08.14.06 · stat-mech / statistical-field-theory

Pointer: matrix models and the topological expansion

shipped3 tiersLean: none

Anchor (Master): Brezin-Itzykson-Parisi-Zuber 1978 genus expansion of the one-matrix model; link to random triangulated surfaces and 2D quantum gravity (Itzykson-Drouffe Vol. 2)

Intuition Beginner

In the large- $N$ vector model the number $N$ counted spin components. Here $N$ counts something different: the size of a square grid of numbers, an $N$ by $N$ matrix. A matrix model is a way of averaging over all such matrices with a chosen weight, the way ordinary statistics averages over single numbers.

The surprise is what the bookkeeping of this average looks like. Each interaction in a matrix model carries two indices, a row label and a column label, not one. So when you draw the diagrams that organise the average, every line splits into a double line, like a flat ribbon with two edges instead of a single thread.

A bundle of ribbons can be glued edge to edge into a surface: a sphere, a doughnut, a doughnut with two holes. The large- $N$ count sorts every diagram by which surface its ribbons tile.

Visual Beginner

The picture shows a single propagator drawn first as one thin line, then as a ribbon with two parallel edges carrying the row and column labels. Several ribbons are then glued along their edges to close up into a surface, and the same set of ribbons is shown tiling a sphere on the left and a one-holed doughnut on the right.

The takeaway: the more holes the surface has, the smaller the contribution at large $N$ . The flat, hole-free surface, the sphere, wins.

Worked example Beginner

Take a closed surface built from flat polygon faces, like a soccer ball or a cube. Count its corners $V$ , its edges $E$ , and its faces $F$ . Euler noticed that the combination $V - E + F$ always comes out the same for surfaces of the same shape.

For a cube: $8$ corners, $12$ edges, $6$ faces, so $V - E + F = 8 - 12 + 6 = 2$ . For a soccer ball with $60$ corners, $90$ edges, $32$ faces: $60 - 90 + 32 = 2$ again. Both are sphere-shaped, and both give $2$ .

Now puncture the surface into a doughnut shape with one hole. A simple doughnut grid has $V - E + F = 0$ . A two-holed shape gives $- 2$ .

What this tells us: the single number $V - E + F$ detects how many holes a surface has. It equals $2$ for the sphere, $0$ for one hole, $- 2$ for two holes, dropping by $2$ per hole. In a matrix model that very number becomes the power of $N$ a diagram carries, so counting holes is the same as counting powers of $N$ .

Check your understanding Beginner

Formal definition Intermediate+

A Hermitian one-matrix model is the probability weight $e^{- N tr V (M)}$ on the space of $N \times N$ Hermitian matrices $M$ , with partition function

Z_{N} = \int d M exp [- N tr V (M)], V (M) = \frac{1}{2} M^{2} + \frac{g}{4} M^{4} + \dots,

where $d M = \prod_{i} d M_{ii} \prod_{i < j} d (Re M_{ij}) d (Im M_{ij})$ is the flat $U (N)$ -invariant measure and the overall factor $N$ in the exponent is the matrix-model analogue of the $1/ N$ scaling that makes the large- $N$ limit of 08.14.03 sensible. Conventions follow Itzykson and Drouffe ^{[Itzykson-Drouffe Vol. 2]} and the original analysis of 't Hooft ^{['t Hooft 1974]}.

The Gaussian part fixes the propagator. Because $M$ is Hermitian, the free two-point function carries a row and a column index at each end,

⟨ M_{ij} M_{k l} ⟩_{0} = \frac{1}{N} δ_{i l} δ_{j k},

so a propagator is drawn as a double line (a ribbon): one strand carries the index $i$ flowing one way, the other carries $j$ flowing back. An interaction vertex from $tr M^{4}$ is a four-ribbon junction that conserves indices around its corner. A ribbon graph (equivalently a fat graph or map) is a Feynman graph in which every edge is a ribbon and the ribbons are glued so the index strands join into closed index loops; each such graph canonically thickens into a compact oriented surface whose genus $g$ is the number of handles.

For a connected ribbon graph with $V$ vertices, $E$ edges (ribbons), and $F$ index loops, the index loops bound faces of a cell decomposition of that surface, so $V - E + F = χ = 2 - 2 g$ is its Euler characteristic. The $N$ -counting is read directly off the weight: each vertex contributes $N$ (from the $N tr V$ in the action), each propagator contributes $1/ N$ , and each closed index loop contributes a free sum $\sum_{i} 1 = N$ . The net power is therefore $N^{V - E + F} = N^{2 - 2 g}$ .

Key theorem with proof Intermediate+

Theorem ('t Hooft genus grading). The connected free energy $F_{N} = ln Z_{N}$ of the Hermitian one-matrix model admits a formal expansion in powers of $1/ N^{2}$ ,

\frac{1}{N ^{2}} F_{N} = g \geq 0 \sum N^{- 2 g} F_{g} (g_{4}, g_{6}, \dots),

in which $F_{g}$ is the generating function of connected ribbon graphs of genus exactly $g$ , each weighted by its symmetry factor and the coupling constants of $V$ . The leading term $F_{0}$ collects the planar ( $g = 0$ ) graphs, those that thicken to a sphere.

Proof. Expand $e^{- N tr V_{int} (M)}$ in the interaction couplings and evaluate each term by Wick contraction against the Gaussian weight $e^{- \frac{N}{2} tr M^{2}}$ . By the index structure of the propagator $⟨ M_{ij} M_{k l} ⟩_{0} = N^{- 1} δ_{i l} δ_{j k}$ , a Wick contraction joins ribbons strand to strand; a fully contracted term is therefore a ribbon graph with closed index loops. A connected ribbon graph with $V$ vertices and $E$ edges has each vertex supplying one factor of $N$ , each edge one factor of $N^{- 1}$ , and each closed index loop one factor of $N$ from the unconstrained index sum, giving total weight $N^{V - E + F}$ where $F$ is the number of index loops. The thickened surface of the graph has the index loops as its faces, the ribbons as its edges, and the vertices as its vertices, so $V - E + F$ equals its Euler characteristic $χ = 2 - 2 g$ . Summing $ln Z_{N}$ keeps only connected graphs, so $F_{N} = \sum_{conn. ribbon graphs} N^{2 - 2 g} (weight)$ . Collecting graphs by genus and factoring the common $N^{2}$ yields the stated series. The reorganisation is formal in the couplings: convergence is a separate analytic question settled, for the quartic model, by the orthogonal-polynomial solution. $□$

Bridge. This genus grading builds toward every structural statement below, and the foundational reason a single integer $g$ controls the whole hierarchy is that the double-line propagator turns index bookkeeping into a count of faces on a surface. This is exactly the large- $N$ saddle mechanism of 08.14.03 seen through a different lens: there a parameter multiplying the action made the saddle exact, and here that same overall $N$ identifies powers with $N^{2 - 2 g}$ genus by genus, so the planar sector is the matrix analogue of the leading saddle. The construction generalises the vector $1/ N$ series of 08.14.03, where the only grading was the loop order of a single auxiliary field, to a grading by a topological invariant of surfaces; the central insight, that the perturbative expansion of a $U (N)$ integral is dual to a sum over triangulated surfaces, appears again in the Advanced section, where summing the planar graphs is reread as a continuum limit of random surfaces and 2D gravity. Putting these together, the bridge is the identification of Feynman bookkeeping with surface topology.

Exercises Intermediate+

Exercise 6 (hard, short-answer).

The quartic one-matrix model is solved by introducing orthogonal polynomials $P_{n} (λ)$ for the weight $e^{- N V (λ)}$ and recursion coefficients $R_{n}$ . Sketch why this diagonalises the problem and how the genus expansion re-emerges from a continuum limit of the recursion.

Hint

Reduce $Z_{N}$ to eigenvalues via the Vandermonde, then expand the Vandermonde in orthogonal polynomials.

Answer

Diagonalising $M$ trades $d M$ for the eigenvalue measure with a Vandermonde squared $\prod_{i < j} (λ_{i} - λ_{j})^{2}$ . Writing the Vandermonde as a determinant of orthogonal polynomials $P_{n}$ collapses $Z_{N}$ to a product $\prod_{n} h_{n}$ of their norms, and the ratios $R_{n} = h_{n} / h_{n - 1}$ obey a finite-difference recursion (the string equation). Setting $x = n / N$ and expanding $R_{n}$ in $1/ N$ around a smooth function $R (x)$ turns the recursion into a differential equation whose successive $1/ N^{2}$ corrections reproduce $F_{0}, F_{1}, F_{2}, \dots$ . This is the orthogonal-polynomial route of Brezin, Itzykson, Parisi, and Zuber ^{[BIPZ 1978]}.

Rubric: full credit for Vandermonde-to-orthogonal-polynomials and the continuum limit of the recursion giving the genus series.

Advanced results Master

The genus grading is more than bookkeeping: it makes the perturbative series of a $U (N)$ matrix integral coincide, term by term, with a sum over triangulated surfaces. Dual to a ribbon graph of genus $g$ is a triangulation (more generally a polygonalisation) of the same surface, with one polygon per vertex of the graph. A model with a single $tr M^{k}$ interaction generates surfaces glued from $k$ -gons, and the free energy $F_{g}$ is the generating function for genus- $g$ surfaces tiled by such polygons, weighted by area. The matrix integral is thus a discrete model of fluctuating geometry, and the sum over $g$ is a sum over topologies, the discrete precursor of summing over worldsheet topologies in string theory ^{[Di Francesco-Ginsparg-Zinn-Justin 1995]}.

The Gaussian model already exhibits the mechanism in its purest form. At leading order in $1/ N$ the eigenvalue density of $M$ converges to the Wigner semicircle $ρ (λ) = \frac{1}{2 π} 4 - λ^{2}$ on $[- 2, 2]$ , and the planar connected moments $\frac{1}{N} ⟨ tr M^{2 k} ⟩ \to C_{k}$ are the Catalan numbers counting planar (non-crossing) pairings. The interacting quartic model is solved exactly at every genus by the orthogonal-polynomial method of Brezin, Itzykson, Parisi, and Zuber: the partition function reduces to norms $h_{n}$ of polynomials orthogonal for $e^{- N V}$ , and the ratios $R_{n} = h_{n} / h_{n - 1}$ satisfy the discrete string equation $R_{n} (1 + g (R_{n - 1} + R_{n} + R_{n + 1})) = n / N$ for the quartic coupling ^{[BIPZ 1978]}. Its continuum limit, where $g$ approaches the critical value at which planar surfaces proliferate, produces a singularity in $F_{0}$ of the form $(g_{c} - g)^{2 - γ}$ with the string susceptibility exponent $γ = - \frac{1}{2}$ for pure 2D gravity, and the double-scaling limit, in which $N \to \infty$ and $g \to g_{c}$ are correlated so that all genera contribute, packages the full topological sum into a Painlevé equation ^{[Di Francesco-Ginsparg-Zinn-Justin 1995]}.

The same double-line idea governs $U (N)$ lattice gauge theory and the planar limit of large- $N$ Yang-Mills: 't Hooft's original motivation was that the planar sector might be analytically tractable while still capturing confinement, with $1/ N$ as the only available small parameter at strong coupling ^{['t Hooft 1974]}. A dedicated statistical-field-theory entry on these surface sums and their continuum gravity limit is forward-pointed from here; this unit fixes only the combinatorial backbone they all share.

Synthesis. Putting these together, the foundational reason the matrix integral computes geometry is that the double-line propagator identifies index bookkeeping with the face structure of a surface, so the perturbative series is dual to a sum over triangulated topologies graded by genus. This is exactly the large- $N$ principle of 08.14.03 sharpened: there an overall factor of $N$ made a single saddle exact, and here the same factor grades the entire expansion by $N^{2 - 2 g}$ , so the planar sector is the matrix analogue of the leading saddle and the higher genera are its topological fluctuations. The central insight is that the orthogonal-polynomial solution turns this surface sum into a difference equation whose continuum limit is 2D quantum gravity, and the string susceptibility $γ = - \frac{1}{2}$ generalises the critical exponents of an ordinary field theory to a theory of fluctuating geometry. The bridge across all of it is one identification: a $U (N)$ matrix average is dual to a statistical sum over discrete surfaces, and the inverse rank $1/ N^{2}$ is the coupling that weights their handles.

Full proof set Master

Proposition (Euler-characteristic power counting). Let $Γ$ be a connected ribbon graph arising from the Hermitian one-matrix model with action $N tr V (M)$ and propagator $⟨ M_{ij} M_{k l} ⟩_{0} = N^{- 1} δ_{i l} δ_{j k}$ , with $V$ vertices, $E$ ribbons, and $F$ closed index loops. Then the $N$ -dependence of the amplitude of $Γ$ is $N^{V - E + F} = N^{2 - 2 g}$ , where $g$ is the genus of the closed oriented surface $Σ_{Γ}$ obtained by thickening $Γ$ .

Each vertex of $Γ$ comes from one factor of the interaction, which carries an explicit $N$ from the prefactor $N tr V$ , contributing $N^{V}$ . Each ribbon is one Wick contraction, each carrying the propagator factor $N^{- 1}$ , contributing $N^{- E}$ . The remaining index structure is a product of Kronecker deltas joined head to tail around the closed index loops; each closed loop is a free index summed over $1, \dots, N$ with no further constraint, contributing one factor $N$ , hence $N^{F}$ . The total power is $N^{V - E + F}$ . The thickening $Σ_{Γ}$ is the CW complex with one $0$ -cell per vertex, one $1$ -cell per ribbon, and one $2$ -cell per index loop (each index loop bounds a face), so its Euler characteristic is $χ (Σ_{Γ}) = V - E + F$ . For a closed oriented surface $χ = 2 - 2 g$ with $g$ the number of handles. Therefore the amplitude scales as $N^{2 - 2 g}$ . $□$

Proposition (planar Gaussian moments are Catalan). In the Gaussian model $V (M) = \frac{1}{2} M^{2}$ , the leading large- $N$ value of $m_{2 k} := \frac{1}{N} ⟨ tr M^{2 k} ⟩$ equals the Catalan number $C_{k} = \frac{1}{k + 1} (k 2 k)$ .

By Wick's theorem $⟨ tr M^{2 k} ⟩$ is a sum over pairings of the $2 k$ matrices around the trace. Each pairing is a ribbon graph with a single $2 k$ -valent vertex ( $V = 1$ ), $k$ ribbons ( $E = k$ ), and $F$ index loops; by the previous proposition its $N$ -power is $N^{1 - k + F} = N^{2 - 2 g}$ , so $F = k + 1 - 2 g$ . Dividing by $N$ for $m_{2 k}$ shifts the leading power to $N^{1 - 2 g}$ , maximised at $g = 0$ , where $F = k + 1$ . The genus-zero pairings are exactly the non-crossing (planar) pairings of $2 k$ points on a circle, and the number of non-crossing pairings of $2 k$ points is the Catalan number $C_{k}$ . Hence $m_{2 k} \to C_{k}$ as $N \to \infty$ , with $O (1/ N^{2})$ corrections from $g \geq 1$ . These moments are those of the semicircle density $ρ (λ) = \frac{1}{2 π} 4 - λ^{2}$ , since $\int_{- 2}^{2} λ^{2 k} ρ (λ) d λ = C_{k}$ . $□$

Connections Master

This unit is the matrix-index sequel to the large- $N$ vector limit of 08.14.03: both rest on an overall factor of $N$ multiplying the action, but the vector model grades its $1/ N$ series by the loop order of a single auxiliary field, while the matrix model grades $N^{2 - 2 g}$ by the genus of a surface. The planar sector here is the exact analogue of the leading saddle there, and the orthogonal-polynomial solution replaces the saddle-plus-Gaussian-fluctuation route used for the vector model.

The genus expansion forward-points to a planned dedicated statistical-field-theory entry on random triangulated surfaces and 2D quantum gravity, where the double-scaling limit packages the full topological sum into a Painlevé equation; that entry will take the combinatorial backbone fixed in this pointer as its starting datum rather than re-deriving the $N^{2 - 2 g}$ counting.

The semicircle law and Catalan moments derived here connect to the broader theory of eigenvalue statistics of large random matrices, the same Wigner ensemble that underlies level-spacing universality; the planar moments computed in the Full proof set are the entry point to that circle of ideas and to free probability, where non-crossing pairings replace ordinary independence.

Historical & philosophical context Master

't Hooft proposed the $1/ N$ expansion in 1974 as a route to strong-interaction dynamics: treating the number of colours $N$ of $U (N)$ gauge theory as large and organising Feynman diagrams in double-line notation, he showed the leading contributions are the planar graphs and that the suppression of a diagram is governed by the genus of the surface it tiles ^{['t Hooft 1974]}. The reinterpretation as a sum over surfaces, and the exact solution of the simplest matrix integrals, came with Brezin, Itzykson, Parisi, and Zuber in 1978, who used orthogonal polynomials to compute the genus expansion of the one-matrix model in closed form and identified the matrix integral as a generating function for triangulated surfaces ^{[BIPZ 1978]}. In the late 1980s this combinatorics was recognised as a discrete model of two-dimensional quantum gravity, and the double-scaling limit tied the topological sum to the Painlevé I equation; the program is surveyed by Di Francesco, Ginsparg, and Zinn-Justin ^{[Di Francesco-Ginsparg-Zinn-Justin 1995]}. Itzykson and Drouffe develop the matrix-model and random-surface material in the second volume of their text ^{[Itzykson-Drouffe Vol. 2]}.

Bibliography Master

@article{tHooft1974Planar,
  author  = {'t Hooft, Gerard},
  title   = {A Planar Diagram Theory for Strong Interactions},
  journal = {Nuclear Physics B},
  volume  = {72},
  year    = {1974},
  pages   = {461--473}
}

@article{BIPZ1978,
  author  = {Br\'ezin, \'Edouard and Itzykson, Claude and Parisi, Giorgio and Zuber, Jean-Bernard},
  title   = {Planar Diagrams},
  journal = {Communications in Mathematical Physics},
  volume  = {59},
  year    = {1978},
  pages   = {35--51}
}

@article{DiFrancesco1995,
  author  = {Di Francesco, Philippe and Ginsparg, Paul and Zinn-Justin, Jean},
  title   = {2D Gravity and Random Matrices},
  journal = {Physics Reports},
  volume  = {254},
  year    = {1995},
  pages   = {1--133}
}

@book{ItzyksonDrouffe1989Vol2,
  author    = {Itzykson, Claude and Drouffe, Jean-Michel},
  title     = {Statistical Field Theory, Volume 2},
  publisher = {Cambridge University Press},
  year      = {1989}
}

Prerequisites

08.14.03

Tier anchors

beginner: Counting how many ways a ribbon glues into a surface; Euler's formula V - E + F = 2 - 2g
intermediate: Double-line / ribbon graphs and the genus grading of a U(N) matrix integral; 't Hooft 1974
master: Brezin-Itzykson-Parisi-Zuber 1978 genus expansion of the one-matrix model; link to random triangulated surfaces and 2D quantum gravity (Itzykson-Drouffe Vol. 2)

References

't Hooft — A Planar Diagram Theory for Strong Interactions · Nucl. Phys. B 72 (1974) 461, the 1/N expansion and double-line notation
Brezin, Itzykson, Parisi, Zuber — Planar Diagrams · Commun. Math. Phys. 59 (1978) 35, genus expansion of the one-matrix model by orthogonal polynomials
Itzykson and Drouffe — Statistical Field Theory, Vol. 2 · matrix models, random surfaces, and 2D gravity
Di Francesco, Ginsparg, Zinn-Justin — 2D Gravity and Random Matrices · Phys. Rep. 254 (1995) 1, review of matrix models and topological expansion

Estimated time

beginner: 10m
intermediate: 26m
master: 48m