betaFeedback

WorksMathematics

Estimating Hessian Log-Determinants without Forming the Hessian Matrix

Abstract

The log-determinant of a positive-definite Hessian matrix is used in uncertainty assessment, Bayesian model comparison, and experimental design. Computing this quantity can be difficult when function values are available but derivatives are not. We combine simultaneous perturbation stochastic approximation with stochastic Lanczos quadrature (SPSA-SLQ) to estimate the log-determinant without forming the Hessian. Possibly noisy function values suffice for fixed-data and expected Hessians, including Fisher information matrices. We derive a local bound on log-determinant error due to Hessian-vector product approximation, quadrature, and averaging random probes. Three numerical studies examine accuracy and computational cost. In a 3002-parameter example, SPSA-SLQ gives a relative root-mean-square error below 1% and uses 65% less peak GPU memory than a basic full-matrix estimator. A sensor-placement study also compares improved full-matrix methods and evaluates layout selection.

Paper details

vlnc:W_9t2snzaq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
arXiv nonexclusive license
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…