[Paper Review] A Condition Number for Hamiltonian Monte Carlo

Ian Langmore, M. Dikovsky+3 more|arXiv (Cornell University)|May 23, 2019

Markov Chains and Monte Carlo Methods27 references4 citations

TL;DR

This paper introduces a novel condition number for Hamiltonian Monte Carlo (HMC) that quantifies sampling efficiency in multivariate Gaussian targets by combining spectral and Schatten norms. It shows this condition number governs the number of leapfrog steps needed for efficient sampling and uses it to analyze preconditioning techniques, deriving asymptotic laws for inverse Wishart matrices and providing burn-in heuristics.

ABSTRACT

Hamiltonian Monte Carlo is a popular sampling technique for smooth target densities. The scale lengths of the target have long been known to influence integration error and sampling efficiency. However, quantitative measures intrinsic to the target have been lacking. In this paper, we restrict attention to the multivariate Gaussian and the leapfrog integrator, and obtain a condition number corresponding to sampling efficiency. This number, based on the spectral and Schatten norms, quantifies the number of leapfrog steps needed to efficiently sample. We demonstrate its utility by using this condition number to analyze HMC preconditioning techniques. We also find the condition number of large inverse Wishart matrices, from which we derive burn-in heuristics.

Motivation & Objective

• To develop a quantitative, intrinsic measure of sampling difficulty in HMC that depends on the target distribution’s scale structure.
• To define a condition number based on spectral and Schatten norms that captures the number of leapfrog steps required for efficient sampling in multivariate Gaussians.
• To analyze HMC preconditioning techniques using this condition number, comparing standardization, variational inference, and low-rank updates.
• To derive the asymptotic distribution of the condition number for inverse Wishart matrices, enabling practical burn-in heuristics.
• To connect HMC efficiency to random matrix theory, providing a foundation for generalization beyond Gaussian targets.

Proposed method

• Proposes a condition number κ that combines the largest and smallest eigenvalues of the covariance matrix, weighted by their contributions to integration error.
• Uses the leapfrog integrator to model HMC dynamics and derives an error bound dependent on the step size and eigenvalue spectrum.
• Applies the Schatten p-norm to account for all eigenvalues, not just the extreme ones, to better reflect cumulative integration error.
• Derives an asymptotic expression for the condition number of the inverse Wishart ensemble, linking it to the sample covariance matrix.
• Uses the condition number to evaluate and compare preconditioning strategies, including component-wise standardization and variational inference-based transforms.
• Employs random matrix theory to analyze the limiting distribution of κ under high-dimensional asymptotics.

Experimental results

Research questions

• RQ1 How can a condition number be defined that quantitatively measures the difficulty of sampling from a multivariate Gaussian using HMC?
• RQ2 What is the relationship between the eigenvalue spectrum of the covariance matrix and the number of leapfrog steps required for efficient sampling?
• RQ3 How does the proposed condition number compare to traditional spectral condition numbers in predicting HMC efficiency?
• RQ4 What are the asymptotic properties of the condition number when preconditioning with sample covariance matrices?
• RQ5 Can the condition number guide effective preconditioning strategies, including low-rank updates, for ill-conditioned Gaussian targets?

Key findings

• The proposed condition number κ, based on spectral and Schatten norms, provides a precise measure of the number of leapfrog steps needed for efficient HMC sampling in multivariate Gaussians.
• The condition number differs from the standard spectral condition number by accounting for all eigenvalues, which collectively contribute to integration error.
• For inverse Wishart matrices, the condition number asymptotically follows a known law derived from random matrix theory, enabling burn-in heuristics.
• Component-wise standardization is not always superior to preconditioning via variational inference, and both can be outperformed by low-rank updates in certain spectral regimes.
• Spectra with only a few large eigenvalues and many small ones are particularly detrimental to HMC efficiency, and low-rank preconditioners are shown to be effective in such cases.
• The condition number provides a rigorous link between HMC efficiency and random matrix theory, offering a framework for generalization beyond Gaussian targets.