Edge-Triangle Random Graph Model

• Edge-Triangle Model is a dense random graph ensemble defined by fixed edge and triangle densities, using a graphon variational principle to study entropy maximization.
• The model uncovers a rich phase diagram with multipodal structures, continuous and discontinuous transitions, and symmetry-breaking phenomena in network ensembles.
• Comparative analysis shows that edge-independent models underestimate triangle counts and clustering, underscoring the model’s importance in capturing higher-order network structure.

Searching arXiv for the cited papers and closely related work on the edge–triangle model. The edge-triangle model is a dense random-graph ensemble in which the primary observables are edge density and triangle density. In its microcanonical form, one fixes target values for these two subgraph densities and studies the asymptotic entropy of labeled graphs realizing them; in the dense limit, the problem becomes a constrained variational problem over graphons, that is, symmetric measurable functions $g:[0,1]^2\to[0,1]$. The model is a canonical setting for analyzing phase structure, multipodality, and symmetry breaking in constrained network ensembles, and it also serves as a useful foil for edge-independent generative models, which cannot sustain high triangle density under bounded overlap without effectively memorizing a graph (Kenyon et al., 2017).

1. Variational definition in the dense limit

For a simple graph $G$ on $n$ vertices, the edge density and triangle density are

$$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$

The microcanonical ensemble consists of all labeled graphs with prescribed $e(G)$ and $t(G)$, up to the $O(1/n)$ rounding inherent in finite $n$. The associated entropy density is defined by a large-$n$ limit of the normalized logarithm of the number of such graphs. One formulation is

$$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$

In the dense regime, Chatterjee–Varadhan and Radin–Sadun imply a graphon variational principle. Writing

$G$0

and

$G$1

one has

$G$2

An equivalent formulation in the $G$3 notation uses

$G$4

with the graphon entropy written as

$G$5

Whenever the supremum is attained at a unique graphon $G$6, up to measure-preserving relabelings, almost all large graphs with densities $G$7 look like $G$8 (Neeman et al., 2022).

2. Feasible region and multipodal phase organization

The admissible region of edge and triangle densities is

$G$9

The upper boundary $n$0 is the Erdős–Rényi curve. The lower boundary $n$1 is Razborov’s piecewise-quadratic “scalloped triangle,” with cusps at

$n$2

The phase diagram described in the large-deviation literature contains four infinite families of interior phases together with one additional $n$3-phase (Kenyon et al., 2017).

Phase family	Symmetry	Defining feature
$n$4	$n$5 equivalent podes	Two edge-probabilities; touches the lower boundary at the cusp $n$6
$n$7	One special pode plus $n$8 equivalent podes	Four block-probabilities plus a size parameter $n$9
$$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$0	Two equivalent small podes plus $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$1 equivalent large podes	Five block-probabilities plus $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$2
$$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$3	Bipodal but not in the $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$4-family	Occurs in the region $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$5

Here a “pode” is a block in a finite partition of $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$6 on which the entropy-maximizing graphon is constant. The reported synthesis of sampling, local stability analysis, and variational arguments yields a conjectured phase diagram with fourteen distinct phases together with a classification of which neighboring-phase boundaries are continuous and which are discontinuous (Kenyon et al., 2017).

3. Bipodal graphons and explicit formulas near the Erdős–Rényi curve

With two constraints, the dominant structures are often bipodal. A bipodal graphon is specified by block size $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$7 and three link-probabilities $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$8: $$e(G)=\frac{\#(\text{edges in }G)}{\binom n2},\qquad t(G)=\frac{\#(\text{triangles in }G)}{\binom n3}.$$9 For this ansatz,

$e(G)$0

$e(G)$1

and

$e(G)$2

Hence, within the bipodal approximation,

$e(G)$3

A particularly important regime lies just below the Erdős–Rényi curve $e(G)$4. In the symmetric bipodal phase, denoted phase II in the symmetry-breaking analysis, the maximizer has

$e(G)$5

and the stationarity relation

$e(G)$6

gives

$e(G)$7

Thus the graphon has lower within-block density $e(G)$8 and higher across-block density $e(G)$9. Below a critical curve $t(G)$0, the maximizer remains bipodal but becomes asymmetric: $t(G)$1 and $t(G)$2, with two mirror-image solutions related by exchanging the two blocks. The cited analysis refers to this phase III as “crystalline,” in contrast to the “fluid-like” symmetric phase II (Radin et al., 2016).

4. Symmetry breaking, local stability, and transition morphology

The symmetry-breaking transition is conveniently described by fixing $t(G)$3, maximizing over $t(G)$4 at fixed $t(G)$5, and studying the one-dimensional function

$t(G)$6

By symmetry, $t(G)$7 is even about $t(G)$8, so one expands

$t(G)$9

In the symmetric phase, $O(1/n)$0, so $O(1/n)$1 is locally maximizing. Along the critical curve $O(1/n)$2, one has $O(1/n)$3, and the numerical analysis reported for the bipodal problem finds $O(1/n)$4. This yields a pitchfork bifurcation: $O(1/n)$5 Near the point $O(1/n)$6, the critical curve has the cubic expansion

$O(1/n)$7

so $O(1/n)$8 departs from the Erdős–Rényi curve with a cubic cusp (Radin et al., 2016).

Beyond the bipodal transition, the global phase diagram exhibits both continuous and discontinuous neighboring-phase boundaries. The continuous cases identified in the large-network phase analysis are $O(1/n)$9 for $n$0, together with $n$1 and $n$2 along $n$3. All other neighboring-phase boundaries shown in the cited phase diagram are discontinuous. The proof strategy combines Euler–Lagrange equations, divergence-of-multipliers arguments ruling out spurious continuity, Hessian tests for $n$4 stability, and a Landau–Anderson power-series analysis of the pitchfork at $n$5 (Kenyon et al., 2017).

5. The symmetric bipodal phase and its breakdown

A later existence result establishes a genuine symmetric bipodal phase in the edge-triangle model. There is an open region $n$6 in the $n$7-plane containing the line segment $n$8, $n$9, such that for every $n$0 the unique entropy-maximizing graphon is symmetric bipodal. In particular, for $n$1 near $n$2 and $n$3, the optimizer is the two-block graphon with

$n$4

The same work also proves that symmetric bipodality fails below a critical edge density

$n$5

Specifically, for any $n$6 and $n$7 with $n$8 sufficiently small, the symmetric bipodal graphon is not the entropy maximizer (Neeman et al., 2022).

The threshold $n$9 is obtained from the stability condition

$$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$0

which reduces to

$$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$1

The proof of the positive result near $$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$2 uses a spectral decomposition

$$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$3

showing that the mass is concentrated in the first eigenmode, together with the pointwise Euler–Lagrange condition

$$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$4

A cost-benefit expansion in the parameters

$$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$5

then proves uniqueness of the symmetric bipodal maximizer for small perturbations of $$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$6. For $$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$7, the breakdown is shown by constructing an explicit 3-block graphon with the same edge and triangle densities but strictly larger entropy than any symmetric bipodal competitor. An intrinsic order parameter separating the symmetric bipodal region $$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$8 from the oversaturated region $$B(e,t)= \lim_{\delta\to0}\;\lim_{n\to\infty}\; \frac1{n^2}\,\ln\#\{G\in G_n:\;e(G)\in[e\pm\delta],\;t(G)\in[t\pm\delta]\}.$$9 is

$G$00

on $G$01, $G$02, whereas on $G$03, $G$04 (Neeman et al., 2022).

6. Relation to edge-independent models and triangle-density limitations

The edge-triangle model imposes triangle density as a constraint and thereby encodes explicit higher-order structure. By contrast, edge-independent graph models sample each edge independently from a probability matrix

$G$05

with

$G$06

This framework includes Erdős–Rényi, stochastic block models, Chung–Lu, Kronecker graphs, variational graph autoencoders, and NetGAN/CELL-style models. To exclude trivial memorization, one introduces the expected volume

$G$07

and the expected overlap

$G$08

For such models,

$G$09

but under bounded overlap this is sharply limited: $G$10 The Erdős–Rényi case shows the $G$11 scaling is tight. More generally, the same spectral argument bounds simple $G$12-cycle counts, and the global clustering coefficient satisfies

$G$13

when $G$14 and the denominator concentrates appropriately (Chanpuriya et al., 2021).

This contrast resolves a common misunderstanding. The edge-triangle model is not merely an Erdős–Rényi graph with a tuned parameter, nor is it equivalent to an independent-edge generator with a fitted probability matrix. In empirical comparisons on seven real-world graphs, all edge-independent approaches were reported to systematically underestimate triangle counts and clustering coefficients when overlap was low; as overlap increased, triangle counts rose as predicted by the theory, but diversity fell. Degree-matching baselines such as the Odds-Product variants CCOP and HDOP exactly reproduce degree sequences at any overlap and often produce more triangles than CELL at the same overlap level, while TSVD was described as remarkably robust at very low overlap. Even so, characteristic path lengths remained underestimated in all edge-independent models. The stated conclusion is that generating graphs with realistic high clustering at low overlap requires edge-dependent or auto-regressive models such as GraphRNN, rather than edge independence alone (Chanpuriya et al., 2021).