MuJoCo: A physics engine for model-based control

🎯 Pitch

Unlike complex matrix models where Wick pairings glue faces into orientable surfaces, real Gaussian matrices allow twisted edge identifications that produce nonorientable topology—projective planes and Klein bottles appear in the same calculation. This contribution reveals that the asymptotic fluctuations of real ensembles depend on a second-order noncrossing condition evaluated under both relative face orientations, a structural refinement absent from the complex theory, and illustrates the full dictionary between trace cumulants, surface connectivity, and the permutation premap encoding that makes the proofs tractable.

1. Executive Summary

This appendix paper illustrates and explains the topological and combinatorial constructions underlying the proofs in the main paper on real second-order freeness, using diagrams of surface gluings to build intuition for how Gaussian random matrix calculations are represented via permutations and premaps (permutations encoding both a face gluing and its orientable double cover, allowing nonorientable surfaces). The document walks through concrete examples — including Ginibre, GOE, and Wishart matrix models — to show how expected trace products decompose into sums over surface gluings, how cumulants correspond to connected surfaces, and how asymptotic highest-order terms correspond to noncrossing diagrams on spheres. The central pedagogical insight is that real matrices admit twisted edge identifications which produce nonorientable surfaces (projective planes, Klein bottles) that do not appear in the complex case, establishing that the asymptotic fluctuations of real ensembles depend on noncrossing conditions evaluated under both relative orientations of the constituent faces — a structural difference from the complex setting where only one relative orientation matters.

2. Context and Motivation

The Core Problem: How Do You Actually Calculate with Real Random Matrices?

The fundamental problem this paper addresses is making the proofs in the main paper intelligible. The main paper (Redelmeier, 2011) establishes a substantial theoretical result: real second-order freeness — a characterization of the asymptotic joint fluctuations of traces of products of certain real random matrix ensembles. The proofs in that paper rely heavily on combinatorial encodings of surface gluings using permutations, specifically a formalism called premaps that extends the well-established map/hypermap encodings (from Cori, 1975; Zvonkin, 1998; Lando and Zvonkin, 2004) to handle nonorientable surfaces. However, these permutation-based proofs are notoriously opaque to readers encountering them for the first time. The combinatorial manipulations are precise but give little geometric intuition for why the formulas take the form they do.

This paper, presented as an appendix, fills a specific pedagogical gap: there is no existing resource that walks the reader through the geometric intuition behind the permutation encodings of nonorientable surface gluings as they appear in real random matrix calculations. The author states this explicitly in the introduction:

"we attempt to provide some intuition for how the topological constructions are represented in the permutations, and describe the pictures which motivated the various proofs."

The gap is not in the existence of the permutation formalism itself — Tutte (1984) developed a similar encoding for unoriented maps, and work by Lando and Zvonkin (2004) covers oriented surfaces extensively — but in the application of these encodings to real Gaussian random matrix calculations and, critically, in the translation between matrix index constraints, surface topology, and permutation cycles that forms the backbone of the main paper's proofs.

Why This Problem Matters: Real Matrices Are Everywhere, but Their Theory Is Harder

The distinction between real and complex random matrices carries substantial practical consequence. Real random matrices arise naturally in applications where the underlying data or symmetries are real-valued:

Statistics and machine learning: Wishart matrices $X^T X$ , where $X$ has real Gaussian entries, form the cornerstone of multivariate statistics (sample covariance matrices, PCA, regression diagnostics). GOE matrices model real symmetric Hamiltonians in disordered systems and serve as null models in network analysis.
Wireless communications: Real Ginibre matrices appear in the capacity analysis of multiple-input multiple-output (MIMO) channels when the channel coefficients are real-valued (as in certain in-phase/quadrature decomposition settings).
Quantum chaos and condensed matter: Real symmetric matrices (GOE) describe time-reversal-invariant quantum systems, while complex Hermitian matrices (GUE) describe systems with broken time-reversal symmetry. The symmetry class matters for the spectral statistics.

However, the moment and cumulant calculations for real ensembles are structurally harder than for complex ensembles. The reason, which this paper illustrates in detail, is topological:

In the complex case, Gaussian entries satisfy $\mathbb{E}[Z^2] = \mathbb{E}[\overline{Z}^2] = 0$ and $\mathbb{E}[Z\overline{Z}] = 1$ . When applying Wick's theorem (the multivariate Gaussian moment factorization) to a product of complex matrix entries, paired terms must couple an $X$ (untransposed) with an $X^*$ (conjugate transpose). In the topological picture, this constraint forces all edge identifications between faces to be untwisted — the faces are glued together with matching orientations, like sewing two pieces of fabric right sides together. Consequently, all surfaces produced in complex matrix calculations are orientable. The combinatorial encodings (maps, hypermaps) developed by Cori and Zvonkin assume orientability, and they work perfectly for the complex case.

In the real case, Gaussian entries satisfy $\mathbb{E}[f^2] = 1$ (where $f \sim \mathcal{N}(0,1)$ ). The Wick pairing can couple an $X$ with another $X$ , or an $X^T$ with another $X^T$ , not just $X$ with $X^T$ . In the topological picture, this permits twisted edge identifications — gluing faces together with a half-twist, like constructing a Möbius strip. Surfaces containing twisted gluings become nonorientable: projective planes (Euler characteristic 1), Klein bottles (Euler characteristic 0), and connected sums involving them. As the exercises in Lando and Zvonkin (2004), Chapter 3, note:

"calculations with real matrices include nonorientable surfaces."

This is the central object of study, and it creates two intertwined difficulties that the paper's permutation formalism must address:

Encoding nonorientable surfaces combinatorially. The standard map formalism (an ordered triple of permutations $(\sigma, \alpha, \phi)$ on edge-ends representing vertices, edges, and faces) assumes a globally consistent orientation — counterclockwise cyclic ordering around every vertex. A nonorientable surface has no such global orientation. The solution, which the paper illustrates, is to work in the orientable double cover: construct two preimages of every face (a "front" with positive orientation and a "back" with reversed orientation), then encode twisted gluings as identifications between fronts and backs. This produces a permutation $\gamma_+ \gamma_-^{-1}$ that satisfies the definition of a premap — a permutation whose cycles may mix positive and negative integers, but for which every cycle has a distinct "mirror" cycle containing the same integers with reversed signs and reverse order.
Handling relative orientations of faces. Even when the surface is a sphere (the highest-order asymptotic contribution), the relative orientation of the constituent faces matters in the real case. Two faces glued with the same orientation versus opposite orientation produce different pairings. The asymptotic fluctuations therefore depend on counting noncrossing pairings that work under both relative orientations of the faces — a condition absent from the complex case, where only one relative orientation is compatible with the matrix model's pairing constraints.

The practical consequence is that the asymptotic second-order structure (fluctuations, covariances of trace products) of real ensembles differs from the complex case, even when the complex ensemble is the "complexification" of the real one (e.g., complex Ginibre vs. real Ginibre). Understanding these differences requires mastering the permutation encoding, which is exactly what this appendix paper teaches by example.

Prior Approaches and Where They Fall Short

The paper builds on several established frameworks, each of which provides part of the necessary machinery but is insufficient on its own for the real matrix setting:

1. Topological surface interpretations of matrix integrals (Zvonkin, 1997; Lando and Zvonkin, 2004). These works establish the connection between Gaussian matrix moment calculations and map enumeration: applying Wick's theorem to a product of matrix entries produces a sum over pairings, each of which defines a surface gluing of the polygonal faces representing the traces. The order of a term in $1/N$ depends on the Euler characteristic of the glued surface. This framework is comprehensive for the oriented case but assumes edges come with a direction (e.g., complex matrices force untwisted gluings). The real case, where edges may be glued with or without a twist, falls outside the standard map enumeration paradigm because the resulting surfaces may be nonorientable.

2. Permutation encodings of maps and hypermaps (Cori, 1975; Zvonkin, 1998). These provide the combinatorial backbone: a map on an oriented surface is encoded by two permutations $\sigma$ (vertices) and $\alpha$ (edges) on the set of edge-ends, with the face permutation recovered as $\phi = \sigma^{-1} \alpha^{-1}$ . Hypermaps generalize this to hyperedges (edges incident to more than two vertices) by adding a third permutation $\pi$ and interleaving the roles of vertices, hyperedges, and faces. However, both map and hypermap encodings presuppose an orientable surface. The cyclic orders around vertices must be consistently definable. For a nonorientable surface, the encoding breaks down because "counterclockwise" cannot be globally defined.

3. Tutte's encoding of unoriented maps (Tutte, 1984). Tutte developed a permutation-based description of maps on possibly nonorientable surfaces by assigning each edge two ends with opposite orientations (positive and negative) and defining permutations that may mix signs. The paper's premap formalism (Section 3, main paper, and Section 3 here) is closely related to Tutte's construction. However, Tutte's work is written at a high level of abstraction within graph theory and does not connect to random matrix calculations, trace product expansions, or asymptotic analysis. The translation from "a surface gluing representing a Wick pairing" to "a premap encoding the constraints on summed indices" is left entirely to the reader.

4. Annular noncrossing permutations and second-order freeness (Mingo and Nica, 2004). This work establishes the combinatorial characterization of asymptotic fluctuations for complex random matrices: the limiting covariances are given by annular noncrossing permutations (or partitions) satisfying compatibility conditions with the matrix models. The framework is complete for the complex case but does not address the extra conditions — particularly the relative orientation constraints — that appear for real matrices. The main paper extends this framework to the real setting; this appendix paper illustrates why the extension is needed and how the extra conditions emerge from the topology.

5. Existing real second-order freeness results. Prior to the main paper, the asymptotic real second-order freeness of certain real ensembles (GOE, real Wishart) was known through independent calculations, but a unified framework comparable to Mingo and Nica's annular noncrossing formalism for the complex case did not exist. The main paper provides that framework. This appendix exists because the proofs in the main paper, while rigorous, are permutation-theoretic and opaque: a reader encountering $\gamma_+ \gamma_-^{-1} \pi^{-1} \gamma_- \gamma_+^{-1}$ has no visual anchor for what this composition means in terms of the surface.

The key shortfall across all prior work is the missing pedagogical layer: no prior document draws the pictures, labels the edge-ends, traces the vertex boundaries, and shows step-by-step how a matrix computation becomes a surface becomes a permutation. This paper fills exactly that gap.

How This Paper Positions Itself

This paper is explicitly an appendix and companion document, not a standalone contribution. It does not introduce new theorems or new combinatorial machinery. Instead, it translates the main paper's permutation-based proofs into a visual language of face gluings and surface topology, then uses that visual language to explain why the permutations take the specific forms they do.

The positioning is layered:

Section 2 (“Matrix Calculations as Face Gluings”) serves as a Rosetta Stone: given a concrete trace product (a product of two traces involving $X$ , $X^T$ , and deterministic matrices $Y_k$ ), the paper walks through constructing the polygonal faces, applying Wick's theorem to the Gaussian entries, and gluing paired edges — showing both untwisted and twisted identifications in the same example. The result is a surface (in this case, a Klein bottle) whose vertices correspond to new trace products of the $Y_k$ matrices. This section bridges the gap between the algebraic manipulation (tracing indices through summation constraints) and the topological construction (gluing faces along edges).
Section 3 (“Cartography on Unoriented Surfaces”) explains the permutation encoding itself, starting from the familiar oriented case (maps and hypermaps, Examples 3.1 and 3.2) and then extending to the nonorientable case via the double cover construction. The key pedagogical move is showing that the premap encoding is not an arbitrary abstraction but a direct translation of the double cover picture: $\gamma_+$ encodes the front faces (positive orientation), $\gamma_-^{-1}$ encodes the back faces (reversed orientation), and the product $\gamma_+ \gamma_-^{-1}$ encodes all faces of the covering space. The vertex permutation is recovered as $\gamma_{-}^{-1} \pi \gamma_+$ , derived by tracing boundary cycles in the double cover — a formula whose geometric origin is clear from the diagrams but would be mysterious from the algebra alone.
Sections 4–8 then apply this apparatus to the specific matrix models (Ginibre, GOE, Wishart) and to the structural properties that matter for freeness: cumulants (connected surfaces), centring (exclusion of disconnected components), genus/Euler-characteristic expansions (highest-order = spheres = noncrossing diagrams), and the spoke-diagram structure that emerges for two-trace cumulants of alternating algebras.

The paper's intended audience is someone who has read (or is reading) the main paper and needs geometric intuition for the permutation formulas. It does not replace the main paper's proofs but makes them accessible by providing the pictures the author had in mind while writing them. The fact that this document exists as a separate arXiv submission (rather than being integrated into the main paper) reflects the reality that mathematical papers often strip away the scaffolding that led to the proofs; this appendix is that scaffolding, preserved and explained.

3. Technical Approach

3.1 Reader Orientation

This paper is fundamentally a pedagogical exposition and geometric companion to the permutation-based proofs in the main paper on real second-order freeness. Rather than introducing new theoretical machinery, it builds a visual bridge between three representations of the same mathematical object: matrix index calculations (algebraic), surface gluings of polygonal faces (topological), and permutations on signed edge-ends (combinatorial). The system being explained is not a computational pipeline but a translation protocol that converts a trace product involving Gaussian random matrices into (1) a sum over gluings of oriented polygons into possibly-nonorientable surfaces, (2) an encoding of each such gluing as a premap — a permutation whose cycles may mix positive and negative integers, representing the orientable double cover of the surface — and (3) an evaluation of the trace product in terms of the Euler characteristic of the surface and trace products of the deterministic matrices appearing on its vertices. The problem this solves for the reader is making the main paper's abstract permutation formulas (\gamma_+ \gamma_-^{-1}, \gamma_-^{-1} \pi \gamma_+) visually and geometrically meaningful by showing exactly which surface features each permutation cycle tracks.

3.2 Big-picture Architecture (Diagram in Words)

The translation protocol described in this paper has four major components, applied sequentially to a given trace product expression:

Face construction (Section 2, main paper Lemma 3.1): Each trace in the input expression becomes a polygonal face. The edges of the polygon correspond to occurrences of the Gaussian random matrix X or its transpose X^T, alternating with vertices labelled by the deterministic matrices Y_k. The cyclic order of edges around the polygon follows the cyclic order of matrix factors in the trace. Each edge-end is assigned a labelled index variable (\iota_k^+, \iota_k^-), with the direction (arrow) determined by whether the matrix appears transposed.
Edge pairing via Wick's theorem (Section 2): The expectation of a product of Gaussian entries factorises into a sum over pairings of the Gaussian random variables. Each pairing specifies which pairs of edges are identified by equating their index variables. An edge identification may be untwisted (pairing an X-edge with an X^T-edge, preserving local orientation) or twisted (pairing an X with another X, or X^T with X^T, reversing local orientation). The faces are glued along paired edges to form a topological surface.
Permutation encoding via the orientable double cover (Section 3): The surface gluing is encoded by constructing its orientable double cover: each original face produces two preimage faces — a "front" with positive orientation and a "back" with reversed orientation. The faces, edges (or hyperedges), and vertices of the cover are each represented by a permutation on the set of signed edge-ends (positive integers for fronts, negative integers for backs). The face permutation is \gamma_+ \gamma_-^{-1}, where \gamma_+ encodes the front faces and \gamma_-^{-1} encodes the back faces. The edge pairing defines a permutation \pi whose cycles contain the signed edge-ends that were identified (with sign patterns reflecting whether identifications are twisted). The vertex permutation is recovered as \gamma_-^{-1} \pi \gamma_+, derived by tracing the boundary of each region enclosed by faces and edges in the cover.
Evaluation (Sections 5–8): The surface's Euler characteristic \chi(\gamma, \pi) determines the order in 1/N of the term. Connected surfaces correspond to cumulants; disconnected surfaces are partitioned by the moment-cumulant formula. Centring (subtracting means) enforces that no face is disconnected from all others, implemented by inclusion-exclusion. The leading asymptotic order (N^0) comes from surfaces with \chi = 2 (spheres), which correspond to noncrossing pairings drawn on the faces. For real matrices, the noncrossing condition must be satisfied under both relative orientations of the constituent faces, producing a counting problem absent in the complex case.

Information flows strictly forward: an input trace product → faces → edge pairing → surface gluing → premap encoding → Euler characteristic → asymptotic contribution. The paper's contribution is making every step in this flow visible and interpretable through concrete worked examples, so the reader can map any algebraic step in the main paper's proofs to a geometric operation on surfaces.

3.3 Roadmap for the Deep Dive

First, the single most fundamental example — the face gluing produced by applying Wick's theorem to an expected product of two traces involving X, X^T, and deterministic Y_k matrices (Section 2). This example introduces faces, twisted vs. untwisted edge identifications, vertex recovery, and the Euler characteristic in a concrete, visual setting. Understanding this example is prerequisite for all subsequent abstraction.
Second, the permutation encoding of surface gluings, progressing from the familiar oriented case (maps and hypermaps with \sigma, \alpha, \phi) to the nonorientable case via the double cover construction (\gamma_+, \gamma_-, \pi, and the premap formalism). The crucial geometric insight is that the vertex permutation \gamma_-^{-1} \pi \gamma_+ is derived by walking the boundaries of regions in the covering space, and this derivation is shown step-by-step.
Third, the explicit expansion formula that all matrix models in the main paper satisfy (Section 4), interpreted as a set of rules for constructing surfaces given faces and a matrix model. The Ginibre case (\delta_\varepsilon encoding transpose structure) and GOE case (all pairings allowed) are contrasted to show how the matrix model's Gaussian moment structure determines which edge identifications are permitted.
Fourth, the structural properties that build on the basic gluing: how cumulants select connected surfaces (Section 5), how centring subtracts terms with disconnected components via inclusion-exclusion (Section 6), how asymptotic scaling is governed by Euler characteristic and noncrossing diagrams (Section 7), and how free independence between algebras forces spoke diagrams for two-trace cumulants (Section 8). Each property has a geometric interpretation that the permutation formalism makes precise.

3.4 Detailed, Sentence-based Technical Breakdown

This paper is a pedagogical exposition that translates the main paper's permutation-based proofs into a visual language of surface gluings. The core idea is that a matrix expectation involving Gaussian entries naturally produces a sum over pairings of those Gaussian entries (via Wick's theorem), and each pairing defines a surface gluing whose topology controls the algebraic value of the term.

The Face Construction: From Traces to Polygons

The starting point for every calculation in this paper is a product of traces containing Gaussian random matrices. The fundamental operation appears in Lemma 3.1 of the main paper: the constraints on the indices appearing in a trace have a cyclic structure. This cyclic structure is geometrically represented by constructing a polygonal face for each trace in the expression.

How faces are built (Section 2, explicit example). Consider the expression:

$E\left[\mathrm{tr}\left(XY_1XY_2X^TY_3Y_4X^TY_5\right) \cdot \mathrm{tr}\left(X^TY_6XY_7XY_8\right)\right]$

where X is an M \times N matrix with i.i.d. \mathcal{N}(0, 1/N) entries, and Y_1, \ldots, Y_8 are random matrices independent of X. The expression contains two traces, so we construct two faces: one 5-sided face (one trace has five X/X^T factors, hence five edges) and one 3-sided face (three X/X^T factors, three edges).

Let k index the occurrences of X or X^T in order through the expression. The k-th occurrence of the Gaussian matrix is assigned two index variables: \iota_k^+ and \iota_k^-, corresponding to the row and column indices of the matrix entry (which is which depends on whether the matrix appears transposed or not). These indices appear as edge labels on the polygonal faces:

If the k-th occurrence is X (untransposed), the edge is oriented with \iota_k^+ at the tail and \iota_k^- at the head — the index \iota_k^+ is the row index, \iota_k^- the column index.
If the k-th occurrence is X^T (transposed), the edge is oriented in the opposite direction — \iota_k^+ goes to the column index and \iota_k^- to the row index.

Between consecutive edges of a face sit the deterministic matrices Y_k, which label the vertices of the polygon. The cyclic order of edges around each face follows the cyclic order of factors in the corresponding trace.

What Figure 1 shows: The two faces for the example above. The first face (a 5-gon, or pentagon) has edges k = 1, 2, 3, 4, 5 with X at k=1,2 (untransposed) and X^T at k=3,4,5 (transposed), with vertices Y_1, Y_2, Y_3, Y_4, Y_5 between them. The second face (a 3-gon, or triangle) has edges k = 6, 7, 8 with X^T at k=6, X at k=7, and X at k=8, with vertices Y_6, Y_7, Y_8. The arrows on the edges indicate the direction: arrows pointing into the face correspond to transposed edges; this directional information is stored in the function \varepsilon: [n] \to \{+1, -1\}, where \varepsilon(k) = +1 for untransposed X and \varepsilon(k) = -1 for transposed X^T.

Design choice — why edges instead of vertices: The edge-centred construction (faces are cycles of Gaussian matrix entries, vertices are deterministic matrices) follows the structure of the Wick expansion: it is the X entries, not the Y_k entries, whose Gaussian moments produce pairings. The Y_k are "spectator" matrices whose indices are carried along by the identifications forced by the X pairings. The faces capture the input constraints (cyclic trace structure); the edge gluings capture the Wick pairings (constraints from Gaussian moments); the vertices capture the output constraints (what trace products of Y_k survive after summing over paired indices).

Edge Identification: Untwisted vs. Twisted Gluings

The expectation over the Gaussian entries f_{ij} \sim \mathcal{N}(0, 1) of X is evaluated using Wick's theorem (also called Isserlis' theorem for the real case). For centred Gaussian variables, the expectation of a product factorises into a sum over all complete pairings of the factors, where each pair contributes the covariance \mathbb{E}[f_{ab}f_{cd}] = \delta_{ac}\delta_{bd}. In the matrix setting, the factors are the individual entries X_{ij} = \frac{1}{\sqrt{N}}f_{ij}, and pairing two entries forces their index pairs to match.

The crucial distinction: real vs. complex Gaussian moments.

In the complex case, \mathbb{E}[Z_{ij}\overline{Z}_{kl}] = \frac{1}{N}\delta_{ik}\delta_{jl} but \mathbb{E}[Z_{ij}Z_{kl}] = \mathbb{E}[\overline{Z}_{ij}\overline{Z}_{kl}] = 0. Therefore, every Wick pairing must couple an untransposed Z with a conjugate-transposed \overline{Z} — one factor from an edge labelled Z and the other from an edge labelled Z^*. In the topological picture, this forces every edge identification to be untwisted: the front sides of both faces face the same way at the seam, like sewing two pieces of fabric right sides together.

In the real case, \mathbb{E}[f_{ij}f_{kl}] = \delta_{ik}\delta_{jl} — the only constraint is that both factors are real normal variables. A pairing can couple X_{ij} with another X_{kl} (both untransposed), or X_{ji} with X_{lk} (both transposed, since X^T_{ji} = X_{ij}), or X_{ij} with X_{lk} (one untransposed, one transposed). The three possibilities have different topological effects:

Untwisted identification (pairing X with X^T): The indices (\iota_k^+, \iota_k^-) from the X edge are equated to (\iota_\ell^-, \iota_\ell^+) from the X^T edge: \iota_k^+ = \iota_\ell^- and \iota_k^- = \iota_\ell^+. Geometrically, the edges are glued with matching orientation — arrow heads meet arrow tails, as shown in Figure 2.
Twisted identification (pairing X with X, or X^T with X^T): The indices are equated as \iota_k^+ = \iota_\ell^+ and \iota_k^- = \iota_\ell^- (both X), or analogously for both X^T. Geometrically, the edges are glued with a half-twist — arrow heads meet arrow heads, as shown in Figure 3. In the resulting surface, moving along a path that crosses this seam reverses the local sense of "up" (orientation), making the surface nonorientable.

This distinction is the paper's foundational topological insight: complex matrix calculations produce only orientable surfaces; real matrix calculations may produce nonorientable surfaces via twisted identifications. The nonorientable surfaces (projective planes, Klein bottles, connected sums involving them) have lower Euler characteristics than spheres, so they contribute to lower-order terms in 1/N. The GOE (real symmetric Gaussian) case particularly highlights this: calculating \mathbb{E}[\mathrm{tr}(T^2)] for T a GOE matrix (Example 4.2) yields two terms — one from untwisted gluing (a sphere, Euler characteristic 2, contributing 1) and one from twisted gluing (a projective plane, Euler characteristic 1, contributing N^{-1}), as shown in Figure 12.

Vertex Recovery: Walking the Boundaries of the Glued Surface

Once the faces are glued along paired edges, the resulting surface has vertices formed where the corners of the original polygons meet. At each vertex, the deterministic matrices Y_k that appear at those corners are multiplied together in the order imposed by walking around the vertex on the surface.

Why vertices matter: After summing over all indices (each summation over 1, \ldots, N contributes a factor N to the term), the surviving contribution is a product of traces — one trace per vertex — of products of the Y_k matrices. Each trace contributes a factor N, so the total power of N in the term is (\text{number of vertices}) - (\text{number of edges}) + (\text{number of faces}) = \chi, the Euler characteristic of the surface. Higher Euler characteristic (more vertices) means higher order in N.

How vertices are read off (Figures 4 and 5). For a given pairing (call it \rho, a perfect matching on the set of edge indices), the glued surface is formed. Walking around each vertex boundary in the glued surface yields a cyclic sequence of Y_k matrices. However, a critical subtlety arises from the relative orientations of the faces meeting at that vertex: some corners of the original faces may be upside-down relative to others. Since reversing the orientation of a face segment reverses the order in which the Y_k appear, an upside-down corner contributes Y_k^T instead of Y_k. The paper describes this as:

"We can imagine that the transposes of the matrices are written on the backs of the faces."

If we look at the same vertex from the opposite side of the surface, the matrix product is the transpose of the original product — but a trace is invariant under transposition (\mathrm{tr}(AB) = \mathrm{tr}((AB)^T) = \mathrm{tr}(B^TA^T)), so the value is well-defined regardless of which side we read from.

Example: In Figure 5, the vertex produced by the gluing in Figure 4 is shown from both sides. Walking the boundary yields \mathrm{tr}(Y_1 Y_3^T Y_6 Y_5^T Y_7^T). Reading from the opposite side gives the transpose \mathrm{tr}(Y_7 Y_5 Y_6^T Y_3 Y_1^T), which is equal.

The Orientable Double Cover and the Premap Encoding

The permutation formalism developed in Section 3 is designed to encode the surface gluing without drawing it, enabling algebraic manipulation. The key challenge is that the standard map/hypermap encoding (Cori, 1975; Zvonkin, 1998) requires a globally consistent orientation — "counterclockwise" must be well-defined at every vertex. Nonorientable surfaces lack this property.

The solution: work in the orientable double cover. The double cover is constructed by creating two preimages of every point on the surface, corresponding to the two possible local orientations. Points in the cover are close if they are close in the original surface and have the same orientation. For a nonorientable surface, the cover is a connected orientable surface (a double cover of the projective plane is a sphere; of the Klein bottle is a torus). For an orientable surface, the cover is two disconnected copies of the original.

From faces to the cover. Each face of the original surface produces two preimage faces: a front (positive orientation) and a back (negative orientation). The k-th edge on the front face is labelled with the positive integer k; the corresponding edge on the back face is labelled with the negative integer -k, and its orientation is reversed (what was clockwise on the front is counterclockwise on the back — or equivalently, the order of edges around the back face is the reverse of the order around the front face, viewed from the same side of the surface).

This yields the fundamental face permutation of the cover. Let \gamma be the permutation describing the original faces (cycles of positive integers). On the front, edges appear in the order given by \gamma, so the front faces are encoded by \gamma_+ (the same as \gamma but with all signs positive). On the back, edges appear in reverse order, encoded by \gamma_-^{-1}, where \gamma_- has the same cycle structure as \gamma but with all signs negative, and the inverse accounts for the reversed orientation. The full set of faces in the cover is given by:

$\gamma_+ \gamma_-^{-1}$

where \gamma_+ = \gamma is the permutation of the front faces (all entries positive), \gamma_- is the same permutation with all signs flipped to negative, and \gamma_-^{-1} is its inverse.

What this product means operationally: \gamma_+ \gamma_-^{-1} is a permutation on the set of signed integers \{\pm 1, \pm 2, \ldots, \pm n\} whose cycles enumerate the faces of the covering space. Each cycle is a sequence of signed integers encountered when walking clockwise around one face of the cover. Because a face in the cover may be a "front" or a "back" (or may traverse portions of both), a cycle of \gamma_+ \gamma_-^{-1} may contain both positive and negative integers.

Why this is a premap: A permutation \tau on \{\pm 1, \ldots, \pm n\} is called a premap if, for every cycle (a_1, a_2, \ldots, a_m) of \tau, there is a distinct mirror cycle (-a_m, \ldots, -a_2, -a_1) containing the same integers with signs flipped and order reversed. The product \gamma_+ \gamma_-^{-1} satisfies this property automatically: the mirror of a front face is the corresponding back face, and vice versa. The factor of 1/2 that appears in the Euler characteristic formula (below) accounts for this doubling.

Edge pairing in the cover. An edge identification in the original surface (pairing edge k with edge \ell) becomes a pairing of the corresponding edges in the cover, but the sign pattern depends on whether the identification is twisted:

Untwisted (X with X^T): front glues to front (k with \ell in the same sign) and back glues to back (-k with -\ell). The edge permutation \pi contains a cycle (k, \ell) and its mirror (-k, -\ell).
Twisted (X with X, or X^T with X^T): front glues to back (k with -\ell) and back glues to front (-k with \ell). The edge permutation \pi contains a cycle (k, -\ell) and its mirror (\ell, -k).

Thus, the sign pattern of paired integers in a cycle of \pi directly encodes whether the edge identification is twisted: same-sign integers = untwisted; opposite-sign integers = twisted. The permutation \pi is itself a premap — each cycle has a mirror cycle.

Vertex recovery in the cover (the critical formula). The paper derives that vertices of the glued surface correspond to the boundary cycles formed by faces and edges in the cover. The derivation, though the paper gives it in compressed form, has clear geometric content:

Walking along the boundary of a face in the cover, the encounter with an edge-end is represented by following the face permutation \gamma_+ (or \gamma_-) to the next edge-end, then crossing the edge to the paired edge-end via \pi^{-1}. So one step around a vertex is given by following \gamma_+ then \pi^{-1} — but the edge-end labelling convention requires a conjugation.
Specifically, the edge-end at the corner clockwise of the k-th occurrence of X or X^T is associated with the integer k on the front. After crossing an edge and arriving at a back-face corner, the corresponding edge-end would be labelled -m if it were on the front, but the convention for labelling back faces by negative integers means the contribution from the back is \gamma_-^{-1}(-m). Tracking through this yields the vertex permutation:

$\text{vertices} = (\gamma_-^{-1} \pi^{-1} \gamma_+)^{-1} = \gamma_+^{-1} \pi \gamma_-$

The paper states the vertex boundaries are given by \gamma_-^{-1} \pi \gamma_+ (the inverse of this appears when reading the trace clockwise). The text notes:

"The traces read off the vertices are read clockwise, so we will be using the inverse of the permutation describing the vertices: \gamma_-^{-1} \pi \gamma_+."

Computing \gamma_-^{-1} \pi \gamma_+ geometrically: This permutation is computed by conjugating \pi by \gamma_+ and \gamma_-^{-1} — one takes a cycle of \pi, follows \gamma_+ backwards from its entries to find the corresponding face-edge positions, then follows \gamma_-^{-1} forward to place them in the back-face ordering. Each cycle of the result corresponds to one vertex, and the sequence of Y_k at that vertex (with transposes as appropriate) is read off by identifying which edge-end indices appear in the cycle.

Example 3.3 (nonorientable hypermap): The paper works through the permutation calculation for a hypermap on a nonorientable surface. Given \gamma = (1,2,3,4,5)(6,7) (two faces) and \pi = (1,7,-6)(6,-7,-1)(2,5,-3)(3,-5,-2)(4)(-4) (representing hyperedges, some with twists — indicated by mixed signs), it computes:

$\gamma_-^{-1} \pi \gamma_+ = (1,-6,7,5)(-5,-7,6,-1)(2,-3,-4)(4,3,-2)$

Each cycle corresponds to one vertex. The mirror pairs appear as cycles with opposite signs and reversed order: (1,-6,7,5) mirrors (-5,-7,6,-1).

The Euler characteristic formula. From the cover, the Euler characteristic of the original surface is half the Euler characteristic of the cover:

$\chi(\gamma, \pi) := \frac{1}{2}\#(\gamma_+ \gamma_-^{-1}) + \frac{1}{2}\#(\pi) + \frac{1}{2}\#(\gamma_-^{-1} \pi \gamma_+) - |I|$

where \#(\cdot) denotes the number of cycles in the permutation (faces, edges, and vertices of the cover, respectively), |I| is the number of edges (half the number of edge-ends, since each original edge has two ends, each producing two signed preimages), and the factors of 1/2 correct for the doubling in the cover.

For the Klein bottle example (Example 3.4): \chi = 2 + 4 + 2 - 8 = 0, which is the Euler characteristic of a Klein bottle. For the sphere example in Figure 15 (the noncrossing gluing of the faces from Figure 1): \chi = 2 + 4 + 4 - 8 = 2, confirming a sphere.

The General Expansion Formula (Section 4)

The central formula that all matrix models in the main paper satisfy is:

$E\left[\mathrm{tr}_\gamma\left(X^{(\varepsilon(1))}_{\lambda_1} Y_1, \ldots, X^{(\varepsilon(n))}_{\lambda_n} Y_n\right)\right] = \sum_{\pi \in \mathrm{PM}^c(\pm[n])} N^{\chi(\gamma, \delta_\varepsilon \pi \delta_\varepsilon) - 2\#(\gamma)} f^c(\pi) \; E\left[\mathrm{tr}_{\gamma_-^{-1} \delta_\varepsilon \pi^{-1} \delta_\varepsilon \gamma_+ / 2}(Y_1, \ldots, Y_n)\right]$

Defining each symbol:

\gamma: the permutation encoding the input faces — its cycles correspond to the traces in the expression.
X^{(\varepsilon(k))}: the k-th Gaussian matrix factor, where \varepsilon(k) = +1 for untransposed X, \varepsilon(k) = -1 for transposed X^T.
\delta_\varepsilon: a fixed-point-free involution that encodes the transpose structure: it pairs k with -k for untransposed occurrences, and pairs k and -k cross-wise for transposed occurrences (reflecting the index reversal \iota_k^+ \leftrightarrow \iota_k^-). Specifically, \delta_\varepsilon has a 2-cycle (k, \varepsilon(k) \cdot k) for each k — for \varepsilon(k) = +1, the cycle is (k, -k); for \varepsilon(k) = -1, the cycle is also (k, -k) (the effect of the sign appears when \delta_\varepsilon is composed with \pi).
\pi: a pairing (an element of \mathrm{PM}^c(\pm[n]), the set of pairings on the signed integers that are compatible with the matrix model — see below), encoding the edge identifications.
\mathrm{PM}^c(\pm[n]): the set of "compatible perfect matchings" on \{\pm 1, \ldots, \pm n\}. Compatibility conditions depend on the matrix model: for Ginibre, \pi must be a *-pairing (pairing positive with negative); for GOE, any pairing is allowed.
f^c(\pi): a weight factor depending on the pairing and the matrix model (often 1).
N^{\chi(\gamma, \delta_\varepsilon \pi \delta_\varepsilon) - 2\#(\gamma)}: the power of N (matrix size) contributed by this term, determined by the Euler characteristic of the surface defined by faces \gamma and edge identifications \delta_\varepsilon \pi \delta_\varepsilon, adjusted by 2\#(\gamma) (the number of faces times 2, a normalisation from the trace normalisation).
\mathrm{tr}_{\gamma_-^{-1} \delta_\varepsilon \pi^{-1} \delta_\varepsilon \gamma_+ / 2}(Y_1, \ldots, Y_n): the product of traces of the Y_k matrices arising from the vertices of the glued surface. The subscript \gamma_-^{-1} \delta_\varepsilon \pi^{-1} \delta_\varepsilon \gamma_+ / 2 is the vertex permutation (halved, since the cover produces two copies of each vertex).

What this formula computes operationally:

Input: a product of traces containing deterministic matrices Y_k and Gaussian random matrices X (or X^T), with a specified transpose structure \varepsilon.
Sum: over all pairings \pi of the signed edge-ends that are compatible with the matrix model. For each pairing:
- Compose \pi with \delta_\varepsilon on both sides (\delta_\varepsilon \pi \delta_\varepsilon) to account for the transpose structure — this adjusts the sign patterns so that the pairing correctly represents the index equalities from Wick's theorem.
- Compute the Euler characteristic \chi(\gamma, \delta_\varepsilon \pi \delta_\varepsilon) using the formula from Section 3.
- Compute the vertex permutation \gamma_-^{-1} \delta_\varepsilon \pi^{-1} \delta_\varepsilon \gamma_+ / 2 to determine the surviving Y_k trace products.
- Weight by f^c(\pi) and N^{\chi - 2\#(\gamma)}.
Output: the expected value, expressed as a sum of products of expected traces of the Y_k matrices, with explicit powers of N.

Why this form: This formula is the precise combinatorial encoding of "draw faces, apply Wick pairings, glue edges, read off vertices, compute Euler characteristic." The conjugation by \delta_\varepsilon is the algebraic representation of the rule that the index constraints from pairing X with X^T differ from pairing X with X. The composition \delta_\varepsilon \pi \delta_\varepsilon translates the abstract pairing into the correct sign patterns for the premap edge permutation. The factor N^{\chi - 2\#(\gamma)} embodies the fundamental principle that each vertex trace contributes a factor N (from summing over a free index in \{1, \ldots, N\}), each edge identification reduces the number of free indices by two (costing a factor N^{-1} relative to a face), and the net power is the Euler characteristic.

Contrasting Ginibre and GOE:

Ginibre (Example 4.1): For real Ginibre matrices Z, the entries are i.i.d. real normal. The expansion uses \delta_\varepsilon to enforce directional compatibility: edges for Z and Z^T have their index assignments reversed, and \delta_\varepsilon encodes this reversal. A pairing \pi is compatible if, after conjugation by \delta_\varepsilon, it satisfies a *-pairing condition — roughly, paired integers must have opposite signs (front-to-back pairing) because an untwisted gluing corresponds to Z with Z^T. The example in Section 4 computes the expected value of a specific two-trace product, showing that one pairing (the noncrossing *-pairing shown in Figure 11) contributes a term with \chi = 2 (a sphere) to produce a trace product \mathrm{tr}(Y_1 Y_3 Y_5^T Y_7^T) \mathrm{tr}(Y_2) \mathrm{tr}(Y_4 Y_6^T) \mathrm{tr}(Y_8) multiplied by N^{-2}.

GOE (Example 4.2): For GOE matrices T (real symmetric, T_{ij} = T_{ji} \sim \mathcal{N}(0, (1+\delta_{ij})/N)), the expansion admits all pairings — both untwisted and twisted — because T = T^T means there is no directionality to the edges. Calculating \mathbb{E}[\mathrm{tr}(T^2)] with two edges on a single face: the untwisted pairing (sphere) gives \chi = 2, contributing 1; the twisted pairing (projective plane) gives \chi = 1, contributing N^{-1}. The result, 1 + N^{-1}, comes from summing over all elements of \mathrm{PM}(\pm[2]) — the four pairings on four signed edge-ends, of which some are excluded for Ginibre but all are allowed for GOE.

Connectedness and Cumulants (Section 5)

The moment-cumulant formula organises the terms in the expansion by their connectedness. A surface is connected if it cannot be split into two nonempty sub-surfaces with no edges crossing between them. The moment-cumulant relationship is:

A moment (expected value of a product of traces) is a sum over all pairings, whether the resulting surface is connected or disconnected.
A cumulant k_r(\ldots) is a sum over only connected pairings — those that produce a single connected surface.

If a surface has multiple connected components (each component involving some subset of the original faces), then the faces in each component are multiplied together in a product of cumulants on smaller numbers of traces. The algebraic statement of the moment-cumulant formula for matrix trace expectations is that the disconnected terms are exactly those that can be partitioned by the face-connectivity of the surface.

Example: Figure 13 shows a surface with two connected components: one component involves the first and third traces (two faces), the other involves the second trace (one face). This term appears in the product k_2(\text{trace 1, trace 3}) \cdot k_1(\text{trace 2}), not in the joint cumulant k_3(\text{trace 1, trace 2, trace 3}). Figure 14 shows a connected surface involving all three faces, which does contribute to the joint cumulant k_3.

The geometric interpretation is that cumulants select for topologically connected surfaces, exactly analogous to how classical cumulants select for connected diagrams in perturbative expansions (connected Feynman diagrams, connected moments in probability theory). This connection is not incidental — the moment-cumulant formula for matrix trace expectations is the same combinatorial structure that appears in the relationship between moments and cumulants of any family of random variables, applied here with "connected" meaning "the surface formed by the Wick pairing has one connected component."

Centring and the Principle of Inclusion-Exclusion (Section 6)

When the matrices inside the traces are centred (their means are subtracted), the cumulant expansion acquires additional constraints. The centred cumulant is defined by expanding the centred traces as alternating sums of uncentred traces with subsets of terms taken at their means:

$k_r\left(\mathrm{tr}(\mathring{A}_1 \cdots \mathring{A}_{p_1}), \ldots, \mathrm{tr}(\mathring{A}_{p_1+\cdots+p_{r-1}+1} \cdots \mathring{A}_p)\right) = \sum_{K \subseteq [p]} (-1)^{|K|} \prod_{k \in K} \mathbb{E}[\mathrm{tr}(A_k)] \; \cdot \; k_r\left(\mathrm{tr}\left(\prod_{k \in [1,p_1]\setminus K} A_k\right), \ldots\right)$

What the centred condition means geometrically: A term A_k that appears "disconnected" in the surface gluing — that is, the face region corresponding to that term's matrix factor is part of a connected component consisting only of that term, with no edges connecting it to any other term — corresponds exactly to a factor of \mathbb{E}[\mathrm{tr}(A_k)] in the expansion. The alternating sum \sum_K (-1)^{|K|} in the inclusion-exclusion formula subtracts out all terms where any term is disconnected, then compensates for over-subtraction, leaving only terms where no term is disconnected.

The paper illustrates this with a specific example (Section 6, Figure 14) where the terms at positions \{2, 4, 5, 7\} are disconnected. This configuration appears in the inclusion-exclusion sum for every subset K \subseteq \{2, 4, 5, 7\} with appropriate sign, but the net contribution is that only configurations with K = \{2, 4, 5, 7\} survive — all terms with at least one disconnected term cancel. The geometric upshot is: centred cumulants correspond to surfaces where every face is connected to at least one other face via edge identifications, which for two-trace cumulants forces the two faces to be connected to each other.

Genus Expansion, Asymptotics, and Noncrossing Diagrams (Section 7)

The power of N in each term is N^{\chi - 2\#(\gamma)}. As N \to \infty, terms with larger Euler characteristic dominate. The classification of compact connected surfaces states:

Sphere: \chi = 2 (highest possible for a connected surface)
Connected sum of g tori: \chi = 2 - 2g (orientable, genus g)
Connected sum of k projective planes: \chi = 2 - k (nonorientable, demigenus k)

Disconnected surfaces have Euler characteristics that are sums of the components' Euler characteristics. The maximum possible Euler characteristic for a connected surface is 2 (the sphere). For r traces (faces), the highest-order terms have Euler characteristic 2r (r disjoint spheres) — but for cumulants, connectedness is required, so the highest-order cumulant terms have \chi = 2 (a single sphere).

Noncrossing diagrams encode spherical gluings. A surface gluing that produces a sphere can be drawn as a noncrossing pairing on the boundary of a disc (or on the boundaries of multiple discs whose interiors are connected by noncrossing chords). Crossings in the pairing correspond to handles (genus) in the surface: a single crossing forces the surface to have genus at least 1, lowering the Euler characteristic. Therefore, the leading-order (N^0) contributions to cumulants come precisely from noncrossing pairings.

The real case: two relative orientations matter. Figure 15 shows a noncrossing pairing on the two faces from Figure 1 that produces a sphere. However, note that the triangular face on the right has been flipped over relative to the pentagonal face — its back is visible. In this relative orientation, all edge identifications are untwisted (compatible with the Ginibre *-pairing condition). What about the opposite relative orientation, where the triangle's front is visible?

The asymptotic contribution for real matrices sums over all noncrossing pairings that work under both relative orientations of the constituent faces. Figure 17 shows the same faces and pairing as Figure 11, but with the right face flipped over a horizontal axis — on its back, transposes of the matrices on the front are shown. The pairing is checked to be a *-pairing (and noncrossing) under this relative orientation as well, so it counts toward the asymptotic contribution.

Example 7.2 (asymptotic covariance of Ginibre traces): To compute \lim_{N \to \infty} k_2(\mathrm{Tr}(ZZZ^TZ^T), \mathrm{Tr}(ZZ^TZZ^T)), the paper constructs the two faces (Figure 11) and counts annular noncrossing *-pairings that connect the two faces under both relative orientations. The counting yields 4 pairings in each orientation, totalling 8, giving \lim k_2 = 8. The geometric constraint — that the pairing must connect two faces with at least two chords between them, those chords must connect adjacent pairs of edges to adjacent pairs to avoid crossings, and the chords must be *-pairings — is satisfied by exactly four configurations per orientation.

Spoke Diagrams and Freeness (Section 8)

The final section applies the preceding machinery to real second-order freeness: the asymptotic vanishing of certain cumulants when the centred terms come from cyclically alternating algebras (algebras \mathcal{A} and \mathcal{B} such that terms in a trace alternate between \mathcal{A} and \mathcal{B}).

The geometric constraint from alternating algebras. When terms alternate (e.g., A_1 B_1^T A_2 B_2^T \cdots), any edge connecting a term to its immediate neighbour on the same face would pair factors from the same algebra. This is forbidden by the cyclic alternating condition (which requires that no two adjacent terms on a face belong to the same algebra). Combined with centring (no term can be disconnected), this forces any connecting edge to go to a non-neighbouring term on the same face, or to a term on the other face in a two-trace cumulant.

Why non-neighbouring connections on the same face are problematic: A chord connecting two non-adjacent positions on a face creates two regions separated by the chord. Any term inside one region cannot connect to terms in the other region without crossing the chord (which would create handles and lower the Euler characteristic below the leading order). In the asymptotic limit (noncrossing required), this forces terms inside the enclosed region to either connect to each other (creating a smaller enclosed region, recursively) or be disconnected — but disconnected terms are forbidden by centring. The only noncrossing resolution is that all connections go between the two faces, not within a single face.

The spoke diagram structure (Figure 18). For a cumulant of two traces, each trace being a cyclic product of alternating terms A_k and B_k^T, the asymptotic noncrossing condition forces a structure where each term is connected to exactly one term on the other face, and the connections form a pattern resembling spokes on a wheel. Figure 18 shows an example: the terms A_1, A_2, A_3 on one face connect respectively to B_2^T, B_3^T, B_1^T on the other face. The specific spoke pattern depends on the cyclic offsets between the two traces.

Multiplicative factorisation over spokes. Each spoke — connecting a pair of terms, say A_k on face 1 and B_\ell^T on face 2 — contributes a factor equal to the asymptotic connected correlation of those two specific terms. Geometrically, the region between two consecutive spokes is a disc bounded by segments of the two faces and the two spoke edges. The possible noncrossing pairings within that disc that connect the A_k interval to the B_\ell^T interval must be summed over — but these are precisely the diagrams that contribute to \mathbb{E}[\mathrm{tr}(A_k B_\ell^{(\pm 1)})] - \mathbb{E}[\mathrm{tr}(A_k)] \mathbb{E}[\mathrm{tr}(B_\ell^{(\pm 1)})]. The subtraction removes the disconnected configuration where A_k and B_\ell^T are not connected to each other (Figure 19 shows the connected disc-noncrossing diagrams on a single spoke).

The cumulant as a product over spokes:

$\lim_{N \to \infty} k_2(\mathrm{Tr}(\mathring{A}_1 \cdots \mathring{A}_r), \mathrm{Tr}(\mathring{B}_1^{(\pm 1)} \cdots \mathring{B}_s^{(\pm 1)})) = \sum_{\text{spoke diagrams}} \prod_{\text{spokes } (k, \ell)} \left(\mathbb{E}[\mathrm{tr}(A_k B_\ell^{(\pm 1)})] - \mathbb{E}[\mathrm{tr}(A_k)] \mathbb{E}[\mathrm{tr}(B_\ell^{(\pm 1)})]\right)$

Example 8.1 (concrete computation with three matrix models): The paper computes an asymptotic two-trace cumulant involving three different matrix models:

A_1 = ZZZ^T (Ginibre, colour 1)
A_2 = T^2 (GOE, colour 2)
A_3 = W^2 (Wishart, colour 3)
B_1 = W^5, B_2 = ZZZ^T, B_3 = T^4

Only one spoke diagram (k=1 — the reversed spoke arrangement, corresponding to the bottom-centre diagram in Figure 1 of the main paper) accommodates the constraint that the Ginibre-generated terms A_1 and B_2^T connect to each other. For each spoke, the contribution is computed using known moment formulas for the three matrix models (e.g., \mathbb{E}[\mathrm{tr}(T^6)] - \mathbb{E}[\mathrm{tr}(T^2)]\mathbb{E}[\mathrm{tr}(T^4)] = 5 - 1 \cdot 2 = 3 for the GOE spoke; \mathbb{E}[\mathrm{tr}(W^7)] - \mathbb{E}[\mathrm{tr}(W^2)]\mathbb{E}[\mathrm{tr}(W^5)] = 429 - 2 \cdot 42 = 345 for the Wishart spoke; 2 for the Ginibre spoke). The product is 2 \cdot 3 \cdot 345 = 2070, giving the asymptotic cumulant. This example demonstrates the concrete predictive power of the formalism: once the combinatorial structure (spoke diagram constraints) is understood, the actual value reduces to elementary products of known single-trace moments.

4. Key Insights and Innovations

Innovation 1: Twisted Edge Identifications as the Conceptual Bridge Between Real Gaussian Moments and Nonorientable Topology

The paper's most fundamental conceptual move is not the permutation formalism itself, but the identification of twisted edge identifications as the precise mechanism by which real Gaussian matrix calculations produce nonorientable surfaces — and, conversely, why complex Gaussian calculations are restricted to orientable topology. This insight is the intellectual linchpin that makes the entire extension from complex to real second-order freeness coherent.

Prior work (Zvonkin, 1997; Lando and Zvonkin, 2004) had already established the topological interpretation of complex matrix integrals as map enumeration problems on orientable surfaces. The orientability constraint was baked into the formalism — edges carry a direction (from untransposed to transposed, or Z to Z^*), and Wick pairings couple oppositely directed edges, producing only untwisted gluings. The exercises to Lando and Zvonkin (2004, Chapter 3) note that "calculations with real matrices include nonorientable surfaces," but prior to the main paper, no unified combinatorial framework had been developed to handle those nonorientable surfaces in the context of asymptotic freeness.

What distinguishes this paper's contribution is that it diagnoses exactly where the orientability constraint breaks and provides the geometric language to understand the breakage. The distinction is in the Gaussian moment structure: \mathbb{E}[f^2] = 1 for real standard normals means a pairing can couple X with X (both untransposed) or X^T with X^T (both transposed), not just X with X^T. The paper's Figures 2 and 3 crystallize this: an untwisted gluing (Figure 2) matches arrow heads to tails, preserving local orientation across the seam; a twisted gluing (Figure 3) matches heads to heads, reversing orientation and creating a Möbius-strip-like nonorientability in the resulting surface.

This is a fundamental diagnostic concept, not an incremental refinement. It answers why the complex theory is simpler (pairings are constrained to be untwisted by the vanishing of \mathbb{E}[Z^2] and \mathbb{E}[\overline{Z}^2]), why the real theory is richer (twisted pairings produce projective planes, Klein bottles, and their connected sums), and why the GOE calculation \mathbb{E}[\mathrm{tr}(T^2)] = 1 + N^{-1} (Example 4.2, Figure 12) splits into two terms — the sphere from untwisted gluing and the projective plane from twisted gluing. The sphere term has Euler characteristic 2 and dominates asymptotically; the projective plane term has Euler characteristic 1 and is suppressed by N^{-1}. The complex analogue would have only the sphere term, since the twisted pairing is impossible for complex Gaussians.

The significance extends beyond a single example. This diagnostic concept provides a unified geometric language for understanding which terms survive asymptotically in real matrix cumulants (only untwisted, noncrossing pairings on spheres), which are subleading (twisted pairings, orientable higher-genus surfaces), and how the real and complex asymptotic fluctuations differ (the real case must count noncrossing pairings under both relative orientations of faces, not just one, because twisted pairings are excluded at leading order but the relative orientation constraint persists). Without this conceptual bridge, the permutation formulas in the main paper — \delta_\varepsilon \pi \delta_\varepsilon, the premap conditions on \pi, the double-cover vertex formula \gamma_-^{-1} \pi \gamma_+ — would appear as unmotivated algebraic manipulations rather than the direct combinatorial translations of "which edge gluings are twisted and which aren't."

Innovation 2: The Premap as a Combinatorial Double Cover for Nonorientable Surfaces in Matrix Calculations

The paper introduces (and explains) the premap formalism — a permutation-based encoding of surface gluings on possibly-nonorientable surfaces via their orientable double covers — as the combinatorial backbone for real random matrix moment and cumulant calculations. While the premap concept itself has precursors in Tutte (1984) for graph theory, and the double cover construction is standard in algebraic topology (Hatcher, 2002, pp. 234–235), the paper's innovation is applying and operationalizing premaps specifically for the trace product expansions that arise in random matrix freeness theory.

Prior to this work, the standard combinatorial tool for matrix calculations was the map or hypermap formalism (Cori, 1975; Zvonkin, 1998; Lando and Zvonkin, 2004): a triple of permutations (\sigma, \alpha, \phi) on edge-ends encoding vertices, edges, and faces of an oriented surface gluing. This formalism implicitly assumes a globally consistent counterclockwise orientation. When the surface is nonorientable — as occurs with twisted edge pairings in real matrix calculations — the map formalism breaks down because "counterclockwise" is not globally defined. Tutte (1984) developed an encoding for unoriented maps by assigning each edge two ends with opposite signs and defining permutations that may mix signs, but this work was written within pure graph theory and was not connected to random matrix trace expansions, asymptotic 1/N scaling, or the specific pairing structures (\delta_\varepsilon-conjugation, compatibility conditions for Ginibre vs. GOE) that matter for freeness.

What this paper does that is conceptually novel for the random matrix community is to show that the premap encoding — specifically, the face permutation \gamma_+ \gamma_-^{-1}, the edge permutation \pi with sign patterns encoding twist status, and the vertex permutation \gamma_-^{-1} \pi \gamma_+ — is the exact algebraic translation of the geometric double cover construction. Every element of the encoding has a direct geometric interpretation:

\gamma_+ encodes the front faces of the cover (positive orientation), \gamma_-^{-1} encodes the back faces (reversed orientation), and their product \gamma_+ \gamma_-^{-1} enumerates all faces of the covering space. The premap condition (every cycle has a mirror cycle with signs flipped and order reversed) corresponds geometrically to the fact that the cover is a genuine surface where every face has a front and back.
The sign pattern in a cycle of \pi — paired integers with the same sign versus opposite signs — encodes whether the corresponding edge identification is untwisted or twisted. This is not an arbitrary convention; it follows directly from whether front glues to front (untwisted, same sign) or front glues to back (twisted, opposite sign) in the cover.
The Euler characteristic formula \chi = \frac{1}{2}\#(\gamma_+ \gamma_-^{-1}) + \frac{1}{2}\#(\pi) + \frac{1}{2}\#(\gamma_-^{-1} \pi \gamma_+) - |I| is the standard V - E + F but computed in the cover and halved (since each original feature has two preimages).

Example 3.3 and Figure 9 are particularly instructive: a nonorientable hypermap is drawn with twists in its edges (shown as crossings in the diagram), and the resulting vertex computation via \gamma_-^{-1} \pi \gamma_+ is shown to produce cycles that correspond to boundary walks in the diagram — the vertex (1, -6, 7, 5) can be traced explicitly by following faces and hyperedges in the figure. This anchors the algebra in geometry: the reader who encounters \gamma_-^{-1} \pi \gamma_+ in the main paper can now visualize it as "walk around a vertex in the double cover, flipping between front and back when crossing a twisted edge."

The significance of this innovation is that it transforms the premap from an opaque combinatorial gadget into a transparent and teachable tool. A researcher who understands the geometric origin of the premap formulas can reason about the structure of matrix calculations without performing algebraic manipulations — they can visualize the face gluing, check whether edge identifications are twisted, trace vertex boundaries in the double cover, and directly read off the resulting trace products and Euler characteristic. This representational shift is what makes the main paper's proofs intelligible, and it is a genuine pedagogical contribution to the random matrix literature.

Innovation 3: The Two-Orientation Condition as the Structural Distinction Between Real and Complex Asymptotic Fluctuations

Beyond the mechanics of encoding nonorientable surfaces, the paper articulates a structural condition that distinguishes real from complex asymptotic fluctuations: for real matrices, the noncrossing pairings contributing to the leading-order cumulants must be valid under both relative orientations of the constituent faces. This condition is entirely absent in the complex case, where only one relative orientation is compatible with the *-pairing constraints, and it is the ultimate source of the different asymptotic second-order structure for real vs. complex ensembles.

The geometric origin of this condition is subtle. The highest-order asymptotic terms (N^0 for cumulants) come from spherical surface gluings with untwisted edge identifications. An untwisted gluing requires that the two faces meeting at each edge have their local orientations aligned — the fronts face the same way at the seam. For a given pairing to produce an UNTWISTED gluing on a spherical surface, it must be possible to orient all faces consistently. But a single face can be embedded in the plane in two ways: "front up" or "front down" (flipped over). For a given pairing of edges between two faces, the pairing may be realizable as a noncrossing, untwisted gluing when face A is front-up and face B is front-up, but NOT when face B is flipped over — because flipping the face reverses the roles of X and X^T on its edges (the transpose of what was on the front appears on the back), and the pairing may no longer satisfy the *-pairing condition (pairing untransposed with transposed) in that orientation.

Figure 11 and Figure 17 demonstrate this concretely. The pairing shown in Figure 11 works as a *-pairing when the right face is in its original orientation. Figure 17 shows the same faces and the same pairing, but with the right face flipped over a horizontal axis — on its back, transposes of the matrices on the front appear. The pairing is checked to also work as a *-pairing in this flipped orientation. The asymptotic contribution from this pairing is counted because it survives under both relative orientations.

This contrasts sharply with the complex setting. For complex Ginibre matrices, the moment constraints (\mathbb{E}[Z^2] = 0) force every edge identification to be untwisted and to pair a Z with a Z^*. The *-pairing condition is so restrictive that a pairing that works on both relative orientations in the real case would only work on one relative orientation in the complex case, because flipping a face swaps which edges are starred and which aren't. Consequently, the complex asymptotic fluctuations are given by counting noncrossing *-pairings on the faces in their original orientation only. The real case requires the union of pairings that work in both orientations — a genuinely different counting problem.

This insight is fundamentally a theoretical diagnostic, not an algorithmic improvement. It explains why real second-order freeness differs from complex second-order freeness even when the real ensemble is the "real version" of the complex one (real Ginibre vs. complex Ginibre). The difference is not in the 1/N expansion's structure (both have genus expansions) or in the noncrossing condition (both require noncrossing for leading order), but in the compatibility condition that selects which noncrossing pairings count. Mingo and Nica (2004) characterized complex second-order freeness entirely in terms of annular noncrossing permutations satisfying a *-condition; this paper shows that the real analogue requires annular noncrossing permutations satisfying a *-condition under both cyclic directions (both relative orientations), which is a strictly stronger condition that filters out some pairings that would contribute in the complex case.

Example 7.2 makes this quantitative: computing \lim_{N \to \infty} k_2(\mathrm{Tr}(ZZZ^TZ^T), \mathrm{Tr}(ZZ^TZZ^T)) for real Ginibre produces 8 contributing pairings (4 in each relative orientation). The complex analogue would produce a different number (likely fewer), and the asymptotic covariance would differ accordingly. The example thus serves as a verification that the two-orientation condition is not a minor technical nuance but a quantitatively consequential constraint on the asymptotic structure.

Innovation 4: The Spoke Diagram Decomposition as a Unified Geometric Framework for Real Second-Order Freeness

Section 8 crystallizes the preceding topological and combinatorial machinery into a spoke diagram factorization that characterizes the asymptotic two-trace cumulants of cyclically alternating algebras. While the spoke diagram structure itself is inherited from the complex case (it appears in the annular noncrossing formalism of Mingo and Nica, 2004), the paper's innovation is showing geometrically how the alternating algebra condition, centring, and noncrossing constraints jointly force the spoke structure — and how the real case's two-orientation condition integrates into this factorization.

The geometric reasoning is a study in constraint propagation. Start with two faces (one per trace), each a cycle of alternating terms A_k and B_k^T. The alternating condition forbids pairing adjacent terms on the same face (they belong to the same algebra). Centring forbids any term from being disconnected. The noncrossing condition (required for asymptotic leading order) means that any chord connecting two non-adjacent terms on the same face would isolate the terms between them, forcing those interior terms to be either interconnected (creating a smaller noncrossing structure that recursively forces isolation of still-smaller sets) or disconnected (forbidden by centring). The only resolution to avoid disconnection while respecting noncrossing is that all connections go between the two faces. Once all connections cross between faces, the noncrossing condition forces them to form a nested or spoke-like arrangement — any "crossing" of inter-face connections would create a handle (genus) and drop the Euler characteristic below 2.

The paper's Figures 18 and 19 make this visual. Figure 18 shows a complete spoke diagram: each A_k connects to one B_\ell^T, and the connections do not cross. Figure 19 zooms in on a single spoke — the region between two consecutive spoke edges, which is topologically a disc, and shows the noncrossing pairings within that disc that connect the A_k segment to the B_\ell^T segment. The contribution of that spoke is \mathbb{E}[\mathrm{tr}(A_k B_\ell^{(\pm 1)})] - \mathbb{E}[\mathrm{tr}(A_k)] \mathbb{E}[\mathrm{tr}(B_\ell^{(\pm 1)})] — the connected correlation of those two terms, with the disconnected product subtracted because centring excludes the configuration where A_k and B_\ell^T are in separate connected components.

The multiplicative factorization over spokes is what makes this framework practically computable: rather than enumerating all possible surfaces connecting the two traces (which is combinatorially explosive), one enumerates the possible spoke patterns (which are few, constrained by the cyclic orders of the traces) and multiplies pre-computed spoke contributions. This reduces the asymptotic second-order freeness calculation to elementary single-trace moments of the constituent matrix models.

What elevates this from a merely useful decomposition to an intellectual innovation is how it unifies the treatment of different matrix ensembles within a single framework. Example 8.1 demonstrates this directly: a single two-trace cumulant involves terms generated by three different matrix models — Ginibre (ZZZ^T), GOE (T^2), and Wishart (W^2). Each spoke's contribution is computed using the moment formulas appropriate to its matrix model (e.g., \mathbb{E}[\mathrm{tr}(T^6)] - \mathbb{E}[\mathrm{tr}(T^2)]\mathbb{E}[\mathrm{tr}(T^4)] = 3 for the GOE spoke; 429 - 2 \cdot 42 = 345 for the Wishart spoke; 2 for the Ginibre spoke). The factorization 2 \cdot 3 \cdot 345 = 2070 gives the asymptotic cumulant. The framework thus handles mixed ensembles where different algebras are generated by different random matrix models — a level of generality that goes beyond treating each ensemble in isolation.

The design choice that makes this work is the separation of concerns: the topology (spoke diagram constraints, noncrossing conditions) depends only on the algebra structure (alternating, cyclically alternating) and centring; the specific matrix model (Ginibre, GOE, Wishart) affects only the per-spoke correlation factors. This modularity means that proving real second-order freeness for a new ensemble reduces to computing its single-trace moment formulas and verifying they satisfy the necessary conditions — the topological and combinatorial structure is already handled by the premap framework.

5. Experimental Analysis

Evaluation Methodology

Dataset. This paper is a pedagogical appendix and does not introduce new experimental datasets. The running examples throughout are constructed ad hoc to illustrate the combinatorial and topological constructions: the main running example (Section 2, Figures 1–5) uses an expected product of two traces involving a Gaussian random matrix X and eight deterministic matrices Y_1, \ldots, Y_8; Example 4.1 uses two traces with a real Ginibre matrix Z and eight Y_k matrices; Example 4.2 uses E[tr(T^2)] for a GOE matrix T; Section 5 uses three-trace cumulant examples; Section 7 calculates covariances of Ginibre trace products; Example 8.1 computes an asymptotic two-trace cumulant involving Ginibre, GOE, and Wishart matrices. All examples are chosen to demonstrate specific structural features (twisted vs. untwisted gluings, connectedness, noncrossing conditions, spoke diagram constraints) rather than to establish empirical benchmarks.
Base model(s). This paper is not a machine learning paper and does not involve trained models. The relevant objects are random matrix ensembles: real Ginibre matrices (Z, with i.i.d. N(0, 1/N) entries), GOE matrices (T, real symmetric with T_{ij} = T_{ji} \sim N(0, (1+\delta_{ij})/N)), and Wishart matrices (W = X^T X where X has i.i.d. N(0, 1/N) entries). These are standard ensembles in random matrix theory, chosen because they are the three canonical real ensembles whose asymptotic freeness properties are analyzed in the main paper.
Metrics. The quantities computed throughout are expected values of products of traces (moments) and their connected versions (cumulants), both at finite N (exact expected values with explicit powers of 1/N) and in the N \to \infty limit (asymptotic leading-order contributions). For normalized traces (\mathrm{Tr} = \frac{1}{N}\mathrm{tr}), cumulants of products of traces are the natural objects for second-order freeness; the asymptotic values (remainder after extracting the N^0 term) are computed by counting noncrossing pairings satisfying compatibility conditions. The paper computes these values combinatorially — by enumerating valid pairings and applying the premap Euler characteristic formula — rather than by simulation.
Baselines. The paper's "baselines" are theoretical expectations from known formulas. In the complex case, Mingo and Nica (2004) provide the annular noncrossing permutation characterization; for real matrices, the main paper's formulas (Equation at the start of Section 4 here) provide the sum-over-pairings expansion. The paper verifies its combinatorial counts against these formulas by explicit enumeration: for Ginibre (Example 4.1), the *-pairings on the faces are enumerated and checked for compatibility with \delta_\varepsilon; for GOE (Example 4.2), all pairings (both twisted and untwisted) are enumerated; for Example 7.2, annular noncrossing *-pairings are counted by inspection under both relative orientations. The explicit numerical results (1 + N^{-1} for GOE E[tr(T^2)]; 1 for the asymptotic Ginibre moment E[tr(ZZZ^TZ^T)]; 2 for E[tr(ZZ^T Z Z^T)]; 8 for their asymptotic covariance; 2070 for the three-ensemble cumulant in Example 8.1) serve as consistency checks that the combinatorial framework reproduces correct values.
Generation budget / compute accounting. Not applicable in the conventional ML sense. The "cost" of a calculation is measured by the Euler characteristic \chi(\gamma, \pi) of the surface produced by a given face structure \gamma and edge pairing \pi. The power of N in each term is N^{\chi - 2\#(\gamma)}, so terms with larger Euler characteristic (more vertices relative to edges and faces) dominate as N \to \infty. The highest Euler characteristic for a connected surface is \chi = 2 (the sphere), which produces the leading N^0 term for cumulants. Nonorientable surfaces (projective plane, \chi = 1; Klein bottle, \chi = 0; etc.) and higher-genus orientable surfaces (torus, \chi = 0; etc.) contribute to subleading orders. This "Euler characteristic accounting" is the combinatorial analogue of measuring compute scaling: higher \chi = higher order in N = dominant asymptotic term.
Cross-validation / statistical protocol. All calculations in this paper are exact combinatorial enumerations, not statistical estimates. There is no train/test split, no sampling, no confidence intervals. The "validation" consists of checking that the enumerated pairings satisfy the required combinatorial constraints (noncrossing, *-pairing, compatibility with \delta_\varepsilon, premap conditions) and that the resulting trace products and Euler characteristics are consistent with known results. For instance, the Klein bottle example (Example 3.4) yields \chi = 2 + 4 + 2 - 8 = 0, which is the correct Euler characteristic for a Klein bottle, confirming the counting. Example 8.1 verifies the spoke diagram factorization by multiplying known single-trace moments and obtaining 2070, which can be checked against independent computation.

Main Quantitative Results

Surface Gluings and Euler Characteristics for the Running Example (Section 2)

The central experimental demonstration in Section 2 is that a specific trace product expectation — E[tr(X Y_1 X Y_2 X^T Y_3 Y_4 X^T Y_5) · tr(X^T Y_6 X Y_7 X Y_8)], with X having i.i.d. N(0, 1/N) entries — decomposes into a sum over pairings \rho of the eight Gaussian matrix occurrences. For each pairing, the faces shown in Figure 1 are glued according to whether each paired edge produces an untwisted (Figure 2) or twisted (Figure 3) identification.

Headline result for one specific pairing: The pairing shown in Figure 4 (corresponding to \pi = (1,-7)(-1,7)(2,-4)(-2,4)(3,-6)(-3,6)(5,8)(-5,-8), with some twisted and some untwisted identifications) produces a connected, nonorientable surface. The vertex computation (Example 3.4) yields \gamma_-^{-1} \pi \gamma_+ = (1,-3,6,-5,-7)(7,5,-6,3,-1)(2,-8,-4)(4,8,-2), with the halved version \gamma_-^{-1} \pi \gamma_+ / 2 = (1,-3,6,-5,-7)(2,-8,-4) giving two vertices. The Euler characteristic is \chi = 2 + 4 + 2 - 8 = 0. Since the surface is connected and nonorientable with \chi = 0, it is a Klein bottle. The vertices contribute the trace product shown in Figure 5: tr(Y_1 Y_3^T Y_6 Y_5^T Y_7^T) (and its transpose, which is equal). The term thus appears at order N^{\chi - 2\#(\gamma)} = N^{0 - 4} = N^{-4} (since \#(\gamma) = 2 traces), multiplied by the expected value of that vertex trace product.

This single example demonstrates the entire pipeline: face construction → edge pairing → surface gluing → vertex recovery → Euler characteristic → contribution to the expected value. The fact that the surface is a Klein bottle rather than a sphere is a direct consequence of the twisted edge identifications — a sphere would require \chi = 2, which is impossible here because the vertex count (2) minus edge count (8) plus face count (2) gives 2 - 8 + 2 = -4 in the cover, halved to -2, but the paper's formula yields 0 because of the different accounting for faces and vertices in the cover.

Ginibre vs. GOE Pairing Enumeration (Section 4, Examples 4.1 and 4.2)

Example 4.1 (Real Ginibre): For the expression E[tr(Z Y_1 Z Y_2 Z^T Y_3 Z^T Y_4) · tr(Z Y_5 Z^T Y_6 Z Y_7 Z^T Y_8)] with real Ginibre Z, the faces are shown in Figure 11. The pairing marked in the figure corresponds to \rho = (1,7)(2,3)(4,6)(5,8) on the unsigned indices. After conjugation by \delta_\varepsilon (which encodes the transpose structure), the pairing becomes \pi = \delta_\varepsilon \rho \delta_\varepsilon = (1,-7)(-1,7)(2,3)(-2,-3)(4,-6)(-4,6)(5,8)(-5,-8). Paired integers of the same sign represent untwisted identifications; paired integers of opposite sign represent twisted ones.

The vertex computation yields \gamma_-^{-1} \pi \gamma_+ = (1,3,-5,-7)(7,5,-3,-1)(2)(-2)(4,-6)(6,-4)(8)(-8), which after halving gives four vertices: (1,3,-5,-7), (2), (4,-6), and (8). The Euler characteristic is \chi = 2 + 4 + 4 - 8 = 2, so this term corresponds to a sphere (connected, orientable since \chi = 2). The contribution is E[tr(Y_1 Y_3 Y_5^T Y_7^T) tr(Y_2) tr(Y_4 Y_6^T) tr(Y_8)] N^{-2}. For the special case Y_k = I_N for all k, this gives the asymptotic moment lim_{N \to \infty} E[tr(Z Z Z^T Z^T)] = 1, which is verified by the fact that only one pairing (the noncrossing *-pairing shown) contributes at leading order.

Example 4.2 (GOE): For E[tr(T^2)] with T a GOE matrix, a single face with two edges is constructed. The possible pairings on \{\pm 1, \pm 2\} that are compatible with GOE (all pairings allowed) are:

The untwisted pairings: (1,2)(-1,-2) and (1,-1)(2,-2) (these produce a sphere, \chi = 2)
The twisted pairings: (1,-2)(-1,2) and its mirror (these produce a projective plane, \chi = 1)

The paper states (Figure 12 shows these two surfaces): "These two surfaces correspond to the two elements of PM(\pm[2]) \cap P_2(\pm[2]), that is, (1,2)(-1,-2) and (1,-2)(-1,2) respectively." The untwisted term contributes 1 (from \chi = 2, giving N^{2-2} = N^0 = 1), and the twisted term contributes N^{-1} (from \chi = 1, giving N^{1-2} = N^{-1}). The total expected value is E[tr(T^2)] = 1 + N^{-1}, shown in Figure 12. This is a classic result in random matrix theory, reproduced here to illustrate how the premap framework handles both orientable and nonorientable contributions within a unified summation.

For the slightly more involved expression E[tr(T Y_1 T Y_2)] with deterministic Y_1, Y_2, the paper reports: "we get E(tr(T Y_1 T Y_2)) = E(Y_1)E(Y_2) + E(Y_1 Y_2^T)N^{-1}; the matrices Y_1 and Y_2 appear in different vertices in the sphere, but appear in the same vertex but opposite orientation in the projective plane." This illustrates how vertex structure — which matrices are multiplied together in a trace — depends on whether the gluing is untwisted or twisted.

Asymptotic Covariance Counting for Real Ginibre (Section 7, Examples 7.1 and 7.2)

Example 7.1 (Asymptotic moments): For the 4-gon face (left face in Figure 11, representing tr(Z Y_1 Z Y_2 Z^T Y_3 Z^T Y_4)), there are three possible pairings on the edges (Figure 16): the leftmost (noncrossing, pairing adjacent edges), the centre (crossing), and the rightmost (noncrossing, pairing adjacent edges). Only the centre and right pairings are *-pairings (pairing Z with Z^T), so only one term (the rightmost, since the centre pairing is crossing and drops to lower order) contributes to the asymptotic moment:

$\lim_{N \to \infty} E[\mathrm{tr}(Z Y_1 Z Y_2 Z^T Y_3 Z^T Y_4)] = E[\mathrm{tr}(Y_1 Y_3) \mathrm{tr}(Y_2) \mathrm{tr}(Y_4)]$

For the special case Y_k = I_N, this yields lim E[tr(Z Z Z^T Z^T)] = 1 — exactly one noncrossing *-pairing contributes.

For the 4-gon on the right of Figure 11 (representing tr(Z Y_5 Z^T Y_6 Z Y_7 Z^T Y_8)), the left and right pairings in Figure 16 (but NOT the centre one) are *-pairings, so TWO noncrossing *-pairings contribute:

$\lim_{N \to \infty} E[\mathrm{tr}(Z Y_5 Z^T Y_6 Z Y_7 Z^T Y_8)] = E[\mathrm{tr}(Y_5 Y_7) \mathrm{tr}(Y_6) \mathrm{tr}(Y_8)] + E[\mathrm{tr}(Y_5) \mathrm{tr}(Y_6 Y_8) \mathrm{tr}(Y_7)]$

For Y_k = I_N, this gives lim E[tr(Z Z^T Z Z^T)] = 2.

These counts demonstrate the central mechanism: asymptotic moments are given by counting noncrossing *-pairings on the faces. The difference between 1 and 2 reflects the different arrangement of transposes on the two faces — the left face has Z, Z, Z^T, Z^T while the right face has Z, Z^T, Z, Z^T.

Example 7.2 (Asymptotic covariance): To compute lim_{N \to \infty} k_2(Tr(Z Z Z^T Z^T), Tr(Z Z^T Z Z^T)), the paper constructs both faces (Figure 11) and counts annular noncrossing *-pairings connecting the two faces, under BOTH relative orientations.

For the original relative orientation (Figure 11): the pairing must connect the two faces with at least two chords (since the cumulant requires connectedness). To be noncrossing, the chords must pair an adjacent pair of edges on one face with an adjacent pair on the other. By inspection, there are 4 such pairings.

For the flipped relative orientation (Figure 17, where the right face is flipped over a horizontal axis, showing transposes on its back): the pairing shown in Figure 11 is checked to also be a *-pairing in this orientation. Again by inspection, there are 4 noncrossing *-pairings connecting the two faces.

Total: 4 + 4 = 8. The paper states:

"Thus, \lim_{N \to \infty} k_2(Tr(Z Z Z^T Z^T), Tr(Z Z^T Z Z^T)) = 8."

This is the key quantitative demonstration of the two-orientation condition for real matrices: the asymptotic covariance is the sum of noncrossing *-pairing counts under both relative orientations. In the complex analogue (complex Ginibre), only one relative orientation would be compatible with the *-pairing constraint, so the count — and thus the asymptotic covariance — would differ.

Spoke Diagram Factorization and Mixed-Ensemble Cumulant (Section 8, Example 8.1)

The paper's most elaborate quantitative demonstration is Example 8.1, which computes the asymptotic two-trace cumulant:

$\lim_{N \to \infty} k_2\left(\mathrm{Tr}(\mathring{A}_1 \mathring{A}_2 \mathring{A}_3), \mathrm{Tr}(\mathring{B}_1 \mathring{B}_2 \mathring{B}_3)\right)$

where:

A_1 = Z Z Z^T (Ginibre, colour 1), A_2 = T^2 (GOE, colour 2), A_3 = W^2 (Wishart, colour 3)
B_1 = W^5, B_2 = Z Z Z^T, B_3 = T^4

The paper states that only one spoke diagram (the k=1 reversed spoke pattern, corresponding to the bottom-centre diagram in Figure 1 of the main paper) accommodates the constraint that the two Ginibre-generated terms A_1 and B_2^T connect to each other. The spokes pair:

A_1 (Ginibre) with B_2^T (Ginibre)
A_2 (GOE) with B_3^T (GOE)
A_3 (Wishart) with B_1^T (Wishart)

Each spoke contributes a connected correlation (moment minus product of means):

Ginibre spoke (A_1 = Z Z Z^T and B_2^T = (Z Z Z^T)^T): The paper states, referring to Figure 19, that there are exactly two diagrams on the disc connecting these terms, so the contribution is E[tr(Z Z Z^T (Z Z Z^T)^T)] - E[tr(Z Z Z^T)] E[tr((Z Z Z^T)^T)]. Using the asymptotic moment formulas (the Ginibre moment E[tr(Z Z Z^T Z^T Z^T Z)] — note that transposing B_2 and multiplying, (Z Z Z^T)^T = Z Z^T Z^T — yields 2, and the individual traces vanish or are subtracted), the paper reports: "the contribution of this spoke is [...] = 2" (the specific moment formula is said to equal 2 by inspection of the diagrams on the spoke).
GOE spoke (A_2 = T^2 and B_3^T = (T^4)^T = T^4): Using known GOE moment formulas (cited as "Remark 5.13 of the main paper"), the paper computes: E[tr(T^6)] - E[tr(T^2)] E[tr(T^4)] = 5 - 1 · 2 = 3.
Wishart spoke (A_3 = W^2 and B_1^T = (W^5)^T = W^5): Using known Wishart moment formulas (also from Remark 5.13 of the main paper), the paper computes: E[tr(W^7)] - E[tr(W^2)] E[tr(W^5)] = 429 - 2 · 42 = 345.

The asymptotic cumulant is the product over the three spokes:

$\lim_{N \to \infty} k_2(\ldots) = 2 \cdot 3 \cdot 345 = 2070$

This computation is significant because it demonstrates the modularity of the framework: the topological constraints (spoke diagram, one valid pattern) separate from the ensemble-specific moment computations. Once the spoke structure is determined, the cumulant reduces to multiplying independently computed single-trace connected correlations — and these can use known results from the random matrix theory literature (e.g., moment formulas for GOE, Wishart, and Ginibre ensembles) without requiring any new combinatorial enumeration of the full two-trace surface gluings.

The paper verifies the spoke diagram constraint qualitatively: "By inspection, we can see that this is the only spoke diagram that will accommodate a diagram on A_1 and B_2^T, the two members of the algebra generated by Z." This "inspection" refers to the geometric constraint that the two Ginibre faces must be connected by a chord in the noncrossing annular diagram, and only one spoke pattern places them opposite each other in the cyclic order.

Critical Assessment

Does the paper demonstrate what it claims to demonstrate?

The paper's central claim is pedagogical: that the permutation formulas in the main paper can be understood geometrically as encoding surface gluings, and that working through concrete examples makes this translation transparent. The experiments — by which we mean the worked examples and their quantitative verification — support this claim with substantial qualification.

What the examples successfully demonstrate:

The face gluing → surface topology pipeline works concretely. The running example (Sections 2–3) walks from a specific trace product, through a specific Wick pairing, to a specific surface (Klein bottle, \chi = 0), to a specific vertex trace product. Every step is shown visually and verified combinatorially: the Euler characteristic is computed and matches the expected topology, and the vertex traces are read off and shown to be consistent. This is a genuine achievement in exposition — it anchors the abstract permutation algebra in a single, fully-worked case.
The distinction between untwisted and twisted gluings is geometrically clear. Figures 2 and 3 leave no ambiguity about what "twisted" means in the surface gluing: arrow head to arrow tail (untwisted) versus arrow head to arrow head (twisted). The consequences — untwisted gluings preserve surface orientability; twisted gluings create nonorientability — are demonstrated by the two terms in Example 4.2 (sphere vs. projective plane, Figure 12) and by the Klein bottle in the running example.
The premap encoding is geometrically grounded in the double cover. Section 3's progression from oriented maps (Examples 3.1, 3.2) to the nonorientable double cover (Example 3.3, Figure 9) makes the derivation of \gamma_+ \gamma_-^{-1}, the sign patterns in \pi, and the vertex formula \gamma_-^{-1} \pi \gamma_+ visually traceable. The fact that \gamma_-^{-1} \pi \gamma_+ produces the correct vertex cycles is verified in Example 3.3 by tracing the boundaries in Figure 9.
The asymptotic counting is checked against known results. Example 7.1 reproduces the correct asymptotic moments for Ginibre traces (1 and 2 for the two 4-gons; Example 7.2 produces 8 for the covariance; Example 8.1 multiplies known moment differences to produce 2070. These numbers are not derived from first principles here — they import known results (e.g., E[tr(T^6)] = 5, E[tr(W^7)] = 429, etc.) from the main paper's Remark 5.13.

Significant weaknesses and missing validations:

No systematic enumeration or cross-checking. The paper works through individual pairings and individual surfaces — one Klein bottle, one sphere in Example 4.1, one projective plane in Example 4.2, one spoke diagram in Example 8.1. It does NOT perform a complete enumeration of all pairings for any nontrivial expression and verify that the sum reproduces a known result via an independent method (e.g., simulation or direct algebraic computation). The numbers 1, 2, 3, 8, 345, 2070 are stated without showing the full enumeration that produces them. A reader cannot verify from this paper alone that no pairings were missed or that the combinatorial counts are correct — they must either trust the author's inspection or redo the enumeration themselves. This is a significant gap for a paper whose purpose is to make the proofs "intelligible."
The "inspection" method for counting asymptotic pairings is not rigorous. Example 7.2 states that there are 4 annular noncrossing *-pairings in each relative orientation "by inspection." No systematic counting method is presented, no enumeration is listed, and no verification (e.g., by a small computer program) is provided. The reader who wants to confirm that the count is 4, not 3 or 5, has no recourse except to attempt the enumeration themselves. Similarly, Example 8.1's claim that "only one spoke diagram" works is stated as "by inspection" — but the constraints on spoke diagrams for cyclic alternating algebras on two faces with 3 terms each are not systematically analyzed; the paper simply asserts that the specific spoke pattern shown is the unique one.
The paper borrows heavily from unverified external results. The GOE moment E[tr(T^6)] = 5, the Wishart moments E[tr(W^7)] = 429 and E[tr(W^2)] = 2, E[tr(W^5)] = 42, and the Ginibre spoke contribution of 2 are all cited from "Remark 5.13 of the main paper" without derivation. This is understandable for a pedagogical appendix, but it means the paper does not stand alone as an experimental validation — its quantitative correctness depends on results from another document that the reader may not have access to or may not have verified.
No exploration of edge cases or failure modes. What happens when a pairing produces a surface with boundary (i.e., is not a complete pairing of all edges)? The paper only considers complete pairings (perfect matchings). What happens when the number of edges is odd (making a complete pairing impossible — the expectation would be zero)? These are not discussed. What happens when the vertex trace product involves matrices whose expected values are not known — how would one complete the computation? The paper gives no guidance on practical computation when Y_k are nontrivial random matrices rather than identity.
No connection to numerical simulation. All examples are combinatorial. The paper never simulates random matrices (e.g., generating 100,000 real Ginibre matrices of size N = 100 and computing sample covariances to verify that the asymptotic formula k_2 = 8 is approached). Such numerical validation would strongly reinforce the pedagogical message — it would show that the combinatorial framework correctly predicts empirical finite-N behaviour, converging to the asymptotic value as N grows. The absence of simulation is understandable given the paper's purpose (it is an appendix to a theoretical paper, not an empirical study), but it leaves a gap between the combinatorial algebra and the actual random matrix phenomena the algebra is meant to describe.
The difficulty estimation for the "experiments" is entirely implicit. Unlike a typical experimental paper, there is no discussion of which examples were chosen, how many pairings were tried, whether any pairings were found that violated expected patterns, or whether the "inspection" method ever produced incorrect counts that had to be corrected. The paper presents a polished final product — the successful examples — without showing the process of arriving at them. This is standard for mathematical exposition but limits the paper's value as an "experimental" validation.

Summary assessment:

The paper succeeds as a pedagogical Rosetta Stone — it translates specific instances of the main paper's permutation formulas into visual, traceable geometric operations. The examples are well-chosen to illustrate the key phenomena (twisted gluings, nonorientable surfaces, two-orientation asymptotic counting, spoke diagram factorization). However, as experimental validation, it is thin. It verifies that the framework produces correct and consistent numbers for a handful of cases, but does so without systematic enumeration, without independent cross-checks (simulation, exhaustive listing of all pairings), and with heavy reliance on external moment formulas. The quantitative results demonstrate internal consistency of the framework, not empirical correctness. The paper does not contain experiments in the standard sense — it contains worked examples that serve as checks on the combinatorial formalism's coherence.

6. Limitations and Trade-offs

The Permutation Formalism Requires Substantial Technical Investment to Use Independently

The assumption or constraint. This paper presents the premap formalism as a pedagogical translation of the main paper's proofs, but it does not provide a standalone "user's manual" for applying the formalism to new calculations. The reader who wants to compute the asymptotic two-trace cumulant of a novel trace product involving real Gaussian matrices must independently extract the operational rules from a collection of worked examples, without systematic guidance on how to handle edge cases, how to enumerate valid pairings efficiently, or how to verify that their enumeration is complete. The paper acknowledges its role as an appendix:

"This document is intended as an appendix to [6] ... in which we attempt to provide some intuition for how the topological constructions are represented in the permutations, and describe the pictures which motivated the various proofs."

The consequence. A practitioner attempting to apply this framework to a new matrix model or a new trace product faces several non-obvious hurdles. (1) The rules for which pairings π are compatible with a given matrix model, encoded through δ_ε-conjugation and sign-pattern constraints, are demonstrated through examples but never stated as an explicit algorithm. (2) The enumeration of noncrossing pairings for asymptotic counting relies on "by inspection" methods (e.g., Example 7.2 asserts there are 4 pairings without listing them), providing no systematic procedure. (3) The handling of multiple relative orientations — which orientations to check, how to verify that all relevant orientations have been considered, and how to avoid double-counting — is illustrated for two faces but not generalized. A researcher working with three or more faces, or with a matrix model not treated here (e.g., real Wishart without centring), would need to reconstruct the general rules from the main paper's more formal treatment, undermining this appendix's purpose as a standalone pedagogical tool. Mistakes in applying the formalism — a miscounted pairing, a missed relative orientation, an incorrect Euler characteristic — have no built-in cross-check within the paper's framework.

What evidence exists in the paper. Every quantitative result in the paper is produced by inspection or by citing external moment formulas. Example 7.2: "The pairing must connect the two faces, so there must be at least two pairings between the faces, which must pair an adjacent pair of edges to another adjacent pair of faces to avoid crossings. We find that there are 4 contributing pairings in the relative orientation shown in Figure 11 and 4 in the relative orientation shown in Figure 17." The phrase "We find that there are 4" is the entirety of the enumeration — no listing, no verification, no algorithm. Example 8.1 similarly asserts: "By inspection, we can see that this is the only spoke diagram that will accommodate a diagram on A_1 and B_2^T." The inspection is not shown, and the systematic constraints on spoke diagrams are not derived.

Mitigation status. The paper does not attempt to mitigate this limitation. It explicitly positions itself as providing intuition and pictures, not as a self-contained computational manual. A reader who needs to perform new calculations must either reverse-engineer the general principles from the examples (and verify them against the main paper's formal definitions) or consult the main paper directly, which provides the rigorous permutation definitions but lacks the geometric pictures. This gap between geometric intuition and operational procedure is fundamental to the paper's design — it teaches what the formulas mean, not how to compute with them efficiently.

The Two-Orientation Asymptotic Condition Is Not Validated Against Finite-N or Simulated Data

The assumption or constraint. The central structural claim distinguishing real from complex second-order freeness — that asymptotic fluctuations depend on noncrossing pairings counted under both relative orientations of the constituent faces — rests entirely on combinatorial reasoning within the permutation formalism. At no point does the paper compare its asymptotic formulas (e.g., lim k_2 = 8 or = 2070) to empirical estimates obtained from simulating actual random matrices at finite N and observing convergence as N grows.

The consequence. The two-orientation condition is derived from the structure of the premap encoding (specifically, from the requirement that pairings be *-pairings under both the original and flipped orientations of each face). However, the encoding itself is a mathematical model of the matrix calculation — a claim that the premap Euler characteristic expansion correctly reproduces the expected trace products for real Gaussian matrices. Without simulation, there is no independent verification that this encoding is correct. It is possible that the encoding misses contributions (e.g., from non-Gaussian corrections, from finite-N effects not captured by the asymptotic limit, or from the specific normalization conventions used) or that the two-orientation counting rule has exceptions not covered by the examples. The asymptotic results 1, 2, 8, and 2070 are internally consistent within the combinatorial framework, but a reader who doubts the framework's correspondence to actual random matrix behaviour has no empirical evidence to evaluate. This is particularly acute for the nonorientable surface contributions (twisted pairings), which have no analogue in the better-validated complex case and whose combinatorial rules are therefore less tested by prior literature.

What evidence exists in the paper. No simulations, no numerical experiments, no empirical convergence plots. The paper contains no data tables comparing theoretical asymptotic values to sample moments from generated random matrices. All quantities are computed combinatorially and verified only for internal consistency (e.g., the Euler characteristic of the Klein bottle example is 0, which is the known Euler characteristic of a Klein bottle — a topological consistency check, not an empirical validation of the matrix formula). The paper does not even cite prior simulation work that might corroborate the asymptotic formulas.

Mitigation status. The paper does not acknowledge this as a limitation — it is a pure mathematics appendix and operates entirely within the combinatorial proof paradigm. The validation of the premap encoding's correctness is presumed to reside in the main paper's formal proofs, not in numerical experiment. For a mathematically inclined reader who trusts the main paper's proof structure, this is not a gap; for a practitioner who wants empirical confirmation that the formulas describe real matrices, it is a significant absence. The paper would be strengthened — particularly as pedagogy — by including even a brief simulation showing that k_2(Tr(ZZZ^T Z^T), Tr(ZZ^T Z Z^T)) for N = 50, 100, 200 approaches 8, anchoring the abstract counting in observable phenomena.

The Asymptotic Framework Provides No Information on Convergence Rates or Finite-N Corrections

The assumption or constraint. All asymptotic results in the paper — the moments in Example 7.1, the covariance in Example 7.2, the mixed-ensemble cumulant in Example 8.1 — are stated as limits as N → ∞. The paper does not derive or discuss the subleading terms in 1/N, does not characterize how quickly the finite-N cumulants approach their asymptotic values, and does not discuss at what matrix size the asymptotic formulas become practically reliable approximations.

The consequence. For a practitioner using these formulas in applied settings — say, approximating the covariance of two trace statistics of a Wishart matrix with N = 100 — the paper provides no guidance on whether the N → ∞ limit is accurate to within 1%, 10%, or 50% at that size. The subleading contributions come from lower-Euler-characteristic surfaces: the projective plane (χ = 1, suppressed by N^{-1}), the Klein bottle and torus (χ = 0, suppressed by N^{-2} for cumulants), and so on. The paper identifies these surfaces (Example 3.4 produces a Klein bottle; Example 4.2 produces a projective plane) and notes their Euler characteristics, which determine their N-scaling, but does not compute their actual coefficient magnitudes. A Klein bottle term with a large combinatorial coefficient (many contributing pairings) could be non-negligible at moderate N even though it is formally O(N^{-2}) relative to the leading term. Without estimates of these coefficients, the practitioner cannot assess finite-N error. Additionally, the nonorientable character of the subleading terms (projective planes, Klein bottles) has no analogue in the complex case, so prior empirical experience with convergence rates from complex random matrix theory does not directly transfer.

What evidence exists in the paper. The expansion in Section 4 explicitly shows that each term is weighted by N^{χ - 2#(γ)}, and the Euler characteristics of example surfaces are computed (sphere χ = 2, projective plane χ = 1, Klein bottle χ = 0). Example 4.2 gives the full finite-N expression E[tr(T^2)] = 1 + N^{-1} for the GOE case, showing both the leading and first subleading term. This is, however, the simplest possible case (one trace, two edges, trivial Y_k) — it does not generalize to the multi-trace, multi-edge expressions that are the paper's actual subject. The paper does not state or compute finite-N corrections for any of the more complex examples (the Klein bottle in Section 2, the Ginibre two-trace expressions, the mixed-ensemble cumulant).

Mitigation status. The main paper presumably contains the full 1/N expansion, since its formal definition (the sum over all compatible pairings π) is an exact finite-N formula. The appendix paper's exclusive focus on asymptotic leading-order terms and a few illustrative subleading examples (the N^{-1} term in Example 4.2) is a choice of pedagogical scope, not an inherent limitation of the formalism. However, the paper does not signal to the reader that the asymptotic results are truncated expansions, does not indicate how many subleading orders exist, and does not provide any tools or recipes for computing them. A reader who needs finite-N accuracy must return to the main paper and sum the full pairing expansion — a task the appendix gives no practical guidance for.

Difficulty Estimation Is Treated as Implicit and Uncosted — The Reader Must Identify Which Pairings Contribute

The assumption or constraint. The combinatorial framework requires enumerating all compatible pairings π for a given trace product and matrix model, computing the Euler characteristic χ(γ, δ_ε π δ_ε) and vertex permutation for each, and summing the contributions. The paper treats this enumeration as feasible and implicitly assumes that the reader can identify which pairings are "compatible" (satisfying the matrix model's constraints, e.g., *-pairing for Ginibre) and which are the relevant ones for a given asymptotic order (noncrossing, connected, untwisted). No discussion is provided of the computational complexity of this enumeration, the number of pairings that must be checked, or how to prune the search space.

The consequence. For a trace product with n Gaussian matrix occurrences, the total number of complete pairings on 2n signed edge-ends is (2n - 1)!!, which grows super-exponentially. For n = 8 (the running example in Section 2), (2n - 1)!! = 15!! = 2,027,025 possible pairings — a number small enough for computer enumeration but large enough that manual inspection is impossible. For n = 20, the number exceeds 10^{15}. The paper's examples all involve small n (4 edges for the asymptotic moment calculations in Example 7.1; 8 edges for the covariance in Example 7.2; 6 terms per trace in Example 8.1), and the asymptotic counting is done "by inspection" rather than by systematic enumeration. The paper gives no guidance on how the pairing space is pruned — which constraints (noncrossing, *-pairing, connectedness, spoke diagram structure) reduce the enumeration from a factorial-scale problem to a tractable one, and in what order those constraints should be applied. A practitioner facing a larger expression would not know, from this appendix alone, whether the needed enumeration is a five-minute manual task, an hour-long computer search, or computationally intractable.

What evidence exists in the paper. Every example involves a small enough number of edges that the author can claim results "by inspection." Example 7.1 (asymptotic moments of a 4-edge face) enumerates the three possible pairings of 4 points (Figure 16) and rejects the crossing one — a trivial enumeration. Example 7.2 (asymptotic covariance, 8 edges total) asserts 4 pairings in each orientation by a geometric adjacency argument but does not enumerate them. The full set of compatible pairings for the Section 2 example (8 edges, 2,027,025 total pairings) is not enumerated; only one specific pairing is shown. The paper never discusses the size of the pairing space, how many pairings survive each constraint, or whether any example required non-trivial search.

Mitigation status. The paper does not acknowledge this as a limitation. It presents the enumeration as unproblematic because the examples are small. For the main paper's purposes (proving asymptotic freeness theorems), the precise enumeration of all pairings is handled by algebraic manipulation of premaps (conjugation, cycle counting) rather than by brute-force search, and the formal proofs establish which pairings contribute to which order without enumerating them individually. The appendix, however, teaches by example rather than by general algebraic manipulation, and does not bridge the gap between "inspect the small example" and "apply the general theorem." A reader who internalizes the geometric intuition but faces a larger calculation may find that the geometric method does not scale — they need the algebraic shortcuts from the main paper, which the appendix was meant to make accessible but does not fully operationalize.

The Framework Is Demonstrated Exclusively on Hand-Constructed Toy Examples with Trivial Deterministic Matrices

The assumption or constraint. Every worked example in the paper either sets the deterministic matrices Y_k to identity (Y_k = I_N for all k), or leaves them as symbolic placeholders whose expected traces are not actually evaluated. The most complex explicit computation — Example 8.1 — uses known single-trace moments (E[tr(T^6)] = 5, E[tr(W^7)] = 429, etc.) imported from the main paper's Remark 5.13, without showing how those moments themselves are computed within the premap framework. There is no example where the Y_k are nontrivial random matrices with their own internal structure, and no demonstration of how the framework handles nested Gaussian expectations (e.g., Y_k themselves being products of Gaussian matrices whose expectations must also be evaluated).

The consequence. The paper's pedagogical claim is that it shows "how a matrix calculation involving Gaussian random matrices can be interpreted in terms of face gluings." But the matrix calculations it actually shows are restricted to (a) extracting the Y_k vertex structure from a single layer of Gaussian expectation, and (b) evaluating the result when Y_k = I. A practitioner wanting to compute something like E[tr(X A X^T B) · tr(X C X^T D)] where A, B, C, D are themselves GOE or Wishart matrices must perform a nested or iterated application of the framework: first expand the X expectation to obtain trace products of A, B, C, D, then expand those expectations using the appropriate ensemble rules. The paper gives no example of this nesting, no discussion of how vertex products from one expansion become face structures for the next, and no indication of whether the framework modularizes cleanly under iteration or whether cross-terms between the two layers require special handling. The spoke diagram factorization in Example 8.1 sidesteps this by using pre-computed moments — it does not show how the GOE moment E[tr(T^6)] = 5 is derived from face gluings.

What evidence exists in the paper. All examples with nontrivial Y_k stop at the vertex permutation stage: the output is a product of expected traces of Y_k matrices (e.g., E[tr(Y_1 Y_3^T Y_6 Y_5^T Y_7^T) tr(Y_2) tr(Y_4 Y_6^T) tr(Y_8)] in Section 2), and the expectation over Y_k is left unevaluated. The paper states in Section 4: "An example calculation involving Wishart matrices and another involving several independent matrices are given in the main paper" — referring the reader elsewhere for the nested case. The examples that produce numerical answers all set Y_k = I_N (Examples 4.1, 7.1, 7.2) or import external moments (Example 8.1). No example shows a fully worked nested expectation.

Mitigation status. The paper is aware of this limitation and defers to the main paper for more complex examples. The statement that examples "are given in the main paper" is an acknowledgement, albeit a brief one, that this appendix does not cover the full scope of the formalism's application. For a reader whose interest is specifically in the topological interpretation of the permutation formulas (the paper's stated purpose), the restriction to single-layer Gaussian expectations with trivial Y_k is defensible — it keeps the topology clean and the diagrams legible. However, a reader who takes the paper as a tutorial on using the framework will encounter a significant gap between the worked examples and realistic calculations, and is given no bridge across that gap except the main paper itself.

7. Implications and Future Directions

How This Work Changes the Landscape

This appendix paper does not introduce new theorems or new mathematical machinery — it is explicitly a pedagogical companion to the main paper on real second-order freeness. As such, its contribution to changing the landscape is not through novel results, but through an intellectual reframing of how researchers access and understand the premap formalism that underpins the main paper's proofs. The magnitude is not paradigm-shifting in the sense of altering the theoretical structure of random matrix theory, but it is a substantial methodological contribution to the practice of reading, teaching, and applying real second-order freeness results.

The specific reframing is foundational: it converts the premap formalism from a formal algebraic apparatus — permutations γ_+ γ_-^{-1}, γ_-^{-1} π γ_+, sign-pattern constraints on π, Euler characteristic formulas — into a visual, geometric, traceable language of surface gluings. Prior to this work, a reader encountering the main paper's proofs faced a wall of permutation algebra whose geometric meaning was implicit at best. The exercises in Lando and Zvonkin (2004, Chapter 3) noted that real matrix calculations include nonorientable surfaces, but no document showed step-by-step how a specific Wick pairing translates into a specific nonorientable surface, how the Euler characteristic emerges from the number of vertices, edges, and faces in the double cover, or how the vertex permutation γ_-^{-1} π γ_+ corresponds to walking the boundary of a vertex in the glued surface. This appendix fills that gap by providing the geometric scaffolding underneath the algebra.

The most significant methodological shift is the elevation of twisted edge identifications to the status of a first-class diagnostic concept. Section 2's Figures 2 and 3 do not merely illustrate untwisted vs. twisted gluings — they establish a unified causal chain: real Gaussian moments (E[f²] = 1) → both X-with-X and X-with-X^T Wick pairings permitted → twisted gluings possible → nonorientable surfaces (projective planes, Klein bottles) enter the sum → the asymptotic counting problem now involves relative orientations. This chain is present in the main paper's formal definitions but is obscured by the algebraic encoding δ_ε π δ_ε. By making it visual, the appendix gives researchers a mental model for diagnosing whether a new matrix model or trace structure will produce nonorientable contributions and how those contributions scale with N.

The work also reconciles a latent contradiction in the random matrix theory literature: the standard map/hypermap encodings (Cori, Zvonkin, Lando-Zvonkin) work beautifully for complex ensembles, producing systematic genus expansions and noncrossing characterizations, but the extension to real ensembles had been piecemeal and ensemble-specific. Prior to the main paper, results for real Wishart, real GOE, and real Ginibre existed as separate calculations using different combinatorial approaches. The premap formalism provides a unified language — and this appendix makes that unified language teachable. The paper demonstrates (in Example 8.1) that Ginibre, GOE, and Wishart algebras can be treated within a single spoke-diagram factorization, with ensemble-specific differences isolated to per-spoke moment computations. This modularity was implicit in the main paper; the appendix makes it explicit and visually compelling.

A subtle but important consequence: this work makes nonorientable surface topology a practical working tool for random matrix theorists, not just a theoretical curiosity. The fact that the Klein bottle (χ = 0) appears in the running example (Section 2) and that the projective plane (χ = 1) appears in the GOE calculation (Example 4.2) means that researchers can now expect to encounter these surfaces in real matrix calculations and can estimate their contribution magnitudes from the Euler characteristic. This shifts nonorientable topology from an exotic footnote to a core part of the real random matrix analyst's toolkit.

In terms of research directions, this work marginally strengthens the case for further investigation of real second-order freeness — not because it proves new results, but because it lowers the barrier to entry for researchers who might otherwise be deterred by the main paper's formal abstraction. The tractability of the geometric approach (draw faces, pair edges, check twists, walk vertices, compute Euler characteristic) suggests that new real matrix models — beyond the canonical Ginibre/GOE/Wishart trio — could be analyzed within the premap framework without requiring novel combinatorial theory, provided the researcher can translate the model's moment structure into pairing constraints. Conversely, the paper's silence on simulation-based validation makes the empirical testing of the premap expansions a more attractive direction than it might otherwise appear — someone needs to check these asymptotic formulas against finite-N simulated data, and the appendix's worked examples provide concrete test cases (e.g., lim k₂ = 8 for the Ginibre covariance in Example 7.2) against which simulations can be compared.

Follow-Up Research This Work Enables

Systematic enumeration and computational verification of the premap pairing expansions for nontrivial trace products. Each quantitative result in the appendix (the 8 pairings in Example 7.2, the 2070 cumulant in Example 8.1) is obtained "by inspection" — manual counting of noncrossing pairings under compatibility constraints. This manual approach does not scale beyond the toy examples presented, and it provides no protection against human counting errors. A natural and technically straightforward follow-up is to implement a computer enumeration of the premap pairing space: for a given input trace product (expressed as faces γ and matrix model constraints), generate all complete pairings on 2n signed edge-ends (or all noncrossing subset for asymptotic enumeration), filter by the matrix model's compatibility conditions (e.g., *-pairing for Ginibre, all pairings for GOE), compute χ(γ, δ_ε π δ_ε) and the vertex permutation for each, and sum the contributions. This would serve three purposes: (1) validate the paper's manual counts by explicit enumeration of all pairings; (2) extend the enumeration to larger expressions (e.g., n = 12, 16, 20 edges) where manual inspection is impossible; (3) provide empirical finite-N results by evaluating the full sum (including subleading surfaces) at specific N and comparing to Monte Carlo simulations of actual random matrices. A strong follow-up would report: the exact number of contributing pairings at each Euler characteristic for a set of benchmark trace products, the convergence of the truncated sum to the asymptotic limit as N increases, and the empirical bias and variance of sample cumulants from simulated random matrices compared to the theoretical predictions.

Numerical simulation studies testing convergence rates of real vs. complex asymptotic formulas. The two-orientation condition that distinguishes real from complex asymptotic fluctuations (Section 7) is derived entirely within the combinatorial formalism and is never tested against simulated random matrix data. A direct empirical investigation would generate real and complex Ginibre matrices at sizes N = 20, 50, 100, 200, 500, compute sample covariances of the trace products from Example 7.2 (Tr(Z Z Z^T Z^T) and Tr(Z Z^T Z Z^T)) over many independent trials, and compare the empirical convergence to the theoretical asymptotic values — 8 for real Ginibre (from the two-orientation count) vs. the complex Ginibre prediction (which must be computed from the analogous one-orientation count). This experiment would test two claims simultaneously: that the real asymptotic formula correctly predicts finite-N behaviour, and that the real and complex limits are genuinely different (not an artifact of the combinatorial model). The paper's Example 4.2 provides a simpler test case: E[tr(T²)] = 1 + N⁻¹ for GOE predicts a specific finite-N curve; simulation can verify the N⁻¹ coefficient and check for higher-order corrections not captured by the projective plane term. A strong follow-up would plot empirical vs. theoretical values for 4-6 benchmark expressions, with error bars from simulation, and report at what N the asymptotic formula achieves 1% and 5% relative accuracy.

Extension of the premap formalism to real random matrix ensembles with non-Gaussian entries or with correlations across entries. The entire premap construction in this paper depends on the Wick pairing structure of Gaussian moments: the expectation of a product of Gaussian entries factorizes into a sum over complete pairings, each contributing a product of covariance factors. For non-Gaussian real random matrices (e.g., matrices with uniformly distributed entries, heavy-tailed distributions, or entries with known higher cumulants), the moment expansion involves more complex cumulant structures — not just pairings, but general partitions. The premap's translation from a partition of edge-ends to a surface gluing (with each block of the partition corresponding to a hyperedge identifying all edge-ends in that block) is natural — the paper already handles hyperedges in Example 3.2 — but the mapping from the distribution's cumulants to the weight fᶜ(π) and the Euler characteristic scaling of higher-order cumulants is not addressed here. A research program that characterizes the genus expansion for non-Gaussian real matrix models in terms of the premap framework, analogous to the 1/N expansion for matrix models with arbitrary potentials, would generalize the main paper's results beyond the Gaussian case. The starting point would be to translate the known diagrammatic expansion for non-Gaussian matrix models (e.g., the 1/N expansion for matrix models with polynomial potentials, which produces maps with vertices of degree > 3 from the higher cumulants) into the premap language and determine whether the two-orientation asymptotic condition persists beyond the Gaussian case.

Investigation of the spoke diagram constraints for mixed-algebra cumulants with more than two traces. Section 8 characterizes the asymptotic two-trace cumulants of cyclically alternating algebras in terms of spoke diagrams, with the factorization into per-spoke connected correlations demonstrated for r = 3 terms per trace in Example 8.1. The paper does not discuss cumulants of three or more traces, which would correspond to surfaces with three or more faces. The geometric constraint propagation that forces spoke diagrams for two faces — alternating condition + centring + noncrossing forces all connections to go between faces, which forces a noncrossing interlacement pattern — does not obviously generalize: with three faces, connections between face A and face B can be interleaved with connections between face A and face C in multiple topologically distinct ways. Characterizing the possible noncrossing gluing patterns for r > 2 faces of cyclically alternating terms, and determining whether the factorization into per-spoke connected correlations extends (and if so, in what form — presumably a product over "channels" connecting specific face pairs), would complete the combinatorial picture of real second-order freeness for multi-trace cumulants. This is a natural extension of the spoke diagram analysis, made tractable by the geometric intuition that the appendix cultivates: one can literally draw the faces and attempt to connect them with noncrossing chords, observe when the constraints force isolation or crossing, and classify the viable patterns.

Development of a user-facing computational toolkit for premap-based random matrix calculations. The gap between the paper's geometric pedagogy and its manual "by inspection" counting method is addressable through software. A symbolic computation package (in Python, SageMath, or Mathematica) that takes as input a trace product expression and a matrix model specification, constructs the face permutation γ and the matrix-model-specific constraints on π (via δ_ε-conjugation for transpose structure, compatibility rules for Ginibre/GOE/Wishart), enumerates all compatible pairings (or the noncrossing subset for asymptotic computation), computes the Euler characteristic and vertex permutation for each, and outputs the full 1/N expansion up to a specified order, would convert the premap formalism from a theoretical tool into a practical one. The appendix's worked examples provide an ideal test suite for such a toolkit: the Klein bottle example (Section 2) checks the full enumeration for a specific pairing; Examples 4.1-4.2 check compatibility filtering; Examples 7.1-7.2 check asymptotic noncrossing enumeration; Example 8.1 checks spoke diagram identification and multiplicative factorization. A strong implementation would reproduce all the paper's manual counts and extend them to expressions where manual counting is infeasible.

Empirical investigation of finite-N correction terms from nonorientable surfaces in real random matrix statistics. The paper establishes theoretically that twisted edge identifications produce nonorientable surfaces (projective planes, Klein bottles) that contribute to subleading orders in 1/N, and Example 4.2 demonstrates this concretely with the projective plane term N⁻¹ in E[tr(T²)]. However, the paper does not quantify the practical significance of these nonorientable corrections — are they typically negligible at moderate N, or do they have large combinatorial coefficients that make them practically important even when formally subleading? A simulation study comparing real and complex ensembles with matched leading-order asymptotics but different subleading structure (e.g., real Ginibre vs. complex Ginibre, where the complex ensemble has no twisted pairings and thus no nonorientable subleading surfaces) could measure the magnitude and sign of the nonorientable contribution at finite N. This would answer a practical question for applied random matrix users: when can you safely use the complex asymptotic formulas (which are simpler and better-documented) for real matrix problems, and when do you need the full real formalism with its nonorientable corrections? The paper's Example 4.2 shows that the nonorientable correction to E[tr(T²)] is exactly N⁻¹ with coefficient 1 — a small effect at N = 100 (1%) but potentially meaningful at N = 10 (10%). Extending this measurement to more complex trace products would map the regime where the real/complex distinction matters practically.

Practical Applications and Downstream Use Cases

Teaching and curriculum development for advanced random matrix theory courses. The appendix's primary practical impact is in the classroom and seminar room. Random matrix theory is taught in mathematics, statistics, physics, and engineering departments, but the existing pedagogical literature (e.g., Lando and Zvonkin, 2004; Mingo and Speicher, 2017) covers the complex/orientable case thoroughly while the real/nonorientable extensions remain confined to research papers with high formalism barriers. This appendix provides a self-contained geometric introduction to the topological interpretation of real Gaussian matrix calculations. The worked examples — especially the progression from Figure 1 (faces) through Figure 4 (surface gluing) to Figure 5 (vertex trace product) in Section 2, and the parallel progression from oriented maps (Examples 3.1-3.2) to the nonorientable double cover (Example 3.3) in Section 3 — can serve directly as lecture material or as assigned reading in a graduate course. The paper's explicit enumeration of pairings, its labelling of edge-ends with signed integers, and its step-by-step permutation computations provide templates that students can follow for homework exercises (e.g., "compute E[tr(X Y₁ X^T Y₂ X Y₃ X^T Y₄)] by drawing faces, enumerating pairings, and computing vertex traces"). The concrete numerical results (1 + N⁻¹ for GOE, 8 for the Ginibre covariance, 2070 for the mixed-ensemble cumulant) give students checkable targets. This teaching impact is amplified by the paper's self-contained nature — it does not require the main paper to be read first, making it accessible as a standalone introduction to the combinatorial topology of random matrices.

Diagnostic tool for applied random matrix users encountering real data. In multivariate statistics, signal processing, and machine learning, practitioners routinely compute trace statistics of sample covariance matrices (real Wishart), Gram matrices, and real symmetric matrices drawn from real data. The theoretical justification for using asymptotic formulas (e.g., for hypothesis tests based on linear spectral statistics) often implicitly assumes the complex case, where the 1/N expansion is well-characterized by orientable map enumeration. This appendix provides a framework for diagnosing when the real case differs: if the statistic involves traces where Gaussian entries could be paired X-with-X (not just X-with-X^T), then nonorientable contributions enter at subleading orders, and the complex-asymptotic approximation may have errors at O(1/N) that are absent in the complex case. Example 4.2 gives the simplest diagnostic: E[tr(T²)] for a real symmetric matrix has an extra N⁻¹ term relative to its complex Hermitian analogue. A practitioner computing the variance of a linear spectral statistic of a real Wishart matrix can, in principle, use the premap framework to identify which terms in the variance expansion receive nonorientable corrections and estimate their magnitude using the Euler characteristic scaling. While the current paper does not provide a ready-to-use formula for arbitrary statistics, it provides the conceptual lens — twisted edge identifications = nonorientable surfaces = subleading corrections — that allows a practitioner to recognize when complex-case formulas are likely to be insufficient and to seek out the real-case results from the main paper.

Verification target for simulation-based random matrix inference. In applications where theoretical asymptotic formulas are unavailable or untrusted (e.g., for novel test statistics, non-Gaussian matrix ensembles, or finite-N critical regimes), simulation is the standard fallback. This paper provides a set of exact combinatorial predictions that can serve as ground truth for validating simulation code. For instance, generating real Ginibre matrices at N = 100 and computing sample covariances of tr(Z Z Z^T Z^T) and tr(Z Z^T Z Z^T) over 10,000 trials should produce values converging to the theoretical 8 as the number of trials increases. If the simulation converges to a different number, either the simulation is buggy (wrong ensemble, wrong normalisation, incorrect centring) or the theoretical prediction is wrong — either outcome is valuable. The numerical predictions in the paper — 1, 2, 8, 1+N⁻¹, 2070 — are specific, parameter-free, and testable with standard random matrix generation code. This makes the paper useful as a validation suite for simulation software, analogous to unit tests in software engineering. A researcher implementing a general random matrix simulation framework can use these examples to verify that their Ginibre, GOE, and Wishart generators are correct and that their cumulant estimation code properly handles centring and normalization.

When to Prefer This Method

This paper is a pedagogical exposition, not a methodological contribution — it does not propose a new method to compare against alternatives. The premap formalism and the topological surface-gluing approach it illustrates are the only systematic frameworks available for real second-order freeness calculations at the time of the main paper, and the appendix does not benchmark different approaches. The decision it illuminates is not "which method to use," but rather when the additional complexity of the real case (twisted pairings, nonorientable surfaces, two-orientation condition) is necessary versus when the simpler complex-case formulas suffice. The paper provides clear diagnostic criteria:

Prefer the full real premap / double-cover framework when: the random matrix ensemble is real (entries are real Gaussian, not complex Gaussian), and the trace product being evaluated involves matrices and their transposes in a pattern that admits X-with-X pairings (i.e., the number of untransposed and transposed Gaussian factors permits same-type pairing). Example 4.1 demonstrates this for real Ginibre; Example 4.2 demonstrates it for GOE. In such cases, nonorientable surfaces enter the 1/N expansion, the asymptotic counting must consider both relative orientations of faces (Section 7), and using complex-Ginibre or complex-Wishart formulas will produce incorrect subleading terms and potentially incorrect asymptotic limits.
The complex-case formalism (oriented maps, *-pairings only) suffices when: the ensemble is genuinely complex (entries satisfy E[Z²] = 0, forcing all pairings to be untwisted), or the real ensemble's trace product structure is such that X-with-X pairings are combinatorially impossible (e.g., all Gaussian factors appear as X with no X^T, or in strictly alternating X, X^T, X, X^T patterns that exhaust all X factors with X^T partners). In the latter case, the real and complex calculations coincide because the pairing constraints from the trace product itself exclude twisted identifications, regardless of the Gaussian moment structure.
The GOE calculation in Example 4.2 provides the simplest decision rule: if E[f²] = 1 (real Gaussian) and the trace product has fewer factors than needed to force all pairings to be untwisted, the real formalism is required to capture all contributions; the size of the real/complex discrepancy is at least O(1/N). If only the leading asymptotic order (N⁰) is needed and the required limit is known to be identical in the real and complex cases, the complex formalism may be used as a shortcut, but the paper provides no general theorem guaranteeing when the leading orders coincide — this must be checked case by case.