.
## 1. I ntroduction
Suppose $k$ is a local field, ${\mathrm G}$ is a connected reductive $k$ -group, ${\mathrm G}'$ is a subgroup of ${\mathrm G}$ containing the derived group, and $\pi$ is a smooth, irreducible, complex rep- resentation of ${\mathrm G}(k)$. In an earlier work [1], the authors showed that for many choices of ${\mathrm G}$, the restriction $\Res^{{\mathrm G}(k)}_{{\mathrm G}'(k)}\pi$ decomposes without multiplicity.
A number of years ago, in the process of identifying situations where multi- plicity one did not hold, one of us discovered an example of a depth-zero super- cuspidal representation of $\GU(2d,2d)$, a $k$ -quasisplit group, whose restriction to $\SU(2d,2d)$ decomposes with multiplicity two, and the other formulated a con- jecture in the form of a reciprocity law involving enhanced Langlands param- eters. In this paper, we present both the example and the conjecture, together with some consequences of the latter, and a verification of some of those con- sequences. Besides these, the paper proves several results by elementary means involving classical groups where multiplicity one holds.
A complete analysis of decomposition of the unitary principal series for $\U(n,n)$ and its restriction to $\SU(n,n)$ was done by Keys [6], who also phrased his results in terms of “ reciprocity ” theorems for $R$ -groups; in particular, he found cases of multiplicity greater than one.
After presenting our conjecture (§2), we give some of the heuristics behind it. In the formulation of the conjecture, we have considered a more general situation than that of a subgroup. We consider ${\mathrm G}_1$ and ${\mathrm G}_2$ to be two connected reductive groups over a local field $k$, and $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$ a $k$ -homomorphism that is a central isogeny when restricted to their derived subgroups, allowing us to “ restrict ” rep- resentations of ${\mathrm G}_2(k)$ to ${\mathrm G}_1(k)$. Since under such a homomorphism $\lambda$, the image of ${\mathrm G}_1(k)$ is a normal subgroup of ${\mathrm G}_2(k)$ with abelian quotient, all the irreducible representations of ${\mathrm G}_1(k)$ which appear in this restriction problem for a given ir- reducible representation of ${\mathrm G}_2(k)$ appear with the same multiplicity. In §3, we verify that for our conjectural multiplicity, this relationship does indeed hold. We show (§4) that if the conjecture is true for tempered representations, then via the Langlands classification it holds for all representations.
Our conjecture (for $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$ a $k$ -homomorphism), implies multiplicity one in situations where Langlands parameters for ${\mathrm G}_1$ have abelian component groups.
| |
180603635/3
|
MULTIPLICITY UPON RESTRICTION TO THE DERIVED SUBGROUP
This allows one to ‘restrict’ representations of $\pi_0(Z_{{\widehat{\mathrm G}}_1}(\varphi_1))$ to representations of $\pi_0(Z_{{\widehat{\mathrm G}}_2}(\varphi_2))$, giving rise to the restriction functor
$$
\lambda_\star:
K_0(\pi_0(Z_{{\widehat{\mathrm G}}_1}(\varphi_1)))
\rightarrow
K_0(\pi_0(Z_{{\widehat{\mathrm G}}_2}(\varphi_2))),
$$
where $K_0(H)$ denotes the Grothendieck group of finite-length representations of a group $H$.
The formulation of our conjecture below presumes that the local Langlands correspondence involving enhanced Langlands parameters has been achieved, giving rise to a bijection between enhanced Langlands parameters and the set of isomorphism classes of irreducible admissible representations of all pure inner forms of quasi-split groups. This will be needed for both of the groups ${\mathrm G}_1$ and ${\mathrm G}_2$; it is possible on the other hand that one could reverse this role, and use the con- jectural multiplicity formula to construct an enhanced Langlands parametrization for ${\mathrm G}_2$, knowing it for ${\mathrm G}_1$.
- Conjecture1. (a) Let ${\mathrm G}_1$ and ${\mathrm G}_2$ be two connected reductive groups over a local field $k$ and $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$, a $k$ -homomorphism that is a central isogeny when restricted to their derived subgroups. For $i=1,2$, let $\pi_i$ be an irreducible admissible representa- tion of ${\mathrm G}_i(k)$ with Langlands parameter $\varphi_i$. Let
$$
\mm(\pi_2,\pi_1)
:= \dim \hom_{{\mathrm G}_1(k)}[\pi_1, \lambda^\star \pi_2]
= \dim \hom_{{\mathrm G}_1(k)}[\lambda^\star \pi_2, \pi_1].
$$
- Then $\mm(\pi_2,\pi_1) = 0$ unless $\varphi_1 = {}^L \lambda \circ \varphi_2$. (b) Let ${\mathrm G}^{{\textup{qs}}}_1$ and ${\mathrm G}^{{\textup{qs}}}_2$ be two connected reductive quasi-split groups over a local field $k$ and $\lambda: {\mathrm G}^{{\textup{qs}}}_1\rightarrow {\mathrm G}^{{\textup{qs}}}_2$, a $k$ -homomorphism that is a central isogeny when restricted to their derived subgroups. Let $\varphi_1$ and $\varphi_2$ be Langlands parameters associated to the groups ${\mathrm G}^{{\textup{qs}}}_1$ and ${\mathrm G}^{{\textup{qs}}}_2$ with $\varphi_1
= {}^L \lambda \circ \varphi_2$, and let$\chi_i$be characters of their component groups$\pi_0(Z_{\widehat{{\mathrm G}}_i}(\varphi_i))$. Then, if$\hom_{\pi_0(Z(\varphi_2))}[\chi_2, \lambda_\star \chi_1]$is nonzero, the characters$\chi_i$define pure inner forms${\mathrm G}_i$of${\mathrm G}^{{\textup{qs}}}_i$together with a$k$-homomorphism,$\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$, as discussed earlier. Then if$\pi_i = \pi(\varphi_i,\chi_i)$are the corresponding irreducible admissible representations of${\mathrm G}_i(k)$, we have
$$
\mm(\pi_2,\pi_1)
= \dim \hom_{\pi_0(Z(\varphi_2))}[\chi_2, \lambda_\star \chi_1].
$$
The main heuristic for the conjectural multiplicity is the following.
- (1) For any $L$ -packet $\{\pi \}$ on any reductive group ${\mathrm G}(k)$ defined by a param- eter $\varphi$, thus $\{\pi\} = \{\pi_{(\varphi, \chi)}\}$ where one takes those characters $\chi$ of the component group which have a particular restriction to $Z(\widehat{{\mathrm G}})^{W_k}$ defining the group ${\mathrm G}(k)$ assumed to be a pure inner form of a fixed quasi-split group ${\mathrm G}^{{\textup{qs}}}$,
$$
\sum _{\chi} \chi(1) \Theta(\pi_{(\varphi, \chi)})
$$
- is a stable distribution on ${\mathrm G}(k)$. Here, for any admissible representation $\pi$ we are letting $\Theta(\pi)$ denote its character, regarded as a distribution on ${\mathrm G}(k)$.
- (2) For a homomorphism $\lambda: {\mathrm G}_1\rightarrow {\mathrm G}_2$ of reductive groups over $k$ which is an isogeny when restricted to their derived subgroups, the pullback of a stable distribution on ${\mathrm G}_2(k)$ is a stable distribution on ${\mathrm G}_1(k)$.
| |
180603635/9
|
MULTIPLICITY UPON RESTRICTION TO THE DERIVED SUBGROUP
- As an abstract group, $E_1$ inherits a direct product decomposition from $R_E^\times \cong k_E^\times \times (1+P_E)$. Thus, $E_1$ is a direct product of a cyclic group (of order coprime to $p$) and a pro- $p$ -group $A$, implying that $E_1 / E_1^r$ is cyclic if and only $A/A^r$ is cyclic. But this latter quotient is trivial if $r$ is coprime to $p$, and is cyclic if $k=\mathbf{Q}_p$ ($p$ odd). (c) This follows from the previous two parts of the theorem.
## 7. A n example of multiplicity upon restriction
Let $\varpi$ be a uniformizer of $k$, $E/k$ an unramified quadratic extension, $R_k$ and $R_E$ the rings of integers in $k$ and $E$, and $\mathfrak{f}$ and $\mathfrak{f}_E$ the residue fields. Let $V$ be a $4d$ -dimensional hermitian space over $E$, with hyperbolic basis $\{e_1, f_1, \dots, e_{2d}, f_{2d}\}$. Thus, $\langle e_i,f_i\rangle = 1$ for all $1\leq i \leq 2d$, and all the other products being $0$. Let $\U(V)$ be the corresponding unitary group. Define the lattice $\mathcal{L}$ in $E$ by
$$
\mathcal{L} =
\Span_{R_E}\{
e_1,f_1,\ldots,e_d,f_d,
\varpi e_{d+1},f_{d+1},\ldots, \varpi e_{2d}, f_{2d}
\}.
$$
Clearly, $\mathcal{L}^\vee
:=
\{v\in V | \text{$\langle v,\ell\rangle\in R_E$ for all $\ell\in\mathcal{L}$}\}$ is given by
$$
\mathcal{L}^\vee =
\Span_{R_E}\{
e_1,f_1,\ldots,e_d,f_d,
e_{d+1},\varpi^{-1} f_{d+1},\ldots, e_{2d}, \varpi^{-1} f_{2d}
\}.
$$
Observe that
$$
\varpi \mathcal{L}^\vee \subseteq \mathcal{L} \subseteq \mathcal{L}^\vee,
$$
and $\mathcal{L}^\vee/\mathcal{L}$ and $\mathcal{L}/\varpi\mathcal{L}^\vee$ are $2d$ -dimensional hermitian spaces over $\mathfrak{f}_E$ with natural hermitian structures. For example, given two elements $\ell_1$ and $\ell_2$ in $\mathcal{L}^\vee$ with images $\overline\ell_1$ and $\overline\ell_2$ in $\mathcal{L}^\vee/\mathcal{L}$, the hermitian structure on $\mathcal{L}^\vee/\mathcal{L}$ is defined by having $\langle \overline\ell_1,\overline\ell_2\rangle$ as the image of $\varpi\langle\ell_1,\ell_2\rangle$ (which belongs to $R_E$) in $\mathfrak{f}_E$.
Define $K=\U(\mathcal{L})$ to be the stabilizer of the lattice $\mathcal{L}$ in $\U(V)$, i.e., $\U(\mathcal{L}) = \{g \in \U(V) | \text{$g\ell \in \mathcal{L}$for all$\ell\in\mathcal{L}$} \}$. If an element of $\U(V)$ preserves $\mathcal{L}$, then it clearly preserves $\mathcal{L}^\vee$ and $\varpi\mathcal{L}$, giving a map $\U(\mathcal{L}) \longrightarrow \U(2d,\mathfrak{f})\times \U(2d,\mathfrak{f})$. Similarly, we have a map $\SU(\mathcal{L}) \longrightarrow {\textup{S}}(\U(2d)\times \U(2d))(\mathfrak{f})$.
Let $g_0\in \GU(V)$ be defined by (for $i \leq d$)
$$
e_i \mapsto e_{d+i},
\quad
f_i \mapsto \varpi^{-1} f_{d+i},
\quad
e_{d+i} \mapsto \varpi^{-1} e_i,
\quad
f_{d+i} \mapsto f_i,
$$
Clearly, $g_0$ has similitude factor $\varpi^{-1}$, and $g_0\mathcal{L} = \mathcal{L}^\vee$. Therefore, we have
$$
g_0\U(\mathcal{L})g_0^{-1} =\U(\mathcal{L}^\vee).
$$
Thus conjugation by $g_0$ induces an isomorphism of $\U(\mathcal{L})$ into $\U(\mathcal{L}^\vee)$, making the following diagram commute:
$$
\begin{xy}
\xymatrix{
\U(\mathcal{L}) \ar[r]^{g_0} \ar[d] & \U(\mathcal{L}^\vee) \ar[d] \\
\U(2d,\mathfrak{f})\times\U(2d,\mathfrak{f}) \ar[r]^j &
\U(2d,\mathfrak{f})\times\U(2d,\mathfrak{f})
}
\end{xy}
$$
where $j(x,y) = (y,x)$.
Theorem 10. Let $\rho$ be any irreducible cuspidal representation of $\U(2d)(\mathfrak{f})$ such that $\rho\not\cong\rho\chi$, where $\chi$ is a quadratic character of $\U(2d)(\mathfrak{f})$ trivial on $\SU(2d)(\mathfrak{f})$. Let $\sigma := \infl(\rho\otimes\rho\chi)$ denote the inflation of $\rho\otimes\rho\chi$ from $(\U(2d)\times\U(2d))(\mathfrak{f})$ to $\U(\mathcal{L})$ and let
| |
180603635/8
|
JEFFREY D. ADLER AND DIPENDRA PRASAD
either $(\OO_n, \GO_n)$ or $(\Sp_n, \GSp_n)$. In the next section, we will see that multiplic- ity $\leq 1$ also holds for $(\U_n, \GU_n)$. Gee and Ta ¨ ı bi [3] * Proposition 8.2.1 show that multiplicity $\leq 1$ holds for the pair $(\SO_n,\GSO_n)$ if $k$ has characteristic zero.
## 6. G eneralities on restriction to unitary and special unitary groups
Let $E/k$ denote a separable quadratic extension of nonarchimedean local fields, $N= N_{E/k}$ the norm map from $E^\times$ to $k^\times$, and $E_1$ the kernel of this map.
Let $B$ denote a nondegenerate $E/k$ -hermitian form on some $E$ -vector space $V$ of some dimension $r$. Then we can form algebraic groups $\SU(V,B)$, $\U(V,B)$, and $\GU(V,B)$ whose $k$ -points consist respectively of the elements of $\SL(r,E)$ that preserve $B$; the elements of $\GL(r,E)$ that preserve $B$; and the elements of $\GL(r,E)$ that preserve $B$ up to a scalar in $k^\times$. The group $\GU(V,B)$ comes equipped with a map $\mu \colon \GU(V,B) \longrightarrow \GL_1$ called the similitude character. We will write our algebraic groups as $\SU(r)$, $\U(r)$, and $\GU(r)$ when $V$ and $B$ are understood.
If $G$ is a group, $H$ is a subgroup, and $G/Z(G)H$ is cyclic, then every irreducible representation of $G$ restricts to $H$ without multiplicity. How far can we exploit this fact?
Theorem9. Let $p$ be the residual characteristic of $k$.
- (a) All irreducible representations of $\GU(r)(k)$ decompose without multiplicity upon restriction to $\U(r)(k)$. Such a restriction is irreducible when $r$ is odd, and has at most two components when $r$ is even. (b) All irreducible representations of $\U(r)(k)$ decompose without multiplicity upon re- striction to $\SU(r)(k)$ when $r$ is coprime to $p$, or $k=\mathbf{Q}_p$ ($p$ odd). (c) All irreducible representations of $\GU(r)(k)$ decompose without multiplicity upon restriction to $\SU(r)(k)$ when $r$ is odd and coprime to $p$.
- Proof. Let $\mu \colon \GU(r) \longrightarrow \GL(1)$ denote the similitude character. Clearly the group $\GU(r)$ contains the scalar matrices $eI_r$ for all $e\in E^\times$, and for such matrices the similitude is $N_{E/k}(e)$. Therefore, the image under $\mu$ of the center of $\GU(r)(k)$ is $N_{E/k}(E^\times)$, so $\mu$ thus gives an isomorphism
$$
\frac{\GU(r)}{Z(\GU(r))\U(r)}
\stackrel{\sim}{\longrightarrow}
\frac{\Image(\mu)}{N(E^\times)}.
$$
- A scalar $a\in k^\times$ is a similitude for some linear transformation $g$ of $V$ if and only if for all $v,w \in V$, we have that $B(gv,gw) = a\cdot B(v,w)$. That is, $B$ and $a\cdot B$ are equivalent Hermitian forms. It is known that two Hermitian forms over a non-archimedean local field $k$ are equivalent if and only if their discriminants, which are elements of $k^\times/ N(E^\times)$, are the same. Therefore, $B$ and $aB$ are equivalent if and only if $\disc B = a^r \disc B$ in $k^\times/N(E^\times)\cong \mathbf{Z}/2$. Thus, if $r$ is even, then $B$ and $aB$ are equivalent for $a$ an arbitrary element of $k^\times$, but if $r$ is odd, then $a$ must lie in $N(E^\times)$. Thus,
$$
\frac{\GU(r)}{{\mathrm Z}(\GU(r)) \U(r)}
\cong
\mathbf{Z}/2
\quad \text{or} \quad
\{1\}.
$$
- (b) Let $R_E$ and $P_E$ denote the ring of integers and prime ideal for $E$. The deter- minant character gives us an isomorphism
$$
\det \colon \frac{\U(r)(k)}{Z(\U(r))(k)\SU(r)(k)}
\stackrel{\sim}{\longrightarrow}
\frac{E_1}{(E_1)^r}.
$$
| |
231002570/5
|
Table 4. Results of the subjective severity and identity evaluation experiments by 3 SLPs. S2S means similarity to source, S2T means similarity to target, and S-T means comparison of source to target. Boldface indicates better severity preservation.
Table5. Mean opinion scores (MOS) from the naturalness experi- ment from 7 native listeners.
that differences in severity could influence the evaluation of speaker identity.
## 5.2.2. Naturalness
The results of the naturalness test are shown in Table 5. Consistent with [2,1,29], the high correlation between the severity labels and the GT naturalness ratings ($r=0.88$) shows that the listeners can- not differentiate well between these aspects. Overall, the PPG-GST is rated as less natural than the PPG. However, this is not unexpected due to the following reasons: (1) The better preservation of severity is perceived as lower naturalness. (2) The reference encoder in the PPG-GST directly interacts with the noisy filterbank during recon- struction, leading to noisier sounding examples.
## 5.3. Analysis of the impact of speech pathology on x-vectors
Figure 3 shows the impact of speech pathology on the similarity of x-vectors. As expected, when looking only at the T1 speakers (T1), there is a high similarity as we are comparing the same speakers with the same identity. Introducing the T2 speakers into the comparison, the similarity decreases (T1+T2) but it is still much higher than the non-target scores. This shows that pathology impacts embeddings.
Table 6 quantifies the impact with EERs, and also shows the impact of speech pathology on the speaker-level. We observe that adding the T2 utterances always makes the EER higher. Including all speakers in the analysis, an EER of 8.87% can be maintained with even the T2 distribution. When looking at the results on the speaker level, the worst EER (16.57%) is achieved for PIIM, and the best EER is achieved for PHNF (3.70%). Interestingly, there does not seem to be a clear pattern in the increase in severity, and the increase in EER. The lack of this relation seems to be consistent with the
Table6. Equal error rates (%) of T1 and T1-T2 distributions.
finding of [30], which reports that the effect of speech pathology lies in the range of other factors in speaker verification, such as the quality of the microphone.
We can conclude that the x-vectors are impacted by the speech pathology, however, the speaker still remains identifiable. Therefore, the pre-operative (T1) x-vector should contain sufficient information about the identity. We acknowledge that the current analysis has some limitations. First, we only have parallel data from low-to-mid severity speakers, therefore it might be that the extracted embed- dings are affected for high severity speakers. Furthermore, as we use utterances from the same recording to calculate the T1 EER, the estimation of T1 identification performance is likely somewhat opti- mistic.
## 5.4. Summary of Results and Comparative Analysis of PPG and PPG+GST
In the previous sections, we have provided extensive evaluation of the PPG and PPG+GST systems. The baseline PPG system demon- strates better performance in terms of speaker identity conversion, as noted in the subjective identity evaluation. However, this assess- ment could be influenced by the severity of speech pathology. Fur- thermore, the PPG system falls short of preserving the severity of the speech, especially when compared to the PPG+GST system. Fi- nally, as we have seen by the PER results, it tends to overenhance the speech, potentially leading to a less realistic output.
On the other hand, the proposed PPG+GST system preserves the severity of speech well during conversion as shown by the objective P-ESTOI, PER and subjective evaluations. In terms of speaker iden- tity conversion, however, the PPG+GST system performs slightly less well than the PPG system. In addition, the PPG+GST system may be prone to introducing noise in the output, which can detract from its naturalness. In summary, we find that while PPG+GST still needs further work, however, its superior ability to preserve speech severity makes PPG-GST a more appropriate choice for mid-to-high severity voices.
## 6. CONCLUSION
In this paper, we presented a new voice conversion system which
| |
231002570/2
|
both GST and PPG, allows us to model the time-invariant (sever- ity) and time-variant (intelligibility) aspects of the speech pathology separately, preserving the features on both levels.
In addition, we perform a more rigorous evaluation of severity, and speaker identity, using NKI-OC-VC, a newly collected oral can- cer speech dataset, which contains parallel pathological and healthy recordings from the same speaker. Compared to the UASpeech cor- pus [7], the dataset contains utterance-level stimulus that is more appropriate to evaluate aspects of severity that might be only present in longer stretches of speech. Using this dataset, we aim to revisit an assumption of H2P-VC – whether pathology impacts x-vectors. This assumption holds key importance. If a pathology indeed changes x-vectors notably, it directly impacts the effectiveness of our VC model.
Furthermore, we also seek to reevaluate the current method of choosing source speakers for VC, which is based on matching sever- ity ratings between the source and target speakers. We hypothesise that the same severity rating does not always mean the same type of voice changes.
The primary contributions of our paper are the following:
- (1) We propose a new model for the second stage of voice conver- sion (VC) in [1], which improves the preservation of speech severity while being able to convert identity. (2) Using a unique, parallel oral cancer speech dataset, NKI-OC- VC, we present a comprehensive and precise evaluation of our VC system by speech-language pathologists (SLPs). (3) We reevaluate the assumption of voice pathology’s impact on speaker’s embedding and its implications on the accuracy of our VC model. (4) We ask speech-language pathologists (SLPs) to examine our current practice of matching severity ratings for source speaker selection and its potential impact on the VC model’s accuracy.
## 2. DATASET
We have collected the NKI-OC-VC dataset for the VC task, com- prising Dutch pathological speech from 16 oral cancer (OC) speak- ers (10 male) who had undergone surgery for tongue tumours and 5 healthy control speakers (2 male).
We collected data from the subjects at a maximum of three time points (stages): before the surgery (T1), after the surgery (T2), and six months after the surgery (T3). The recordings themselves took place in the patient room during scheduled speech therapy sessions.
The subjects were asked to read the standard Dutch text “ Jorinde en Joringel” [16] during the recording session. The total duration of all speech recordings is approximately 2.5 hours. One recording session (speaker/stage) is 5 minutes on average. Each speaker re- cites 92 sentences in total per stage. In some cases, patients felt the experiment was difficult, in that case, we prematurely stopped the experiment.
The speech was recorded with the Roland R-09HR field recorder at 44.1 kHz sampling frequency and 24-bit depth. This was later downsampled to 16 kHz and quantised to 16-bit depth. For the ex- periments, we have partitioned the 92 sentences into training (78 utterances), development (7 utterances), and test sets (7 utterances).
In order to obtain ratings for the overall severity of the speech, the speech of each speaker was rated by five Dutch SLPs on a 5-point Likert scale, where 5 meant healthy, and 1 meant severe. The ratings had a very high interrater correlation ($r=0.88$). The obtained rat- ings were averaged per speaker which we treat as the severity of the speaker for the rest of the paper. The available stages and the severity of each speaker are indicated in Table 1.
Table1. Speaker breakdown of the NKI-OC-VC dataset. The sever- ity of the speaker according to 5 SLPs is noted in the parentheses. The patients marked with an asterisk (*) had to end the experiment prematurely, therefore we do not have all the 92 utterances from them.
## 3. PROPOSED FRAMEWORK
Given a speech sample from a source OC speaker (T2 or T3), and a speech sample from a target speaker before the surgery (T1), the VC aims to change the speaker identity of the speaker so that it resem- bles the T1 speaker and preserves the severity of the source T2 or T3 speech. In the following sections, we describe the Baseline and Proposed systems for the H2P-VC task.
## 3.1. Baseline: PPG
For the baseline, we follow the phonetic posteriorgram-based (PPG) recognition-synthesis approach implemented in S3PRL-VC 1. A pho- netic posteriorgram (PPG) is a feature map derived from an auto- matic speech recogniser that represents the posterior probabilites of phonetic units over time frames of the speech. We make adaptations so that the system is able to deal with Dutch speech [17].
In order to extract appropriate PPG features, we first train a Dutch Conformer E2E model ASR on the standard training set of the Corpus Gesproken Nederlands (CGN) dataset [18]. CGN contains Dutch recordings spoken by 1185 female and 1678 male speakers (age range 18-65 years old) from all over the Netherlands and Flan- ders. The Conformer models parameters were taken from [19,20]: 12 encoder layers and 6 decoder layers, all with 2048 dimensions; the attention dimension is 512 and the number of attention heads is 8; the convolution subsampling layer in the encoder has 2-layer CNNs with 256 channels, stride with 2, and a kernel size of 3. A default conv kernel size of 31 was used. Subword units with a vocabulary
| |
231002570/4
|
Table 2. Results of P-ESTOI-based severity evaluation. Rows include the source speaker. Target speaker is always RDH-VL. Boldface indicates higher correlation, and better severity preservation.
Table3. Comparison of phoneme error rate (PER%) of the two sys- tems. Boldface indicates when the PER is closer to GT. Italics indi- cate when the system overenhances the speech over the GT.
[1], which is a variant of the standard MOS test but with incre- ments of 0.5. This change is made in order to avoid lower precision of our naturalness test with high severity speech. For this test, we chose 7 source speakers from our dataset, each with differing levels of speech severity to capture a wide range. We created 7 conver- sion pairs in total using each source speaker, and targeting the same speaker from the RDH-VL dataset. This approach was taken due to evidence that severity and naturalness are often conflated in lis- tener perceptions [2,1,29]. The test included 153 utterances in total, with 7 VC utterances per conversion pair and per system, alongside their corresponding 7 source ground truth utterances and 6 utterances from the target RDH-VL speaker. These utterances were rated by 7 native Dutch listeners.
## 4.3. Analysis of the impact of speech pathology on the x-vectors
To assess the impact of speech pathology on x-vectors, we create three separate distributions of similarity scores by calculating the co- sine similarity of x-vectors. First, we do same-speaker pre-operative comparisons of the utterances (T1). Second, we do same-speaker different-severity comparisons (T1+T2). Finally, we do different- speaker comparisons (non-target). By separating these distributions, it is possible to quantify the impact of severity on the x-vectors. To quantify this, we calculate the Equal Error Rate (EER) for each group, a common performance measure in speaker verification stud- ies. In the context of our H2P-VC task, the EER quantifies the ro- bustness of the x-vector extraction in the presence of speech pathol- ogy. Specifically, a high EER might indicate that the x-vectors are significantly influenced by the severity of the pathology.
## 5. RESULTS AND DISCUSSION
## 5.1. Objective evaluations
## 5.1.1. Severity: P-ESTOI
Table 2 shows the results of the P-ESTOI experiment. We find that the correlation with the ground truth P-ESTOI scores $r_{GT}$ are higher for the PPG-GST-based system, which indicates that the PPG-GST- based system has a better ability to capture the severity of the speech.
## 5.1.2. Severity: PER
The PER results in Table 3 show that the proposed PPG-GST sys- tem has a mean PER that is closer to the GT than the PPG. We can observe that the PER is only closer in the case of speaker $PHNF_{T2}$ and $PRVM_{T2}$, and $RQOF_{T3}$. $PHNF_{T2}$ and $PRVM_{T2}$ are both speakers with relatively healthy ratings (4+), where we would expect severity-related modelling of the PPG-GST to be less beneficial. The difference in the case of $RQOF_{T3}$ is negligible. Furthermore, we observe that the PPG on average has a tendency to make the speech more intelligible than the ground truth, the average PER being nearly 3% better than in the GT case. We conclude that the PPG-GST has a better severity preservation ability than the PPG based on the results of the PER experiment.
## 5.2. Subjective evaluations
## 5.2.1. Severity and speaker identity
The left side of Table 4 shows the results of the severity experiment. The raters found that the severity of the GT source and target speaker sounded fairly different (S-T column), even though they were noted to have similar severity of speech. We conclude that using severity labels alone is insufficient for the source speaker selection based on this result. Therefore, for evaluating the severity preservation prop- erty, we should look at the S2S column rather than the S2T column. We find that for nearly all speakers the proposed model had a higher similarity to the source speakers’ speech severity, with the exception of $PRVM_{T2} \rightarrow PHNF_{T2}$ where performance was on par. We conclude the PPG-GST outperforms the baseline model in severity preservation.
The right side of Table 4 shows the results of the speaker iden- tity conversion experiment. We find that with the exception of the conversion pair $PGAF_{T2}\rightarrow PHNF_{T2}$, (S-T column) the raters rated the two speakers as different. Furthermore, the proposed sys- tem does not seem to achieve the same level of speaker identity con- version as the PPG system but outperforms the PPG in the case of the $PGAF_{T2} \rightarrow PHNF_{T2}$, and is on par in case of the $PRVM_{T2} \rightarrow PHNF_{T2}$ conversion pair.
The challenges in evaluating speaker identity, as outlined in [1], warrant further discussion. Even though we have acquired a parallel dataset of healthy and pathological speech from the same speaker, the precise severity of the speech pathology could not exactly be matched, despite our best efforts. As a result, it is still possible
| |
231002570/1
|
# IMPROVING SEVERITY PRESERVATION OF HEALTHY-TO-PATHOLOGICAL VOICE CONVERSION WITH GLOBAL STYLE TOKENS
In healthy-to-pathological voice conversion (H2P-VC), healthy speech is converted into pathological while preserving the identity. The paper improves on previous two-stage approach to H2P-VC where (1) speech is created first with the appropriate severity, (2) then the speaker identity of the voice is converted while preserving the severity of the voice. Specifically, we propose improvements to (2) by using phonetic posteriorgrams (PPG) and global style tokens (GST). Furthermore, we present a new dataset that contains parallel recordings of pathological and healthy speakers with the same iden- tity which allows more precise evaluation. Listening tests by expert listeners show that the framework preserves severity of the source sample, while modelling target speaker’s voice. We also show that (a) pathology impacts x-vectors but not all speaker information is lost, (b) choosing source speakers based on severity labels alone is insufficient.
IndexTerms — voice conversion, pathological speech, oral can- cer speech, autoencoder
## 1. INTRODUCTION
Healthy-to-pathological voice conversion (H2P-VC) is a new voice conversion (VC) task, where healthy speech is converted so that it resembles characteristics of a speech pathology, while preserv- ing the identity of the original speaker. H2P-VC could play a cru- cial role in several applications. For example, H2P-VC could be used to help patients mentally prepare for the changes in their fu- ture voice due to medical interventions [1, 2, 3]. By listening to the post-treatment voice, patients can better understand and adapt to the potential alterations, reducing anxiety and leading to a better quality of life [4]. Additionally, H2P-VC could be also used as a data augmentation technique for improving automatic speech recog- nition (ASR) systems [5,6]. Moreover, H2P-VC holds potential for speaker anonymisation of clinically collected speech samples, mak- ing it valuable for protecting speaker identity in sensitive contexts, ensuring confidentiality while preserving severity that might be re- quired for further analysis.
H2P-VC’s main challenge is the lack of appropriate, parallel data, which leads to difficulties during evaluation, and means that only non-parallel VC techniques are applicable. The lack of paral- lel healthy and pathological data from the same speaker means that we do not have the ground truth (GT) for the task, which makes the evaluation of the speech complicated. Furthermore, most work on H2P-VC has been done using the UASpeech corpus [7], making it challenging to adequately model and evaluate severity features that might be apparent at the prosodic level.
Fig. 1. Outline of our proposed approach for severity-preserving voice conversion.
Therefore, current approaches to H2P-VC resort to non-parallel VC techniques. These techniques either take an existing pathological speech sample [2] or create a synthetic pathological speech sample [1], which is then converted to a new speaker’s identity. Recently, [1] proposed a two-stage model for the H2P-VC task. The first stage focused on capturing time-variant aspects of the pathological speech using the Voice Transformer Network (VTN) [8]. This led to the pro- duction of speech mimicking the severity of the pathological speech but, unfortunately, resulted in the loss of source speaker identity. During the second stage, [1] attempted to reconstruct the original identity of the healthy speaker using a VQVAE-based model [9] from the crank toolkit [10]. Although the framework successfully altered the source speaker’s characteristics, the result did not accu- rately resemble the target speaker’s characteristics. The work hy- pothesised that these shortcomings were partially due to evaluation limitations, as [1] lacked access to parallel GT. An additional issue was that the second stage of the VC adversely affected the severity features of the speech.
To address these limitations, our study’s primary focus is en- hancing the second stage of the method presented in [1]. We pro- pose a new VC model leveraging phonetic posteriorgram (PPG) and global style tokens (GST) shown in Figure 1. Many PPG-based VC systems were among the best in the VCC2020 challenge [11], how- ever, they have a tendency to remove the severity features of the speech [12,13], which is undesired in H2P-VC. Our approach aims to improve the preservation of the severity while effectively convert- ing to the target speaker’s voice. GSTs have been shown to be able to control prosody in text-to-speech (TTS) [14], and in VC [15]. Using
| |
231002570/3
|
size of 5000 were used as basic units. After training, phonetic poste- riorgram (PPG) features are extracted using the Conformer encoder.
As for incorporating speaker-specific information, we use the pre-trained ECAPA-TDNN x-vector checkpoint from the Speech- Brain toolbox with default parameters 2 [21]. The Taco2-AR archi- tecture used is described in [17]. Due to the scarcity of training data, the NKI-OC-VC dataset (around 2 hours), we train the model with a curated (RDH-VL) subset of the Mozilla Common Voice dataset 3 [22] alongside the NKI-OC-VC dataset. The RDH-VL dataset con- tains 12.3 hours of speech from a Flemish Dutch speaker.
## 3.2. Proposed: PPG + GST
While PPGs are highly effective speaker-independent features, us- ing PPGs and x-vectors only has several shortcomings with patho- logical speech. First, PPGs primarily capture time-variant features of pathological speech, and are unable to model time-invariant as- pects. Second, if the ASR used to extract the PPGs can recognise pathological speech very well, the approach of PPG can lead to en- hanced speech [12,13]. X-vectors can indeed model time-invariant aspects but x-vectors are trained only with healthy data, therefore their performance is likely insufficient.
In order to address this, we propose a global style-token (GST) based reference encoder to the architecture which we use as a sever- ity encoder. GST has been proposed in [14] as a way to control var- ious aspects of prosody, such as speed, style of speaking, and style of singing [23,24]. It has been successfully used as a style transfer technique also in VC [15]. We think that a mixture of using PPG to model linguistic information, x-vector to model severity informa- tion, and GST to model severity information is an efficient way to learn severity-related features from our dataset.
For the implementation of the GST reference encoder used as the severity encoder, we follow the default parameters of [14] with the following exceptions: instead of adding the style embedding, we concatenate the style embedding, the x-vector and the encoder’s hidden states, and add a linear projection layer. Furthermore, as the reference encoder leaks noise from the samples, we enhanced the speech with ConvTASNet using a public implementation 4 [25].
HifiGAN was used as a neural vocoder for both systems [26]. We followed an open-source implementation 5. The training data of the HifiGAN was the RDH-VL dataset.
## 4. EXPERIMENTAL SETUP
In the next sections, we present the methodology to evaluate three as- pects of the systems: severity preservation, naturalness, and speaker identity. Due to a large number of possible conversion pairs, we only evaluate a limited set of converted speech. The selection of these sets is explained in their respective section. The audio samples can be found online 6,
## 4.1. Objective evaluation metrics
## 4.1.1. Severity: P-(E)STOI
We use P-(E)STOI as an objective severity measure as it has been demonstrated to work well for the objective evaluation of dysarthric speech, and used previously to evaluate the severity of synthetic speech samples [27]. We calculate the P-(E)STOI scores for each pathological ground truth (GT) utterance, using the corresponding utterance of the control speakers as a healthy reference. These scores are then averaged to provide a speaker-level score. We repeat this with a limited set of the VC utterance: we convert one post-operative (T2 if available) stage of each speaker to the target speaker in the RDH-VL dataset. We compare the obtained P-(E)STOI scores for each VC system to the ground truth samples by calculating the Pear- son’s correlation ($r_{GT}$).
Before doing our evaluation, we performed a preliminary analy- sis to ensure that P-ESTOI is indeed appropriate for severity evalu- ation, by calculating the P-(E)STOI scores on the GT samples first, and calculating the correlation between the scores and the severity labels. We found that both methods have a high correlation with the severity labels but the P-ESTOI had higher ($r = 0.82$) than the P- STOI ($r = 0.69$), therefore we will only consider P-ESTOI for the rest of our work.
## 4.1.2. Severity: Phoneme error rate (PER)
We use the phoneme error rate (PER) as another measure for evaluat- ing the severity objectively. We use a publicly available implementa- tion 7 phoneme recogniser which was trained on the Common Voice dataset [22]. We used phonemizer to acquire phonetic transcriptions [28]. For the PER evaluation experiment, we convert the T2 (or T3 if T2 is not available) stage of each speaker to the target speaker in the RDH-VL dataset, and we calculate the PER per speaker and per system.
## 4.2. Subjective evaluation experiments
## 4.2.1. Severity and speaker identity
In evaluating speaker similarity, we adapt the protocol outlined in [2] and incorporate an additional question about the similarity of speech severity. During the listening test, listeners are presented with two samples (either converted or ground truth), and are asked to judge (a) whether the two samples are from the same speaker, (b) whether the two samples have the same severity of speech. Listeners are asked to rate the similarity between the two voices on a 4-point Likert scale, where 1 means ’not similar at all’ and 4 means ’very similar’. For analysis, we converted these Likert scale ratings into percentages, i.e., a rating of ’1’ was considered as 0%, and ’4’ as 100%. Due to the time constraints of the test, we limit our analysis to 5 speaker pairs here. The test was done by 3 Dutch SLPs. For the listening ex- periment, we have prioritised pairs of source and target speakers who have similar severity. Choosing speakers with similar severity allows us to investigate whether the current practice of matching speakers based on their severity ratings is an appropriate approach for this VC task. Furthermore, compared to the protocol in [1], this setup has the advantage that the speaker identity can also be evaluated with expert listeners.
## 4.2.2. Naturalness
In order to evaluate the naturalness of the VC samples, we carried out a mean opinion score (MOS) test. We followed the setup in
| |
230607070/5
|
which is equipped with a norm
$$
\|w\|_X:=\|w\|+\|w_t\|
$$
where
$$
\|w\|:= \sup_{(x,t)\in\mathbb R\times[0,T]}|w(x,t)|.
$$
We note that (2.9) implies that
$$
\begin{array}{l}
\text{supp} w_j\subset\{(x,t)\in\mathbb R\times[0,T] : |x|\leq t+R\}\\
\Longrightarrow\text{supp} w_{j+1}\subset\{(x,t)\in\mathbb R\times[0,T] : |x|\leq t+R\}.
\end{array}
$$
The following lemma provides us a priori estimate.
Proposition Let $w\in C(\mathbb R\times[0,T])$ and supp $w\subset\{(x,t)\in\mathbb R\times[0,T]:|x|\leq t+R\}$. Then
$$
\|L'(|w|^p)\|\leq C\|w\|^p(T+R)
$$
where $C$ is a positive constant independent of $T$ and $\varepsilon$.
Proof. The proof of Proposition 3.1 is completely same as the one of Proposition 3.1 in Morisawa, Sasaki and Takamura [9]. $\Box$
Let us continue to prove Theorem 2.1. Set
$$
M:=\sum_{\alpha=0}^2\|f^{(\alpha)}\|_{L^\infty(\mathbb R)} + \sum_{\beta=0}^1\|g^{(\beta)}\|_{L^\infty(\mathbb R)}.
$$
First we note that $\|w_1\| \leq M\varepsilon$ by (2.10). (3.1) and (3.3) yield that
$$
\|w_{j+1} \| \leq M \varepsilon + C \|w_j\|^p(T+R)
$$
because it is trivial that
$$
|\overline{L'}(v)| \leq L'(|v|).
$$
Therefore the boundedness of $\{w_j\}$, i.e.
$$
\|w_j\| \leq 2M\varepsilon \quad(j\in\mathbb N),
$$
follows from
$$
C(2M\varepsilon)^p (T+R)\leq M\varepsilon.
$$
Assuming (3.5), one can estimate $\|w_{j+1}-w_j\|$ as follows.
$$
\begin{split}
\|w_{j+1}-w_j\|
& =\|\overline{L'}(|w_{j}|^p-|w_{j-1}|^p) \| \leq \|{L'}(||w_{j}|^p-|w_{j-1}|^p |) \| \\
&\leq 2^{p-1}p\|{L'}((|w_{j}|^{p-1}+|w_{j-1}|^{p-1})|w_{j}-w_{j-1}|) \| \\
&\leq 2^{p-1}pC(\|w_{j}\|^{p-1}+\|w_{j-1}\|^{p-1})(T+R)\|w_{j}-w_{j-1}\| \\
&\leq 2^{p}pC(2M\varepsilon)^{p-1}(T+R)\|w_{j}-w_{j-1}\|.
\end{split}
$$
| |
230607070/7
|
The convergence of the sequence $\{w_j\}$ to $w$ in the closed subspace of $X$ satisfying $\|w\|, \|w_t\| \leq 2M\varepsilon$ is established by (3.5), (3.6), (3.7), and (3.8), which follow from
$$
2^{p+1}pC(2M)^{p-1}\varepsilon^{p-1}(T+R) \leq 1.
$$
Therefore the statement of Theorem 2.1 is established with
$$
\varepsilon_1:=(2^{p+2}pC(2M)^{p-1}R)^{-1/(p-1)},\quad
C_1:= 2^{p+1}pC(2M)^{p-1}
$$
because $R\leq (2C_1)^{-1}\varepsilon^{-(p-1)}$ holds for $0<\varepsilon\leq\varepsilon_1$. $\Box$
## Proof of Theorem 2.2
Following Rammaha [10], set
$$
H(t):=\int^t_0(t-s)ds\int^{s+R}_{s+R_0}u(x,s)dx
$$
where $R_0$ is some fixed point with $0<R_0<R$. We may assume that there exists a point $x_0\in(R_0,R)$ such that $f(x_0)>0$ because of the assumption () and of a possible shift of $x$ -variable.
Then it follows that
$$
\begin{array}{ll}
H''(t)
&\displaystyle=\int^{t+R}_{t+R_0}u(x,s)dx\\
&\displaystyle=\frac{\varepsilon}{2}\int^{t+R}_{t+R_0}
\left\{f(x+t)+f(x-t)+\frac{1}{2}\int_{x-t}^{x+t}g(y)dy\right\}dx+\frac{1}{2}F(t),
\end{array}
$$
where
$$
F(t):=\int^{t+R}_{t+R_0}dx\int_0^tds\int_{x-t+s}^{x+t-s}|u_x(y,s)|^pdy.
$$
By virtue of (2.15) and (4.1), we have that
$$
H''(t)\geq \frac{\varepsilon}{2}\int^{t+R}_{t+R_0}f(x-t)dx \geq2C_f\varepsilon
$$
where
$$
C_f:=\frac{1}{4}\int^R_{R_0}f(y)dy>0.
$$
Integrating this inequality in [$0,t$] twicely and noting that $H'(0)=H(0)=0$, we have
$$
H(t)\geq C_f\varepsilon t^2.
$$
| |
230607070/2
|
We note that, even if $|u_x|^p$ is replaced with $|u_t|^p$, (1.2) still holds. Such a result is due to Zhou [15] for the upper bound of $T(\varepsilon)$, and Kitamura, Morisawa and Takamura [6] for the lower bound of $T(\varepsilon)$. As model equations to ensure the optimality of the general theory for nonlinear wave equations by Li, Yu and Zhou [7, 8], the nonlinear term $|u_t|^p$ is sufficient to be studied except for the “ combined effect” case. See Morisawa, Sasaki and Takamura [9] and Kido, Sasaki, Takamatsu and Takamura [5] for this direction with a possibility to improve the general theory. But it is quite meaningful to deal with also $|u_x|^p$ because their proofs are technically different from each others. Moreover, there is no result on its blow-up boundary due to lack of the monotonicity of the solution, while the one for $|u_t|^p$ is well- studied by Sasaki [12, 13], and Ishiwata and Sasaki [1, 2]. See Remark 2.1 below. It is also remarkable that it can be studied if the nonlinear term has a special form of both $u_t$ and $u_x$. See Sasaki [14] for this direction. Our research may help us to study the blow-up boundary for the equation in (1.1) near future. This paper is organized as follows. In the next section, the preliminaries are intro- duced. Moreover, (1.2) are divided into two theorems. Section 3 is devoted to the proof of the existence part, the lower bound of $T(\varepsilon)$, of (1.2). The main strategy is the iteration method for the system of integral equations for $(u,u_x)$, which is essentially due to Kita- mura, Morisawa and Takamura [6]. They employed it for the system of integral equations for $(u,u_t)$ to construct a classical solution of the wave equation with nonlinear term $|u_t|^p$, which is originally introduced by John [3]. In the section 4, following Rammaha [10, 11]. we prove the blow-up part, the upper bound of $T(\varepsilon)$, of (1.2). We note that the method to reduced to $u$ -closed integral inequality by Zhou [15] for the nonlinear term $|u_t|^p$ cannot be applicable to (1.1) because a time delay appears in the reduced ordinary differential inequality. Rammaha [10, 11] overcomes this difficulty by employing weighted functionals along with the characteristic direction in studyng two or three dimensional wave equations with nonlinear terms of spatial derivatives.
## Preliminaries and main results
Throughout this paper, we assume that the initial data $(f,g)\in C_0^2(\mathbb R)\times C^1_0(\mathbb R)$ satisfies
$$
\mbox{\rm supp}f,\ \mbox{supp}g\subset\{x\in\mathbb R: |x|\le R\},\quad R\geq1.
$$
Let $u$ be a classical solution of (1.1) in the time interval $[0,T]$. Then the support condition of the initial data, (2.1), implies that
$$
\mbox{supp}\ u(x,t)\subset\{(x,t)\in\mathbb R\times[0,T]:|x|\leq t+R\}.
$$
For example, see Appendix of John [4] for this fact.
It is well-known that $u$ satisfies the integral equation;
$$
u(x,t)=\varepsilon u^0(x,t)+L(|u_x|^p)(x,t),
$$
where $u^0$ is a solution of the free wave equation with the same initial data,
$$
u^0(x,t):=\frac{1}{2}\{f(x+t)+f(x-t)\}+\frac{1}{2}\int_{x-t}^{x+t}g(y)dy,
$$
| |
230607070/6
|
Therefore the convergence of $\{w_j\}$ follows from
$$
\|w_{j+1}-w_j\| \leq \frac{1}{2}\|w_j-w_{j-1}\|
$$
provided (3.5) and
$$
2^{p}pC(2M\varepsilon)^{p-1}(T+R) \leq \frac{1}{2}
$$
are fulfilled.
First we note that $\|(w_1)_t\| \leq M\varepsilon$ by (2.13). Assume that (3.5) and (3.6) are fulfilled. Since (3.2) and (3.3) yield that
$$
\begin{split}
\|(w_{j+1})_t\|
&\leq M\varepsilon+ \|{L'}(p|w_j|^{p-2}w_j(w_j)_t)\| \\
&\leq M\varepsilon + \|{L'}(p|w_j|^{p-1}|(w_j)_t|)\| \\
&\leq M\varepsilon + pC\|w_j\|^{p-1}(T+R)\|(w_j)_t\| \\
&\leq M\varepsilon + pC(2M\varepsilon)^{p-1}(T+R)\|(w_j)_t\|, \\
\end{split}
$$
the boundness of $\{(w_j)_t\}$, i.e.
$$
\|(w_j)_t\|\leq 2M\varepsilon,
$$
follows as long as it is fulfilled that
$$
pC(2M\varepsilon)^{p-1} (T+R)\leq 1.
$$
Assuming (), one can estimate $\{(w_{j+1})_t-(w_j)_t\}$ as follows. Noting that
$$
\begin{split}
&||w_j|^{p-2}w_j-|w_{j-1}|^{p-2}w_{j-1}| \\
&\leq
\left\{
\begin{array}{ll}
(p-1)2^{p-2}(|w_j|^{p-2}+|w_{j-1}|^{p-2})|w_j-w_{j-1}| & \mbox{for}\ p\ge2,\\
2|w_j-w_{j-1}|^{p-1} & \mbox{for}\ 1<p<2,
\end{array}
\right.
\end{split}
$$
we have
$$
\begin{array}{l}
\|(w_{j+1})_t-(w_j)_t\|
= \|L'(p|w_j|^{p-2}w_j(w_j)_t -p|w_{j-1}|^{p-2}w_{j-1}(w_{j-1})_t) \| \\
\leq p \|L'(|w_j|^{p-1} | (w_j)_t -(w_{j-1})_t |) \|
+ p\|L'(| |w_j|^{p-2}w_j -|w_{j-1}|^{p-2}w_{j-1}||(w_{j-1})_t |) \| \\
\leq pC \|w_j\|^{p-1}(T+R) \|(w_j)_t -(w_{j-1})_t \| \\
\quad + \left\{
\begin{array}{ll}
L'(p(p-1)2^{p-2}(|w_j|^{p-2}+|w_{j-1}|^{p-2})|w_j-w_{j-1}||(w_{j-1})_t | & \mbox{for}\ p\geq2,\\
L'(2p|w_j-w_{j-1}|^{p-1}|(w_{j-1})_t |) & \mbox{for}\ 1<p<2
\end{array}
\right. \\
\leq pC \|w_j\|^{p-1} (T+R) \|(w_j)_t -(w_{j-1})_t \| \\
\quad + \left\{
\begin{array}{ll}
p(p-1)2^{p-2}C(\|w_j\|^{p-2}+\|w_{j-1}\|^{p-2})\|w_j-w_{j-1}\|\|(w_{j-1})_t \|& \mbox{for}\ p\geq2,\\
2pC\|w_j-w_{j-1}\|^{p-1}\|(w_{j-1})_t \| & \mbox{for}\ 1<p<2
\end{array}
\right. \\
\leq pC (2M\varepsilon)^{p-1} (T+R) \|(w_j)_t -(w_{j-1})_t \| + O\left(\displaystyle\frac{1}{2^{j\min(p-1,1)}}\right).
\end{array}
$$
Therefore, we obtain the convergence of $\{(w_j)_t\}$ provided
$$
pC(2M\varepsilon)^{p-1} (T+R)\leq \frac{1}{2}.
$$
| |
230607070/4
|
Proposition Assume that $(f,g)\in C_0^2(\mathbb R)\times C_0^1(\mathbb R)$. Let $w$ be a $C^1$ solution of () in which $u_x$ is replaced with $w$. Then,
$$
u(x,t):= \int^x_{-\infty}w(y,t)dy
$$
is a classical solution of (1.1) in $\mathbb R\times[0,T]$.
Proof. This is easy along with the computations above in this section. $\Box$
Our results are divided into the following two theorems.
Theorem 2.1 Assume (2.1). Then, there exists a positive constant $\varepsilon_1=\varepsilon_1(f,g,p,R)>0$ such that a classical solution $u\in C^2(\mathbb R\times[0,T])$ of (1.1) exists as far as $T$ satisfies
$$
T \leq C_1\varepsilon^{-(p-1)}
$$
where $0<\varepsilon\leq\varepsilon_1$, and $C_1$ is a positive constant independent of $\varepsilon$.
Theorem 2.2 Assume (2.1) and
$$
f(x), g(x)\geq0,\ \mbox{and}\ f(x)\not\equiv0.
$$
Then, there exists a positive constant $\varepsilon_2=\varepsilon_2(f,p,R)>0$ such that any classical solution of (1.1) in the time interval $[0,T]$ cannot exist as far as $T$ satisfies
$$
T>C_2\varepsilon^{-(p-1)}
$$
where $0<\varepsilon\leq\varepsilon_2$, and $C_2$ is a positive constant independent of $\varepsilon$.
The proofs of above theorems are given in following sections.
## Proof of Theorem 2.1
According to Proposition 2.1, we shall construct a $C^1$ solution of (2.9) in which $u_x=w$ is the unknown function. Let $\{w_j\}_{j\in \mathbb N}$ be a sequence of $C^1(\mathbb R\times[0,T])$ defined by
$$
\left\{
\begin{array}{l}
w_{j+1}=\varepsilon u_x^0+\overline{L'}(|w_j|^p), \\
w_1=\varepsilon u_x^0.
\end{array}
\right.
$$
Then, in view of (2.12), $(w_j)_t$ has to satisfy
$$
\left\{
\begin{array}{l}
(w_{j+1})_t=\varepsilon u_{xt}^0+{L'}(p|w_j|^{p-2}w_j(w_j)_t), \\
(w_1)_t=\varepsilon u_{xt}^0,
\end{array}
\right.
$$
so that the functional space in which $\{w_j\}$ converges is
$$
X:=\{w\in C^1(\mathbb R\times[0,T]): \|w\|_X<\infty, \mbox{supp}w \subset \{(x,t)\in\mathbb R\times[0,T] : |x|\leq t+R\}\}
$$
| |
230607070/1
|
# The lifespan of classical solutions of one dimensional wave equations with semilinear terms of the spatial derivative
Takiko Sasaki ∗, Shu Takamatsu †, Hiroyuki Takamura ‡
This paper is devoted to the lifespan estimates of small classical solutions of the initial value problems for one dimensional wave equations with semilinear terms of the spatial derivative of the unknown function. It is natural that the result is same as the one for semilinear terms of the time-derivative. But there are so many differences among their proofs. Moreover, it is meaningful to study this problem in the sense that it may help us to investigate its blow-up boundary in the near future.
## Introduction
In this paper, we consider the initial value problems;
$$
\left\{
\begin{array}{ll}
u_{tt}-u_{xx}=|u_x|^p
&\mbox{in}\quad \mathbb R\times(0,T),\\
u(x,0)=\varepsilon f(x),\ u_t(x,0)=\varepsilon g(x),
& x\in\mathbb R,
\end{array}
\right.
$$
where $p>1$, and $T>0$. We assume that $f$ and $g$ are given smooth functions of compact support and a parameter $\varepsilon>0$ is “ small enough”. We are interested in the lifespan $T(\varepsilon)$, the maximal existence time, of classical solutions of (1.1). Our result is that there exists positive constants $C_1,C_2$ independent of $\varepsilon$ such that $T(\varepsilon)$ satisfies
$$
C_1\varepsilon^{-(p-1)}\le T(\varepsilon) \le C_2\varepsilon^{-(p-1)}.
$$
∗ Department of Mathematical Engineering, Faculty of Engineering, Musashino University, 3-3-3 Ari- ake, Koto-ku, Tokyo 135-8181, Japan./ Mathematical Institute, Tohoku University, Aoba, Sendai 980- 8578, Japan. e-mail: t-sasaki@musashino-u.ac.jp. † Master course, Mathematical Institute, Tohoku University, Aoba, Sendai 980-8578, Japan. email: shu.takamatsu.r8@dc.tohoku.ac.jp ‡ Mathematical Institute, Tohoku University, Aoba, Sendai 980-8578, Japan. e-mail: hi- royuki.takamura.a1@tohoku.ac.jp.
| |
230607070/3
|
and a linear integral operator $L$ for a function $v=v(x,t)$ in Duhamel’s term is defined by
$$
L(v)(x,t):=\frac{1}{2}\int_0^tds\int_{x-t+s}^{x+t-s}v(y,s)dy.
$$
Then, one can apply the time-derivative to (2.3) to obtain
$$
u_t(x,t)=\varepsilon u_t^0(x,t)+L'(|u_x|^p)(x,t)
$$
and
$$
u_t^0(x,t)=\frac{1}{2}\{f'(x+t)-f'(x-t)+g(x+t)+g(x-t)\},
$$
where $L'$ for a function $v=v(x,t)$ is defined by
$$
L'(v)(x,t):=\frac{1}{2}\int_0^t\{v(x+t-s,s)+v(x-t+s,s)\}ds.
$$
Therefore, $u_t$ is expressed by $u_x$. On the other hand, applying the space-derivative to (2.3), we have
$$
u_x(x,t)=\varepsilon u_x^0(x,t)+\overline{L'}(|u_x|^p)(x,t)
$$
and
$$
u_x^0(x,t)=\frac{1}{2}\{f'(x+t)+f'(x-t)+g(x+t)-g(x-t)\},
$$
where $\overline{L'}$ for a function $v=v(x,t)$ is defined by
$$
\overline{L'}(v)(x,t):=
\frac{1}{2}\int_0^t\{v(x+t-s,s)-v(x-t+s,s)\}ds.
$$
Remark 2.1 In view of (2.9), it is almost impossible to obtain a point-wise positivity of $u_x$. This fact prevents us from studying its blow-up boundary as stated in Introduction.
Moreover, applying one more time-derivative to (2.9) yields that
$$
u_{xt}(x,t)=\varepsilon u_{xt}^0(x,t)+{L'}(p|u_x|^{p-2}u_xu_{xt})(x,t)
$$
and
$$
u_{xt}^0(x,t)=\frac{1}{2}\{f''(x+t)-f''(x-t)+g'(x+t)+g'(x-t)\}.
$$
Similarly, we have that
$$
u_{tt}(x,t)=\varepsilon u^0_{tt} + |u_x|^p(x,t) + \overline{L'}(p|u_x|^{p-2}u_xu_{xt})(x,t).
$$
Therefore, $u_{tt}$ is expressed by $u_x,u_{xt}$ and so is $u_{xx}$.
First, we note the following fact.
| |
230607070/9
|
that
$$
\begin{split}
F(t)
&\geq \int_0^{t-R_1}\frac{t-s}{t}ds\int_{s+R_0}^{s+R}(y-s-R_0)|u_x(y,s)|^pdy\\
&\ +\int_{t-R_1}^t\frac{t-s}{t}ds\int_{s+R_0}^{2t-s+R_0}(y-s-R_0)|u_x(y,s)|^pdy\\
&\ +\int_{t-R_1}^t2(t-s)ds\int_{2t-s+R_0}^{s+R}\frac{y-s-R_0}{2t}|u_x(y,s)|^pdy\\
&= \frac{1}{t}\int_0^t(t-s)ds\int_{s+R_0}^{s+R}(y-s-R_0)|u_x(y,s)|^pdy.\\
\end{split}
$$
In this way, (2.15), (4.1) and the estimate of $F(t)$ above yield that
$$
H''(t)\geq \frac{1}{2}F(t)\geq \frac{1}{2t}\int_0^t(t-s)ds\int_{s+R_0}^{s+R}(y-s-R_0)|u_x(y,s)|^pdy
\quad\mbox{for}\ t\geq R_1.
$$
Moreover, it follows from (2.2), integration by parts and H ¨ o lder’s inequality that
$$
\begin{split}
|H(t)|
&=\left|\int_0^t(t-s)ds\int_{s+R_0}^{s+R}\partial_y(y-s-R_0)u(y,s)dy\right|\\
&=\left|\int_0^t(t-s)ds\int_{s+R_0}^{s+R}(y-s-R_0)u_x(y,s)dy\right|\\
&\leq \int_0^t(t-s)ds\int_{s+R_0}^{s+R}(y-s-R_0)|u_x(y,s)|dy\\
&\leq \left(\int_0^t(t-s)ds\int_{s+R_0}^{s+R}(y-s-R_0)|u_x(y,s)|^p dy\right)^{1/p}I(t)^{1-1/p},\\
\end{split}
$$
where
$$
I(t):=\int_0^t(t-s)ds\int_{s+R_0}^{s+R}(y-s-R_0)dy=\frac{1}{4}t^2(R-R_0)^2=t^2R_1^2.
$$
Hence we obtain that
$$
H''(t)\geq \frac{1}{2}R_1^{-2(p-1)}t^{1-2p}|H(t)|^p\quad\mbox{for}\ t\geq R_1.
$$
Therefore, the argument in Rammaha [10] can be applied to (4.2) and (4.4) to ensure that there exist positive constants $\varepsilon_2=\varepsilon_2(f,p,R)$ and $C_2$ independent of $\varepsilon$ such that a contradiction appears provided
$$
T>C_2\varepsilon^{-(p-1)}
$$
holds for $0<\varepsilon \leq \varepsilon_2$. The proof is now completed. $\Box$
## Acknowledgement
The first author is partially supported by the Grant-in-Aid for Young Scientists (No. 18K13447), Japan Society for the Promotion of Science. The third author is partially supported by the Grant-in-Aid for Scientific Research (A) (No. 22H00097), Japan Society for the Promotion of Science.
| |
230607070/8
|
On the other hand, $F(t)$ can be rewritten as
$$
F(t)=\int_0^tds\int_{t+R_0}^{t+R}dx\int_{x-t+s}^{x+t-s}|u_x(y,s)|^pdy.
$$
From now on, we assume that
$$
t \geq R_1:=\frac{R-R_0}{2}>0.
$$
Then, inverting the order on $(y,x)$ -integral, for $0 \leq s \leq t-R_1$, we have that
$$
\begin{split}
&\int_{t+R_0}^{t+R}dx\int_{x-t+s}^{x+t-s}|u_x(y,s)|^pdy\\
&=\left(\int_{s+R_0}^{s+R}\int_{t+R_0}^{y+t-s}
+\int_{s+R}^{2t-s+R_0}\int_{t+R_0}^{t+R}
+\int_{2t-s+R_0}^{2t-s+R}\int_{y-t+s}^{t+R}\right)|u_x(y,s)|^pdxdy\\
&\geq \int_{s+R_0}^{s+R}dy\int_{t+R_0}^{y+t-s}|u_x(y,s)|^pdx.
\end{split}
$$
Similarly, for $t-R_1\leq s \leq t$, we also have that
$$
\begin{split}
&\int_{t+R_0}^{t+R}dx\int_{x-t+s}^{x+t-s}|u_x(y,s)|^pdy\\
&=\left(\int_{s+R_0}^{2t-s+R_0}\int_{t+R_0}^{y+t-s}
+\int_{2t-s+R_0}^{s+R}\int_{y-t+s}^{y+t-s}
+\int_{s+R}^{2t-s+R}\int_{y-t+s}^{t+R}\right)|u_x(y,s)|^pdxdy\\
&\geq \int_{s+R_0}^{2t-s+R_0}dy\int_{t+R_0}^{y+t-s}|u_x(y,s)|^pdx
+\int_{2t-s+R_0}^{s+R}dy\int_{y-t+s}^{y+t-s}|u_x(y,s)|^pdx.
\end{split}
$$
Hence we obtain that
$$
\begin{split}
F(t)
&\geq \int_0^{t-R_1}ds\int_{s+R_0}^{s+R}(y-s-R_0)|u_x(y,s)|^pdy\\
&\ +\int_{t-R_1}^tds\int_{s+R_0}^{2t-s+R_0}(y-s-R_0)|u_x(y,s)|^pdy\\
&\ +\int_{t-R_1}^tds\int_{2t-s+R_0}^{s+R}2(t-s)|u_x(y,s)|^pdy.
\end{split}
$$
Therefore it follows from (4.3) and
$$
1=\frac{y-s-R_0}{y-s-R_0} \geq \frac{y-s-R_0}{R-R_0} \geq \frac{y-s-R_0}{2t}
$$
| |
180802589/5
|
$$
\begin{aligned}
\int_{3}\tilde{\Gamma}^{(m)}(2,3)X_{\rm pp}(3)\tilde{\Gamma}^{(n)}(3,1)=-\int_{\hat{3}}\Bigg(\int_{\xi_{n_{3}}^{-}}^{-\epsilon_{0}}\frac{\text{d}\xi_{n_{3}}}{2|\xi_{n_{3}}|}\frac{\rho(\xi_{n_{3}})}{\rho_{n_{3}}}+\int_{\epsilon_{0}}^{\xi_{n_{3}}^{+}}\frac{\text{d}\xi_{n_{3}}}{2|\xi_{n_{3}}|}\frac{\rho(\xi_{n_{3}})}{\rho_{n_{3}}}\Bigg)\tilde{\Gamma}^{(m)}(2,3)\tilde{\Gamma}^{(n)}(3,1)\ .
\end{aligned}
$$
Here the cut-off should be chosen sufficiently small such that all quantities that converge in the limit $\xi\rightarrow0$ can be approximated by their values at the Fermi level for $|\xi|\leq\epsilon_{0}$.
The next step is to calculate the change in $\Gamma$ when the cut-off is reduced to $\epsilon$, i.e.,
$$
\Delta\Gamma=\Gamma_{\epsilon}-\Gamma_{\epsilon_{0}},
$$
where $(\epsilon_{0}-\epsilon)/\epsilon_{0}\ll1$. Using Eqs. (27) and (32), we have
$$
\begin{aligned}
\Delta\Gamma(2,1)=&-\int_{\hat{3}}\int_{\epsilon}^{\epsilon_{0}}\frac{\text{d}\xi_{n_{3}}}{|\xi_{n_{3}}|}\tilde{\Gamma}(2,\hat{3}){\Gamma}(\hat{3},1)\\
=&-\ln\left(\frac{\epsilon_{0}}{\epsilon}\right)\int_{\hat{3}}\tilde{\Gamma}(2,\hat{3}){\Gamma}(\hat{3},1)\\
=&-\ln\left(\frac{\epsilon_{0}}{\epsilon}\right)\int_{\hat{3}}\tilde{\Gamma}(2,\hat{3})\tilde{\Gamma}(\hat{3},1)\\
&+\ln^{2}\left(\frac{\epsilon_{0}}{\epsilon}\right)\int_{\hat{3}\,\hat{4}}\tilde{\Gamma}(2,\hat{3})\tilde{\Gamma}(\hat{3},\hat{4})\tilde{\Gamma}(\hat{4},1)\\
&-\ln^{3}\left(\frac{\epsilon_{0}}{\epsilon}\right)\int_{\hat{3}\,\hat{4}\,\hat{5}}\tilde{\Gamma}(2,\hat{3})\tilde{\Gamma}(\hat{3},\hat{4})\tilde{\Gamma}(\hat{4},\hat{5})\tilde{\Gamma}(\hat{5},1)\\
&\pm\dots.
\end{aligned}
$$
Note that (35) and the first line of (36) differ by the tilde on top of $\Gamma(\hat 3,1)$. From here we obtain the RG flow equation
$$
\frac{\partial\Gamma(2,1)}{\partial\ln(\epsilon_{0}/\epsilon)}=\frac{\partial\Delta\Gamma(2,1)}{\partial\ln(\epsilon_{0}/\epsilon)}=-\int_{\hat{3}}\Gamma(2,\hat{3})\Gamma(\hat{3},1).
$$
Due to the onsite interaction being repulsive, the node- less $s$ -wave solution is suppressed. Thus, we only need to focus on the subspace of nodal superconducting states, i.e., those with sign-changing form factors of the gap- function. We achieve that by splitting $\Gamma$ as follows:
$$
\begin{aligned}
&\Gamma=\Gamma_{0}+\Gamma_{1}+\Gamma_{01}, \\
&\Gamma_{0}=U_{0}\tilde{\Gamma}^{(1)}+U_{0}^{2}\tilde{\Gamma}^{(1)}*\tilde{\Gamma}^{(1)}+U_{0}^{3}\tilde{\Gamma}^{(1)}*\tilde{\Gamma}^{(1)}*\tilde{\Gamma}^{(1)}+\dots, \\
&\Gamma_{1}=U_{0}^{2}\tilde{\Gamma}^{(2)}+U_{0}^{3}\tilde{\Gamma}^{(3)}+U_{0}^{4}\left(\tilde{\Gamma}^{(2)}*\tilde{\Gamma}^{(2)}+\tilde{\Gamma}^{(4)}\right)+\dots, \\
&\Gamma_{01}=U_{0}^{3}\left(\tilde{\Gamma}^{(1)}*\tilde{\Gamma}^{(2)}+\tilde{\Gamma}^{(2)}*\tilde{\Gamma}^{(1)}\right)+\dots.
\end{aligned}
$$
Here, $\Gamma_{0}$ is the subspace of the nodeless $s$ -wave, $\Gamma_{1}$ the one of nodal superconductivity and orthogonal to $\Gamma_{0}$, and $\Gamma_{01}$ describes the interaction between these two. Since we are in the weak-coupling limit, we take only the lowest relevant order in $U_{0}$. With the nodeless $s$ -wave being suppressed, one obtains for the singlet channel
$$
\Gamma^s(2,1)\approx U_{0}^{2}\tilde{\Gamma}^{s,(2)}(2,1)=U_{0}^{2}\Gamma^{(2b)}(2,1)\
$$
and for the triplet channel
$$
\Gamma^t(2,1)\approx U_{0}^{2}\tilde{\Gamma}^{t,(2)}(2,1)=U_{0}^{2} \Gamma^{(2c)}(2,1)\ .
$$
The diagram $\Gamma^{(2a)}$ is not present as it is not divergence- free.
We now transform the integral over the Fermi surface into a sum over discrete points suited to be numerically evaluated. Since we work in the limit $T\rightarrow 0$, we can restrict the scattering processes to the ones where the electrons start and end on the Fermi surface. Thus, we can rescale $\Gamma$ with a momentum dependent scaling factor given by
$$
g(\hat{2},\hat{1}):=\sqrt{\rho_{n_{2}}\rho_{n_{1}}\frac{\bar{v}_{n_{2},F}\bar{v}_{n_{1},F}}{v_{F}(\hat{2})v_{F}(\hat{1})}\frac{\ell(\hat{2})\ell(\hat{1})}{S_{n_{2},F}S_{n_{1},F}}}\Gamma(\hat{2},\hat{1})\ ,
$$
where $\ell(\hat{i})$ is the length associated with the discrete Fermi surface point with index $i$. We substitute an or- thonormal eigensystem of $g$, i.e.,
$$
\begin{aligned}
&\sum_{n_{1}}\sum_{\hat{k}_{1}}g(\hat{2},\hat{1})\psi_{\nu}(\hat{1})=\lambda_{\nu}\psi_{\nu}(\hat{2}),\\
&\sum_{n_{1}}\sum_{\hat{k}_{1}}\psi_{\nu}(\hat{1})\psi_{\eta}(\hat{1})=\delta_{\nu\eta},
\end{aligned}
$$
in Eq. (37). The RG flow equation then becomes
$$
\frac{\partial\lambda_{\nu}}{\partial\ln(\epsilon_{0}/\epsilon)}=-\lambda_{\nu}^{2},
$$
where, due to the rescaling of $\Gamma$, the eigenvalues $\lambda_{\nu}$ coin- cide with the eigenvalues of $g$. Thus, the RG flow equa- tion is solved by a simple diagonalization of $g$. Also, we see from Eq. (44) that all $\lambda_{\nu}$ renormalize independently and that Eq. (44) is solved by
$$
\lambda_{\nu}=\frac{\lambda_{\nu}^{0}}{1+\lambda_{\nu}^{0}\ln(\epsilon_{0}/\epsilon)},
$$
where $\lambda_{\nu}^{0}$ is the eigenvalue obtained by using the cut-off $\epsilon_{0}$. Once the cut-off $\epsilon$ is reduced to
$$
\epsilon^{*}=\epsilon_{0}\,e^{-1/\lambda_{\nu}^{0}},
$$
$\lambda_{\nu}$ diverges.
The leading instability is given by the first appearing divergence, i.e., by the most negative eigen- value $\lambda_{\text{min}}$ and $\epsilon^{*}$ is identified with the critical tempera- ture. Note that $\lambda_{\nu}^{0}$, when taken to the appropriate order
| |
180802589/7
|
honeycomb lattice with its two-atomic unit cell is the only multiband system we will discuss in this paper. Using Eq. (6), the orbital factor as defined in Eq. (8) needs to be modified for the case of longer-ranged interactions,
$$
\begin{aligned}
M_{1}&(1,2,3,4)=\\
&\sum_{ij}h_{1,ij}(\vec{k}_{2}-\vec{k}_{3})u_{n_{1},i}^{*}(\vec{k}_{1})u_{n_{2},j}^{*}(\vec{k}_{2})u_{n_{3},j}^{{\phantom{*}}}(\vec{k}_{3})u_{n_{4},i}^{{\phantom{*}}}(\vec{k}_{4}).
\end{aligned}
$$
Thus, we obtain the vertex function as
$$
\Gamma^{s}(2,1)\approx U_{0}^{2}\left(\Gamma^{(2b)}(2,1)+\alpha\frac{M_{1}(2,\bar{2},\bar{1},1)}{W} \right)
$$
for the singlet channel. The corresponding expression for the triplet channel, $\Gamma^t(2,1)$, is the same after replacing $\Gamma^{(2b)}$ by $\Gamma^{(2c)}$. The factor $\alpha$ will be used in the following to control the nearest-neighbor interaction strength.
## RESULTS
In the following, we will discuss the extended Hubbard model for three paradigmatic 2D lattices and the cor- responding superconducting ground states. All results obtained within the WCRG method are asymptotically correct, i.e., for interactions $U_0 \rightarrow 0$ which we refer to as “ weak-coupling limit. ” We further emphasize that our analysis is limited to the Cooper channel with vanish- ing total momentum of the Cooper pairs [18]; the limi- tation to superconductivity is reasonable as this is the only weak-coupling instability of the Fermi sea. Other competing instabilities such as magnetism can also occur at weak interactions due to nesting (e.g. on the square lattice with $t_2=0$) but this is not a generic feature. Nest- ing causes these strong-coupling phenomena to emerge at weak couplings, thus we neglect them here. Although the results of the WCRG method are only exact in the limit $U\rightarrow0$, we will argue below (see the section “ Summary of results and discussion ”) that they provide a guiding principle for stronger-correlated phases and materials.
What can the method do – and what cannot? The WCRG only finds superconducting solutions, as men- tioned before. Moreover, it always finds a superconduct- ing instability. In contrast to mean-field theories, where “ you only can get what you put in ”, the WCRG method finds the leading instability without any bias towards one superconducting state or another. The underlying lat- tice determines the point group (in this paper, $D_4$ for the square and $D_6$ for the hexagonal lattices) and super- conducting instabilities must transform according to an irrep of this point group. Details about the irreps will be discussed in the subsections about the specific lattices. How meaningful it is to find a superconducting ground state is dictated by the corresponding coupling strength $V_{\rm eff}$ which is a measure of the critical temperature $T_c$. If $V_{\rm eff}$ is practically zero, so is $T_c$ – apparently it is not of much interest to discuss superconductors with a $T_c$ of a few $\mu$ K. Thus we will in addition to the superconducting ground states also show the effective coupling strengths $V_{\rm eff}$. Moreover, a superconducting instability can be frag- ile if another superconducting solution is close-by or even almost degenerate. In this case, tiny changes of parame- ters can cause a phase transition from one superconduct- ing phase to another; more experiment-oriented, small variations in strain, pressure, or temperature might have such effects. Thus we will also investigate the relative difference between the second-lowest eigenvalue and the lowest eigenvalues $\lambda_{\rm min}$ of the normalized vertex function $g$.
As mentioned before, the WCRG was introduced by Raghu et al. [18] based on early work of Kohn and Lut- tinger [7]. Given the interest to study unconventional su- perconductivity on 2D lattices, some of our results have been derived previously. For the sake of completeness, we will show them nevertheless. Moreover, these results will allow direct comparison with Ref. 18 (and other papers in the literature) and also serve for benchmarking pur- poses. In particular, the results for the square lattice for $\alpha=0$ will be vital for the discussion in the section about “ Numerical development and performance analysis ”.
Given the longstanding puzzle in understanding the pairing mechanism of the copper-oxide superconduc- tors [8, 37], there is enormous interest in $d$ -wave super- conductors. In the past years there has been, however, a growing interest in topological superconductivity. Be- sides the fundamental interest in this exotic state of mat- ter, the discovery of topological and Chern insulators [38 – 40], the electronic cousins of topological superconductors, has renewed the research interest in the past decade. The by far most important aspect of topological superconduc- tors is, however, the ability to trap Majorana zero modes at their vortex cores or at their boundaries – then referred to as chiral Majorana modes – to other superconducting or normal-state systems including vacuum. Topological superconductivity naturally occurs due to the following mechanism. Superconducting instabilities could be de- generate – this might either happen due to fine-tuning of parameters or generically for two-dimensional irreps such as $E$ on the square lattice. Whenever there is a degen- eracy, a quantum mechanical system can choose an arbi- trary superposition within this subspace, including com- plex superpositions of the type $\Delta(\vec k) \sim \psi_1 \pm i \psi_2$. With the exception of an ordinary $s$ -wave state, all higher- angular momentum superconducting states possess nodes in the superconducting gap functions. These zeros lead to a decrease of the condensation energy which is pro- portional to $|\Delta(\vec k)|^2$. The aforementioned complex su- perposition causes, however, a full gap and thus maxi- mizes the condensation energy. That is, for purely en- ergetic reasons, nature might prefer gap functions of the form $\Delta(\vec k) \sim \psi_1 \pm i \psi_2$ if $\psi_1$ and $\psi_2$ are degenerate. The price nature has to pay for this energy maximization is
| |
180802589/12
|
ment as for the stability of the chiral topological states in the two dimensional irreps holds: the superconducting state is particularly stable when nodes can be avoided. Note, however, that for purely repulsive electron-electron interactions this case is only possible when there are at least two Fermi surface pockets.
For larger values of $t_{2}/t_{1}$, the Fermi surface will move into the nodal lines of the $f$ -wave form factor of $B_{1}$ again. The geometry of the Fermi surface in this region is cap- tured in the last column for $B_{1}$ in Fig. 7. However, this $f$ -wave state still dominates, since it is the state that maximizes the superconducting order parameter close to the van Hove points.
In the following, we will discuss the effect of longer- ranged interactions. The form factor of (54) can be writ-
FIG. 8. (a) Phase diagram $t_{2}/t_{1}$ vs. band filling $n$ for the triangular lattice. The van Hove filling $n_{vH}$ is drawn as a red dotted line. For fillings $n<0.5$, the effective interaction, $V_{\text{eff}}$, becomes very small, such that the difference between different symmetries reaches the level of numerical noise. (b) Effective interaction, $V_{\text{eff}}$, as a function of the band filling $n$ and the second neighbor hopping $t_{2}/t_{1}$. Regions between cal- culated points have been interpolated. (c) Difference $\delta_\lambda$ of the two lowest eigenvalues corresonding to the superconduct- ing ground state and the solution that is closest to it.
FIG.9. Form factor of the superconducting instability on the Fermi surface in the extended Brillouin zone of the triangular lattice. The lattice parameters are $t_{2}/t_{1}=0.1$, $n=1.83$, and $\alpha=0$. The Brillouin zone is drawn in black and the green dashed lines are the nodal line of the $f$ -wave form factor of irrep $B_{1}$. Blue and orange colors denote different signs of the form factor.
Phase diagram showing the irrep of the leading superconducting instability in the triangular lattice for $0.05\leq\alpha\leq0.2$, $0.7\leq n\leq1.8$ and $-0.3\leq t_{2}/t_{1}\leq0.3$.
| |
180802589/14
|
FIG. 11. (a) Phase diagram $t_{2}/t_{1}$ vs. band filling $n$ for the honeycomb lattice. The van Hove fillings $n_{\text{vH}}$ are drawn as a red dotted line, the filling $n_{D}$ at which the Dirac points are located is drawn as a black dotted line. (b) Effective interac- tion, $V_{\text{eff}}$, as a function of the band filling $n$ and the second neighbor hopping $t_{2}/t_{1}$. Regions between calculated points have been interpolated. (c) Difference $\delta_\lambda$ of the two lowest eigenvalues corresonding to the superconducting ground state and the solution which is closest to it.
written as
$$
\begin{aligned}
&h_1(\vec{k}_{1}-\vec{k}_{2})=\begin{pmatrix}
0 & T(\vec{k}_{1}-\vec{k}_{2}) \\
T^{*}(\vec{k}_{1}-\vec{k}_{2}) & 0
\end{pmatrix},\\
&T(\vec{k})=e^{-ik_{y}}+e^{i\frac{1}{2}k_{y}}\left(e^{i\frac{\sqrt{3}}{2}k_{x}}+e^{-i\frac{\sqrt{3}}{2}k_{x}}\right),
\end{aligned}
$$
with
$$
\begin{aligned}
T(\vec{k}_{1}-\vec{k}_{2})=\frac{1}{\sqrt{3}}\sum_{i}\psi_{i}^{D_{3},1}(\vec{k}_{1})\left(\psi_{i}^{D_{3},1}(\vec{k}_{2})\right)^{*},
\end{aligned}
$$
where the sum runs over the nearest neighbor lattice har- monics, $i\in\{A_{1},A_{2},E\}$. On the honeycomb lattice, these correspond to the following basis functions:
$$
\begin{aligned}
&\psi_{A_{1}}^{D_{3},1}(\vec{k})=\frac{1}{\sqrt{3}}\left[e^{-ik_{y}}+e^{i\frac{1}{2}k_{y}}\left(e^{i\frac{\sqrt{3}}{2}k_{x}}+e^{-i\frac{\sqrt{3}}{2}k_{x}}\right)\right], \\
&\psi_{E,1}^{D_{3},1}(\vec{k})=\frac{1}{\sqrt{6}}\left[2e^{-ik_{y}}-e^{i\frac{1}{2}k_{y}}\left(e^{i\frac{\sqrt{3}}{2}k_{x}}+e^{-i\frac{\sqrt{3}}{2}k_{x}}\right)\right], \\
&\psi_{E,2}^{D_{3},1}(\vec{k})=\frac{1}{\sqrt{2}}\,e^{i\frac{1}{2}k_{y}}\left(e^{i\frac{\sqrt{3}}{2}k_{x}}-e^{-i\frac{\sqrt{3}}{2}k_{x}}\right), \\
&\psi_{A_{2}}^{D_{3},1}(\vec{k})=0
\end{aligned}
$$
Note that the set of nearest-neighbor vectors on the hon- eycomb lattice has lower symmetry ($D_{3}$) than the point group of the lattice itself ($D_{6}$). Also, $D_{3}$ is a subgroup of $D_{6}$, hence, the symmetry of the superconducting order parameters is still fully described by the $D_{6}$ group.
Comparing the possible basis functions of $D_{3}$ and $D_{6}$, we can observe the correspondence
$$
\begin{aligned}
&A_{1}^{D_{3}}\rightarrow A_{1}^{D_{6}},B_{1}^{D_{6}},\\
&A_{2}^{D_{3}}\rightarrow A_{2}^{D_{6}},B_{2}^{D_{6}},\\
&E^{D_{3}}\rightarrow E_{1}^{D_{6}},E_{2}^{D_{6}},
\end{aligned}
$$
i.e., we can expect the nearest-neighbor interaction on the honeycomb lattice to affect only the lowest order basis functions of the irreps $A_{1}$, $B_{1}$, $E_{1}$, and $E_{2}$, similar to the triangular lattice. However, we do not observe changes for $B_{1}$.
The main effect of including the nearest-neighbor in- teractions is the suppression of the chiral singlet states ($E_2$) in favor of the chiral triplet states ($E_{1}$), the $f$ -wave states ($B_{1}$ and $B_{2}$), and a small pocket of $i$ -wave ($A_{2}$) close to the van Hove singularity at $n\approx1.25$, as shown in the phase diagram for $0.1\leq\alpha\leq0.4$ in Fig.. However, the $i$ -wave state is almost degenerate to the chiral singlet state and thus susceptible to small perturbations. As for the case $\alpha=0$, the effective interaction $V_{\rm eff}$ is almost negligible for $n<1$ but particularly strong for large $n$ and finite $t_2/t_1$.
Again we could identify superconducting solutions for all irreps belonging to the $D_6$ symmetry group. As for the triangular lattice case, almost all superconducting groundstates are topologically nontrivial – either they constitute spin-triplet superconductivity or they realize a chiral topological superconductor (or both in case of $E_1$ irrep). The honeycomb lattice is thus as good as the triangular lattice to search for topological superconduc- tivity.
## Summary of results and discussion
For all paradigmatic 2D lattices studied in this section, we find that the effective interaction, $V_{\text{eff}}$, is especially high close to fillings with a van Hove singularity in the density of states, as expected. Also, we find all pairing
| |
180802589/2
|
only odd Chern numbers are of interest since pairs of real Majorana zero modes can recombine into complex Dirac fermions. One can show that odd parity, i.e., spin triplet, superconductors possess odd Chern numbers. That is the reason why there is so much interest in identifying the so- called $p+ip$ or $f+if$ superconducting states in a real ma- terial. More recently, a second notion of “ topology ” has been established: more than ten years after the discovery of topological insulators, there is also considerable inter- est in superconductors which are gapped in their bulk but carry gapless edge modes. Indeed, chiral supercon- ductors are the superconducting analogues of topological insulators (or, to be more precise, of Chern / quantum Hall insulators). Both notions of topology are frequently used in the literature. It is also worth emphasizing that the $p+ ip$ superconductor (with Chern number $C=1$) fits into both categories in contrast to the chiral $d+ id$ state ($C=2$). In this paper, we will distinguish between even and odd parity (or, alternatively, spin singlet and triplet) superconductors; but we will refer to all chiral states as topological superconductors, regardless what their parity or angular momentum is.
Here we study three paradigmatic two-dimensional (2D) lattices and investigate their unconventional super- conducting ground states caused by local and non-local Coulomb interactions. While there are many works in the literature [10 – 28], we are not aware of a systematic inves- tigation of the phase spaces involving both longer-ranged hoppings and longer-ranged repulsive interactions. We restrict ourselves to consider only one orbital per site. Most materials are described by multiple orbitals, even after “ downfolding ” them to an effective tight-binding model. Nonetheless, as a first guess, the one-orbital sce- nario might give a reasonable answer. Moreover, there is recent interest in certain adatom systems such as $X$ /Ge(111) and $X$ /Si(111) ($X$ =Sn,Pb) [29 – 35] with the surface atoms forming a triangular net; Sn/Si(111) has recently even been claimed to show superconductivity at low temperatures [36]. These systems are well-described by one-band Hubbard models.
In this paper, we use a weak coupling renormaliza- tion group (WCRG) approach, developed by Raghu and coworkers [18, 19, 22] built up on earlier work by Kohn and Luttinger [7], to study the various superconducting phases that may arise in paradigmatic 2D lattices from repulsive interaction between the electrons. The paper is organized as follows: in the next section, we give a thor- ough description of the WCRG method. Then we study the superconducting phase diagrams for the square, tri- angular, and honeycomb lattices as paradigms for 2D sys- tems. In the subsequent section, we will discuss how the numerical resolution influences the groundstates found within the WCRG method. In particular, we resolve some ambiguities reported in the litreature. This leads us to discuss how to efficiently reach a high numerical accuracy. The paper ends with a Conclusion.
## METHOD
We use the WCRG approach to calculate the super- conducting instabilities that arise in the Hubbard-model with repulsive interactions. The Hamiltonian of an arbi- trary 2D lattice without spin-orbit coupling is given by
$$
\begin{aligned}
H&=H_{0}+H_{\text{int}}, \\
H_{0}&=-\sum_{n=1}^{n_r}t_{n}\sum_{\langlei,j\rangle_{n}}\sum_{\sigma}c_{i\sigma}^{\dagger}c_{j\sigma}^{{\phantom{\dagger}}}\ ,\\
%-\mu\sum_{i,\sigma}c_{i\sigma}^{\dagger}c_{i\sigma}^{\pd}, \\
H_{\text{int}}&=\sum_{m=0}^{m_r}U_{m}\sum_{\langlei,j\rangle_{m}}\sum_{\sigma,\sigma'}c_{i,\sigma}^{\dagger}c_{j,\sigma'}^{\dagger}c_{j,\sigma'}^{{\phantom{\dagger}}}c_{i,\sigma}^{{\phantom{\dagger}}},
\end{aligned}
$$
where $t_{n}$ and $U_{n}$ denote the $n$ -th nearest-neighbor hop- ping and Coulomb interaction amplitudes, respectively, and $U_0$ the local Hubbard interaction. $c_{i,\sigma}^{\dagger}$ is the cre- ation operator of an electron on site $i$ with spin $\sigma$ and $\langlei,j\rangle_{n}$ are $n$ -th nearest neighbor sites. $n_r$ is the number of unit cells and $m_r$ the range of non-local interactions.
The bandstructure $E(n,\vec{k})$ is obtained by diagonalizing the Bloch matrix $h(\vec{k})$ of the non-interacting system $H_{0}$:
$$
\begin{aligned}
E(n,\vec{k})=&-\sum_{i,j}u^{*}_{i,n}(\vec{k})h_{ji}(\vec{k})u_{n,j}(\vec{k})\\
=&-\sum_{\nu=1}^{n_r}t_{\nu}\varepsilon_{n,\nu}(\vec{k}),
\end{aligned}
$$
where $u_{n,i}(\vec{k})$ is the $i$ -th element of the $n$ -th eigenvector of $h(\vec{k})$ and $n$ is the band-index. In the following, we will drop $n$ for the case of one-band models (square and triangular lattices). The amplitude of the $m$ -th neighbor interaction in momentum space, $V_{m}$, can be conveniently written as
$$
\begin{aligned}
V_{m}&(n_{1},\vec{k}_{1};n_{2},\vec{k}_{2};n_{3},\vec{k}_{2}-\vec{q};n_{4},\vec{k}_{1}+\vec{q})=U_{m}\sum_{i,j}h_{m,ij}(\vec{q})\\
&\times u_{n_{1},i}^{*}(\vec{k}_{1})u_{n_{2},j}^{*}(\vec{k}_{2})u_{n_{3},j}^{{\phantom{*}}}(\vec{k}_{2}-\vec{q})u_{n_{4},i}^{{\phantom{*}}}(\vec{k}_{1}+\vec{q}),
\end{aligned}
$$
where we used $h(\vec{k})=\sum_{m}t_{m}h_{m}(\vec{k})$. For the case of a single orbital per site $h_{0}(\vec{k})$ is the unit matrix and in the one-band case (6) simplifies to
$$
V_{m}(\vec{q})=U_{m}\varepsilon_{m}(\vec{q}).
$$
In this paper, we only consider onsite and nearest- neighbor interactions with amplitudes $U_{0}$ and $U_{1}$, respec- tively. For the sake of clarity, we will neglect $U_{1}$ for the time being and discuss its implication later.
We define an orbital factor $M$ as
$$
\begin{aligned}
M&(n_{1},\vec{k}_{1};n_{2},\vec{k}_{2};n_{3},\vec{k}_{3};n_{4},\vec{k}_{4})=\\
&=\sum_{i}u_{n_{1},i}^{*}(\vec{k}_{1})u_{n_{2},i}^{*}(\vec{k}_{2})u_{n_{3},i}^{{\phantom{*}}}(\vec{k}_{3})u_{n_{4},i}^{{\phantom{*}}}(\vec{k}_{4}).
\end{aligned}
$$
| |
180802589/16
|
FIG.13. Effects of low accuracy of the integration grid, $N_{\text{int}}$, and of the number of patching points, $N_{p}$, on the effective interaction $V_{\text{eff}}$. The irrep with the largest $V_{\text{eff}}$ is the leading superconducting instability. The figures shown are for the square lattice with $t_{2}=\alpha=0$ and $n\approx0.54$. (a) Effective interaction, $V_{\text{eff}}$, of the $B_{2}$ and $E$ irreps as a function of $N_{\text{int}}$. (b) $V_{\text{eff}}$ as a function of $N_{p}$.
square lattice, where the $d_{xy}$ ($B_{2}$) and $p_{x}+ip_{y}$ ($E$) sym- metries are almost degenerate. Figure 13 shows the im- pact of the accuracy of the integration grid and of the number of patching points on the effective interaction $V_{\text{eff}}$ of $B_{2}$ and $E$. For an integration grid with $40\times40$ grid points, the leading superconducting instability is the $B_{2}$ irrep, independent of $N_{p}$. This is true for up to $N_{\text{int}}\approx80$. However, for more dense integration grids, the $E$ repre- sentation becomes the leading instability. Thus, too low accuracy in the integration may clearly lead to wrong results for the symmetry of the superconducting state.
Thus we come back to the earlier discussion about the correct groundstate for the square lattice phase diagram. The discrepancy between our result on the one hand and the findings of Ref. [18] on the other hand (see section “ Results ”) might be understood as an issue of numeri- cal resolution as demonstrated in Fig. 13. Reference [18] finds the leading instability with $B_2$ symmetry, while we find it with $E$ symmetry. As explicitly shown in Fig. 13, increase of $N_p$ does not resolve this issue. Only a suffi- ciently large $N_{\rm int}$ renders the state with $E$ symmetry to be the groundstate.
However, keeping $N_{\text{int}}$ and $N_{p}$ moderately large, also considerably increases the runtime of the routine, which grows $\propto (N_{\text{int}}N_{p})^{2}$. Studying the integrand
$$
\tilde{X}_{ph}(\vec{k},\vec{p})=\sum_{n_{1},n_{2}}X_{\rm ph}(n_{1},\vec{k}+\vec{p};n_{2},\vec{p}),
$$
where $X_{\rm ph}$ is given by Eq. (12) and $\vec{p}$ is the integration variable, one finds that it consists mainly of areas with little change, narrow regions in which divergencies can be found and large areas where it is zero, all depending on the momentum $\vec{k}$. The integration is done in the $p_{x}$ - $p_{y}$ plane. Using a standard quadrature integration routine leaves us with the following dilemma: if we cover the di- vergencies and narrow areas with fast changing integrand with appropriate accuracy, we also add up unecessarily
FIG. 14. Progression of the dynamic grid after 1, 3 and 6 iterations for a starting grid size $N_{0}=10$. First row: square lattice with large $|\vec{k}|$, $t_{2}=0$ and $n\approx0.54$. Second row: square lattice with the same parameters and $|\vec{k}|$ close to zero. Third row: honeycomb lattice with large $|\vec{k}|$, $t_{2}=0$ and $n\approx1.4$. Fourth row: honeycomb lattice with the same system parameters and $|\vec{k}|$ close to zero.
FIG.15. Relative runtime, $T_{n}/T_{0}$, as a function of the number of iterations, $n$, while leaving the effective gridsize constant at $N_{\text{eff}}=1280$. $T_{n}$ denotes the runtime when using $n$ iterations to calculate $g$ from Eq. (41) for a fixed set of parameters in each curve.
many zeros and integrate the slow changing regions of the integrand with much higher accuracy than needed. If we take too low accuracy, we just cut out the regions around the divergencies, where the integrand is very large.
Thus, we have developed a dynamical grid method, a 2D analog the tetrahedron method [47 – 49], which puts a more dense integration grid around the divergencies (more generally, around regions where the gradient of the integrand is large), and ignores areas where the in- tegrand is zero anyway. In this routine, we start with a low-density grid of polygons appropriate for the Brillouin
| |
180802589/6
|
in $U_{0}$, depends on the initial cut-off, $\epsilon_{0}$, such that it can- cels out [18], i.e.,
$$
T_{c}\sim e^{-1/\lambda_{\text{min}}},
$$
which yields for the effective interaction
$$
V_{\text{eff}}=\frac{\lambda_{\text{min}}}{\rho}.
$$
That only negative eigenvalues are relevant in the RG flow reflects the general idea that only an attractive ef- fective interaction between electrons can lead to super- conductivity. However, we see that this may arise under renormalization, starting from a repulsive onsite interac- tion in a many-body system in the same way a matrix with only positive entries can have negative eigenvalues.
The eigenvector $\psi_{\text{min}}$ which corresponds to $\lambda_{\text{min}}$ de- scribes the discrete form-factor of the gap function of the resulting superconducting instability along the Fermi sur- face. Superconducting states can be fully characterized by the irreps of the point group associated with the lat- tice structure [8]. That is, $\psi_{\text{min}}$ must transform according to one of the irreps of the relevant point group (see also discussion in the next section).
If there is no dependence on spin other than the selec- tion of the diagrams due to the Pauli exclusion principle (i.e., there is no spin-orbit coupling) and only Hubbard interactions within the same orbital are considered, we identify via Eqs. (21) and (22) the following relation:
$$
\Gamma^{(2b)}(2,1)=-\Gamma^{(2c)}(\bar{2},1)\ .
$$
That is, all triplet states are already contained in the singlet channel. Symmetrization (anti-symmetrization) according to (17) removes the triplet (singlet) states.
The reader should note that the above discussion about what diagrams contribute to what channel is valid for the single-band case discussed in this paper. Multi-orbital scenarios naturally involve additional inter-orbital inter- actions as well as Hund’s coupling. Then the above dis- cussion needs to be extended, e.g. the diagram 2c from Fig. 1 contributes to both singlet and triplet channels. Moreover, new additional diagrams must be included [22].
## Limit $k_{4}\rightarrow k_{3}$
Special attention has to be taken when handling the integral of $\Gamma$, Eqs. (21) and (22), in the limit of small momentum transfer, $k_{4}\rightarrow k_{3}$, and for $T\rightarrow0$, in which the integrand function of the particle-hole susceptibility, $X_{\rm ph}$, yields for equal band indices, $n_{3}=n_{4}$:
$$
\begin{aligned}
% \lim_{T\rightarrow0}X(3,4)&=\frac{\theta(E(4))-\theta(E(3))}{E(4)-E(3)}, \\
\lim_{k_{4}\rightarrow k_{3}}\lim_{T\rightarrow0}X_{\rm ph}(3,4)&= \lim_{k_{4}\rightarrow k_{3}}\frac{\theta(E(4))-\theta(E(3))}{E(4)-E(3)}
=\delta(E)
\end{aligned}
$$
where $\delta(E)$ denotes the Dirac delta-distribution. Con- sequently, Eq. (22) becomes
$$
\begin{aligned}
\lim_{k_{4}\rightarrow k_{3}}\lim_{T\rightarrow0}\Gamma_{n_{3}=n_{4}}^{(2b)}(2,1)=
\int_{\hat{3}}M(2,\hat{3},\bar{1},\hat{3})M(\hat{3},\bar{2},\hat{3},1).
\end{aligned}
$$
The importance of this limit lies in the correct numer- ical handling of the integrand approaching the delta- distribution. In contrast, an incorrect handling might possibly result in
$$
\begin{aligned}
&\lim_{k_{4}\rightarrow k_{3}}\lim_{T\rightarrow0}\Gamma_{n_{3}=n_{4}}^{(2b)}(2,1)=0,
\end{aligned}
$$
producing different (incorrect) effective couplings $V_{\text{eff}}$.
## Longer range interaction
We now consider also nonlocal interactions; in particu- lar, we include a nearest neighbor Coulomb repulsion to the Hamiltonian,
$$
\begin{aligned}
H_{\text{int}}&=U_{0}\sum_{i}\sum_{\sigma\neq\sigma'}c_{i\sigma}^{\dagger}c_{i\sigma'}^{\dagger}c_{i\sigma'}^{{\phantom{\dagger}}}c_{i\sigma}^{{\phantom{\dagger}}} \\
&+U_{1}\sum_{\langlei,j\rangle}\sum_{\sigma,\sigma'}c_{i\sigma}^{\dagger}c_{j\sigma'}^{\dagger}c_{j\sigma'}^{{\phantom{\dagger}}}c_{i\sigma}^{{\phantom{\dagger}}}.
\end{aligned}
$$
The changes that arise in the vertex function are given by replacing the on-site interaction $U_{0}$ with the full interac- tion $U$, and by possibly adding new diagrams. The latter comes from the fact that non-local interactions may also connect particle lines with equal spin in the Feynman diagrams.
Since we only consider terms up to order $U_{0}^{2}$, we take $U_{1}\propto U_{0}^{2}$ for simplicity. As a consequence, the longer range interactions only appear in $\Gamma^{(1)}$ which is linear in $U$. To be precise, this choice implies that in the limit $U_{0}\rightarrow0$, we change $U_{1}$ such that
$$
\alpha = \frac{U_1 W}{U_0^2}
$$
is kept constant. Here, $W$ is the bandwidth of the non- interacting system $H_{0}$. Thus no new diagrams have to be considered. For single-band systems, we have $M=1$. Thus, the relevant terms of the vertex function are then given by
$$
\Gamma^{s}(2,1)\approx U_{0}^{2}\left(\Gamma^{(2b)}(2,1)+\alpha\frac{\varepsilon_{1}(\vec{k}_{2}-\vec{k}_{1})}{W}\right)
$$
for the singlet channel. Due to the momentum depen- dence [see Eqs. (6) and (7)], longer ranged interactions are handled slightly differently in these vertex functions.
Multiple bands arise only due to spin-orbit coupling (not considered in this paper) or multiple orbitals. The
| |
180802589/4
|
FIG. 2. General Feynman diagram that produces an addi- tional logarithmic divergence, independent of the behavior of $\Gamma^{(n)}$ and $\Gamma^{(m)}$.
advantage that we are able to expand the vertex function in powers of the on-site interaction, $U_{0}$, which yields
$$
\Gamma(2,1)=\sum_{n=1}^{\infty}U_{0}^{n}\Gamma^{(n)}(2,1).
$$
Here, $n$ denotes the number of interaction lines (dashed) in the corresponding Feynman diagrams of $\Gamma^{(n)}$, shown in Fig. 1, which are given by
$$
\begin{aligned}
\Gamma^{(1)}(2,1)&=M(2,\bar{2},\bar{1},1), \\
\Gamma^{(2a)}(2,1)&=-\int_{3}X_{\rm pp}(3)M(2,\bar{2},\bar{3},3)M(3,\bar{3},\bar{1},1),\\
\Gamma^{(2b)}(2,1)&=-\int_{3'}X_{\rm ph}(3,4)M(2,3,\bar{1},4)M(4,\bar{2},3,1)
\\
\Gamma^{(2c)}(2,1)&=\int_{3'}X_{\rm ph}(3,4)M(2,4,3,1)M(3,\bar{2},\bar{1},4)
\end{aligned}
$$
where we introduced the short notations for the inte- grals:
$$
\begin{aligned}
\int_{3}\equiv\sum_{n_{3}}\int\frac{\text{d}k_{3}^{2}}{(2\pi)^{2}},\hspace{2mm} \int_{3'}\equiv\sum_{n_{4}}\int_{3}.
\end{aligned}
$$
Momentum conservation yields $\vec{k}_{4}=\vec{k}_{1}+\vec{k}_{2}+\vec{k}_{3}$ for $\Gamma^{(2b)}$ and $\vec{k}_{4}=\vec{k}_{1}-\vec{k}_{2}+\vec{k}_{3}$ for $\Gamma^{(2c)}$. Note that since we only considered on-site interaction so far, the interaction lines in the diagrams can only connect electron lines of opposite spin.
It is important to note that all diagrams containing $X_{\rm pp}(n,\vec k)$ in their integrand show a logarithmic diver- gence. As discussed before, this is directly connected to the respective Feynman diagram being of the form as shown in Fig. 2,
which can formally be written as
$$
\begin{aligned}
\Gamma(2,1) &= \int_{3}\Gamma^{(m)}(2,3)X_{\rm pp}(3)\Gamma^{(n)}(3,1)\\
&=: \, \Gamma^{(m)}(2,3)*\Gamma^{(n)}(3,1).
\end{aligned}
$$
However, if a vertex function cannot be written as the convolution of two vertex functions with $X_{\rm pp}$, then it must be divergence-free. As examples, let us consider the diagrams shown in Fig. 1: diagram 1 is not of the form shown in Fig. 2 and thus divergence-free. Diagram 2a, on the other side, can be obtained by gluing to- gether two $\Gamma^{(1)}$ diagrams and writing according to (24), $\Gamma^{(2a)} = \Gamma^{(1)}*\Gamma^{(1)}$. It produces a logarithmic divergence. Diagrams 2b and 2c are not of the form $\Gamma^{(m)}*\Gamma^{(n)}$ and also divergence-free. In the following, we use the con- vention that we add a tilde on top of $\Gamma$ if the vertex function is divergence-free. For instance, $\Gamma^{(1)}\equiv \tilde\Gamma^{(1)}$, $\Gamma^{(2b)}\equiv \tilde\Gamma^{(2b)}$, $\Gamma^{(2c)}\equiv \tilde\Gamma^{(2c)}$ etc. From this discussion we learn that we can build the full vertex function $\Gamma(2,1)$ just from non-divergent parts, $\tilde{\Gamma}(2,1)$, by a convolution with $X_{\rm pp}$, which yields
$$
\begin{aligned}
\Gamma=&\sum_{n=1}^{\infty}U_{0}^{n}\tilde{\Gamma}^{(n)}+\sum_{m=1}^{\infty}U_{0}^{m}\tilde{\Gamma}^{(m)}*\sum_{n=1}^{\infty}U_{0}^{n}\tilde{\Gamma}^{(n)}+\dots\\
=&\,\tilde{\Gamma}+\tilde{\Gamma}*\tilde{\Gamma}+\tilde{\Gamma}*\tilde{\Gamma}*\tilde{\Gamma}+\dots=\tilde{\Gamma}+\tilde{\Gamma}*\Gamma.
\end{aligned}
$$
From here, we can obtain the RG flow equation by eval- uating the integral in Eq. (24):
$$
\begin{aligned}
&\int_{3}\tilde{\Gamma}^{(m)}(2,3)X_{\rm pp}(3)\tilde{\Gamma}^{(n)}(3,1)\\
&=\int_{\hat{3}}\int_{\xi_{n_{3}}^{-}}^{\xi_{n_{3}}^{+}}\text{d}\xi_{n_{3}}\frac{\rho(\xi_{n_{3}})}{\rho_{n_{3}}}\frac{1-2f(\xi_{n_{3}})}{-2\xi_{n_{e}}}\tilde{\Gamma}^{(m)}(2,3)\tilde{\Gamma}^{(n)}(3,1),
\end{aligned}
$$
where $\xi_{n_{i}}^{-(+)}$ denote the bottom (top) of the $n_{i}$ -th band with respect to the Fermi level and $\rho_{n_{i}}$ the density of states (DOS) of band $n_{i}$ at the Fermi level. The short notation for the integral is defined as
$$
\int_{\hat{3}}\equiv\sum_{n_{3}}\rho_{n_{3}}\int\frac{\text{d}\hat{k}_{3}}{S_{n_{3},F}}\frac{\bar{v}_{n_{3},F}}{v_{F}(\hat{3})}.
$$
Here, $v_{F}(\hat{i})$ denotes the Fermi velocity of band $n_{i}$ at mo- mentum $\hat{k}_{i}$, $S_{n_{i},F}$ is the total length of the Fermi surface of band $n_{i}$ and
$$
\frac{1}{\bar{v}_{n_{i},F}}=\int\frac{d\hat{k}_{i}}{S_{n_{i},F}}\frac{1}{v_{F}(\hat{i})}.
$$
In the limit of zero temperature the Fermi distribution becomes the Heaviside step-function, $\theta(\xi)$, i.e.,
$$
\lim_{T\rightarrow0}f(\xi)=1-\theta(\xi)=\begin{cases}
1, & \text{if}\xi<0,\\
0, & \text{otherwise}.
\end{cases}
$$
Substituting this limit in Eq. (28) lets us split the energy integral into parts over positive and negative energies, and thus introduce a small cut-off $\epsilon_{0}>0$:
| |
180802589/15
|
Phase diagram showing the irrep of the leading superconducting instability in the honeycomb lattice for $0.1\leq\alpha\leq0.4$, $0.5\leq n\leq1.7$ and $0\leq t_{2}/t_{1}\leq0.4$.
symmetries that are possible within the symmetry group of the respective lattices by varying the nearest-neighbor hopping, $t_{2}$, and nearest-neighbor interaction, $U_{1}$ (quan- tified through $\alpha$). All the phase diagrams we find have large domains of chiral superconducting states: chiral triplet states such as $p_x \pm i p_y$, but the hexagonal lattices can also host chiral singlet states such as $d_{x^2-y^2} \pm i d_{xy}$.
Concerning the magnitude of $T_{c}$ ($\sim V_{\text{eff}}$), our results suggest that chiral superconducting phases with reason- ably high $T_{c}$ may be found in all the studied lattice sys- tems, as shown in Figs. 4, 8, and 11. The honeycomb lattice is the most promising candidate to search for high temperature topological superconductivity.
Our results presented above are only exact in the limit $U\rightarrow0$. Note, however, that comparison with other methods (see for instance Refs. 23, 27, 28, and 42) that yield promising results even for stronger interactions shows that the superconducting instabilities found within the WCRG approach often coincides with those found for stronger interactions.
Quite generally, there is no reason to assume that a weak-coupling instability coincides with a strong- coupling instability. In principle, there might be one or even several phase transitions under increasing cou- pling strength. When we have a closer look, however, at some of the most important material classes, namely cuprates, pnictides, ruthenates but also the honeycomb- lattice Hubbard model systems, we find a remarkable cor- respondence between weak-coupling and strong-coupling instabilities. Let us first consider the cuprates: in the weak-coupling regime, the dominating instability is lo- cated at momentum $(\pi,\pi)$ with superconducting $d_{x^2-y^2}$ symmetry; in the strong-coupling regime, i.e., for the cor- responding spin model, again the dominating instability is located at $(\pi,\pi)$ leading to antiferromagnetic Neel or- der (for an extensive review see Ref. 43). The second prominent example are the pnictides, where the dom- inating superconducting fluctuations are at $(\pi,0)$ and $(0,\pi)$ [44, 45]; the strong-coupling analysis leads to a spin Hamiltonian featuring columnar antiferromagnetic order with magnetic Bragg peaks located at $(\pi,0)$ and $(0,\pi)$ [46]. A similar reasoning is valid for the ruthen- ates and a class of honeycomb-lattice Hubbard models, showing that this line of argument is not exclusive for the square lattice. While there is no exact method at intermediate coupling strength, the above-mentioned ex- amples are so extensively studied that it is widely ac- cepted that there are no (superconducting) intermedi- ate phases and phase transitions from the superconduct- ing weak-coupling phase into another superconducting intermediate-coupling phase do not occur. Thus we can conclude that, at least empirically, the weak-coupling RG method finds superconducting solutions which are not only exact in the limit $U\rightarrow 0$ but which also provide often a guiding principle for stronger-correlated regimes or materials.
## NUMERICAL DEVELOPMENT AND PERFORMANCE ANALYSIS
At some regions in the $t_{2}/t_1$ – $n$ phase space, the second lowest eigenvalue of the scattering amplitude matrix $g$ is very close to the lowest one, see panels c) in Figs. 4, 8, and 11. These regions are particularly prone to the smallest variations in the parameters or perturbations, including variations in the numerical resolution. In other words, it is more challenging to obtain the true groundstate.
In this section, we demonstrate the importance of the accuracy in terms of the number of patching points, $N_{p}$, used to discretize the eigenvalue equation (41), and the number of points for the integration grid, $N_{\text{int}}$, which refers to the numerical evaluation of Eq. (21), i.e., the number of discretization points in each momentum di- mension of the integral in Eq. (23). As an example, we take the regime $0.5\leq n \leq 0.6$ and $t_{2}=\alpha=0$ on the
| |
180802589/10
|
where the sum runs over all nearest-neighbor lattice har- monics, $i\in\{A_{1},A_{2},B_{1},B_{2},E\}$. On the square lat- tice, these correspond to the following basis functions:
$$
\begin{aligned}
&\psi_{A_{1}}^{D_{4},1}(\vec{k})=\cos(k_{x})+\cos(k_{y}), \\
&\psi_{B_{1}}^{D_{4},1}(\vec{k})=\cos(k_{x})-\cos(k_{y}), \\
&\psi_{E,1}^{D_{4},1}(\vec{k})=\sqrt{2}\sin(k_{x}),\hspace{2mm} \psi_{E,2}^{D_{4},1}(\vec{k})=\sqrt{2}\sin(k_{y}), \\
&\psi_{A_{2}}^{D_{4},1}(\vec{k})=\psi_{B_{2}}^{D_{4},1}(\vec{k})=0,
\end{aligned}
$$
where $\psi_{I}^{G,n}(\vec{k})$ denotes the $n$ -th lattice harmonic of ir- rep $I$ in group $G$.
Thus, the nearest-neighbor interaction only affects the irreps with nonzero nearest-neighbor lat- tice harmonics, since the set of all basis functions spans an orthogonal space. It is important to stress that not all superconducting ground states that transform under a given representation are affected in the same way. To il- lustrate this further, we show in Fig. 5 an example where the leading instability is affected by $U_{1}$ (parameters used are $t_{2}/t_{1}=-0.35$ and $n\approx0.853$). Here, for onsite inter- actions $U_{0}$ ($\alpha=0$), the form factor has $d_{x^{2}-y^{2}}$ -wave symmetry ($B_{1}$). Adding nearest-neighbor interaction $U_{1}$ (i.e., $\alpha >0$), the effective interaction of this form fac- tor becomes smaller the larger $\alpha$ gets. The leading in- stability can change its symmetry, when the eigenvalue corresponding to the $d_{x^{2}-y^{2}}$ -wave becomes larger than the next lowest eigenvalue of $g$. In the case shown here, the form factor of the second eigenvalue is also of the $B_{1}$ irrep, but it forms a linear combination of the $d$ - and $i$ -waves. Further increasing $\alpha$ will have no effect on the leading instability anymore. In other words, within the region with $B_1$ symmetry, regions with pure $d$ -wave sym- metry are suppressed by finite $\alpha$ while regions with $i$ - wave symmetry are not.
We have mapped out the full three-dimensional $t_2/t_1$ versus $n$ versus $\alpha$ phase diagram. As discussed in de- tail before, only regions with sufficiently large $V_{\rm eff}$ are of interest. For the sake of clarity, we limited the phase di- agram to a relevant range which is $0.5\leq n\leq1.45$ and $0.05\leq\alpha\leq0.15$, see Fig.. The main changes due to finite $\alpha$ is the appearance of a pocket with $g$ -wave sym- metry (irrep $A_2$) around $n\approx 1.2$ which quickly develops into a large phase slightly above half filling. The chiral superconducting phase (irrep $E$) at lower filling essen- tially persists; in addition, the narrow stripe of $E$ sym- metry emerges between the two $d$ -wave phases. There are also several small pockets or stripes with $B_1$ and $B_2$ symmetry. We have not shown plots for $V_{\rm eff}$ and $\delta_\lambda$ for the only reason that these plots are very similar to the ones shown for $\alpha=0$.
We will conclude this section with a discussion of the results and comparison to the literature. In Ref. 18 (see Fig. 2), the coupling strengths $V_{\rm eff}$ for irreps $B_1$, $B_2$, and $E$ are shown for $\alpha=0$. In Ref. 19 (see Fig. 2), the $\alpha$ – $n$ phase diagram for $t_2=0$ is shown. In contrast to our
FIG. 6. Phase diagram showing the irrep of the leading su- perconducting instability in the square lattice for $0.05\leq\alpha\leq0.15$, $0.4\leq n\leq1.45$ and $-0.7\leq t_{2}/t_{1}\leq0$.
results, in these works the $E$ phase is absent, mostly $d_{xy}$ - wave is found instead (irrep $B_2$). We further note that a large range of our $t_2/t_1$ – $n$ phase diagram Fig. 4 was dis- cussed within a $T$ matrix approach in Ref. 13 (see Fig. 6); remarkably, it coincides with our findings (and disagrees with Ref. [18, 19] regarding the phase with $E$ symmetry, as emphasized in Ref. 19). We believe that the sensitivity of the integration grid might be a candidate to explain this discrepancy; we discuss this in detail at the end of the paper in the Section about “ Numerical development and performance analysis ” and explicitly show that there is a phase transition from $B_2$ to $E$ depending on the res- olution of the integration grid.
| |
180802589/13
|
ten as
$$
\varepsilon_1(\vec{k}_{2}-\vec{k}_{1})=\sum_{i}\psi_{i}^{D_{6},1}(\vec{k}_{1})\psi_{i}^{D_{6},1}(\vec{k}_{2}),
$$
where the sum runs over all nearest neighbor lattice harmonics, $i\in\{A_{1},A_{2},B_{1},B_{2},E_{1},E_{2}\}$. On the tri- angular lattice, these correspond to the following basis functions:
$$
\begin{aligned}
&\psi_{A_{1}}^{D_{6},1}(\vec{k})=\sqrt{\frac{2}{3}}\left[\cos(k_{x})+2\cos\left(\frac{1}{2}k_{x}\right)\cos\left(\frac{\sqrt{3}}{2}k_{y}\right)\right], \\
&\psi_{B_{1}}^{D_{6},1}(\vec{k})=\sqrt{\frac{2}{3}}\left[\sin\left(k_{x}\right)-2\sin\left(\frac{1}{2}k_{x} \right)\cos\left(\frac{\sqrt{3}}{2}k_{y} \right)\right], \\
&\psi_{E_{1},1}^{D_{6},1}(\vec{k})=\frac{2}{\sqrt{3}}\left[\sin(k_{x})+\sin\left(\frac{1}{2}k_{x}\right)\cos\left(\frac{\sqrt{3}}{2}k_{y}\right)\right], \\
&\psi_{E_{1},2}^{D_{6},1}(\vec{k})=2\cos\left(\frac{1}{2}k_{x}\right)\sin\left(\frac{\sqrt{3}}{2}k_{y}\right), \\
&\psi_{E_{2},1}^{D_{6},1}(\vec{k})=\frac{2}{\sqrt{3}}\left[\cos(k_{x})-\cos\left(\frac{1}{2}k_{x}\right)\cos\left(\frac{\sqrt{3}}{2}k_{y}\right)\right], \\
&\psi_{E_{2},2}^{D_{6},1}(\vec{k})=2\sin\left(\frac{1}{2}k_{x}\right)\sin\left(\frac{\sqrt{3}}{2}k_{y}\right), \\
&\psi_{A_{2}}^{D_{6},1}(\vec{k})=\psi_{B_{2}}^{D_{6},1}(\vec{k})=0.
\end{aligned}
$$
The phase diagram for $0.05\leq\alpha\leq0.2$, $-0.3\leq t_{2}/t_{1}\leq0.3$, and $0.7\leq n\leq1.8$ is shown in Fig.. Without nearest-neighbor interactions, the leading superconduct- ing instabilities are mostly of triplet type ($B_{1}$) or chi- ral topological states ($E_{1}$ and $E_{2}$). The biggest change happens for small $\alpha$: especially the chiral singlet states (irrep $E_2$) for $n<1$ are suppressed, whereas new do- mains of chiral triplet states appear (irrep $E_1$) and the pocket with extended $s$ -wave ($A_1$) becomes large. We note, however, that there are no dramatic changes in the range $0.1<\alpha<0.2$ in the phase diagram. In the regime of moderate negative $t_2$ changes due to $\alpha$ are almost ab- sent; in contrast, for positive moderate $t_2$ several small pockets with $f$ -wave, $i$ -wave or $E_{1/2}$ symmetry are in- duced. In addition, we observe that the $i$ -wave phase for positive $t_2$ around $n\approx 1.4$ becomes stabilized and in- creased towards smaller $t_2$. At $\alpha=0.2$, the $i$ -wave phase is even present at $t_2=0$.
As for the square lattice, we could identify supercon- ducting solutions for all irreps belonging to the $D_6$ sym- metry group. It is important to stress that essentially all unconventional superconducting phases on the triangular lattice are either of chiral topological type or spin-triplet states, which are also topologically nontrivial as discussed in the introduction. While one could have guessed this already from the list of irreps, we find that the chance to obtain a topologically non-trivial state on the triangular lattice is extremely high.
## Honeycomb lattice
The honeycomb lattice is the third paradigmatic lat- tice we are studying. Like the square lattice it is bi- partite. Like the triangular lattice it is hexagonal. Un- like both square and triangular lattices, it features a two-atomic unit cell yielding two bands. Both bands touch each other at special points, the so-called Dirac points, located at the corners of the Brillouin zone. In the vicinity of these points, the dispersion relation cor- responds to the spectrum of a Dirac Hamiltonian. The tight-binding band structure $E(\vec{k})$ involving nearest and second-nearest-neighbor hoppings ($n_r=2$) is given by Eq. (5) with
$$
\begin{aligned}
&\varepsilon_{1}(\vec{k})=\pm\sqrt{3+2\cos(\sqrt{3}k_{x})+4\cos\bigg(\frac{\sqrt{3}}{2}k_{x}\bigg)\cos\left(\frac{3}{2}k_{y}\right)}, \\
&\varepsilon_{2}(\vec{k})=2\cos(\sqrt{3}k_{x})+4\cos\bigg(\frac{\sqrt{3}}{2}k_{x}\bigg)\cos\bigg(\frac{3}{2}k_{y}\bigg)\ .
\end{aligned}
$$
The upper (lower) band corresponds to the plus (minus) sign in Eq. (68). Examples of Fermi surfaces are similar to those shown in the right column of Fig. 8. In contrast to the other lattices, now we deal with two orbitals per unit cell and $M(1,2,3,4)$ appearing in the vertex functions. The symmetry group of the honeycomb lattice still is $D_{6}$ and the discussion of the triangular lattice applies. Thus we can look up the corresponding basis functions in Fig. 7.
The superconducting $t_2/t_1$ – $n$ phase diagram, Fig. 11, is somewhat different from the previous cases: now we have two bands and for $|t_{2}/t_{1}|<1/3$ the half-filled case, $n=1$, corresponds to a completely filled lower band. For $t_{2}=0$, the honeycomb lattice for $0<n<1$ is some- what similar to the triangular lattice for $0<n<2$. In the honeycomb lattice, we find that due to the noncon- stant orbital factors, $M(1,2,3,4)$, superconductivity for $n\lesssim 0.6$ is suppressed (i.e., $V_{\rm eff}$ is very small), and chiral topological singlet states (irrep $E_{2}$) become the leading instability close to the Dirac filling, $n=1$. As for the triangular lattice, we find that mostly spin-triplet phases ($f$ -wave with irreps $B_1$ or $B_2$) or chiral topological phases ($p+ip$ or $d+id$ with irreps $E_1$ or $E_2$, respectively) are present. In particular, the region around van Hove filling is dominated by chiral $d$ - and $p$ -wave superconductivity. For large $t_2/t_1$, also a pocket with extended $s$ -wave sym- metry (irrep $A_1$) shows up. In agreement with van Hove fillings, the effective coupling strength is much larger for fillings $n>1$ and finite $t_2/t_{1}$. Especially the chiral $p$ - wave state for large $n$ and $t_{2}$ shows an extraordinarily large effective coupling strength.
Eventually we consider the effect of longer-ranged in- teractions, $\alpha>0$. The form factor of Eq. (54) can be
| |
180802589/1
|
# Unconventional superconductivity in the extended Hubbard model: Weak-coupling renormalization group
Sebastian Wolf, 1
Thomas L. Schmidt, 2 and
Stephan Rachel 1
We employ the weak-coupling renormalization group approach to study unconventional supercon- ducting phases emerging in the extended, repulsive Hubbard model on paradigmatic two-dimensional lattices. Repulsive interactions usually lead to higher-angular momentum Cooper pairing. By con- sidering not only longer-ranged hoppings, but also non-local electron-electron interactions, we are able to find superconducting solutions for all irreducible representations on the square and hexag- onal lattices, including extended regions of chiral topological superconductivity. For the square, triangular and honeycomb lattices, we provide detailed superconducting phase diagrams as well as the coupling strengths which quantify the corresponding critical temperatures depending on the bandstructure parameters, band filling, and interaction parameters. We discuss the sensitivity of the method with respect to the numerical resolution of the integration grid and the patching scheme. Eventually we show how to efficiently reach a high numerical accuracy.
## INTRODUCTION
Superconductivity is amongst the oldest known quan- tum liquids on earth. More than 100 years after its discovery it has maintained its fascination to experi- mental and theoretical researchers. Most elemental su- perconductors such as Pb and Hg are well explained within the framework of the celebrated Bardeen – Cooper – Schrieffer (BCS) theory [1] where the pairing of elec- trons into Cooper pairs is due to electron-phonon inter- action. There are, however, several materials that be- have differently. In particular, for the important ma- terial classes of copper-oxide [2] and iron-based [3] high- temperature superconductors, phonon-mediated pairing is very unlikely. Also topological superconductors, super- conducting analogs of quantum Hall-type systems, seem to have an unconventional pairing mechanism [4 – 6]. In- stead, strong electron correlations which arise from re- pulsive interactions are believed to be crucial for the for- mation of Cooper pairs in these systems.
An attractive effective interaction which is needed for the formation of Cooper pairs can be caused by a repul- sive interaction, as first pointed out by Kohn and Lut- tinger [7]. Usually, an attractive interaction between elec- trons yields a superconducting state exhibiting $s$ -wave symmetry (i.e., angular momentum $\ell=0$), which pro- vides a full superconducting gap and nodes are absent. Contrary to that, superconducting states stemming from repulsive interactions typically prefer higher angular mo- mentum pairing ($\ell > 0$) symmetries. These are labeled, for instance, as $p$ -, $d$ -, $f$ -wave symmetric corresponding to $\ell=1,2, 3$, respectively. Higher angular momentum pairing causes sign changes in the superconducting gap leading to nodes [7], which are absent for $s$ -wave super- conductors. Quite generally, superconducting states can be fully classified by virtue of the irreducible represen- tations (irreps) of the symmetry group of the underly- ing lattice [8]. That is, the superconducting gap func- tion must transform according to one of the irreps; and the basis functions of the irreps correspond to $s$ -, $p$ -, $d$ - wave or higher $\ell$ symmetry. These symmetries refer to the “ orbital ” wavefunction in real or momentum space. The total Cooper pair wavefunction, which is always an- tisymmetric under exchange of electrons, is a product of this orbital and the spin wavefunctions. Since two elec- trons are involved, only spin singlet (totally antisymmet- ric) or spin triplet pairing (totally symmetric) is possible. Correspondingly, the orbital wavefunction must be either symmetric ($\ell=0, 2, 4, \ldots$) with even parity or antisym- metric ($\ell = 1, 3, 5, \ldots$) with odd parity.
Nodes in the superconducting gapfunction correspond to zeros in the condensation energy, i.e., they reduce it. Intuitively, one can understand this as the mentioned en- ergy gain is proportional to the number of Cooper pairs or the average size of the gap. An opportunity to get rid of the nodes naturally presents itself when the supercon- ducting groundstate is degenerate. For the scenario of degenerate gapfunctions, say, $\psi_{1}$ and $\psi_{2}$, one can avoid nodes by forming complex superpositions, $\psi_{1}\pm i\psi_{2}$. The resulting superconducting gap and the condensation en- ergy are node-free. However, the system has to choose one of the chiralities, either $\psi_{1}+ i\psi_{2}$ or $\psi_{1}- i\psi_{2}$, and by doing so break spontaneously time-reversal symmetry. The resulting time-reversal broken superconductor is chi- ral [4]: it features chiral edge modes circulating around the sample edge and it is characterized by a topological invariant, the first Chern number also known as TKNN invariant [9] (which reveals the mentioned connection to the quantum Hall effect). This exotic state is referred to as topological superconductivity.
Traditionally, a superconductor is considered to be “ topological ” when it has the ability to trap Majorana zero modes at its vortex cores. For instance, part of the activities to understand the superfluid phases of He-3 are motivated by this idea. It turns out that the number of Majorana zero modes is given by the Chern number; thus
| |
180802589/9
|
FIG.4. a) Phase diagram $t_{2}/t_{1}$ vs. band filling $n$ of the square lattice. The van Hove filling $n_{\text{vH}}$ is shown as a red dotted line. Note that there is also perfect nesting for $t_{2}/t_{1}=0$ at half filling. Due to the inherent particle-hole symmetry, it is sufficient to consider only negative values for $t_{2}/t_{1}$. b) Effective interaction, $V_{\text{eff}}$, as a function of the band filling $n$ and the second neighbor hopping $t_{2}/t_{1}$. c) Difference $\delta_\lambda$ of the two lowest eigenvalues corresonding to the superconducting ground state and the solution which is closest to it. Large differences indicate stable phases; the smaller the difference, the more fragile the superconducting groundstate.
Hove singularity appears as a red dotted line; it is given by
$$
n_{\text{vH}}=\int_{\xi^{-}}^{\mu_{\text{vH}}}\rho(E)\text{d}E,\hspace{2mm}\mu_{\text{vH}}=\begin{cases}
4t_{2}, &\text{if}\,|t_{2}|<0.5,\\
t_{1}^{2}/t_{2}, &\text{if}\,|t_{2}|\geq0.5.
\end{cases}
$$
Note that for $|t_{2}|\geq0.5$, also (imperfect) nesting oc- curs at the van Hove fillings, i.e., Fermi surfaces which, consists of points corresponding to other points on the Fermi surface under translation by a reciprocal nest- ing vector $\vec Q_{n,1}=(0,2\arccos[-t_{1}/2t_{2}])$ and $\vec Q_{n,2}=(2\arccos[-t_{1}/2t_{2}],0)$. The nesting affects $75\,\%$ of the Fermi surface, whereas perfect nesting occurs only in the limit $t_{2}/t_{1}\rightarrow\infty$.
As discussed before, within the WCRG approach we always find superconducting solutions. Thus it is im- portant to know which of these are particularly sta-
FIG. 5. (a) Effective interaction, $V_{\text{eff}}$, of the most negative eigenvalue of the irrep $B_{1}$ as a function of the filling $n$ for different nearest-neighbor interaction strengths $\alpha$. The filling $n\approx0.853$, at which the other figures are shown, is indicated as a red, dashed line. (b)-(d) Plot of the form factor corre- sponding to the eigenvalues plotted in (a). (e)-(g) Form factor of the pure $d$ - and $i$ -wave for (e) and (f), respectively, and of the linear combination of the two in (g). (h)-(j) Plot of the form factors (e)-(g) now shown on the Fermi surface.
ble and occur at sufficiently high temperature to be of experimental interest. The effective coupling strength $V_{\rm eff}$ can be directly related to the critical temperature, $T_c \sim \exp{[-1/(\rho V_{\rm eff})]}$. In Fig. b) we show $V_{\rm eff}$ for the same parameters as the phase diagram above. Of inter- est are the regions which correspond to bright colors. In particular, the white, yellow, and orange spots indicate high critical temperatures. As expected, they are con- centrated at van Hove fillings. Note that also the lighter blue regions are interesting and experimentally relevant. Amongst the ground states with promisingly large $V_{\rm eff}$ are those with irreps $B_1$ and $E$, the latter realizing topo- logical superconductivity.
The WCRG approach allows to test the stability of a solution with respect to variation of bandstructure pa- rameters, band filling or variation of interactions. Quite generally, we observe that phases are particularly sta- ble when the lowest eigenvalue $\lambda_{\text{min}}$ of the normalized two-particle vertex $g$ is separated by the second-lowest eigenvalue by a large gap $\delta_\lambda$. In other words, if $\delta_\lambda$ is large, there is no way to destabilize the superconducting ground state. If $\delta_\lambda$ is, however, very small or approaching zero, phase transitions can happen even for the smallest perturbations. In Fig. 4 (c), we plot $\delta_\lambda$. We observe that large $\delta_\lambda$ regions mostly coincide with large $V_{\rm eff}$ regions in Fig. 4 (b).
The second part of this section about the square lat- tice is devoted to the effect of nearest-neighbor Coulomb repulsions, Eq. (54), to the phase diagram, Fig. 4 (a). The form factor in Eq. (56) can be written as
$$
\varepsilon_1(\vec{k}_{2}-\vec{k}_{1})=\sum_{i}\psi_{i}(\vec{k}_{1})\psi_{i}(\vec{k}_{2})\ ,
$$
| |
190610545/2
|
network of quantum computation.
In the remaining part of this paper, we will follow the aforementioned road map to formulate a geometric picture of disentangled representations, which turns out to be a fibre bundle structure and it can be regarded as a gauge theory of disentangled representations.
## III. G EOMETRY OF DISENTANGLED REPRESENTATIONS
## A. Geometry of the evolution of mixed quantum states
To give the geometry of disentangled representation, we need first understand the geometry of the evolution of mixed quantum, which is an analogue of the geometry of disentan- gled representations.
We have two types geometric structures for the evolution of mixed quantum states as in [5][6][7]. It can be proved that they are essentially the same. So our formulation to this geometric structure will be an integrated version of these two structures.
Mixed quantum states can be represented by density op- erators denoted by $\rho$, which are self-adjoint nonnegative unit trace operators. The space of density operators of a quantum system with a Hilbert space $\mathcal{H}$ is denoted by $D(\mathcal{H})$. A density operator evolves according to a von Neumann equation and remains in an orbit of the left conjugation action of the unitary group of $\mathcal{H}$. There exists a one-to-one correspondence between the orbits and the spectra of $\rho$ with $k$ non-increasing nonnegative eigen values $\sigma={p_1,p_2,...,p_k}$. We denote the orbit corresponding to the spectrum $\sigma$ as $D(\sigma)$.
Assuming $\rho_0,\rho_1$ are two isospectral density operators on $D(\sigma)$ and $H$ is a Hamiltonian operator on $\mathcal{H}$ that evolves $\rho_0$ to $\rho_1$ with an evolutional trajectory curve $\rho$, which satisfies:
$$
\begin{split}
i\dot{\rho}&=[H,\rho]\\
\rho(t_0)&=\rho_0\\
\rho(t_1)&=\rho_1\\
\end{split}
$$
We can define the H-distance between $\rho_0,\rho_1$ to be the path integral of the uncertainty of $H$ along $\rho$ given by
$$
D_H(\rho_0,\rho_1)=\int_{t_0}^{t_1}\sqrt{Tr(H^2\rho)-Tr(H\rho)^2}dt
$$
Physically the distance is just the energy needed to evolve $\rho_0$ to $\rho_1$ using $H$. We can then define the dynamic distance between $\rho_0$ and $\rho_1$ as
$$
D(\rho_0,\rho_1)=\min_{H}D_H(\rho_0,\rho_1)
$$
The dynamic distance is a proper distance among isospectral states, which is also unitary invariant so that
$$
D(U\rho_0 U^\dagger,U\rho_1 U^\dagger)=D(\rho_0,\rho_1)
$$
Standard purification of $D(\mathcal{H})$ is denoted by $\Psi \in \mathcal{S}(\mathcal{H}\otimes C^{k*})$ and the surjective map $\pi: \mathcal{S}(\mathcal{H}\otimes C^{k*})\rightarrow D, \pi(\Psi)=\Psi\Psi^\dagger$. Let $P(\sigma)=Diag(\sigma)$ taking $\sigma$ as its diagonal and $\mathcal{S}(\sigma)$ be the subspace of those $\Psi$ that satisfies $P(\sigma)=\Psi^\dagger\Psi$, then $\mathcal{S}(\sigma)$ is a principle fibre bundle over $\mathcal{D}(\sigma)$ with a right acting gauge group $\mathcal{U}(\sigma)$ so that for $u \in \mathcal{U}(\sigma)$, we have
Fig.1. Geometric structure of the evolution of mixed quantum states. $\rho_0$, $\rho_1$ and $\rho_1^{'}$ are all isospectral density operators on $\mathcal{D}(\sigma)$. The different curves on $\mathcal{D}(\sigma)$ are different evolution trajectories between $\rho_0$ and $\rho_1$, or different quantum algorithms to transform $\rho_0$ to $\rho_1$. The geodesic $\rho$ is the optimal algorithm with a computational complexity that equals to the dynamical distance between $\rho_0$ and $\rho_1$. Points on the fibre $\pi^{-1}(\rho_0)$ and $\pi^{-1}(\rho_1)$ are different purification of $\rho_0$ and $\rho_1$ respectively. The two horizontally lifted geodesics in $\mathcal{S}(\sigma)$ of the geodesic $\rho$ are physically equivalent.
$$
\begin{split}
\Psi\Psi^\dagger &= \rho \\
\Psi^\dagger\Psi &= P(\sigma) \\
u\cdot \Psi &= \Psi u \\
uP(\sigma)&=P(\sigma)u
\end{split}
$$
We can then define the vertical and horizontal bundles as $V\mathcal{S}(\sigma)=Ker d\pi$ and $H\mathcal{S}(\sigma)=V\mathcal{S}(\sigma)^\perp$. Vectors in $H\mathcal{S}(\sigma)$ and $V\mathcal{S}(\sigma)$ are called horizontal and vertical vectors respectively. A curve in $\mathcal{S}(\sigma)$ is called horizontal if is velocity vectors are all horizontal. For every curve $\rho$ in $\mathcal{D}(\sigma)$ and $\Psi_0$ in the fibre over the initial state $\rho_0$, there is a unique horizontal lift of $\rho$ in $\mathcal{S}(\sigma)$ starting from $\Psi_0$. The horizontal and vertical bundles define a connection on $\mathcal{S}(\sigma)$.
Now we have two Riemannian structures. One the Rieman- nian structure on $\mathcal{D}(\sigma)$ which defines the dynamical distance between isospectral density operators as the minimal energy needed to evolve $\rho_0$ to $\rho_1$. If the evolution process is regarded as a quantum computation process, the dynamical distance is just the minimal computational complexity to achieve the quantum algorithm $U$ that satisfies $\rho_1=U\rho_0U^\dagger$.
The other geometric structure is the Riemannin structure on $\mathcal{S}(\sigma)$ defined by the connection on it. Accordingly for two isospectral density operators $\rho_0,\rho_1$ we have geodesics between them on $\mathcal{D}(\sigma)$ and geodesics between points on the fibres $\pi^{-1}(\rho_0)$ and $\pi^{-1}(\rho_1)$. About the relationship between these two Riemannian structures, we only point out that the horizontal lift of a geodesic in $\mathcal{D}(\sigma)$ to $\mathcal{S}(\sigma)$ is also a geodesic in $\mathcal{S}(\sigma)$ and they have the same length. The fibre bundle structure is illustrated in Fig. 1.
In order to find the relationship with disentangled repre- sentations in deep networks, here we also give an explicit construction of the gauge field $\mathcal{U}(\sigma)$. In [7] another fibre bundle of mixed states, the generalized Hopf fibration, was
| |
190610545/4
|
Geometric structure of disentangled representations. If $\mathbf{z}$ is the potentially optimal disentangled representation of data $\mathbf{x}$, the curve connection $\mathbf{x}$ and $\mathbf{z}$ represents the bidirectional optimal encoder/decoder. The two-step curve connecting $\mathbf{x}$, $f(z)$ and $g(x)$ is the encoder/decoder that encodes $\mathbf{x}$ as $\mathbf{z}'=f(\mathbf{z})$ and then decodes (generates) data $g(\mathbf{x})$ from $\mathbf{z}'$, where $g(\mathbf{x})$ has the same distribution of $\mathbf{x}$. Or equivalently the two-step curve connecting $\mathbf{g'(x)}$, $z$ and $x$ represents the same encoder/decoder procedure. These two pictures differ at if we choose the input of the encoder or the output of the decoder as our reference.
The fibre bundle structure of disentangled representations can be constructed by establishing a direct correspondence with the evolution of mixed states as follows: (1) Since the generative model is only interested in distribu- tions of $\mathbf{z},\mathbf{x}$, we can set $p(\mathbf{x})$ and $p(\mathbf{z})$ correspond to $\rho_0$ and $\rho_1$ respectively. We may have different distributions of $\mathbf{z}$ because we can have different representations of $\mathbf{x}$, but we have only a fixed $p(\mathbf{x})$. (2) For a given $p(\mathbf{x})$ or $p(\mathbf{z})$, the correspondent fibre bundle of them are just the different data $\mathbf{x}$ and $\mathbf{z}$ that have the same distributions $p(\mathbf{x})$ or $p(\mathbf{z})$. So the base space of the fibre bundle is the distribution space of data and the total space is the data itself. For example, for a given distribution $p(\mathbf{z})$ in base space, its fibre bundle is exactly the data $\mathbf{z}'$ that is obtained by $\mathbf{z}'=f(\mathbf{z})$, where $f$ is a function given in Theorem 1. For data $\mathbf{x}$, we have a similar structure where we denote the transformation function as $g$ instead of $f$. (3) Following the idea of the computational complexity of quantum computation or the network complexity of deep networks, we have an optimal generative deep network $G$ that can generate the distribution $p(\mathbf{x})$ from $p(\mathbf{z})$. It should be point out that since the generative deep network to accomplish $P(\mathbf{x}|\mathbf{z})$ can only accomplish the task by transforming data samples of $p(z)$ to data samples of $p(x)$, the generative deep network is indeed working on the fibre bundle.
Given this construction, the geometric structure of disentan- gled representations can be scratched as in Fig. 2.
The main argument of [1] is that since we can generate data $\mathbf{x}$, to be more exact the distribution of $\mathbf{x}$, from two entangled data $\mathbf{z}$ and $\mathbf{z}'$ with the same generative model, so we can not have an absolute criteria to judge which representation, $\mathbf{z}$ or $\mathbf{z}'$, is disentangled. This puzzling situation gives a contradictive picture about disentangled representations.
What kind of new understanding this geometric picture can bring us about disentangled representations? Are the two rep- resentations equivalent and indistinguishable? Is $\mathbf{x}$ entangled with $\mathbf{z}$ or $\mathbf{z}'$? We think we are facing a similar problem as the famous twins paradox in relativity. The paradox stems from the fact that when we talk about representation of data we only care about distributions of data, i.e. the base space of the fibre bundle, but the encoder/decoder works on data itself, i.e. the fibre bundle or the total space.
(1)Are the two representations $\mathbf{z}$ and $\mathbf{z}'$ identical? The answer really depends on if we only consider the genera- tor/decoder or the complete encoder/decoder process. If we only consider the generator or the decoder, both the repre- sentations $\mathbf{z}$ and $\mathbf{z}'$ give the same horizontally lifted curve as shown in Fig. 2. So physically they are identical according to the idea of gauge theory. But if we also consider the encoder, then obviously they are not equivalent since they correspond to different curves. So both $\mathbf{z}$ and $\mathbf{z}'$ are disentangled repre- sentations for the output data of the generative model, but $\mathbf{z}$ is a disentangled representation of the input data of the encoder, while $\mathbf{z}'$ is an entangled description of the input data of the encoder. So they are not identical just as the twins are not in identical situation in the twin paradox.
(2)Can we discriminate $\mathbf{z}$ or $\mathbf{z}'$? Yes. Intuitively from Fig. [?], $\mathbf{z}$ and $\mathbf{z}'$ lead to different encoder/decoder curves so that by checking the complete circle of encoder/decoder, we can find the difference of $\mathbf{z}$ and $\mathbf{z}'$. Basically they have different computational complexities.
(3)Which representation is the disentangled representation of $\mathbf{x}$, $\mathbf{z}$ or $\mathbf{z}'$? This is exactly the same question as in the twins paradox, which one of the twin brothers is younger? Of course we have an absolute answer, the one experienced acceleration is younger. In fact, the problem of time has a very close relationship with deep networks and the fundamental reason that the twin brothers experienced different time is that their trajectories have different computational complexities. Here we face the same situation, the encoder/decoder proce- dures of $\mathbf{z}$ and $\mathbf{z}'$ have different computational complexities. According to our definition of disentangled representations based on computational complexity, the one with the smaller complexity is a better disentangled representation. In this case, the disentangled representation is $\mathbf{z}$.
## IV. C ONCLUSIONS
Geometrization of deep networks is a programme we pro- posed for the interpretability of deep networks and deep learning systems. By bridging physics and deep networks it aims to bring new interpretations for both fields. In this work, by comparing the fibre bundle structure of the evolution of mixed quantum states, we constructed the fibre bundle structure of disentangled representations, which can be re- garded as a gauge theory of deep networks. This geometric formulation help us to propose a computational complexity based definition of disentangled representations and connect a recent puzzle about disentangled representations with the twins paradox in relativity. We hope this new geometric picture can be integrated in the programme of geometrization of deep networks and helps to build more connections between deep
| |
190610545/1
|
# Gauge theory and twins paradox of disentangled representations
## I. I NTRODUCTION
Achieving a disentangled representation of data is a key issue in machine learning systems. But what’s the essence of disentangled representations? How can we define it? Can we guarantee to reach disentangled representations in machine learning systems? In a recent paper [1] it was claimed that unsupervised learning of disentangled representations is fun- damentally impossible without inductive biases on both the model and the data. This work advocated lots of discussions on disentangled representations.
But till today even a commonly accepted formal definition of disentangled representation is not available yet. We attribute this situation to that we are still lacking of an understanding of the interpretability of deep learning systems. A fully un- derstanding of disentangled representations is only available with a clear interpretability of deep networks. This is to say, we need to examine the disentangled representation problem in a mathematical framework of deep learning systems so that we can define and analysis the properties of disentangled representations.
In this paper we will try to formulate disentangled repre- sentations from a geometric point of view following the pro- gramme of geometrization of deep networks [2][3][4], which was proposed as a framework for the interpretability of deep learning systems. The key idea of this geometrization programme is that, deep networks are physical so that deep networks and physics should have the same mathematical picture. Therefore the well known geometrization of physics can be adapted in the field deep learning as a framework to understand deep learning systems. This idea has shown some promising results by finding the correspondence between deep learning system and physics, for example we find different geometric structures in deep networks which correspond to the geometric structures in general relativity, quantum mea- surement, quantum computation and quantum gravity.
To formulate disentangled representations, we will explore a new geometric structure, a fibre bundle structure. We will see disentangled representations can be understood as a gauge theory which has a counterpart in the evolution of mixed quantum states. By comparing the geometric structures of disentangled representations and the mixed quantum states, we can give a geometric picture for disentangled representations. We will also use this perspective to give a negative argument to the conclusion of [1].
## II. G EOMETRIZATION OF DEEP NETWORKS
Geometrization of deep networks aims to formulate deep networks and deep learning systems by geometrizaiton. That’s to say deep networks can be described by geometric structures and the properties of deep learning systems are determined by their correspondent geometric pictures. The underlining reason for us to propose this idea is that we believe deep networks are physical in the meaning that deep networks are efficient descriptors of our physical world and our physical world is generated by deep networks, to be more specific, by deep networks of quantum computations according to computational university. So the geometrization of physics, which is one of the most fundamental thought in physics, stems from the geometrization of deep networks. So the global picture of geometrization of deep networks is that, we can derive our physical world with all its rules from deep networks. On the other hand, ideas from physics research can be borrowed for the interpretability of deep networks.
Currently the main components of geometrization of deep networks include: (1) The rule of deep networks can be formulated as a least action principle. (2) The action here is defined by an information metric based complexity of deep networks. (3) The basic geometric framework of deep systems is the Riemannian geometry. We point out here this is compatible with the idea that deep networks can be understood as differential equations or optimal control problems since they can also be formulated on Riemannian geometry. (4) There exist a dictionary between the concepts of deep networks and physics, just as the dictionary between physics and geometry. (5) Important properties of deep learning systems such as convergence, generalization, disentanged representation, network architecture search, knowledge distillation and network pruning, can all be understood geometrically. (6) The structure and laws of our physical world can be derived from deep networks so that we also have a better interpretability of physics. For example now we already have a rough idea how spacetime, material and their interactions can be understood as emergent phenomena from the deep
| |
190610545/3
|
introduced where the gauge group is a left acting group $\mathcal{V}(\sigma)$ with $v \in \mathcal{V}(\sigma)$ satisfies
$$
\begin{split}
v\cdot \Psi &= v\Psi \\
v\rho_0 &=\rho_0 v
\end{split}
$$
In it was shown that if $\rho_0=U_0 Diag(\sigma) U_0^\dagger$ with $m_i$ be the multiplicites of the eigenvalues of $\rho_0$, then the gauge group is given by
$$
V(\sigma)=Ad_{U_0}\otimes_i U(m_i)
$$
So the gauge group $U(\sigma)$ is in fact just $\otimes_i U(m_i)$. This means the function of the gauge group is just to mix the eigenvectors of $\rho_0$ with the same eigen value by an unitary operator. This in fact can be understood to entangle those eigenvectors with the same eigenvalue.
From a computation point of view, $\rho_0,\rho_1$ are the input and output data of a computation process and the computation process is given by $\rho$. The geodesic on $\mathcal{D}(\sigma)$, which defines the dynamical distance between $\rho_0$ and $\rho_1$, is the optimal algorithm to accomplish the computation. The fibres $\pi^{-1}(\rho_0)$, $\pi^{-1}(\rho_0)$ are different ways to represent the input and output data. Curves on the fibre bundle to connect points in $\pi^{-1}(\rho_0)$ and $\pi^{-1}(\rho_0)$ are different implementations of the algorithm $\rho$.
According to the basic idea of gauge theory, different gauges are just different equivalent descriptions of the same physics. All the horizontal geodesics, the two horizontally lifted geodesics as shown in Fig. 1, on the fibre bundle are physically equivalent realizations of the optimal algorithm and they have exactly the same physical meaning.
## B. Geometric picture of disentangled representations
Given the geometric picture of the evolution of mixed quan- tum states, which can be regarded as a geometric description of the most fundamental computation system, now we can scratch the geometric picture of disentangled representations of deep network.
The first concern about disentangled representations is, how to define disentangled representations? What’s the fun- damental nature of disentangled representations? Usually we think a disentangled representation should have the property that each data generative factor should be represented by a single or a small set of dimensions. There is also effort to define disentangled representations from the transformation properties of our world [8]. The transformations that change only a subset of properties of the underlying world state while keeping all other properties invariant are regarded as disentangled representations of data. We can see here the word transformation is in fact just an information processing process or a computation algorithm. So is it possible to derive the concept of disentangled representation purely from a computation point of view?
Our suggestion is to put the disentangled representation and the computational complexity on the same foot. A disentangled representation of certain data is just the optimal algorithm (with the minimal computational complexity) to generate that data. Or in other words, a disentangled representation is just the natural by-product of an optimization on the com- putational complexity. For example, in the aforementioned computational complexity of the evolution of mixed quantum states, the optimal algorithm to generate $\rho_1$ from $\rho_0$ is a Hamiltonian evolution of the quantum state with the minimal energy expense. From the quantum computational complexity or the Hamiltonian complexity point of view, the quantum circuit to achieve the optimal quantum algorithm consists of a sequence of simple quantum gates so that at each stage of the computation procedure, we only entangle information from a few quantum bits. In quantum computation, the complexity of a pure quantum state $|\psi\rangle$ is the minimal complexity of all quantum algorithms that can generate $|\psi\rangle$ from a product state $|00...0\rangle$. The optimal algorithm is in fact the optimal disentangled representation of the state $|\psi\rangle$.
From these perspective, in deep network based information processing systems, we can define the optimal deep network as the deep network with the minimal network complexity, where the network complexity is defined by the Fisher information metric. If the deep network is a generative model of data, then naturally it’s also the optimal disentangled representation of data. Now we have an universal definition of disentangled representations and the essence of disentangled representations is just low computational complexity.
In our former work, we have shown that the complexity of deep networks has the same Riemannian structure as the complexity of quantum computation, which is just the Riemannian structure on $\mathcal{D}(\sigma)$ as given above. Now we would like to construct the fibre bundle structure of disentangled representations of deep networks corresponding to the fibre bundle structure of the evolution of mixed quantum state shown in Fig. 1.
Since this work aims to challenge the argument proposed in [1] that disentangled representations can not be obtained in unsupervised learning systems without inductive bias, we here construct the fibre bundle structure of disentangled represen- tations directly based on the formulation of [1].
In [1] their key argument is the Theorem 1, where the formulation of disentangled representations is as follows. (1) The real world data $\mathbf{x}$ is generated by a two-step generative process. The first step is to generate a multivariate latent random variable $\mathbf{z}$ from a distribution $P(\mathbf{z})$. In the second step the observed data $\mathbf{x}$ is sampled from a conditional distribution $P(\mathbf{x}|\mathbf{z})$, where $P(\mathbf{x}|\mathbf{z})$ is achieved by a deep network. The goal of representations learning is to find a representation $r(\mathbf{x})$ of $\mathbf{x}$ that help us to better understand the structure of $\mathbf{x}$.
(2) Theorem 1 of. For $d>1$ and $p(\mathbf{z})=\prod_{i=1}^d p(\mathbf{z}_i)$, there exists an infinite family of bijective functions $f$ so that $\mathbf{z}'=f(\mathbf{z})$ and $p(\mathbf{z})=p(\mathbf{z}')$, where $\mathbf{z}'$ is obtained by a complete entanglement of $\mathbf{z}$.
(3) [1] then argued that since the generative model $P(\mathbf{x}|\mathbf{z})$ can generate data $\mathbf{x}$ from both $\mathbf{z}$ and $\mathbf{z}'$ but $\mathbf{z}'$ is a fully entangled version of $\mathbf{z}$, so there is no way to distinguish the two generative models so that we have no way to define a disentangled representation of $\mathbf{x}$.
| |
230406165/5
|
## Data Processing
Employing webcam eye-tracking during an online, self-paced, se- quential object-interaction task proved to be challenging. The re- sulting raw data required a number of quality assessments and treatments to ensure its utility for analysis. While this paper cen- ters on our empirical findings, the corresponding methodological contribution of this work is not trivial, and we provide a detailed account of our data processing pipeline in Supplementary Materials.
The uncontrolled nature of the online testing environment could give rise to less accurate or spurious gaze predictions. We deter- mined, in a cursory visual inspection of pilot data, that unlike cursor data, gaze data were prone to spatial distortions. That is, while much of the structure of the screen layout of the task was evident in most participants’ gaze data (e.g. many fixations following a pattern shaped like the distribution of targets) these fixation “hot spots” would not necessarily project to the actual target locations - instead they were often shifted or skewed (see Figure 2 and Supplementary Materials for examples). However, if one is primarily interested in which object a person is fixating and when, the exact location of that fixation is mostly irrelevant, and instead you can define and analyze looks to AOIs in relative space. Taking advantage of the fact that our key analyses related to 4 distinct, spatially distributed locations for Pick-up and Drop-off events, we used a data-driven AOI approach. Here, we assumed that participants’ gaze would primarily be driven toward the 4 target locations (NRAOI, NLAOI, FRAOI and FLAOI). Using data from all 50 trials, from all times when the participant had clicked and held the cursor button down, we employed a k-means clustering approach to spatially bin the gaze data into 4 corresponding AOIs (see Figure 2 for a represen- tative participant’s clustering centroids and see Supplementary Materials for additional examples). Thus, our eye-tracking data, while retaining its temporal resolution, was spatially transformed from the 800 x 450 Labvanced coordinate frame into four mutually- exclusive bins: NRAOI, NLAOI, FRAOI and FLAOI. Fourteen of the original 51 subject datasets were rejected because the clustering centroids did not follow the spatial configuration of the AOIs (left targets to the left of right targets, near targets lower on the screen than far targets), suggesting raw gaze data errors beyond a spatial distortion that we could not account for. In the Supplementary Materials, we include a probability density analysis of the accepted clusters in transformed space where we fit bivariate normal distri- butions to each cluster for each participant and demonstrate that, on average, 28.97% of eye-tracking data, if linearly transformed to the Labvanced coordinate frame, would fall within the 80 x 80 pixel AOI it was cluster-assigned to. Importantly, only 0.1% of the eye data risked assignment to any of the 3 non-assigned AOIs in transformed pixel space.
Notably, this approach is not without its risks or limitations. First, by using the data to define the AOIs used for analysis, we run the risk of circularity. Therefore our first test was to ensure that the distribution of looks to each AOI across time matched the time-varying demands of the task. Since our clustering was collapsed across time, this would ensure that the reported looking behaviour was sensitive to the actual task being performed. Second, as discussed in Sections 4 - Results and 6 - Limitations and Future Directions, by only creating four cluster-based AOIs, we lose the ability to detect looks to other areas of the display (e.g. the Home or Fixation targets). Since these other areas were not relevant for the majority of our task, and entirely irrelevant to the key interac- tion events, this was a trade-off we felt was worth making despite the consequence of limiting our approach’s applicability to other task designs or research questions. Third, the spatial accuracy as- sessment of our clustering approach (in Supplementary Materials) highlights the highly effective discrimination between AOIs by hav- ing effectively no chance of a mis-classified eye-gaze, but it also exposes the challenge of noisier eye-data for within -AOI discrim- ination - had we constrained ourselves to transformed eye data that fell within the boundaries of the actual on-screen AOI-objects, we would have lost more than 70% of the data. In Supplementary Materials, we represent our clusters as independent bivariate nor- mal probability density functions to visualize their clear spatial cohesiveness, but we acknowledge that the AOI-binning approach may have limited use where targets are less spatially distributed or the task space necessitates unpredictable, dynamic, numerous and/or densely organized targets.
Beyond the eye data cleaning, various steps (as outlined in Sup- plementary Materials) were taken to ensure participants completed the task correctly. While participants could move the circles about as they pleased, a trial was designed to only advance once all the movements were made. Although various real-time checks were performed to preemptively avoid sequence errors, they still oc- curred and those trials were removed from the analyzed data. Eight further subjects were removed for having a trial count below 50% after trial rejection for sequence (and other) errors. Therefore, 29 datasets were included in the following analyses.
## Segmentation
In order to explore eye-cursor coordination patterns during object interactions, we needed to define and then automatically identify the 8 movements in each trial and the 2 object interactions (Pick-up and Drop-off) within each movement (see Figure 1 D). This first necessitated re-sampling the cursor and eye data to a common sampling frequency of 60 Hz. Most often this meant that the eye data was upsampled while the cursor data was downsampled. Following Lavoie et al.’s approach [27], we considered the object interaction to include the period when the cursor moves toward the object (Reach: onset = cursor approaching Pick-up location + velocity exceed threshold; offset = Transport onset), the period when the cursor drags the object (Transport: onset = object leaving Pick-up location + velocity exceeds threshold; offset = object approaching Drop-off location + velocity drops below threshold) and the period when the cursor moves away from the object (Release: onset = Transport offset; offset = cursor leaving Drop-off location + velocity drops below threshold), as depicted in Figure 1 D. A Pick-up is said to occur at the transition between Reach and Transport while a Drop-off is said to occur at the transition between Transport and Release. We used our custom Gaze and Movement Analysis (GaMA) software in MATLAB to segment our trial data into 8 movements (Moves 1-8) and each movement into the 3 phases (Reach, Transport and Release). In GaMA, spatial and temporal features of the data are used to automatically define key events, like the start of an object Transport (see [52]). We followed the principles of Lavoie et al.
| |
230406165/7
|
Figure3: Average fixation duration to each of the four target AOIs (green and blue squares) across movements when that AOI was task-relevant (yellow bars, when that AOI was a Pick- up or Drop-off site for a movement) or task-irrelevant (grey bars). Relevant AOIs were fixated significantly longer than Irrelevant AOIs. Error bars show 95% confidence intervals of the estimated marginal means.
Critically, therefore, we show that participants’ gaze fixated more on the Task Relevant AOIs than on the Task Irrelevant AOIs (Figure 3 - see Supplementary Materials for a complementary spatial analy- sis). These results align favourably with the early, real-world work of Hayhoe and Land [17,24,25] and give credence to our use of webcam eye-tracking as a method for a preliminary investigation of gaze behaviours during our specific online, screen-based, object interaction task. Our 4 x 2 (Position x Task Relevance [Relevant / Irrelevant]) RMANOVA revealed significant main effects of both Position (F (1.50, 41.97) = 11.6, p <.001) and Task Relevance (F (1.00, 28.00) = 329.2, p <.001), and a significant interaction between the two factors (F (1.83,51.26) = 45.7, p <.001). Post-hoc pairwise follow- ups compared Relevant vs Irrelevant fixation durations at each location - for each location it was fixated more when it was Rele- vant than when it was Irrelevant (all p ’s <.001).
This analysis also allowed us to explore for any specific spatial biases in the eye-tracking data recorded in this task. In general, looks to targets were relatively evenly distributed, except for times when looks to other objects in the environment were mis-classified to spatially-proximal target locations. Specifically, at some times, looks to the Home position may have been categorized as looks to the NRAOI and looks to the Fixation position may have been categorized as looks to the FRAOI. This pattern is also visible in the complementary spatial analysis in Supplementary Materials, where the bivariate normal probability density function of the NRAOI shows more dispersion, likely as a result of also capturing some Home position looks. Along the same lines, in this supplemental analysis the FLAOI has the least dispersion, matching its status as the most isolated task-relevant object. As mentioned, however, the overall lack of spatial specificity is a consequence of our AOI clus- tering but does not appear to add significant noise to our analyses. For a complete analysis of these spatial biases, see Supplementary Materials.
In general, while accounting for the inherent limitations in the design of our study, this analysis offers a demonstration of the sensi- tivity of webcam eye-tracking. The stark differences in looking time driven by the expected pattern of task relevance, complemented by the low risk of assigning gazes to inaccurate clusters (as evi- denced in Supplementary Materials), gave us sufficient confidence to further explore the dynamics of eye-cursor coordination.
## Digital object interactions yield unique, context-specific movement dynamics
As described above, in order to understand the nuances of eye- cursor coordination it is first essential to map the naturally-occurring variations in task demand as indicated by the time spent in each movement and each phase within that movement (Reach, Transport and Release, Figure 4). Thus, we used a 3 x 8 (Phase x Movement) RMANOVA to examine phase duration. Both main effects of Phase (F (1.58, 44.20) = 138.5, p <.001) and Movement (F (3.63, 101.68) = 34.1, p <.001) were significant, as was their interaction (F (5.38,150.53) = 21.0, p <.001). Because we were most interested in learning how changes in each phase might impact eye latencies, we examined how phase values changed across movements. The three follow-up single-factor RMANOVAs (Reach/Transport/Release x Movement) each revealed main effects of Movement (Reach: F (3.74, 104.80) = 24.5, p <.001; Transport: F (3.80, 106.41) = 22.2, p <.001; Release: F (3.13, 87.55) = 29.7, p <.001). These results highlight a general pat- tern of longer Reach phases for movements covering longer screen distances. That is, movements that directly follow a Home visit (i.e. Moves 1, 3, 5, 7) cover more screen distance and elicit longer Reach phases than those that immediately follow a circle movement (i.e. Moves 2, 4, 6, 8) except for Move 1, which has the shortest Home to AOI distance. This finding aligns with the principles of Fitts’ Law.
The single biggest difference between the movement dynam- ics in the digital compared to real-world task is the duration of the Transport phases. Cursor click and drag movements are much faster (around 200 ms) than their real-world counterparts (over 1000 ms from Lavoie et al [27]). Besides a quick final move (M = 0.143 secs) and some slight differences between other movements, the Transport phases are relatively similar in duration (M ’s range from 0.205 to 0.255 secs, see Supplementary Materials for full pairwise analysis). Overall, the relative consistency of the Transport dura- tion across the task again reflects that phase timing is primarily related to movement distance - Transport distance is the same for all movements.
Pairwise comparisons (as reported in Supplementary Materials) between movements for the Release phase also follow a pattern of longer phase durations for movements covering more screen distance (between the movement’s drop-off location and the next destination location), as predicted by Fitts’ Law.
## Eye-cursor coordination during Pick-up interactions resembles the real world, while Drop-off coordination flexibly conforms to the digital context
Our final motivation was to understand eye-cursor coordination during digital object interaction by examining the latencies between the eye and cursor arriving (EAL) and leaving (ELL) the Pick-up and Drop-off sites across the 8 Movements (Figure 4). For each of
| |
230406165/2
|
Jennifer K. Bertrand and
Craig S. Chapman
to explore visuomotor coordination in our specific task, and sec- ond, to quantify if the patterns of eye-cursor coordination would match principles of eye-hand coordination in the real world. Our findings, while preliminary, reveal that both of these are true for this experiment, offering an introduction to entirely new ways of collecting user experience data and suggesting UX design should further explore and consider the tight and environment-invariant principles of visuomotor coordination during target interaction tasks.
## RELATED WORKS
## Real-World Object Interactions
The tight coupling, in both space and time, of eye movements and motor actions has been well-documented. While much of this re- search has involved rigid, paradigmatic tasks, those most relevant to UX leverage technological advances to explore eye-hand coor- dination during self-paced, natural and realistic interactions. For example, the seminal works of Land et al. and Hayhoe measured where people look when making a pot of tea [26] and preparing a sandwich [17]. Despite the complexity of these tasks and the lack of experimental structure, the researchers were able to break down the tasks into their constituent subtasks (e.g. reaching for the kettle, removing the lid, etc.) to reveal remarkably consistent patterns of eye-hand coordination [25].
First, even though there are irrelevant objects throughout the kitchen scenes, the eyes only ever fixate on task-relevant objects [25]. Critically, these are not the most salient (as defined by low level visual properties like contrast) objects in the visual field; rather, gaze only lands upon task-relevant objects. Second, gaze behaves serially, always fixed upon the current object of manipulation, leading the hand to that object, and when the manipulation is almost complete, moving on to the next object without return [25]. In Land et al.’s tea-making task [26], aggregated across all 94 distinct sub-tasks, the dynamics of the gaze and hand display a consistent pattern: participants fixate on the object to be manipulated for about half a second prior to the hand’s initial movement towards that object and remain fixated there until leaving for the next target about half a second before the completion of the current manipulation.
This general pattern of the eye leading motor action has been found in many contexts, extending beyond kitchen activities to less obvious forms of naturalistic, visually-guided interactions like walk- ing [24,38], keyboard typing [4], and music playing [13]. Across these interactions, the exact amount of time the eye leads the hand appears to be at least 500 ms (and up to about 1 second), showing some flexibility for the time [8] or accuracy constraints [40] of the task, or kinematics required for different task contexts [21,39]. Lavoie et al. [27] took a modern approach to investigations of eye-hand coordination during real-world object interaction by com- bining state of the art motion capture and mobile eye-tracking. In perfect alignment with Land et al. [26], Lavoie et al. [27] found that all object interactions involved the eyes fixating the object at least 500 ms prior to the start of the interaction. Then, within 600 ms of the interaction start, the eyes would leave to look ahead to the next area for interaction [27]. This dominant pattern of eye-hand coordination has proven itself highly consistent across real-world object interactions, but, in service of applying real-world findings to digital domains, what happens when we move towards interactions with digital objects?
## Lab-based Digital Interactions
Some visuomotor coordination research trades the complexity of an in-lab kitchen for the control of a computer workstation, treating screen-based, digitally-presented objects as a convenient proxy for real-world objects. Nonetheless, patterns of eye leading hand (or computer cursor or manipulandum) are remarkably consistent in the digital domain. In tracing shapes with a cursor, the eye leads the cursor by 223-295 ms [9], and in distractorless visual search there is a 190 ms lead [2]. A related paradigm of tracking an unpredictable object on a screen shows that the eye lags behind the target object by 24 ms, whereas the hand lags behind it by 108 ms [7]. During simple reaches towards on-screen targets, the eyes arrive about 386 ms before the hand [43], although gaze and hand dynamics are flexible to factors like the target’s visibility before or during the reach [50]. Adaptive gaze behaviour is also shown for two- target sequential reaches: the eye anchors to the first target for an extra 95 ms to ensure the hand’s arrival before continuing to a second target [40]. For visually-guided sequential movements of a manipulandum-handle’s contact with 5 virtual target objects, the gaze arrives at targets 208 ms before contact and leaves 106 ms after contact [3]. Finally, dragging virtual objects about in a pseudo-touchscreen context also elicits gaze-leading-hand patterns [47]. Taken together, this line of research predominantly studies target-directed reaching and following, but not interaction. Even so, in all contexts, the eyes lead the hand (or cursor) by a few hundred milliseconds.
## UX-Focused Digital Interactions
What do eye-cursor coordination patterns look like when digital object interactions are also realistic and ecologically-valid? The HCI domain offers us some answers, however object interactions be- yond simple target clicks remain almost entirely unexplored. Early eye-cursor studies employed search engine results pages (SERPs; [16,34,41]) but mostly considered visuomotor coordination only from a spatial context, measuring the pixel distance between the cursor and eye. Huang et al. [20] did consider timing, finding that the gaze led the cursor by at least 250-, and typically 700 ms during SERP browsing. Of course, with less experimental control, scientists are more likely to find a range of behaviours. Indeed, Smith and colleagues [48] looked at cursor-pointing to graphical user inter- face targets and observed at least three eye-cursor coordination patterns: “eye gaze following the cursor to the target”, “eye gaze leading the cursor to the target”, and less commonly, “eye gaze switching between the cursor and target until the target is reached”. The use of multiple strategies during digital yet ecologically-valid tasks is perhaps best illustrated by Liebling and Dumais [28]. Here subjects performed their regular work duties on their office desktop while their gaze and cursor movements were recorded. To begin making sense of the data, the researchers anchored their analysis to cursor clicks, with 32 classes of click-targets determined by meta- data records. This rich dataset revealed nuanced coordination - in general, the gaze arrived near the click point 100-200 ms earlier
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.