A generalization of variable elimination for separable inverse problems beyond least squares

TL;DR

This paper introduces a generalized variable elimination method for separable inverse problems that extends beyond least squares, enabling faster and more robust solutions for complex, constrained, and non-Gaussian problems.

Contribution

It proposes a novel framework that generalizes variable elimination, applicable to a wider range of inverse problems including those with Poisson likelihoods and bounds.

Findings

Significant speed improvements in exponential sum fitting.

Enhanced robustness in blind deconvolution tasks.

Effective extension to bound-constrained and Poissonian problems.

Abstract

In linear inverse problems, we have data derived from a noisy linear transformation of some unknown parameters, and we wish to estimate these unknowns from the data. Separable inverse problems are a powerful generalization in which the transformation itself depends on additional unknown parameters and we wish to determine both sets of parameters simultaneously. When separable problems are solved by optimization, convergence can often be accelerated by elimination of the linear variables, a strategy which appears most prominently in the variable projection methods due to Golub, Pereyra, and Kaufman. Existing variable elimination methods require an explicit formula for the optimal value of the linear variables, so they cannot be used in problems with Poisson likelihoods, bound constraints, or other important departures from least squares. To address this limitation, we propose a generalization of variable elimination in which standard optimization methods are modified to behave as though a variable has been eliminated. We verify that this approach is a proper generalization by using it to re-derive several existing variable elimination techniques. We then extend the approach to bound-constrained and Poissonian problems, showing in the process that many of the best features of variable elimination methods can be duplicated in our framework. Tests on difficult exponential sum fitting and blind deconvolution problems indicate that the proposed approach can have significant speed and robustness advantages over standard methods.