Approximating the Permanent with Fractional Belief Propagation

We discuss schemes for exact and approximate computations of permanents, and compare them with each other. Specifically, we analyze the Belief Propagation (BP) approach and its Fractional Belief Propagation (FBP) generalization for computing the permanent of a non-negative matrix. Known bounds and conjectures are verified in experiments, and some new theoretical relations, bounds and conjectures are proposed. The Fractional Free Energy (FFE) functional is parameterized by a scalar parameter $\gamma\in[-1;1]$, where $\gamma=-1$ corresponds to the BP limit and $\gamma=1$ corresponds to the exclusion principle (but ignoring perfect matching constraints) Mean-Field (MF) limit. FFE shows monotonicity and continuity with respect to $\gamma$. For every non-negative matrix, we define its special value $\gamma_\in[-1;0]$ to be the $\gamma$ for which the minimum of the $\gamma$-parameterized FFE functional is equal to the permanent of the matrix, where the lower and upper bounds of the $\gamma$-interval corresponds to respective bounds for the permanent. Our experimental analysis suggests that the distribution of $\gamma_$ varies for different ensembles but $\gamma_$ always lies within the $[-1;-1/2]$ interval. Moreover, for all ensembles considered the behavior of $\gamma_$ is highly distinctive, offering an emprirical practical guidance for estimating permanents of non-negative matrices via the FFE approach.

💡 Research Summary

The paper tackles the notoriously hard problem of computing the permanent of a non‑negative matrix, a task that belongs to the #P‑complete class and for which exact algorithms such as Ryser’s formula require O(n 2ⁿ) operations. While the Fully Polynomial‑Randomized Approximation Scheme (FPRAS) of Jerrum, Sinclair and Vigoda provides a provably accurate estimator, its high polynomial degree makes it impractical for many real‑world applications. In this context the authors revisit a graphical‑model perspective: the permanent can be expressed as the partition function of a bipartite factor graph whose binary edge variables encode perfect matchings.

The classical Belief Propagation (BP) approach approximates this partition function by minimizing the Bethe Free Energy (BFE). The BFE depends on a doubly‑stochastic matrix β of edge‑marginals and takes the form
F_BP(β|p)=∑_{ij}