Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Double-Cover-Based Analysis of the Bethe Permanent of Block-Structured Positive Matrices

Binghong Wu, Pascal O. Vontobel

TL;DR

This work addresses the ratio between the exact Gibbs permanent $\mathrm{perm}(\mathbf{A})$ and its Bethe surrogate $\mathrm{perm}_{\mathrm{B}}(\mathbf{A})$ for block-structured positive matrices. It develops a framework based on normal factor graphs, double covers, and cycle-index generating functions to analyze how this ratio behaves when $\mathbf{A}=\mathbf{A}(\mathbf{B},\bm{k},\bm{\ell})$, a block-constant expansion of $\mathbf{B}$. The main technical result provides an asymptotic expression for $\mathrm{perm}(\mathbf{A})/\mathrm{perm}_{\mathrm{B},2}(\mathbf{A})$ as $n\to\infty$, namely $\mathrm{perm}(\mathbf{A})/\mathrm{perm}_{\mathrm{B},2}(\mathbf{A}) \sim \sqrt[4]{\frac{\pi n}{\mathrm{e}}}\cdot \sqrt[4]{\frac{1}{\prod_{i=2}^{m} \mathrm{e}^{\rho_i}(1-\rho_i)}}$, with $\rho_i$ derived from the spectrum of a symmetric matrix $\mathbf{S}(\bm{t},\bm{u})$ at a dominant saddle point. Numerical experiments corroborate the asymptotics and reveal concentration phenomena similar to those observed in previous i.i.d. settings. The results illuminate why Bethe surrogates are reliable in PML-like structured instances and connect the saddle-point parameters to Sinkhorn-type scalings, suggesting future work to tighten links to the full Bethe permanent and practical PML applications.

Abstract

We consider the permanent of a square matrix with non-negative entries. A tractable approximation is given by the so-called Bethe permanent that can be efficiently computed by running the sum-product algorithm on a suitable factor graph. While the ratio of the permanent of a matrix to its Bethe permanent is, in the worst case, upper and lower bounded by expressions that are exponentially far apart in the matrix size, in practice it is observed for many ensembles of matrices of interest that this ratio is strongly concentrated around some value that depends only on the matrix size. In this paper, for an ensemble of block-structured matrices where entries in a block take the same value, we numerically study the ratio of the permanent of a matrix to its Bethe permanent. It is observed that also for this ensemble the ratio is strongly concentrated around some value depending only on a few key parameters of the ensemble. We use graph-cover-based approaches to explain the reasons for this behavior and to quantify the observed value.

Double-Cover-Based Analysis of the Bethe Permanent of Block-Structured Positive Matrices

TL;DR

This work addresses the ratio between the exact Gibbs permanent and its Bethe surrogate for block-structured positive matrices. It develops a framework based on normal factor graphs, double covers, and cycle-index generating functions to analyze how this ratio behaves when , a block-constant expansion of . The main technical result provides an asymptotic expression for as , namely , with derived from the spectrum of a symmetric matrix at a dominant saddle point. Numerical experiments corroborate the asymptotics and reveal concentration phenomena similar to those observed in previous i.i.d. settings. The results illuminate why Bethe surrogates are reliable in PML-like structured instances and connect the saddle-point parameters to Sinkhorn-type scalings, suggesting future work to tighten links to the full Bethe permanent and practical PML applications.

Abstract

We consider the permanent of a square matrix with non-negative entries. A tractable approximation is given by the so-called Bethe permanent that can be efficiently computed by running the sum-product algorithm on a suitable factor graph. While the ratio of the permanent of a matrix to its Bethe permanent is, in the worst case, upper and lower bounded by expressions that are exponentially far apart in the matrix size, in practice it is observed for many ensembles of matrices of interest that this ratio is strongly concentrated around some value that depends only on the matrix size. In this paper, for an ensemble of block-structured matrices where entries in a block take the same value, we numerically study the ratio of the permanent of a matrix to its Bethe permanent. It is observed that also for this ensemble the ratio is strongly concentrated around some value depending only on a few key parameters of the ensemble. We use graph-cover-based approaches to explain the reasons for this behavior and to quantify the observed value.
Paper Structure (21 sections, 10 theorems, 96 equations, 5 figures)

This paper contains 21 sections, 10 theorems, 96 equations, 5 figures.

Table of Contents

  1. Introduction
  2. Background: NFGs, Double Covers, and the Cycle Index
  3. Normal factor graphs and partition sums
  4. NFGs for permanents
  5. Positive Block-Constant Model
  6. Numerical Validation
  7. Conclusion and Outlook
  8. Detailed NFG and Double-Cover Definitions
  9. NFG $\mathsf{N}(\mathbf{A})$ whose partition function equals $\mathop{\mathrm{perm}}\nolimits(\mathbf{A})$
  10. Graph Cover Characterization of the Bethe Approximation
  11. The explicit double cover $\hat{\mathsf{N}}(\mathbf{A})$ and its partition function
  12. Combinatorial Interpretation via the Cycle Penalty
  13. Factor-Graph View of the Trace Kernels $\mathbf{W}$ and $\mathbf{S}$
  14. Counting indeterminates, partitioning selection, and the weighted transition
  15. Symmetrization and the NFG kernel factorization
  16. ...and 6 more sections

Key Result

Lemma 4

Under Assumption assump:general_block and Definition def:S-general, let $(\bm{t},\bm{u})$ be such that $\bm{t},\bm{u}\in\mathbb{R}_{>0}^m$ and $\lambda_1\!\left(\mathbf{S}(\bm{t},\bm{u})\right) < 1$. Define Then

Figures (5)

  • Figure 1: Numerical results for block-structured matrices $\mathbf{A}$, along with theoretical asymptotics, for $n\!=\!5$ and $m\!=\!2$. (See Section \ref{['sec:intro']} for details.)
  • Figure 2: NFGs used to represent $\mathop{\mathrm{perm}}\nolimits(\mathbf{A})$ (left) and $\mathop{\mathrm{perm}}\nolimits_{\mathrm{B},2}(\mathbf{A})$ (right).
  • Figure 3: Numerical results discussed in Section \ref{['sec:numerics']}.
  • Figure 4: Trellis graph with two states over $7$ time steps.
  • Figure 5: OTB/CTB factor-graph derivation of the kernel factorization $\mathbf{W}\mapsto \mathbf{S}$ and the trace representation of length-$h$ closed walks.

Theorems & Definitions (22)

  • Definition 2
  • Definition 3
  • Lemma 4
  • proof
  • Lemma 5
  • proof
  • Proposition 6
  • proof
  • Theorem 7
  • proof
  • ...and 12 more