Identifiability Scaling Laws in Bilinear Inverse Problems

TL;DR

This paper develops a unifying framework for analyzing the identifiability of bilinear inverse problems, deriving scaling laws that relate probability of successful recovery to problem complexity and null space properties.

Contribution

It introduces a flexible approach connecting identifiability in bilinear inverse problems to low-rank matrix recovery, providing deterministic conditions and scaling laws for various signal distributions.

Findings

Scaling laws accurately predict identifiability probability.

Null space characterization aids in understanding problem difficulty.

Numerical experiments validate theoretical predictions.

Abstract

A number of ill-posed inverse problems in signal processing, like blind deconvolution, matrix factorization, dictionary learning and blind source separation share the common characteristic of being bilinear inverse problems (BIPs), i.e. the observation model is a function of two variables and conditioned on one variable being known, the observation is a linear function of the other variable. A key issue that arises for such inverse problems is that of identifiability, i.e. whether the observation is sufficient to unambiguously determine the pair of inputs that generated the observation. Identifiability is a key concern for applications like blind equalization in wireless communications and data mining in machine learning. Herein, a unifying and flexible approach to identifiability analysis for general conic prior constrained BIPs is presented, exploiting a connection to low-rank matrix recovery via lifting. We develop deterministic identifiability conditions on the input signals and examine their satisfiability in practice for three classes of signal distributions, viz. dependent but uncorrelated, independent Gaussian, and independent Bernoulli. In each case, scaling laws are developed that trade-off probability of robust identifiability with the complexity of the rank two null space. An added appeal of our approach is that the rank two null space can be partly or fully characterized for many bilinear problems of interest (e.g. blind deconvolution). We present numerical experiments involving variations on the blind deconvolution problem that exploit a characterization of the rank two null space and demonstrate that the scaling laws offer good estimates of identifiability.