Fundamental limits of Non-Linear Low-Rank Matrix Estimation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Mergny, Pierre, Ko, Justin, Krzakala, Florent, Zdeborová, Lenka
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910356304035840
author Mergny, Pierre
Ko, Justin
Krzakala, Florent
Zdeborová, Lenka
author_facet Mergny, Pierre
Ko, Justin
Krzakala, Florent
Zdeborová, Lenka
contents We consider the task of estimating a low-rank matrix from non-linear and noisy observations. We prove a strong universality result showing that Bayes-optimal performances are characterized by an equivalent Gaussian model with an effective prior, whose parameters are entirely determined by an expansion of the non-linear function. In particular, we show that to reconstruct the signal accurately, one requires a signal-to-noise ratio growing as $N^{\frac 12 (1-1/k_F)}$, where $k_F$ is the first non-zero Fisher information coefficient of the function. We provide asymptotic characterization for the minimal achievable mean squared error (MMSE) and an approximate message-passing algorithm that reaches the MMSE under conditions analogous to the linear version of the problem. We also provide asymptotic errors achieved by methods such as principal component analysis combined with Bayesian denoising, and compare them with Bayes-optimal MMSE.
format Preprint
id arxiv_https___arxiv_org_abs_2403_04234
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Fundamental limits of Non-Linear Low-Rank Matrix Estimation
Mergny, Pierre
Ko, Justin
Krzakala, Florent
Zdeborová, Lenka
Machine Learning
We consider the task of estimating a low-rank matrix from non-linear and noisy observations. We prove a strong universality result showing that Bayes-optimal performances are characterized by an equivalent Gaussian model with an effective prior, whose parameters are entirely determined by an expansion of the non-linear function. In particular, we show that to reconstruct the signal accurately, one requires a signal-to-noise ratio growing as $N^{\frac 12 (1-1/k_F)}$, where $k_F$ is the first non-zero Fisher information coefficient of the function. We provide asymptotic characterization for the minimal achievable mean squared error (MMSE) and an approximate message-passing algorithm that reaches the MMSE under conditions analogous to the linear version of the problem. We also provide asymptotic errors achieved by methods such as principal component analysis combined with Bayesian denoising, and compare them with Bayes-optimal MMSE.
title Fundamental limits of Non-Linear Low-Rank Matrix Estimation
topic Machine Learning
url https://arxiv.org/abs/2403.04234