arXiv subjects

Changji Xu

Publications and source records attributed to Changji Xu.

15 recordsLinked to original sources

arXiv2023

Operator Norm Bounds on the Correlation Matrix of the SK Model at High Temperature

We prove that the two point correlation matrix $ \textbf{M}= (\langle \sigma_i ; \sigma_j\rangle)_{1\leq i,j\leq N} \in \mathbb{R}^{N\times N}$ of the Sherrington-Kirkpatrick model has the property that for every $\epsilon>0$ there exists $K_\epsilon>0$, that is independent of $N$, such that \[ \mathbb{P}\big( \| \textbf{M} \|_{\text{op}} \leq K_{\epsilon}\big) \geq 1- \epsilon \] for $N$ large enough, for suitable interaction and external field parameters $(\beta,h)$ in the replica symmetric region. In other words, the operator norm of $\textbf{M}$ is of order one with high probability. Our results are in particular valid for all $ (\beta,h)\in (0,1)\times (0,\infty) $ and thus complement recently obtained results in \cite{EAG,BSXY} that imply the operator norm boundedness of $\textbf{M}$ for all $\beta<1$ in the special case of vanishing external field.

math-ph↗

arXiv2023

Geometric Barriers for Stable and Online Algorithms for Discrepancy Minimization

For many computational problems involving randomness, intricate geometric features of the solution space have been used to rigorously rule out powerful classes of algorithms. This is often accomplished through the lens of the multi Overlap Gap Property ($m$-OGP), a rigorous barrier against algorithms exhibiting input stability. In this paper, we focus on the algorithmic tractability of two models: (i) discrepancy minimization, and (ii) the symmetric binary perceptron (\texttt{SBP}), a random constraint satisfaction problem as well as a toy model of a single-layer neural network. Our first focus is on the limits of online algorithms. By establishing and leveraging a novel geometrical barrier, we obtain sharp hardness guarantees against online algorithms for both the \texttt{SBP} and discrepancy minimization. Our results match the best known algorithmic guarantees, up to constant factors. Our second focus is on efficiently finding a constant discrepancy solution, given a random matrix $\mathcal{M}\in\mathbb{R}^{M\times n}$. In a smooth setting, where the entries of $\mathcal{M}$ are i.i.d. standard normal, we establish the presence of $m$-OGP for $n=\Theta(M\log M)$. Consequently, we rule out the class of stable algorithms at this value. These results give the first rigorous evidence towards a conjecture of Altschuler and Niles-Weed~\cite[Conjecture~1]{altschuler2021discrepancy}. Our methods use the intricate geometry of the solution space to prove tight hardness results for online algorithms. The barrier we establish is a novel variant of the $m$-OGP. Furthermore, it regards $m$-tuples of solutions with respect to correlated instances, with growing values of $m$, $m=\omega(1)$. Importantly, our results rule out online algorithms succeeding even with an exponentially small probability.

cs.CC↗

arXiv2023

Eigenstate Thermalization Hypothesis for Generalized Wigner Matrices

In this paper, we extend results of Eigenvector Thermalization to the case of generalized Wigner matrices. Analytically, the central quantity of interest here are multiresolvent traces, such as $\Lambda_A:= \frac{1}{N} \text{Tr }{ GAGA}$. In the case of Wigner matrices, as in \cite{cipolloni-erdos-schroder-2021}, one can form a self-consistent equation for a single $\Lambda_A$. There are multiple difficulties extending this logic to the case of general covariances. The correlation structure prevents us from deriving a self-consistent equation for a single matrix $A$; this is due to the introduction of new terms that are quite distinct from the form of $\Lambda_A$. We find a way around this by carefully splitting these new terms and writing them as sums of $\Lambda_B$, for matrices $B$ obtained by modifying $A$ using the covariance matrix. The result is a system of self-consistent equations relating families of deterministic matrices. Our main effort in this work is to derive and analyze this system of self-consistent equations.

math.PR↗

arXiv2022

The Two Point Function of the SK Model without External Field at High Temperature

We show that the two point correlation matrix $ \textbf{M}= (\langle \sigma_i \sigma_j\rangle)_{1\leq i,j\leq N} $ of the Sherrington-Kirkpatrick model with zero external field satisfies \[ \lim_{N\to\infty} \| \textbf{M} - ( 1+\beta^2 - \beta \textbf{G})^{-1} \|_{\text{op}} =0 \] in probability, in the full high temperature regime $\beta < 1$. Here, $\textbf{G}$ denotes the GOE interaction matrix of the model.

math-ph↗

arXiv2022

Spectral Gap Estimates for Mixed $p$-Spin Models at High Temperature

We consider general mixed $p$-spin mean field spin glass models and provide a method to prove that the spectral gap of the Dirichlet form associated with the Gibbs measure is of order one at sufficiently high temperature. Our proof is based on an iteration scheme relating the spectral gap of the $N$-spin system to that of suitably conditioned subsystems.

math.PR↗

arXiv2022

Bulk universality and quantum unique ergodicity for random band matrices in high dimensions

We consider Hermitian random band matrices $H=(h_{xy})$ on the $d$-dimensional lattice $(\mathbb Z/L \mathbb Z)^d$, where the entries $h_{xy}=\overline h_{yx}$ are independent centered complex Gaussian random variables with variances $s_{xy}=\mathbb E|h_{xy}|^2$. The variance matrix $S=(s_{xy})$ has a banded profile so that $s_{xy}$ is negligible if $|x-y|$ exceeds the band width $W$. For dimensions $d\ge 7$, we prove the bulk eigenvalue universality of $H$ under the condition $W \gg L^{95/(d+95)}$. Assuming that $W\geq L^\epsilon $ for a small constant $\epsilon >0$, we also prove the quantum unique ergodicity for the bulk eigenvectors of $H$ and a sharp local law for the Green's function $G(z)=(H-z)^{-1}$ up to ${\mathrm{Im}} \, z \gg W^{-5}L^{5-d}$. The local law implies that the bulk eigenvector entries of $H$ are of order ${\mathrm{O}}(W^{-5/2}L^{-d/2+5/2})$ with high probability.

math.PR↗

arXiv2022

Algorithms and Barriers in the Symmetric Binary Perceptron Model

The symmetric binary perceptron ($\texttt{SBP}$) exhibits a dramatic statistical-to-computational gap: the densities at which known efficient algorithms find solutions are far below the threshold for the existence of solutions. Furthermore, the $\texttt{SBP}$ exhibits a striking structural property: at all positive constraint densities almost all of its solutions are 'totally frozen' singletons separated by large Hamming distance \cite{perkins2021frozen,abbe2021proof}. This suggests that finding a solution to the $\texttt{SBP}$ may be computationally intractable. At the same time, the $\texttt{SBP}$ does admit polynomial-time search algorithms at low enough densities. A conjectural explanation for this conundrum was put forth in \cite{baldassi2020clustering}: efficient algorithms succeed in the face of freezing by finding exponentially rare clusters of large size. However, it was discovered recently that such rare large clusters exist at all subcritical densities, even at those well above the limits of known efficient algorithms \cite{abbe2021binary}. Thus the driver of the statistical-to-computational gap exhibited by this model remains a mystery. In this paper, we conduct a different landscape analysis to explain the algorithmic tractability of this problem. We show that at high enough densities the $\texttt{SBP}$ exhibits the multi Overlap Gap Property ($m-$OGP), an intricate geometrical property known to be a rigorous barrier for large classes of algorithms. Our analysis shows that the $m-$OGP threshold (a) is well below the satisfiability threshold; and (b) matches the best known algorithmic threshold up to logarithmic factors as $m\to\infty$. We then prove that the $m-$OGP rules out the class of stable algorithms for the $\texttt{SBP}$ above this threshold. We conjecture that the $m \to \infty$ limit of the $m$-OGP threshold marks the algorithmic threshold for the problem.

cs.CC↗

arXiv2021

Gardner formula for Ising perceptron models at small densities

We consider the Ising perceptron model with N spins and M = N*alpha patterns, with a general activation function U that is bounded above. For U bounded away from zero, or U a one-sided threshold function, it was shown by Talagrand (2000, 2011) that for small densities alpha, the free energy of the model converges in the large-N limit to the replica symmetric formula conjectured in the physics literature (Krauth--Mezard 1989, see also Gardner--Derrida 1988). We give a new proof of this result, which covers the more general class of all functions U that are bounded above and satisfy a certain variance bound. The proof uses the (first and second) moment method conditional on the approximate message passing iterates of the model. In order to deduce our main theorem, we also prove a new concentration result for the perceptron model in the case where U is not bounded away from zero.

math.PR↗

arXiv2021

Frozen $1$-RSB structure of the symmetric Ising perceptron

We prove, under an assumption on the critical points of a real-valued function, that the symmetric Ising perceptron exhibits the `frozen 1-RSB' structure conjectured by Krauth and Mezard in the physics literature; that is, typical solutions of the model lie in clusters of vanishing entropy density. Moreover, we prove this in a very strong form conjectured by Huang, Wong, and Kabashima: a typical solution of the model is isolated with high probability and the Hamming distance to all other solutions is linear in the dimension. The frozen 1-RSB scenario is part of a recent and intriguing explanation of the performance of learning algorithms by Baldassi, Ingrosso, Lucibello, Saglietti, and Zecchina. We prove this structural result by comparing the symmetric Ising perceptron model to a planted model and proving a comparison result between the two models. Our main technical tool towards this comparison is an inductive argument for the concentration of the logarithm of number of solutions in the model.

math.PR↗

arXiv2019

Distribution of the random walk conditioned on survival among quenched Bernoulli obstacles

Place an obstacle with probability $1-p$ independently at each vertex of $\mathbb Z^d$ and consider a simple symmetric random walk that is killed upon hitting one of the obstacles. For $d \geq 2$ and $p$ strictly above the critical threshold for site percolation, we condition on the environment such that the origin is contained in an infinite connected component free of obstacles. It has previously been shown that with high probability, the random walk conditioned on survival up to time $n$ will be localized in a ball of volume asymptotically $d\log_{1/p}n$. In this work, we prove that this ball is free of obstacles, and we derive the limiting one-time distributions of the random walk conditioned on survival. Our proof is based on obstacle modifications and estimates on how such modifications affect the probability of the obstacle configurations as well as their associated Dirichlet eigenvalues, which is of independent interest.

math.PR↗

arXiv2019

Sharp threshold for the Ising perceptron model

Consider the discrete cube $\{-1,1\}^N$ and a random collection of half spaces which includes each half space $H(x) := \{y \in \{-1,1\}^N: x \cdot y \geq \kappa \sqrt{N}\}$ for $x \in \{-1,1\}^N$ independently with probability $p$. Is the intersection of these half spaces empty? This is called the Ising perceptron model under Bernoulli disorder. We prove that this event has a sharp threshold; that is, the probability that the intersection is empty increases quickly from $\epsilon$ to $1- \epsilon$ when $p$ increases only by a factor of $1 + o(1)$ as $N \to \infty$.

math.PR↗

arXiv2019

Biased random walk conditioned on survival among Bernoulli obstacles: subcritical phase

We consider a discrete time biased random walk conditioned to avoid Bernoulli obstacles on ${\mathbb Z}^d$ ($d\geq 2$) up to time $N$. This model is known to undergo a phase transition: for a large bias, the walk is ballistic whereas for a small bias, it is sub-ballistic. We prove that in the sub-ballistic phase, the random walk is contained in a ball of radius $O(N^{1/(d+2)})$, which is the same scale as for the unbiased case. As an intermediate step, we also prove large deviation principles for the endpoint distribution for the unbiased random walk at scales between $N^{1/(d+2)}$ and $o(N^{d/(d+2)})$. These results improve and complement earlier work by Sznitman [Ann. Sci. Ecole Norm. Sup. (4), 28(3):345--370, 371--390, 1995].

math.PR↗

arXiv2018

Localization for random walks among random obstacles in a single Euclidean ball

Place an obstacle with probability $1-p$ independently at each vertex of $\mathbb Z^d$, and run a simple random walk until hitting one of the obstacles. For $d\geq 2$ and $p$ strictly above the critical threshold for site percolation, we condition on the environment where the origin is contained in an infinite connected component free of obstacles, and we show that for environments with probability tending to one as $n\to \infty$ there exists a unique discrete Euclidean ball of volume $d \log_{1/p} n$ asymptotically such that the following holds: conditioned on survival up to time $n$ we have that at any time $t \in [o(n),n]$ with probability tending to one the simple random walk is in this ball. This work relies on and substantially improves a previous result of the authors on localization in a region of volume poly-logarithmic in $n$ for the same problem.

math.PR↗

arXiv2018

Geometry of the random walk range conditioned on survival among Bernoulli obstacles

We consider a discrete time simple symmetric random walk among Bernoulli obstacles on $\mathbb{Z}^d$, $d\geq 2$, where the walk is killed when it hits an obstacle. It is known that conditioned on survival up to time $N$, the random walk range is asymptotically contained in a ball of radius $\varrho_N=C N^{1/(d+2)}$ for any $d\geq 2$. For $d=2$, it is also known that the range asymptotically contains a ball of radius $(1-\epsilon)\varrho_N$ for any $\epsilon>0$, while the case $d\geq 3$ remains open. We complete the picture by showing that for any $d\geq 2$, the random walk range asymptotically contains a ball of radius $\varrho_N-\varrho_N^\epsilon$ for some $\epsilon \in (0,1)$. Furthermore, we show that its boundary is of size at most $\varrho_N^{d-1}(\log \varrho_N)^a$ for some $a>0$.

math.PR↗

arXiv2017

Poly-logarithmic localization for random walks among random obstacles

Place an obstacle with probability $1-p$ independently at each vertex of $\mathbb Z^d$, and run a simple random walk until hitting one of the obstacles. For $d\geq 2$ and $p$ strictly above the critical threshold for site percolation, we condition on the environment where the origin is contained in an infinite connected component free of obstacles, and we show that the following \emph{path localization} holds for environments with probability tending to 1 as $n\to \infty$: conditioned on survival up to time $n$ we have that ever since $o(n)$ steps the simple random walk is localized in a region of volume poly-logarithmic in $n$ with probability tending to 1. The previous best result of this type went back to Sznitman (1996) on Brownian motion among Poisson obstacles, where a localization (only for the end point) in a region of volume $t^{o(1)}$ was derived conditioned on the survival of Brownian motion up to time $t$.

math.PR↗