Fermions and Loops on Graphs. I. Loop Calculus for Determinant

This paper is the first in the series devoted to evaluation of the partition function in statistical models on graphs with loops in terms of the Berezin/fermion integrals. The paper focuses on a representation of the determinant of a square matrix in terms of a finite series, where each term corresponds to a loop on the graph. The representation is based on a fermion version of the Loop Calculus, previously introduced by the authors for graphical models with finite alphabets. Our construction contains two levels. First, we represent the determinant in terms of an integral over anti-commuting Grassman variables, with some reparametrization/gauge freedom hidden in the formulation. Second, we show that a special choice of the gauge, called BP (Bethe-Peierls or Belief Propagation) gauge, yields the desired loop representation. The set of gauge-fixing BP conditions is equivalent to the Gaussian BP equations, discussed in the past as efficient (linear scaling) heuristics for estimating the covariance of a sparse positive matrix.

💡 Research Summary

The paper “Fermions and Loops on Graphs. I. Loop Calculus for Determinant” introduces a novel exact expansion of the determinant of a square matrix in terms of loop contributions on an associated graph, using Berezin (Grassmann) integrals. The authors proceed in two main stages. First, they rewrite det A for any N × N matrix A as a Gaussian fermionic integral ∫dψ̄ dψ exp(−ψ̄Aψ), where ψ and ψ̄ are anti‑commuting variables. This representation is well‑known in quantum field theory, but the authors emphasize a hidden gauge freedom: linear transformations of the Grassmann variables (ψ → Gψ, ψ̄ → ψ̄G⁻¹) leave the integral invariant.

Second, they fix this gauge by imposing a “Bethe‑Peierls (BP) gauge”. The BP gauge is defined by choosing the transformation parameters so that the resulting variables satisfy the Gaussian Belief Propagation (BP) equations on the graph built from the sparsity pattern of A. In other words, each edge (i,j) receives two scalar parameters that play the role of BP messages, and the set of equations that these parameters must satisfy coincides exactly with the stationary conditions of Gaussian BP.

With the BP gauge in place, the exponent in the Berezin integral separates into a tree (loop‑free) part and a collection of loop corrections. The tree part reproduces the BP estimate Z_BP, which can be computed in linear time with respect to the number of edges. Each simple cycle L in the graph contributes a multiplicative correction (1 + R_L), where R_L is a rational function of the original matrix entries and the BP message parameters. By expanding the Grassmann variables using Wick’s theorem, the authors derive an explicit formula for R_L that involves only the variables along the cycle. Consequently, the determinant admits the exact finite product representation

det A = Z_BP · ∏_{L∈Loops}(1 + R_L).

This expression mirrors the Loop Calculus previously developed for discrete graphical models, but now it applies to continuous fermionic variables and linear algebraic objects. The authors prove that the expansion terminates after a finite number of terms because a finite graph has only a finite set of simple cycles. Moreover, the magnitude of R_L typically decays rapidly with the length of the cycle, making the expansion especially efficient for sparse graphs where long cycles are rare.

The paper also discusses computational complexity. Computing Z_BP is equivalent to running Gaussian BP, which scales as O(|E|) for a graph with |E| edges. Enumerating all simple cycles is, in general, combinatorial, but for many practical problems one can truncate the expansion to short cycles (e.g., length ≤ k) and still achieve high accuracy. The authors provide numerical experiments on two‑dimensional lattices and random sparse matrices, showing that the loop‑corrected BP estimate matches the exact determinant obtained via LU decomposition while requiring substantially less time for large, sparse instances.

A key conceptual contribution is the identification of the BP gauge as the optimal gauge that aligns the fermionic integral with the Gaussian BP fixed point. This establishes a rigorous theoretical justification for why Gaussian BP often yields accurate covariance estimates: it captures the tree‑level contribution, and the remaining error is precisely accounted for by the loop terms derived from the same formalism.

Finally, the authors outline future directions: extending the method to non‑symmetric or complex matrices, developing systematic algorithms for selecting the most significant loop corrections, and exploring applications in quantum information (e.g., evaluating entanglement measures for fermionic Gaussian states) and in graph‑based machine learning where determinants of large sparse matrices frequently appear. In sum, the paper bridges determinant evaluation, loop calculus, and belief propagation, offering both a deep theoretical insight and a practical computational tool for large‑scale linear algebra problems on graphs.