---
title: Loop Equations in Matrix, Gauge & Tensor Models
url: https://www.emergentmind.com/topics/loop-equations
type: topic
---

# Loop Equations in Matrix, Gauge & Tensor Models

Loop equations are exact relations among observables—typically resolvents, traces, or Wilson loops—obtained from integration by parts, infinitesimal changes of variables, or Haar-measure invariance. In different domains they appear as Schwinger–Dyson equations, Ward identities, Nekrasov equations, Makeenko–Migdal equations, or Virasoro/\(\mathcal W\)-constraints, but their common role is to organize full hierarchies of correlation functions rather than isolated moments [1602.01715][2108.07710][2409.03705]. In modern usage, the term covers both finite-dimensional algebraic identities for matrix or tensor integrals and analytic hierarchies on spectral curves, differential systems, and microscopic random-matrix limits.

## 1. General definition and structural form

In random matrix theory, a basic source of loop equations is the eigenvalue measure
\[
d\mu(x_1,\dots,x_N)\propto \prod_{1\le i<j\le N}(x_i-x_j)^2 \prod_{i=1}^N e^{-V(x_i)}\,dx_i,
\]
whose polynomial moments satisfy linear constraints because total derivatives integrate to zero [1909.09372]. Writing \(p_k=\sum_i x_i^k\) and \(p_\mu=\prod_j p_{\mu_j}\), the one-matrix loop-equation generators are
\[
\begin{aligned}
Q_\mu &= \sum_{j=0}^{d} t_{j+1} p_{\mu_1+j}p_{\mu_2}\dots p_{\mu_n} \\
&\quad - \sum_{j=0}^{\mu_1-1} p_j p_{\mu_1-1-j}p_{\mu_2}\dots p_{\mu_n} \\
&\quad - \sum_{i=2}^n \mu_i \ p_{\mu_1+\mu_i-1} \prod_{k\neq j} p_{\mu_k},
\end{aligned}
\]
and they arise from the identity
\[
Q_\mu(X)\Delta(X)^2 e^{-N\operatorname{Tr}V(X)}
= \sum_{i=1}^N \frac{\partial}{\partial x_i}
\left(
x_i^{\mu_1}p_{\mu_2}(X)\cdots p_{\mu_n}(X)\Delta(X)^2 e^{-N\operatorname{Tr}V(X)}
\right).
\]
This is the archetypal Schwinger–Dyson mechanism: loop equations are the vanishing conditions \(E(\mathcal L)=0\) for a distinguished subspace \(\mathcal L\subset\mathcal P_N\) [1909.09372].

A more geometric formulation replaces moments by multidifferentials \(\omega_n^g\) near cuts of a local spectral curve. There, linear loop equations require the symmetric part
\[
\mathcal S f(z)=f(z)+(\iota^*f)(z)
\]
to extend holomorphically near the cuts, while quadratic loop equations require
\[
\mathcal Q_n^g(z;z_I)=
\omega_{n+1}^{g-1}(z,\iota(z),z_I)+
\sum_{J\subseteq I,\ 0\le h\le g}
\omega_{|J|+1}^h(z,z_J)\,
\omega_{n-|J|}^{g-h}(\iota(z),z_{I\setminus J})
\]
to have double zeros at ramification points [1303.5808]. This abstraction isolates the local analytic content common to matrix models, repulsive particle systems, loop models, knot invariants, and Liouville-type theories.

## 2. Finite-\(N\) matrix models and completeness of the equations

In finite-\(N\) matrix models, loop equations are not merely necessary constraints. For one-matrix models with polynomial \(V\), every linear functional on symmetric polynomials that satisfies the loop equations is realized by integrating against an actual random-matrix eigenvalue measure on a suitable integration cycle [1909.09372]. The contour is not restricted to \(\mathbb R\): admissible complex Jordan arcs, or linear combinations of their homology classes, are required in order to capture all solutions. The admissible one-particle homology has dimension
\[
\dim H_1(e^{-V(x)}dx,K)=d,
\]
and the \(N\)-particle space has
\[
\dim H_N=\binom{N+d-1}{N}.
\]
The theorem
\[
\mathbb E:H_N\stackrel{\sim}{\longrightarrow}\mathcal L^\perp
\]
identifies the full solution space of the loop equations with the homology of admissible contours [1909.09372]. A standard source of apparent nonuniqueness is therefore geometric rather than pathological: different contours correspond to different solutions.

For the arbitrary-\(\beta\) two-matrix model, exact loop equations still exist, but the leading spectral object is no longer an algebraic curve. The master loop equation takes the form
\[
\Bigl( y - V_1'(x) + \frac{T}{N} W(x) + \hbar\, \partial_x \Bigr) U_0(x,y)
= \bigl(V_2'(y)-x\bigr) W(x) - P_0(x,y) + N - \frac{T}{N} U_1(x,y;x),
\]
with
\[
\hbar = \frac{T}{N}\left(1-\frac{1}{\beta}\right),
\]
so the \(\beta\)-deformation promotes the classical spectral curve to a quantum spectral curve
\[
E\!\left(x,\,V_1'(x)-\hat y\right)\psi(x)=0,
\qquad
\hat y=\frac{T}{N}\partial_x,
\qquad
[\hat y,x]=\frac{T}{N}
\]
[1106.0332]. In the completely degenerate sector, \(\psi(x)=\prod_{i=1}^N(x-s_i)\), the loop equations reduce to Bethe-type equations
\[
(V_2'(B)-S)e=0,\qquad e^T(V_1'(S)-B)=0,\qquad [S,B]=\frac{T}{N}(I-ee^T),
\]
and higher corrections are computed by a matrix-valued topological recursion [1106.0332].

## 3. Abstract loop equations, differential systems, and topological recursion

A decisive conceptual shift is that loop equations do not require an underlying matrix integral. For any differential system
\[
d\Psi=\Phi\,\Psi
\]
with \(\Psi\) valued in a Lie group and \(\Phi\) a \(\mathfrak g\)-valued 1-form on a Riemann surface, one can define correlators \(W_n\) from the flat transport kernel and the adjoint-flat section
\[
M(\tilde x\cdot E)=\Psi(\tilde x)E\Psi(\tilde x)^{-1},
\]
and these correlators satisfy universal loop equations expressed through Lie-theoretic Casimirs and the characteristic polynomial of \(\rho(\Phi)\) [1602.01715]. The master identity is
\[
\sum_{k=0}^r (-1)^k y^{r-k}\, W_{k;n}(C_k(x);X_1,\dots,X_n)
=
[\epsilon_1\cdots \epsilon_n]\,
\det_\rho\Bigl(y-\bigl(\Phi(x)+\mathcal M_\epsilon(x;X_1,\dots,X_n)\bigr)\Bigr),
\]
which realizes the same formal structure as matrix-model loop equations and \(\mathcal W\)-algebra Ward identities [1602.01715].

For \(2\times 2\) traceless systems, equivalently scalar second-order ODEs, the Christoffel–Darboux kernel
\[
K(x_1,x_2)=\frac{\psi(x_1)\tilde\phi(x_2)-\tilde\psi(x_1)\phi(x_2)}{x_1-x_2}
\]
generates determinantal correlators that satisfy the exact loop equations
\[
-\,P_{n+1}(x;L)
=
W_{n+2}(x,x,L)
+\sum_{J\subset L} W_{1+|J|}(x,J)\,W_{1+n-|J|}(x,L/J)
+\sum_{j=1}^n \frac{d}{dx_j}\,
\frac{W_n(x,L/\{x_j\})-W_n(L)}{x-x_j},
\]
with \(P_{n+1}\) rational in \(x\) and determined by the differential operator \(D(x)\) [0901.3273]. Under a topological \(1/N\)-expansion, these exact identities feed directly into Eynard–Orantin topological recursion [0901.3273].

The abstract formulation makes this universality precise: if \(\omega_n^g\) satisfy solvable linear loop equations and quadratic loop equations, then they are given by the residue formula
\[
\omega_n^g(z_0,z_I)
=
\sum_{\alpha\in \Gamma^{\mathrm{fix}}}
\mathop{\mathrm{Res}}_{z\to \alpha}
K(z_0,z)\,\mathcal E_n^g(z,\iota(z);z_I),
\]
which is exactly topological recursion [1303.5808]. The same logic underlies the passage from local topological-recursion constraints to global Virasoro constraints for the Gromov–Witten theory of \(\mathbb P^1\), where the logarithmic spectral datum
\[
x(z)=z+\frac1z,\qquad y(z)=\ln z
\]
forces regularized open periods \(\mathcal I_b^0\) in addition to Laurent-coefficient functionals \(\mathcal I_b^1\) [1905.01890]. In a different direction, the GUE partition function satisfies a Dubrovin–Zhang loop equation that coincides with the loop equation of the nonlinear Schrödinger Frobenius manifold, and
\[
Z(x,s;\epsilon)=
Z^{\mathrm{NLS}}_{\mathrm{top}}(t;\epsilon)\Big|_{t^{1,m}=(m+1)!s_{m+1},\; t^{2,m}=x\delta_{m,0}}
\]
[2407.19170].

## 4. Microscopic loop equations and universal random-matrix statistics

At microscopic scale, loop equations can become complete characterization principles for limiting point processes. The recent theorem “Loop Equations Characterize Random Matrix Statistics” proves that for every rational \(\beta>0\), the \(\mathrm{Sine}_\beta\) point process is the unique solution of the bulk loop equation hierarchy, and the \(\mathrm{Airy}_\beta\) point process is the unique solution of the edge loop equation hierarchy [2607.07617]. In this setting a point configuration \(P\subset\mathbb R\) is encoded by a particle-generated Nevanlinna function
\[
s(w)=b+cw+\int_{\mathbb R}\left(\frac{1}{x-w}-\frac{x}{1+x^2}\right)\mu(x),
\]
with \(\mu=\sum_{x\in P}\delta_x\). This representation is stable under local-uniform limits and treats bulk and edge in a unified way [2607.07617].

The bulk hierarchy is
\[
E\left[ \left( \frac{2-\beta}{\beta}\,\partial_{w_1}s(w_1) +s(w_1)^2+1 \right) \prod_{j=2}^p s(w_j) \right]
+\frac{2}{\beta}\, E\left[ \sum_{j=2}^p \partial_{w_j} \frac{s(w_1)-s(w_j)}{w_1-w_j} \prod_{\substack{\ell=2\\ \ell\neq j} }^p s(w_\ell) \right] =0,
\]
and the edge hierarchy is the same identity with \(+1\) replaced by \(-w_1\) [2607.07617]. The distinction reflects the difference between bulk \(N^{-1}\)-scaling and soft-edge \(N^{-2/3}\)-scaling.

The corresponding uniqueness theorems are exact:
\[
\textbf{Theorem 1.4:}\quad
\text{If }\beta>0\text{ is rational and }s\text{ is particle-generated and satisfies the bulk hierarchy,}
\]
then the poles of \(s\) have the law of \(\mathrm{Sine}_\beta\); and
\[
\textbf{Theorem 1.8:}\quad
\text{If }\beta>0\text{ is rational and }s\text{ is particle-generated and satisfies the edge hierarchy,}
\]
then the poles of \(s\) have the law of \(\mathrm{Airy}_\beta\) [2607.07617]. The proof linearizes the nonlinear hierarchy using exponential observables, producing deformed Calogero–Moser–Sutherland systems; establishes local laws such as
\[
\mathbb E\bigl[|s(w)-i|^{2q}\bigr]\le \frac{(Cq)^q}{\eta^{2q}}
\]
in the bulk; analyzes the resulting holonomic PDE systems; and reconstructs the full family of resolvent correlation functions [2607.07617].

This yields a direct route to universality. If a microscopic ensemble satisfies approximate loop equations together with tightness and local moment bounds, then every subsequential limit satisfies the exact hierarchy, hence must be \(\mathrm{Sine}_\beta\) or \(\mathrm{Airy}_\beta\) [2607.07617]. For Wigner matrices the approximate bulk loop equation is obtained from
\[
1+zm_N(z)=\frac1N\sum_{i,j=1}^N h_{ij}G_{ji}(z)
\]
and the cumulant expansion
\[
E[hf(h)] = \frac1N E[f'(h)] + \mathcal R,
\]
while for random \(d\)-regular graphs ordinary integration by parts is replaced by a discrete switching calculus [2607.07617]. The paper also notes a sharp limitation: at \(\beta=0\) the bulk hierarchy degenerates and is satisfied by all homogeneous Poisson processes and their mixtures [2607.07617].

## 5. Multi-level, discrete, and generalized eigenvalue models

Loop equations extend beyond one-level eigenvalue gases. For \(\beta\)-corners processes, the natural observables are the levelwise Stieltjes transforms
\[
\mathcal G^j(z)=\sum_{i=1}^j \frac{1}{z-Y_i^j}
\qquad\text{and}\qquad
G_L^j(z)=\sum_{i=1}^j \frac{1}{z-\ell_i^j/L},
\]
and the loop equations become relations among joint cumulants across levels [2108.07710]. In the continuous setting, for the triangular interlacing array \(Y^N\succ \cdots \succ Y^k\), the contour identity
\[
\oint_{\Gamma}   \frac{(z-a_-)(z- a_+)  }{2 \pi i  (z - v)}  \cdot \left[ \frac{ \bigl((z);\mathfrak{M}\bigr)}{2} +\sum_{(r,f) \in \mathfrak{M}} \frac{ \bigl( \mathcal{G}^r(z) ; \mathfrak{M} \setminus \{ (r,f)\} \bigr)}{(z - v^r_f )^2} \right] = 0
\]
encodes the full multi-level Schwinger–Dyson hierarchy [2108.07710]. When \(k=N\), this reduces to the one-level Borot–Guionnet loop equations; the genuinely new feature is the appearance of adjacent-level combinations such as
\[
[\mathcal G^j(z)-\mathcal G^{j-1}(z)]^2.
\]

In the discrete setting, the same role is played by multi-level Nekrasov equations. For \(\theta\neq 1\), analytic functions \(R_1(z)\) and \(R_2(z)\) are constructed from products of the form
\[
\prod_{p = 1}^{j} \frac{z- \ell^{j}_p  -\theta}{z - \ell_p^{j} - 1}
\prod_{p = 1}^{j-1} \frac{z-  \ell_p^{j-1} + \theta - 1}{z -  \ell_p^{j-1}},
\]
and for \(\theta=1\) these multiplicative observables collapse to additive differences of resolvents [2108.07710]. Deformation of the measure by factors
\[
\prod_{j=k}^N\prod_{r=1}^j\prod_{i=1}^{m_j}
\left(1+\frac{t_i^j}{v_i^j-\ell_r^j/L}\right)
\]
turns derivatives with respect to \(t_i^j\) into joint cumulants, giving a systematic cumulant hierarchy [2108.07710].

A different generalization replaces the Vandermonde interaction by an arbitrary difference-type measure,
\[
Z=\int\limits_{-\infty }^{+\infty }\prod_{i=1}^N da_{i}\prod_{i<j}\mu(a_{i}-a_{j})\,e^{-N\sum\limits_{i}V(a_{i})}.
\]
Defining
\[
\gamma(a)=a\frac{d}{da}\log\mu(a),\qquad
R(a)=\frac{1}{2}\,\frac{d}{da}\log\mu(a),
\]
one obtains the exact loop equation
\[
V'\!\left(\frac{d}{dx}\right)\!\mathbb{W}(x) =
\frac{1}{2}\int\limits_{-\infty }^{+\infty }\frac{d\omega}{2\pi}\hat{\gamma}(\omega)\!\int_{0}^{x}\!ds\, \left[ \mathbb{W}(s-i\omega)\mathbb{W}(x-s+i\omega) +\frac{\mathbb{W}(s-i\omega,x-s+i\omega)}{N^{2}} \right],
\]
which is nonlocal because of the imaginary shifts in the spectral parameter [2402.13835]. In the planar limit this framework yields an exact solution for the one-parameter \(\nu\)-models, while at strong coupling it leads to a Wiener–Hopf/Riemann–Hilbert analysis with applications to \(\mathcal N=2^*\) super-Yang–Mills theory and the Hoppe model [2402.13835].

## 6. Gauge theory, tensor models, quivers, and combinatorial formulations

In lattice Yang–Mills theory, loop equations govern Wilson loops
\[
\langle w(C) \rangle = {1\over N} \langle W(C) \rangle,
\qquad
w({\rm tr}(\mathds{1})) = 1,
\]
through Schwinger–Dyson equations and trace relations [2601.04316]. For SU(2), the basic trace relation
\[
{\rm tr}(X){\rm tr}(Y)={\rm tr}(X Y)+{\rm tr}(X Y^\dagger)
\]
becomes
\[
w(C_1) w(C_2)  = {1\over2} w(C_1\, C_2) + {1\over2} w(C_1\, \bar C_2).
\]
A central issue is the determination of a dynamically independent truncated basis for the lattice positivity bootstrap. The 2026 framework distinguishes **direct equations**, generated locally by plaquette-cut and subloop-cut algorithms, from **indirect equations**, which arise only after eliminating leaked higher-length loops [2601.04316]. In 2D, for example, the truncation \(S_{[12]}\) contains 119 canonical loops and 51 direct equations, but enlarging to \(S_{[14]}\) yields 53 independent projected equations, exposing two hidden relations [2601.04316].

Tensor models admit analogous exact identities. Unitary invariants are indexed by colored bipartite bubbles, and reparametrization invariance yields tensor-model loop equations in which ordinary traces are replaced by bubble invariants [1603.00172]. The formalism parallels matrix models, but the basic combinatorial objects are \(D\)-bubbles rather than loops on a line or plane. In tensorial group field theory, the same invariants are lifted from finite tensors to fields \(\Phi(g_1,\dots,g_D)\) on \(D\) copies of a group, often with a closure constraint [1603.00172].

For quiver-based noncommutative geometries, a path integral over Dirac operators produces loop equations that generalize Makeenko–Migdal from lattices to arbitrary graphs [2409.03705]. The spectral action
\[
S(D)=\Tr_H f(D)
=
\sum_{k=1}^d f_k\, \sum_{v\in Q_0}
\sum_{\substack{\gamma\in \Omega_v(Q)\\ \ell(\gamma)=k}} \Tr \hol\gamma
\]
is itself a sum over closed paths, and the exact quiver loop equations relate splitting of Wilson loops at a rooted edge to insertions of generalized plaquettes from the action [2409.03705]. In the triangle-quiver example, the model reduces to the Gross–Witten–Wadia integral
\[
\mathcal Z_{Q,B} = \int_{U(N)} e^{-Nx\Tr(U+U^*)}\, dU,
\]
providing an explicit check of the loop-equation-plus-positivity bootstrap [2409.03705].

Combinatorial and noncommutative master equations also arise in matrix models for random maps. For the 3-state Potts model coupled to 2D gravity, a single noncommutative generating equation
\[
\begin{aligned}
\Phi=1 &+ x_0 \Phi\,(x_0 + c x_1 + c x_2)\, \Phi  + g(x_0 + cx_1 +cx_2)\Delta_0^2 \Phi \\
&+ x_1 \Phi\,(cx_0 +  x_1 + cx_2)\, \Phi  + g(cx_0+x_1+cx_2)\Delta_1^2 \Phi \\
&+ x_2 \Phi\,(cx_0 + c x_1 + x_2)\, \Phi  + g(cx_0+cx_1+x_2)\Delta_2^2 \Phi
\end{aligned}
\]
packages the full planar hierarchy of loop equations for colored boundary words [1911.00032]. Its coefficient extractions yield a closed system determining the fixed-spin spectral curve.

Taken together, these developments show that loop equations are not a single formal identity but a family of exact constraint hierarchies whose meaning depends on the observables and the ambient category: polynomial moments on contour homology classes, multidifferentials on local spectral curves, resolvents on interlacing arrays, Wilson loops on lattices or quivers, bubble observables in tensor models, or microscopic transforms of point processes. Their scope ranges from finite-\(N\) reconstruction theorems to universality at microscopic scale, and from classical Schwinger–Dyson derivations to holonomic PDE systems, topological recursion, and bootstrap formulations [1909.09372][1303.5808][2607.07617].

Source: https://www.emergentmind.com/topics/loop-equations