---
title: Weighted Configuration Model
url: https://www.emergentmind.com/topics/weighted-configuration-model-wcm
type: topic
---

# Weighted Configuration Model

Searching arXiv for recent and foundational papers on the Weighted Configuration Model to ground the article.
to=arxiv_search.search  大发游戏官网  菲律宾申博json
{"query":"Weighted Configuration Model weighted networks null model scale invariance configuration multi-edge model weighted configuration model epidemics weighted hypersoft configuration model", "max_results": 10, "sort_by": "relevance"}
I found several directly relevant arXiv records and will use them to support the article, especially the 2025 scale-invariance analysis, the multi-edge/null-model formulation, the epidemic variant, and hypersoft/generalized extensions.
The **Weighted Configuration Model (WCM)** denotes a family of random weighted-network ensembles obtained by extending the configuration model from unweighted edges to weighted interactions. In the null-model literature, the central construction is a weighted multigraph ensemble with prescribed node strengths, in which edge weights are interpreted as bundles of unit edges and randomized by stub matching; closely related grand-canonical formulations replace exact constraints by expected constraints and yield factorized edge-weight laws [2510.23964] [1404.3697]. The same label is also used in adjacent literatures for degree-dependent weighted stub matchings in epidemic models and for configuration models endowed with i.i.d. edge lengths [1104.5585]. This suggests that the WCM is best understood as a class of related constructions whose common purpose is to isolate the consequences of degree, strength, or weight constraints from higher-order structure.

## 1. Formal strength-preserving ensemble

In its standard null-model form, the WCM is defined on an undirected weighted multigraph \(G\) on \(N\) vertices, without self-loops unless stated otherwise. The edge weight on \((i,j)\) is an integer \(W_{ij}\in\{0,1,2,\dots\}\), the node strength is
\[
s_i=\sum_j W_{ij},
\]
and the total stub count is
\[
2W=\sum_i s_i.
\]

The **microcanonical WCM** is the uniform ensemble over all weighted multigraphs with prescribed strength sequence \(\{s_i\}\). Equivalently, each weight \(W_{ij}\) is split into \(W_{ij}\) unit-weight stubs attached to \(i\) and \(j\), and the \(2W\) stubs are paired uniformly at random. The constraint is exact:
\[
\forall i,\qquad \sum_j W_{ij}=s_i.
\]

If self-loops are forbidden, the state count is
\[
\Omega(\{s_i\})=\frac{(2W)!}{\prod_i s_i!\,\prod_{i<j} W_{ij}!},
\]
so every admissible graph has probability
\[
P(G)=\frac{1}{\Omega(\{s_i\})}
\]
and equivalently
\[
P(G)=\frac{\prod_i s_i!\,\prod_{i<j} W_{ij}!}{(2W)!}.
\]

This formulation makes explicit the original intuition behind the model: weights are treated as multiplicities of parallel unit interactions. In that sense, the WCM is not merely a weighted analogue of the configuration model; it is a multigraph ensemble in which the strength sequence plays the role that the degree sequence plays in the unweighted case [2510.23964].

## 2. Canonical relaxations and the configuration multi-edge model

A canonical or maximum-entropy version replaces exact strength preservation by constraints on expected strengths. Writing the Lagrange multipliers as \(\{\alpha_i\}\), the ensemble probability factorizes as
\[
P(\{W_{ij}\}) \propto \exp\!\left[-\sum_i \alpha_i\Bigl(\sum_j W_{ij}\Bigr)\right]
= \prod_{i<j}\exp[-(\alpha_i+\alpha_j)W_{ij}],
\]
with partition function
\[
Z=\prod_{i<j} Z_{ij},\qquad
Z_{ij}=\sum_{w=0}^{\infty} e^{-(\alpha_i+\alpha_j)w}
=\frac{1}{1-e^{-(\alpha_i+\alpha_j)}}.
\]
Each \(W_{ij}\) is then geometric,
\[
P(W_{ij}=w)=(1-q_{ij})\,q_{ij}^w,\qquad
q_{ij}=e^{-(\alpha_i+\alpha_j)},
\]
and the multipliers are fixed by the \(N\) constraints
\[
\left\langle \sum_j W_{ij}\right\rangle = s_i.
\]

In sparse regimes with large weights, a Chung–Lu approximation replaces the geometric law by a Poisson law with the same mean,
\[
\langle W_{ij}\rangle=\lambda_{ij}\approx \frac{s_i s_j}{2W},
\qquad
P(W_{ij}=w)=e^{-\lambda_{ij}}\frac{\lambda_{ij}^w}{w!},
\]
so that the full ensemble factorizes over pairs [2510.23964].

A closely related grand-canonical formulation for multi-edge networks uses nonnegative integer occupation numbers \(t_{ij}\) on node pairs \(i\le j\), with strengths
\[
s_i(G)=\sum_j t_{ij}.
\]
Maximizing entropy under \(\langle s_i\rangle=\hat s_i\) yields independent Poisson occupations,
\[
P_{GC}(G)=\prod_{i\le j} p_{ij}(t_{ij}),
\]
where
\[
p_{ij}(t)=e^{-\mu_{ij}}\frac{\mu_{ij}^{\,t}}{t!},
\qquad
\sum_j \mu_{ij}=\hat s_i.
\]
Introducing hidden variables \(x_i=e^{-\lambda_i}\), one obtains
\[
\langle t_{ij}\rangle = x_i x_j,\qquad
\mu_{ij}=\beta x_i x_j=\frac{\hat s_i\hat s_j}{\hat T},
\]
with \(\hat T=\sum_i \hat s_i\) and a common normalization \(\beta=1/\hat T\). In this ensemble,
\[
\langle w_{ij}\rangle=\mu_{ij},\qquad
\mathrm{Var}(w_{ij})=\mu_{ij},
\]
and higher moments follow the Poisson law. The same framework gives analytic approximations for ensemble-averaged observables such as node degree, disparity, average neighbor properties, and the occupation-number distribution, typically by multivariate Taylor expansion around the means \(\mu_{ij}\) [1404.3697].

The principal caveat is structural: edges are independent, multiple edges and self-loops may be allowed, higher-order correlations are absent, and grand-canonical enforcement fixes strengths only on average. For observables that depend sensitively on rare large weights, truncated Taylor expansions may also be inaccurate [1404.3697].

## 3. Scale invariance and null-model validity

A central recent result is that the conventional WCM is **scale-dependent** when used for statistical significance in weighted networks. If all original weights are multiplied by a factor \(A>0\), then
\[
s_i\to A s_i,\qquad W\to A W.
\]
In the microcanonical WCM, for \(i\neq j\),
\[
W_{ij}\sim \mathrm{Hypergeometric}(N=2W-1,\;K=s_j,\;n=s_i),
\]
with mean
\[
\mu_{ij}=\frac{s_i s_j}{2W-1}\sim A\cdot\frac{s_i s_j}{2W}
\]
and variance \(\sigma^2_{ij}\sim \mu_{ij}\). In the canonical or Poisson approximation,
\[
\mathrm{Var}[W_{ij}]=\lambda_{ij}\sim A\cdot\frac{s_i s_j}{2W}.
\]

Many weighted observables of practical interest are dimensionless in the weights. For a first-order Taylor expansion around the mean matrix \(W^{avg}_{ij}=s_i s_j/(2W)\),
\[
F(W^R)\approx F(W^{avg})+\sum_{i<j}
\left(\frac{\partial F}{\partial W_{ij}}\right)_{W^{avg}}\delta W_{ij}.
\]
Because \(\left(\partial F/\partial W_{ij}\right)\) scales as \(1/A\) while \(\delta W_{ij}\) scales as \(\sqrt{A}\), the standard deviation of \(F\) over the WCM ensemble scales as \(A^{-1/2}\). As \(A\) grows, the null distribution collapses and the corresponding \(p\)-values tend to zero. Statistical significance therefore depends on the arbitrary unit in which weights are measured, even though in most cases the result should be invariant under that choice [2510.23964].

A scale-invariant alternative separates topology from weights in two steps. First, a binary skeleton is randomized via the unweighted configuration model or its canonical Chung–Lu form, preserving the degree sequence \(\{k_i\}\) in expectation:
\[
P(B_{ij}=1)=p_{ij}=\frac{k_i k_j}{2m}.
\]
Second, for pairs with \(B_{ij}=1\), continuous weights \(S_{ij}>0\) are drawn independently from an exponential law,
\[
P(S_{ij}=s)=\lambda_{ij}e^{-\lambda_{ij}s},
\]
with
\[
\lambda_{ij}=\frac{W k_i k_j}{m s_i s_j},
\qquad
\langle B_{ij}S_{ij}\rangle=\frac{p_{ij}}{\lambda_{ij}}=\frac{s_i s_j}{2W}.
\]
The final weighted graph is
\[
W^R_{ij}=B_{ij}\cdot S_{ij},
\]
and the joint measure is
\[
P(\{B,S\})=P_{CM}(\{B\})\cdot
\prod_{i<j:B_{ij}=1}\lambda_{ij}e^{-\lambda_{ij}S_{ij}}.
\]
Under the rescaling \(\{s_i\}\to A\{s_i\}\), the rates \(\lambda_{ij}\) are unchanged, so the null distribution is invariant. In applications to weighted clustering, eigenvector centrality, and modularity, this two-step model yields unit-independent \(p\)-values; clustering often remains significant, eigenvector centrality typically is not, and modularity significance varies with the network [2510.23964].

## 4. Degree-dependent weights and epidemic thresholds

In epidemic modeling, the WCM has been used in a different but related sense: a random graph with prescribed degree distribution and **degree-dependent edge weights**. Each vertex \(i\) receives an i.i.d. degree
\[
D_i\sim p_k=\Pr(D=k),
\]
and, conditional on \(D_i=k\), its \(k\) stubs receive i.i.d. integer weights \(W_{i1},\dots,W_{ik}\) with
\[
\Pr(W_{ij}=w\mid D_i=k)=q(w\mid k),\qquad w=0,1,2,\dots.
\]
For each weight value \(w\), stubs carrying that value are paired uniformly at random; if the number of \(w\)-stubs is odd, one stub is dropped, which vanishes in the \(n\to\infty\) limit.

If
\[
m_w=\sum_r r\,p_r\,q(w\mid r),
\]
then, in the large-\(n\) approximation, the probability that two vertices of observed degrees \(k\) and \(\ell\) are joined by an edge of weight \(w\) is
\[
\Pr(i\sim j\text{ with weight }w\mid D_i=k,D_j=\ell)
\approx
\frac{k\,q(w\mid k)\,\ell\,q(w\mid \ell)}{n\,m_w},
\]
and summing over \(w\) gives the total connection probability.

The giant-component threshold is expressed through a multi-type branching process with type indexed by weight class. The mean offspring matrix is
\[
M_{w,w'}=
\sum_k (k-1)\,q(w'\mid k)\,
\frac{k\,p_k\,q(w\mid k)}{m_w},
\]
and a giant component appears iff
\[
\rho(M)>1.
\]
In the special unweighted case \(q(w\mid k)=\mathbf{1}_{[w=1]}\), this reduces to the classical condition
\[
\frac{E[D(D-1)]}{E[D]}>1.
\]

For epidemics, if an edge of weight \(w\) transmits independently with probability \(\pi(w)\), the corresponding offspring matrix becomes
\[
M^\pi_{w,w'}=
\sum_k (k-1)\,\pi(w')\,q(w'\mid k)\,
\frac{k\,p_k\,q(w\mid k)}{m_w},
\]
and the basic reproduction number is
\[
R_0=\rho(M^\pi).
\]
Equivalently, one may work with a degree-indexed matrix
\[
m_{d,k}
=(d-1)\sum_{w\ge0}\pi(w)\,q(w\mid d)\,
\frac{k\,p_k\,q(w\mid k)}{\sum_j j\,p_j\,q(w\mid j)},
\]
with
\[
R_0=\rho\bigl((m_{d,k})_{d,k\ge2}\bigr).
\]
If \(q(w\mid d)=q(w)\), this collapses to
\[
R_0=E[\pi(W)]\frac{E[D(D-1)]}{E[D]}.
\]

This formulation isolates the effect of degree–weight correlations. If high-degree vertices tend to have larger weights and \(\pi(w)\) increases with \(w\), the offspring matrix becomes more top-heavy and \(R_0\) increases; if high-degree vertices carry systematically lower weights, \(R_0\) decreases. Empirical fitting therefore requires estimating both \(p_d\) and \(q(w\mid d)\), rather than treating weight and degree as independent attributes [1104.5585].

## 5. Maximum-entropy generalizations and asymptotic properties

The WCM sits inside a broader hierarchy of maximum-entropy ensembles. In the weighted hypersoft configuration model, entropy
\[
S=-\sum_G P(G)\ln P(G)
\]
is maximized not with a fixed degree or strength sequence, but with the empirical distribution of expected degrees and strengths constrained to converge to a target \(\rho^*(\kappa,\sigma)\). The Hamiltonian is
\[
H(G)=\sum_i \bigl[\nu_i k_i(G)+\mu_i s_i(G)\bigr],
\]
so the ensemble measure is
\[
P(G)=\frac{1}{Z}\exp\!\left[-\sum_i (\nu_i k_i(G)+\mu_i s_i(G))\right].
\]
Constructively, for each pair \((i,j)\),
\[
p_{ij}=\frac{1}{1+e^{\nu_i+\nu_j}(\mu_i+\mu_j)},
\]
and, conditional on connection,
\[
w_{ij}\sim \mathrm{Exp}(\text{rate}=\mu_i+\mu_j).
\]
The expected degree and strength satisfy
\[
\kappa_i=E[k_i]=\sum_{j\neq i} p_{ij},
\qquad
\sigma_i=E[s_i]=\sum_{j\neq i}\omega_{ij},
\qquad
\omega_{ij}=E[w_{ij}]=\frac{p_{ij}}{\mu_i+\mu_j}.
\]
For target laws with \(\rho(\kappa)\sim \kappa^{-\gamma}\), \(\gamma>2\), and deterministic scaling \(\sigma=\sigma_0\kappa^\eta\), \(\eta\ge1\), the sparse limit yields Poisson degrees conditioned on \(\kappa\), power-law marginal degree and strength distributions, and superlinear scaling \(s\propto k^\eta\). As a null model, it contains no higher-order structure beyond the imposed degree–strength statistics [2007.00124].

A different asymptotic direction concerns weighted adjacency matrices over arbitrary fields. In that setting, one prescribes a degree sequence, generates a configuration-model multigraph, independently fixes a symmetric nonzero weight matrix \(J_n\in(F^*)^{n\times n}\), and defines
\[
A_n(i,j)=J_n(i,j)\quad\text{if at least one edge survives between }i\text{ and }j,\ i\neq j,
\qquad
A_n(i,i)=0.
\]
Under regularity assumptions on the empirical degree law,
\[
\frac{\mathrm{rank}_F(A_n)}{n}\to \rho=\min_{\alpha\in[0,1]}R_\psi(\alpha)
\]
in probability, where
\[
R_\psi(\alpha)
=
2-\psi(1-\hat\psi(\alpha))-\psi(\alpha)-\psi'(\alpha)(1-\alpha).
\]
The limiting normalized rank depends only on the degree distribution, not on the nonzero weights and not on the field \(F\). This result shows that, for this class of weighted configuration models, certain global linear-algebraic observables are asymptotically insensitive to the specific nonzero edge weights [2508.02813].

## 6. Terminological scope and related weighted configuration models

The expression **weighted configuration model** is not uniform across the literature. In one important usage, the graph is first generated by the configuration model and then given i.i.d. edge lengths \(L_e\) from a common distribution \(F_L\), yielding a weighted graph denoted
\[
\mathrm{CM}_n(\boldsymbol d,L).
\]
For vertices \(u,v\), the graph distance is
\[
d_G(u,v)=\min\{\text{number of edges in a path}\},
\]
the weighted distance is
\[
d_L(u,v)=\min_{\pi:u\to v}\sum_{e\in\pi} L_e,
\]
and the hopcount is the number of edges in the \(L\)-shortest path.

For scale-free degree sequences with exponent \(\tau\in(2,3)\), the weighted-distance behavior depends on whether the associated age-dependent branching process is explosive. Writing
\[
S=\sum_{k=1}^{\infty}F_L^{(-1)}(\exp(-e^k)),
\]
the process is explosive if \(S<\infty\) and non-explosive if \(S=\infty\). In the non-explosive case,
\[
\frac{d_L(u,v)}{2A_n}\xrightarrow{p}1,
\qquad
\frac{d_H(u,v)}{2(\log\log n)/|\log(\tau-2)|}\xrightarrow{p}1,
\]
with
\[
A_n=\sum_{i=1}^{h_n(\tau)}
F_L^{(-1)}\!\bigl(\exp\{-(\tau-2)^{-i}\}\bigr),
\qquad
h_n(\tau)=\left\lfloor\frac{\log\log n}{|\log(\tau-2)|}\right\rfloor.
\]
In the explosive case,
\[
d_L(u,v)\xrightarrow{d}Y^{(1)}+Y^{(2)}.
\]
The same first-order behavior persists in the erased configuration model, where loops and multiple edges are removed [1709.09481].

This terminological breadth is methodologically consequential. In some papers, “weighted” means integer edge multiplicities constrained by node strengths; in others it means degree-dependent edge weights, continuous positive weights assigned after topology generation, or i.i.d. edge lengths used to define first-passage metrics. The common thread is the configuration-model backbone, but the constrained quantities, the sampling measure, and the observables of interest differ substantially.

Source: https://www.emergentmind.com/topics/weighted-configuration-model-wcm