Tensor estimation with structured priors

TL;DR

This paper analyzes rank-one symmetric tensor estimation with structured priors, revealing phase transition behaviors and deriving the minimum-mean-square-error in high-dimensional regimes using a variational approach.

Contribution

It introduces a tractable variational framework for tensor estimation with structured signals, extending understanding of phase transitions and error limits in high dimensions.

Findings

Mutual information expressed as a finite-dimensional variational problem.

Critical SNR decreases with increasing latent space dimensions.

Derived the limiting tensor estimation problem as the ratio approaches zero.

Abstract

We consider rank-one symmetric tensor estimation when the tensor is corrupted by Gaussian noise and the spike forming the tensor is a structured signal coming from a generalized linear model. The latter is a mathematically tractable model of a non-trivial hidden lower-dimensional latent structure in a signal. We work in a large dimensional regime with fixed ratio of signal-to-latent space dimensions. Remarkably, in this asymptotic regime, the mutual information between the spike and the observations can be expressed as a finite-dimensional variational problem, and it is possible to deduce the minimum-mean-square-error from its solution. We discuss, on examples, properties of the phase transitions as a function of the signal-to-noise ratio. Typically, the critical signal-to-noise ratio decreases with increasing signal-to-latent space dimensions. We discuss the limit of vanishing ratio of signal-to-latent space dimensions and determine the limiting tensor estimation problem. We also point out similarities and differences with the case of matrices.