---
title: 'Ferromagnetic ERGMs: Theory and Phase Behavior'
url: https://www.emergentmind.com/topics/ferromagnetic-exponential-random-graph-models-ergms
type: topic
---

# Ferromagnetic ERGMs: Theory and Phase Behavior

Searching arXiv for the cited ferromagnetic ERGM papers to ground the article in current literature.
Ferromagnetic exponential random graph models (ERGMs) are Gibbs measures on labeled graphs in which the presence of specified subgraphs is rewarded by nonnegative interaction parameters. In the standard homogeneous formulation, one fixes finite simple graphs \(G_0,G_1,\dots,G_K\), with \(G_0\) the single edge, defines the homomorphism density \(t(G_j,x)=N_{G_j}(x)/n^{v_j}\), chooses \(\beta_0\in\mathbb R\) and \(\beta_1,\dots,\beta_K\ge 0\), and sets
\[
H(x)=\sum_{j=0}^K \beta_j\,t(G_j,x),\qquad \mathbb P[X=x]\propto \exp(n^2 H(x)).
\]
Equivalently, this is an exponential tilt of \(G(n,q_0)\) with \(q_0=e^{2\beta_0}/(1+e^{2\beta_0})\) [2507.10531]. A sparse fractional-power variant fixes simple graphs \(F_1,\dots,F_m\) and exponents \(\alpha_k\in(0,1)\), defines
\[
H_n(G)=\sum_{k=1}^m \beta_k\,[t(F_k,G)]^{\alpha_k},
\]
and assigns Gibbs weight
\[
P(G)= Z_n^{-1}\exp[n^2 H_n(G)-\alpha e(G)],\qquad \alpha=\log((1-p)/p),
\]
with \(p=p(n)\to 0\) [2208.06397]. Across these formulations, ferromagnetic ERGMs are analyzed through variational free energies, phase structure, large deviations, metastable dynamics, and graphon methods; recent work also gives sharp fluctuation and transport results, as well as block-structured extensions [2601.14170] [2602.16604].

## 1. Definitions and model classes

In the homogeneous ferromagnetic model, the state space is the set of all simple graphs on \([n]\), and the nonlinear terms \(\beta_j t(G_j,x)\) reward the appearance of copies of \(G_j\) when \(\beta_j\ge 0\). The single-edge term \(G_0\) plays the role of an external field, while higher-order terms such as triangles and two-stars enter as positive interactions. The measure may also be written
\[
P_\beta(X=x)=\exp\bigl(n^2H(x)-\psi_n(\beta)\bigr)
\]
or, equivalently,
\[
P_\beta(X=x)=\exp\Bigl(\sum_{j=0}^K \beta_j\,N_{G_j}(x)-\psi_n(\beta)\Bigr),
\]
where \(N_{G_j}(x)=n^{|V(G_j)|}t(G_j,x)\) is the raw homomorphism count [2601.14170].

The sparse fractional-power model is formulated on
\[
\mathcal G_n=\{G:[n]^2\to\{0,1\},\ \mathrm{sym.},\ G_{ii}=0\},
\]
with homomorphism density
\[
t(F_k,G)=n^{-|V(F_k)|}\sum_{\phi:V(F_k)\to[n]}\prod_{(u,v)\in E(F_k)}G_{\phi(u),\phi(v)}.
\]
Here \(p=p(n)\to 0\) gives “vanishing edge density,” so that typical \(e(G)=O(np)\). The key modification is the fractional power \(\alpha_k\in(0,1)\): by taking \(\alpha_k<1\), “the penalization for large subgraph counts grows sublinearly, preventing the abrupt jump to the complete graph seen in the standard \((\alpha_k=1)\) edge–triangle model” [2208.06397].

A further inhomogeneous class is the block-structured edge–triangle ERGM. With a partition \([n]=B_1\cup\cdots\cup B_k\), block sizes \(|B_i|=w_i^{(n)}\sim b_i n\), triangle parameters \(\alpha_{ij\ell}\), and edge parameters \(h_{ij}\), the Hamiltonian is
\[
\mathcal H^{(\mathbf B)}_{n;\boldsymbol\alpha,\mathbf h}(X)
=
\frac1n \sum_{i,j,\ell=1}^k \alpha_{ij\ell} \sum_{\substack{u\in B_i,\;v\in B_j,\;w\in B_\ell\\u<v<w}} X_{uv}X_{vw}X_{uw}
+
\sum_{i,j=1}^k h_{ij} \sum_{\substack{u\in B_i,\;v\in B_j\\u<v}} X_{uv},
\]
with \(\alpha_{ij\ell}=\alpha_{\pi(i)\pi(j)\pi(\ell)}\) for \(\pi\in S_3\) [2602.16604].

## 2. Variational structure and phase regimes

For homogeneous ferromagnetic ERGMs, the basic mean-field functional is
\[
L_\beta(q)=H(q)-I(q),\qquad I(q)=\tfrac12[q\log q+(1-q)\log(1-q)].
\]
Its global maximizers \(M_\beta\subset[0,1]\) govern macroscopic laws of large numbers: any ERGM sample \(X\) satisfies \(t(G,X)\to p^{|E(G)|}\) for each \(G\), for some \(p\in M_\beta\). Glauber-dynamics analysis introduces
\[
\phi_\beta(q)=\frac{e^{2H'(q)}}{1+e^{2H'(q)}},
\]
and any attracting fixed point \(p\) of \(\phi_\beta\) is a strictly concave local maximum of \(L_\beta\). The regimes stated in the recent fluctuation theory are: Dobrushin (very-high-temperature), where \(H''(1)<2\); subcritical (high-temperature), where \(\phi_\beta\) has a unique attracting fixed point; and supercritical (low-temperature), where \(\phi_\beta\) has multiple attracting fixed points, with “large-scale metastability and slow mixing from worst-case starts but rapid (metastable) mixing within each ‘well’.” Phase coexistence occurs where \(|M_\beta|>1\), but one may condition on the well around a given \(p\in M_\beta\) to recover uniqueness [2507.10531].

In the sparse fractional-power model, the partition function is
\[
Z_n=\sum_{G\in\mathcal G_n}\exp[n^2H_n(G)-\alpha e(G)].
\]
The Gibbs variational principle gives
\[
\log Z_n
=
\sup_{\mu\ \mathrm{on}\ \mathcal G_n}
\Bigl\{
n^2\mathbb E_{G\sim\mu}[H_n(G)]
-\alpha\,\mathbb E_\mu[e(G)]
-D_{KL}(\mu\Vert \mathrm{Bern}(p))
\Bigr\}.
\]
Restricting \(\mu\) to product laws \(\mu_Q\) and replacing \(\mathbb E[H_n(G)]\approx H_n(\mathbb E[G])=H_n(Q)\) yields the naïve mean-field free energy
\[
\Psi_n(p;\beta,\alpha):=
\sup_{Q\in[0,1]^{n\times n}}
\left\{
r_{n,p}H_n(Q/p)-\sum_{i<j}I_p(Q_{ij})
\right\},
\]
where \(r_{n,p}=n^2p^\Delta\log(1/p)\) and
\[
I_p(q)=q\log(q/p)+(1-q)\log((1-q)/(1-p)).
\]
If one further restricts to homogeneous \(Q_{ij}=u\), then
\[
\log Z_n \simeq r_{n,p}\sup_{0\le u\le1}\{L(u)-I(u)\},
\]
with
\[
L(u)=\sum_{k=1}^m \beta_k [u^{|E(F_k)|}]^{\alpha_k},\qquad
I(u)=u\log u+(1-u)\log(1-u).
\]
The derivation outline is: apply Varadhan’s lemma to the LDP for homomorphism densities in \(G(n,p)\), exchange supremum and limit, and observe that the extremum is reached on constant-\(Q\) [2208.06397].

## 3. Sparse ferromagnetic structure, upper tails, and nondegeneracy

The sparse theory identifies a two-parameter structural description of upper-tail events and of the corresponding ERGM. For thresholds \(s_1,\dots,s_m>0\), define
\[
U_p(\vec F,\vec s)=\bigcap_k \{G:t(F_k,G/p)\ge 1+s_k\},
\]
and let \(e_n=r_{n,p}\). Under \(n^{-1/(\Delta+1)}\ll p\ll 1\) and \(\beta_k\ge 0\), if \((a,b)\) solves the planar variational problem
\[
\phi(\vec F,\vec s)=\inf_{a,b\ge 0}\left\{\tfrac12 a+b: T_{F_k}(a,b)\ge 1+s_k\ \forall k\right\},
\]
with
\[
T_F(a,b)=P_{F^\ast}(b)+a^{|V(F)|/2}\mathbf 1_{F\ \mathrm{regular}},
\]
then with high probability
\[
G_{n,p}^H \in \bigcup_{(a,b)\in \mathrm{Opt}(\phi)} \mathcal G_n^1(a,b,\xi),
\]
and, in a refined regime,
\[
G_{n,p}^H \in \bigcup_{(a,b)\in \mathrm{Opt}(\phi)} \mathcal G_n^2(a,b,\xi).
\]
Here \(\mathcal G_n^1\) imposes an almost-clique of size \(\simeq (ap^\Delta)^{1/2}n\) plus almost-biclique (hub) of size \(\simeq bp^\Delta n\), and \(\mathcal G_n^2\) imposes a spectral-norm neighborhood of the corresponding weighted “clique-hub” graph \(Q^{I,J}\) [2208.06397].

The same two-parameter family appears in the upper-tail LDP. The stated rate is
\[
\mathbb P\bigl(t(F,G_{n,p})\ge 1+s\bigr)\approx \exp[-e_n\,\phi_F(s)],
\qquad e_n=n^2p^\Delta\log(1/p),
\]
with
\[
\phi_F(s)=\inf\{\tfrac12 a+b:T_F(a,b)\ge 1+s\},
\]
and the joint problem for \((F_1,\dots,F_m)\) reduces to
\[
\phi(\vec F,\vec s)=\inf_{a,b\ge0}\left\{\tfrac12 a+b:T_{F_k}(a,b)\ge 1+s_k\ \forall k\right\}.
\]
Hence “typical ERGMs have two phases: clique-driven (\(b=0\)) or hub-driven (\(a=0\))” [2208.06397].

The analytic input singled out in this work is a stability form of Finner’s generalized Hölder inequality. Classical Finner gives
\[
\int \prod_{A\in\mathcal A}(f_A\circ \pi_A)^{\lambda_A}
\le
\prod_{A\in\mathcal A}\left(\int f_A\right)^{\lambda_A}
\]
under \(\sum_{A\ni v}\lambda_A\le 1\). The stability statement says that if additionally \(\int \prod f_A^{\lambda_A}\ge 1-\epsilon\), then there exist nonnegative \(h_v\) with \(\int h_v=1\) such that
\[
\|f_A-\bigotimes_{v\in A}h_v\|_{L^1}\lesssim \epsilon^c
\]
for every \(A\in\mathcal A\), \(c>0\). Combined with quantitative LDPs, this identifies near-optimizers of the free-energy variational problem as clique-hub graphons [2208.06397].

## 4. Metastability and quantitative fluctuation theory

Ferromagnetic ERGMs are “mixtures of metastable wells which each behave macroscopically like new Erdős–Rényi models themselves, exhibiting the same laws of large numbers for the overall edge count as well as all subgraph counts.” The fluctuation theory developed in 2025 works in the noncritical “phase-uniqueness” case \((|M_\beta|=1)\) or, in phase coexistence, after conditioning on a single well. If \(p\) is the unique density, then for the edge count
\[
\sigma_n^2
=
\left[
1-2p(1-p)\sum_{j=1}^K \beta_j e_j(e_j-1)p^{e_j-2}
\right]^{-1}
\cdot p(1-p)\cdot \binom n2,
\]
and for every \(\epsilon>0\),
\[
\mathrm{Was}\!\left(\frac{E(X)-\mathbb E[E(X)]}{\sigma_n},N(0,1)\right)\lesssim n^{-1/2+\epsilon},
\qquad
\mathrm{Kol}\!\left(\frac{E(X)-\mathbb E[E(X)]}{\sigma_n},N(0,1)\right)\lesssim n^{-1/2+\epsilon}.
\]
For a fixed graph \(G\) with \(v\) vertices and \(e\) edges,
\[
N_G(X)-\mathbb E[N_G(X)]
\approx
2ep^{e-1}n^{v-2}\cdot (E(X)-\mathbb E[E(X)]),
\]
and the same \(n^{-1/2+\epsilon}\) rate holds after the corresponding normalization. For any fixed vertex \(v\), the degree satisfies a CLT with variance
\[
\varsigma_n^2
=
\left[
1-p(1-p)\sum_{j=1}^K \beta_j p^{e_j-2}\sum_{\rho\in V_j} d_\rho(d_\rho-1)
\right]^{-1}
\cdot p(1-p)\cdot (n-1),
\]
and both Wasserstein and Kolmogorov distances are \(\lesssim n^{-1/4+\epsilon}\). Local subgraph counts satisfy analogous CLTs with the same \(n^{-1/4+\epsilon}\) rate [2507.10531].

The main probabilistic mechanism is a combination of Stein’s method for nonlinear exponential families with a new “Hájék-projection” analysis via Glauber-dynamics-based concentration. The global proposition states that for
\[
\hat H_G(X)=N_G(X)-2ep^{e-1}n^{v-2}E(X),
\]
one has, for any \(\zeta\in(0,1)\),
\[
\mathbb P\bigl[|\hat H_G(X)-\mathbb E[\hat H_G(X)]|>n^{v-1.5+\zeta}\bigr]\le \exp(-cn^\zeta),
\]
and in particular \(\mathrm{Var}[\hat H_G(X)]\lesssim n^{2v-3+o(1)}\). The local version gives, for \(\hat H_G^{\rho\to v}(X)=N_G^{\rho\to v}(X)-d_\rho p^{e-1}n^{v-2}\deg_v(X)\),
\[
\mathbb P\bigl[|\hat H_G^{\rho\to v}(X)-\mathbb E[\hat H_G^{\rho\to v}(X)]|>n^{v-1.75+\zeta}\bigr]\le \exp(-cn^{(4/3)\zeta}),
\]
with \(\mathrm{Var}[\hat H_G^{\rho\to v}(X)]\lesssim n^{2v-3.5+o(1)}\) [2507.10531].

These results extend quantitative CLTs to the “full supercritical (low-temperature) regime (including metastable cases),” improve the error to \(O(n^{-1/2+\epsilon})\) for edges and global subgraphs, and provide the first quantitative CLTs for vertex degrees and local subgraph counts [2507.10531].

## 5. Erdős–Rényi comparison and Hamming–Wasserstein geometry

A central structural point is that phase-conditioned ferromagnetic ERGMs are close to \(G(n,p)\) at one scale and far from it at another. In the ferromagnetic regime, the free-energy functional
\[
L_\beta(q)=H(q)-\tfrac12(q\log q+(1-q)\log(1-q))
\]
has a finite set \(M_\beta\) of global maximizers, and the ERGM law is, up to total-variation error \(e^{-\Omega(n^2)}\), a finite mixture of phase measures \(M(n,p)\), \(p\in M_\beta\). For each such \(p\), if \(X\sim M(n,p)\) and \(Y\sim G(n,p)\), then their cut-distance goes to zero, but statistical tests can distinguish the laws with vanishing error probability, forcing
\[
\|M(n,p)-G(n,p)\|_{\mathrm{TV}}\to 1.
\]
The Hamming–Wasserstein distance resolves this discrepancy:
\[
\mathrm{Ham}(x,y)=\sum_{e\in \binom{[n]}2}|x(e)-y(e)|,
\qquad
\mathrm{Was}(\mu,\nu)=\inf_{X\sim\mu,\;Y\sim\nu}\mathbb E[\mathrm{Ham}(X,Y)].
\]
Under \(L_\beta''(p)<0\) and the nondegeneracy assumption that “at least one \(G_j\) is not a disjoint union of edges,” there are constants \(0<c<C\) such that
\[
c\,n^{3/2}\le \mathrm{Was}(M(n,p),G(n,p))\le C\,n^{3/2},
\]
so the distance is \(\Theta(n^{3/2})\) [2601.14170].

The upper bound is obtained by monotone Glauber-dynamics coupling and contraction in a “good set.” The lower bound uses the optimal-coupling characterization together with the conditional edge-update
\[
p_e(X)=\phi(n^2\partial_e H(X)).
\]
The key approximation expresses \(n^2\partial_eH(x)\) through local triangle and wedge counts near \(e\):
\[
n^2\partial_eH(x)
=
\frac{C^\triangle}{n}N_\triangle(x,e)
+
\frac{C^\wedge}{n}N_\wedge(x,e)
+
\text{(almost constant)}
+
o_{\mathbb P}(n^{-3/4}),
\]
where \(C^\triangle\) and \(C^\wedge\) are nonnegative combinations of the model parameters, and at least one is positive whenever the model is nondegenerate. This wedge–triangle linearization yields \(\mathbb E|p_e(X)-p|=\Theta(n^{-1/2})\), and summing over all edges gives the \(n^{3/2}\) lower bound [2601.14170].

The same analysis also gives a bound on the marginal edge probability,
\[
|\mathbb E[X(e)]-p|\lesssim n^{-1},
\]
via a bootstrapping argument. A standard misunderstanding is therefore avoided: cut-distance and laws of large numbers describe the macroscopic resemblance to \(G(n,p)\), whereas Hamming–Wasserstein and total variation detect genuine microscopic separation [2601.14170].

## 6. Block-structured ferromagnetic ERGMs

The block-model extension places ERGMs on a vertex partition with type-dependent interaction parameters. The Gibbs measure
\[
\mathbb P_{n;\boldsymbol\alpha,\mathbf h}^{(\mathbf B)}(X)\propto \exp\bigl(\mathcal H^{(\mathbf B)}_{n;\boldsymbol\alpha,\mathbf h}(X)\bigr)
\]
is pushed forward to the space of \(\mathcal B\)-colored graphons. Writing
\[
I(u)=u\ln u+(1-u)\ln(1-u),
\]
\[
\mathcal I^{(\mathcal B)}(g)=\sum_{i,j=1}^k\int_{\mathcal B_i\times \mathcal B_j} I(g(x,y))\,dx\,dy,
\]
and
\[
U^{(\mathcal B)}_{\boldsymbol\alpha,\mathbf h}(g)
=
\frac16\sum_{i,j,\ell=1}^k \alpha_{ij\ell}\,t^{(\mathcal B)}_{ij\ell}(H_2,g)
+
\frac12\sum_{i,j=1}^k h_{ij}\,t^{(\mathcal B)}_{ij}(H_1,g),
\]
the measures satisfy an LDP with rate function
\[
\mathcal I^{(\mathcal B)}_{\boldsymbol\alpha,\mathbf h}(g)
=
\frac12\mathcal I^{(\mathcal B)}(g)
-
U^{(\mathcal B)}_{\boldsymbol\alpha,\mathbf h}(g)
-
\inf_h\left\{\tfrac12\mathcal I^{(\mathcal B)}(h)-U^{(\mathcal B)}_{\boldsymbol\alpha,\mathbf h}(h)\right\},
\]
and the limiting free energy is
\[
f^{(\mathcal B)}_{\boldsymbol\alpha,\mathbf h}
=
\sup_g\left\{
U^{(\mathcal B)}_{\boldsymbol\alpha,\mathbf h}(g)-\tfrac12\mathcal I^{(\mathcal B)}(g)
\right\}
\]
[2602.16604].

In the ferromagnetic block regime, \(\alpha_{ij\ell}\ge 0\) for all \(i,j,\ell\). By a block-wise Hölder argument, any maximizer may be chosen to be a \(\mathcal B\)-block-constant graphon
\[
g_C(x,y)=\sum_{i,j=1}^k c_{ij}\mathbf 1_{\mathcal B_i}(x)\mathbf 1_{\mathcal B_j}(y),
\qquad C=(c_{ij})\in[0,1]^{k\times k},\ c_{ij}=c_{ji}.
\]
This reduces the graphon variational problem to a finite-dimensional optimization over the matrix entries \(c_{ij}\). The fact that every optimizer is block-constant is identified as the natural analogue of the usual “replica-symmetric” regime in the homogeneous model [2602.16604].

A Dobrushin-type uniqueness criterion is obtained from the map
\[
(S(C))_{ij}
=
\sigma\!\left(
h_{ij}+\sum_{\ell=1}^k b_\ell\,c_{i\ell}c_{j\ell}\alpha_{ij\ell}
\right),
\qquad
\sigma(u)=\frac{e^u}{1+e^u},
\]
for which
\[
\|S(C)-S(C')\|_\infty\le \tfrac12\|\boldsymbol\alpha\|_\infty\,\|C-C'\|_\infty.
\]
Hence \(\|\boldsymbol\alpha\|_\infty<2\) implies that \(S\) is a contraction, so there is a unique fixed point \(C^\star\), which is the unique maximizer of the finite-dimensional variational problem and therefore of the original infinite-dimensional problem. Under the same condition,
\[
\frac{2E_n}{n^2}\xrightarrow[n\to\infty]{\mathrm{(a.s.)}} \sum_{i,j=1}^k b_i b_j\,c^\star_{ij}.
\]
Equivalently, under the unique Gibbs measure on the infinite vertex set, the empirical block-pair edge density converges to the entries of \(C^\star\) [2602.16604].

Source: https://www.emergentmind.com/topics/ferromagnetic-exponential-random-graph-models-ergms