Weingarten Formula in Random Matrix Theory

• Weingarten Formula is a method that computes joint moments of matrix entries from compact Lie groups using representation theory and combinatorial sums.
• It reduces high-dimensional Haar integrals to finite sums indexed by permutations or pair-partitions, with key coefficients called Weingarten functions.
• Its applications span random matrix theory, quantum field theory, and spectral analysis, enabling precise assessments of eigenvalue statistics and diagrammatic evaluations.

The Weingarten formula is a cornerstone of modern random matrix theory and representation theory, providing a systematic, representation-theoretic method for computing moments of matrix entries drawn from compact Lie groups (such as the unitary, orthogonal, and symplectic groups) with respect to the Haar probability measure. Originating in the context of quantum field theory and formalized rigorously through the combinatorial and algebraic advances of Weingarten, Collins, Śniady, Matsumoto, and others, the Weingarten calculus reduces high-dimensional Haar integrals to finite sums indexed by the symmetric group or set-partitions, with coefficients given by the so-called Weingarten functions—central objects encoding the intertwiner structure of the group representation.

1. Haar Integration and the Weingarten Expansion

Let $G$ be a compact matrix group (for example, $U(N)$ or $O(N)$), and $\mu$ its unique normalized Haar measure. Haar-invariant integration is characterized by bi-invariance; for any continuous $f:G\to \mathbb{C}$,

$$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$

for all $g_0 \in G$.

The Weingarten formula expresses joint moments of (possibly conjugated) matrix entries as an explicit sum over combinatorial data (permutations, pairings, or partitions), with coefficients determined by the inverse of a Gram matrix encoding the group’s invariant tensors. For the unitary group $U(N)$, the fundamental formula for the expectation of monomials of $U$ and its conjugates reads

$$\int_{U(N)} U_{i_1 j_1} \cdots U_{i_k j_k} \, \overline{U_{i_1' j_1'}} \cdots \overline{U_{i_k' j_k'}} \, d\mu(U) = \sum_{\sigma, \tau \in S_k} \biggl( \prod_{p=1}^k \delta_{i_p,\,i_{\sigma(p)}'} \prod_{p=1}^k \delta_{j_p,\,j_{\tau(p)}'} \biggr) \mathrm{Wg}_N(\sigma^{-1}\tau),$$

where $U(N)$0 is the symmetric group and $U(N)$1 is the unitary Weingarten function (Collins et al., 2021, Köstenberger, 2021, Ginory et al., 2016).

Analogous expressions exist for the orthogonal and symplectic groups, with sums replaced by pairings or other appropriate indexings, and for symmetric groups and quantum homogeneous spaces via specializations of the general formalism.

2. The Weingarten Function: Definitions and Character Expansions

For $U(N)$2, the Weingarten function $U(N)$3 is most elegantly defined as the matrix inverse of the Gram matrix $U(N)$4,

$U(N)$5

where $U(N)$6 denotes the permutation operator on $U(N)$7 (Collins et al., 2021, Köstenberger, 2021, Procesi, 2020). It follows that $U(N)$8 depends only on the conjugacy class (cycle–type) of its argument.

An equivalent formulation employs representation theory via irreducible characters $U(N)$9 of $O(N)$0 and Schur polynomials $O(N)$1: $O(N)$2 where $O(N)$3 is the dimension of the $O(N)$4 irrep $O(N)$5 (Procesi, 2020, Ginory et al., 2016). This expansion captures the full invariant-theoretic content, and analogous expansions in terms of Jack or zonal polynomials arise in the orthogonal and symplectic cases.

3. Orthogonal, Symplectic, and Symmetric Group Generalizations

For $O(N)$6 and $O(N)$7, the integration domain must respect the reality or quaternionic structure, respectively, leading to moment formulas indexed by pair-partitions (perfect matchings) of $O(N)$8 elements: $O(N)$9 with $\mu$0 the set of pairings and

$\mu$1

where $\mu$2, $\mu$3, and $\mu$4 are the combinatorial and representation-theoretic data of the Gelfand pair $\mu$5 (Collins et al., 2021, Matsumoto, 2010).

Weingarten calculus extends to symmetric group integration, permutation matrices (including centered cases), and further to noncommutative homogeneous spaces under "easiness" assumptions, where the Gram matrix is indexed by set-partitions and its inverse gives integration formulas for broad classes of easy quantum groups (Collins et al., 24 Mar 2025, Banica, 2016).

4. Derivation via Schur–Weyl Duality and Orthogonality Relations

The theoretical underpinning is Schur–Weyl duality, which states that the commutant of the action of $\mu$6 on $\mu$7 is generated by permutations, making the Gram matrix construction canonical. The Haar-average projection operator can be explicitly written using the permutation basis and diagonalized using character theory, with the Weingarten function as its matrix inverse (Köstenberger, 2021, Procesi, 2020).

Orthogonality relations for the Gram matrix lead to recursive or combinatorial interpretations. For example, the unitary Weingarten function satisfies a recursion: $\mu$8 connecting values for $\mu$9 and $f:G\to \mathbb{C}$0 (Collins et al., 2017). Markovian or "Weingarten process" interpretations for the path expansions of the Weingarten function further elucidate the probabilistic structure underpinning these matrices (Nissim, 21 Feb 2025).

5. Asymptotic Analysis and Computational Properties

In the "stable range" ($f:G\to \mathbb{C}$1), the leading asymptotics for the Weingarten function is controlled by Mӧbius or monotone path counts: $f:G\to \mathbb{C}$2 where $f:G\to \mathbb{C}$3 is the minimal transposition length (Collins et al., 2021, Nissim, 21 Feb 2025, Collins et al., 2017). For $f:G\to \mathbb{C}$4 the large-$f:G\to \mathbb{C}$5 limit is uniform over all permutations, with sharp error bounds available (Nissim, 21 Feb 2025). The evaluation of the Weingarten function is #P-hard in general but becomes tractable in large-$f:G\to \mathbb{C}$6 regimes or for low-degree moments.

Special cases, such as the full-cycle permutation, yield closed formulas involving Catalan numbers and rational functions in $f:G\to \mathbb{C}$7, underlying links to planar diagrammatics and free-probability theory (Procesi, 2020).

6. Applications, Extensions, and Computational Implementations

The primary application of the Weingarten formula lies in random matrix theory: one can compute all mixed moments of entries of Haar-distributed unitary (or orthogonal/symplectic) matrices, with extensions to noncommutative and "liberated" groups. This underpins the precise study of eigenvalue statistics, spectral convergence in random covers, and analysis of sampling over symmetric or quantum homogeneous spaces (Collins et al., 2021, Collins et al., 24 Mar 2025, Banica, 2016). The formula is central for Itzykson–Zuber integrals, for diagrammatic evaluations in quantum information, and in cumulant expansions of free-probability theory.

Optimized algorithms using symmetries and “graph presentations” dramatically reduce computational cost, as implemented in packages such as IntHaar, enabling rapid computation of moderate-degree moments (Ginory et al., 2016). The virtual isometry approach further yields recursive convolution formulas, allowing efficient evaluation of Weingarten functions across dimensions and direct treatment of high-degree integrals even in the numerically near-singular regime (Collins et al., 24 Oct 2025).

7. Worked Examples and Explicit Values

For low values of $f:G\to \mathbb{C}$8, Weingarten functions admit compact rational expressions. For $f:G\to \mathbb{C}$9:

• $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$0: $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$1;
• $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$2: $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$3, $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$4;
• $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$5: $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$6, etc. (Collins et al., 2021, Ginory et al., 2016, Köstenberger, 2021).

The orthogonal and symplectic cases are similarly explicit for small $$\int_G f(g_0g)\,d\mu(g) = \int_G f(g)\,d\mu(g), \quad \int_G f(gg_0)\,d\mu(g) = \int_G f(g)\,d\mu(g),$$7 in terms of matchings and zonal polynomials (Matsumoto, 2010).

Application to moments of centered permutation matrices reveals that centering introduces accelerated decay in high moments, with the decay rate determined by the occurrence of singleton blocks in the index-joining partition; this behavior is central to proofs of strong operator-norm convergence in random graph covering models (Collins et al., 24 Mar 2025).

---

For further comprehensive accounts, see the foundational works of Collins and Śniady, Collins and Matsumoto, and the synthesis presented in (Collins et al., 2021, Köstenberger, 2021, Nissim, 21 Feb 2025, Ginory et al., 2016), and (Collins et al., 24 Oct 2025).