---
title: Pearson Assortativity Coefficient Analysis
url: https://www.emergentmind.com/topics/pearson-assortativity-coefficient
type: topic
---

# Pearson Assortativity Coefficient Analysis

Searching arXiv for the cited literature to ground the article in current and primary sources.
The **Pearson assortativity coefficient** is the standard scalar statistic used in network science to quantify degree–degree or attribute–attribute correlation across edges. In its classical form, it is the **Pearson correlation coefficient of the values observed at the two endpoints of a randomly chosen edge**, with Newman’s degree-based formulation treating the endpoint values as degrees or remaining degrees, depending on the convention [1209.4675]. Across the literature, it serves as a global summary of assortative mixing (\(r>0\)), disassortative mixing (\(r<0\)), and neutral mixing (\(r=0\)); however, a substantial body of work shows that this scalar coefficient can conceal structural heterogeneity, be constrained by topology and metadata marginals, and become unreliable in heavy-tailed regimes or in networks with important mesoscopic organization [2406.15428].

## 1. Classical definition and equivalent formulations

In the standard network-science setting, the Pearson assortativity coefficient measures whether vertices with similar values tend to be adjacent. For **degree assortativity** on a simple connected graph \(G=(V,E)\), one common Newman form is
\[
r=\frac{|E|^{-1}\sum\limits_{e_{uv}\in E} k_{u}k_{v}-\left[|E|^{-1}\sum\limits_{e_{uv}\in E} \frac{1}{2}(k_{u}+k_{v})\right]^{2}}{|E|^{-1}\sum\limits_{e_{uv}\in E} \frac{1}{2}(k^{2}_{u}+k^{2}_{v})-\left[|E|^{-1}\sum\limits_{e_{uv}\in E} \frac{1}{2}(k_{u}+k_{v})\right]^{2}},
\]
where \(k_u\) and \(k_v\) are the endpoint degrees [2406.15428]. The same coefficient is frequently written as a Pearson covariance normalized by endpoint variances. In the graph-level notation used for finite graphs,
\[
g(G)=\mathbb E^* d(v_1)d(v_2),\qquad  b(G)=\mathbb E^* d(v_1),\qquad  b'(G)=\mathbb E^* d^2(v_1),
\]
and
\[
r(G)=\frac{g(G)-b^2(G)}{b'(G)-b^2(G)},
\]
with \(\mathbb E^*\) denoting expectation over the endpoints of a uniformly chosen edge [1209.4675].

For scalar nodal attributes \(x:V\to\mathbb R\), the coefficient is likewise the Pearson correlation of the two endpoint values. In the directed-arc sampling notation,
\[
X:=x(U),\qquad Y:=x(V),
\]
and
\[
\rho(x)=\frac{\mathbb{E}[XY]-\mathbb{E}[X]\mathbb{E}[Y]}{\sqrt{\mathrm{Var}(X)}\sqrt{\mathrm{Var}(Y)}}\in[-1,1],
\]
provided the variances are positive [2601.19422]. The adjacency-matrix form used for scalar attributes is algebraically equivalent, and for categorical attributes the same notion is written via the mixing matrix:
\[
\rho(x)=\frac{\mathrm{Tr}(e)-\sum_{p=1}^K a_p b_p}{1-\sum_{p=1}^K a_p b_p},
\]
where \(e_{pq}\) is the edge-type mixing proportion and \(a_p,b_q\) are its marginals [2601.19422].

For binary metadata, the coefficient specializes to the \(\phi\)-coefficient. In undirected networks with \(e_{ij}=e_{ji}\) and \(a_i=b_i\),
\[
r = \frac{\sum_i e_{ii} - a_i^2}{1-\sum_i a_i^2},
\]
and, in terms of edge counts \(m_{00},m_{10},m_{11}\),
\[
r = 1 - \frac{2m_{10}m}{m^2-(m_{00}-m_{11})^2},
\]
which makes explicit that assortativity increases as the number of cross-class edges \(m_{10}\) decreases, while also depending on the balance between within-class edge counts [1908.04588].

A related convention uses **remaining degree** or **excess degree** rather than total degree. For an undirected graph, one chooses a random edge, orients it randomly, and sets
\[
(X_0,X_1)=\big(\deg(u)-1,\deg(v)-1\big)
\]
or the reverse orientation with equal probability, then defines
\[
r(G)=\frac{\mathrm{Cov}(X_0,X_1)}{\sqrt{\mathrm{Var}(X_0)}\sqrt{\mathrm{Var}(X_1)}}.
\]
This convention is standard in Newman-style formulations and is particularly prominent in work on heavy-tailed and weighted settings [2508.04608].

## 2. Interpretation, sign, and attainable range

The standard interpretation is stable across the literature: \(r>0\) indicates **assortative mixing**, \(r<0\) indicates **disassortative mixing**, and \(r=0\) indicates **neutral** or uncorrelated mixing in the Pearson sense [2406.15428]. In degree-based usage, positive assortativity means high-degree vertices tend to connect to high-degree vertices and low-degree vertices to low-degree ones; negative assortativity means high-degree vertices tend to connect to low-degree ones [1209.4675].

A notable refinement concerns the attainable range. One recent analysis argues that the degree assortativity coefficient defined on simple connected graphs belongs to the asymmetric interval
\[
r\in[-1,1)
\]
rather than the frequently cited \([-1,1]\) [2406.15428]. The lower bound is attained by the **star graph**, and that paper states that the star is the **unique tree network** achieving \(r=-1\) [2406.15428]. The argument for non-attainment of \(r=1\) is that equality would require
\[
\sum_{e_{uv}\in E}(k_u-k_v)^2=0,
\]
so every adjacent pair would have equal degree; in a connected graph this forces regularity, but in the regular case the degree-based formula degenerates to \(0/0\), so the coefficient is not meaningfully defined there [2406.15428].

Other papers retain the standard correlation range \([-1,1]\) when variances are positive, especially for generic scalar attributes or directed weighted settings [2601.19422]. This suggests that the precise endpoint discussion depends on the exact graph class and formulation under consideration. The literature is consistent, however, on one caveat: if the denominator vanishes, the Pearson coefficient is undefined. This occurs, for example, in completely homogeneous or regular settings [2406.15428; 1212.6456].

For binary metadata, the full interval \([-1,1]\) is often not attainable even in principle. The \(\phi\)-coefficient is constrained by the marginals:
\[
\phi_{\min} = -\sqrt{\frac{a_0 b_0}{a_1 b_1}}, \qquad \phi_{\max} = \sqrt{\frac{a_0 b_1}{a_1 b_0}},
\]
and in the undirected case \(a_i=b_i\), so imbalance alone can force \(\phi_{\min}>-1\) [1908.04588]. The same paper emphasizes that network structure can restrict the feasible range even further, so interpreting an observed value against the naive interval \([-1,1]\) can be misleading [1908.04588].

## 3. Directed, weighted, multilayer, and generalized variants

In directed networks, assortativity is no longer a single number tied to one degree variable. There are **four degree-degree correlation types**, corresponding to source and target in-/out-degree pairings:
\[
\big(D^\alpha(e_\ast),\,D^\beta(e^\ast)\big),\qquad \alpha,\beta\in\{+,-\},
\]
namely Out/In, In/Out, Out/Out, and In/In [1310.6528]. The directed Pearson coefficient is written as
\[
r_\alpha^\beta(G) = \frac{ \frac{1}{|E|}\sum_{e\in E} D^\alpha(e_\ast)D^\beta(e^\ast) - \left(\frac{1}{|E|}\sum_{e\in E} D^\alpha(e_\ast)\right) \left(\frac{1}{|E|}\sum_{e\in E} D^\beta(e^\ast)\right) }{ \sigma_\alpha(G)\,\sigma^\beta(G) }.
\]
This decomposition is essential because the four pairings can behave differently both structurally and dynamically [1310.6528].

Weighted and directed generalizations replace degrees by endpoint-specific features such as **strength**, and replace uniform edge weighting by edge weights themselves. One general weighted directed Pearson form is
\[
\rho_{X,Y}(G) = \frac{\sum_{i,j \in V} w_{ij} (X_i - \bar{X}_{\rm sou}) (Y_j - \bar{Y}_{\rm tar})}{W \sigma_{X}\sigma_{Y}},
\]
with source-side and target-side weighted means and variances [2101.05389]. When \(X\) and \(Y\) are in-/out-strengths, this yields four directed strength-assortativity types \((\mathrm{out},\mathrm{in})\), \((\mathrm{out},\mathrm{out})\), \((\mathrm{in},\mathrm{out})\), and \((\mathrm{in},\mathrm{in})\) [2101.05389].

A more specific weighted framework introduces a family
\[
r^{(\alpha,\beta)},
\]
where \(\alpha\in\{0,1\}\) toggles **degree** versus **strength**, and \(\beta\in\{0,1\}\) toggles **uniform** versus **weight-based** edge contributions [2201.07502]. In that formulation, weighted assortativity is decomposed into a **connection effect** and an **amplification effect**, and the preferred fully weighted-strength coefficient is \(r^{(1,1)}\) [2201.07502]. The same paper stresses that in weighted settings the correct analogue of excess degree is **excess strength**, not total strength, because subtracting the focal edge weight is edge-specific rather than a constant shift [2201.07502].

Multilayer networks admit a tensorial Pearson generalization. With reduced adjacency tensor \(\mathcal W_\beta^\alpha\), the multilayer coefficient is
\[
\rho(\mathcal{W}_\beta^\alpha) = \frac{\mathcal{M}^{-1} \mathcal{W}_\beta^\alpha Q^\beta Q_\alpha -  \left[ 1/2 \mathcal{M}^{-1} \left( \mathcal{W}_\beta^\alpha Q_\alpha u^\beta + \mathcal{W}_\beta^\alpha Q^\beta u_\alpha \right) \right]^2}{\mathcal{M}^{-1} \left( \mathcal{W}_\beta^\alpha (Q_\alpha)^2 u^\beta + \mathcal{W}_\beta^\alpha (Q^\beta)^2 u_\alpha \right) - \left[1/2 \mathcal{M}^{-1} \left( \mathcal{W}_\beta^\alpha Q_\alpha u^\beta + \mathcal{W}_\beta^\alpha Q^\beta u_\alpha \right) \right]^2},
\]
covering weighted, directed, and layer-pair-restricted settings [1507.04550]. The paper emphasizes that projections, overlays, and selected layer pairs can yield sharply different apparent assortativity values [1507.04550].

Several works generalize the scalar coefficient by refining the domain of edge aggregation rather than altering the Pearson algebra itself. One proposal rewrites global assortativity as a sum of **edge contributions**
\[
\rho_e=\frac{(j-U_q)(k-U_q)}{M\sigma_q^2},
\]
and defines a **universal assortativity coefficient** for a target edge set \(E_{\text{target}}\) by summing \(\rho_e\) over that subset [1212.6456]. Another line introduces **multiscale/local assortativity** by replacing the global stationary edge weighting with a node-centered distribution \(w(i;\ell)\), including personalized PageRank-based localization [1708.01236]. A more recent refinement defines **interior–boundary assortativity profiles** relative to a partition, producing type-restricted Pearson coefficients for edge strata such as \(I\to I\), \(I\to B\), \(B\to I\), and \(B\to B\) [2601.19422].

## 4. Structural limitations and interpretive caveats

A central theme in the literature is that Pearson assortativity is a **global average**. It compresses the full endpoint joint distribution into one scalar, which can be unrepresentative when mixing patterns are heterogeneous across the network [1708.01236]. Synthetic constructions show that multiple networks can share the same global \(r_{\rm global}=0\) while having very different local organization [1708.01236]. Likewise, partition-based analyses show that a scalar coefficient can be close to zero even when one interface-specific component, particularly \(r_{B\to I}\), is strongly negative, because opposite-signed within-type contributions and between-type mean shifts can cancel [2601.19422].

Another limitation is that the feasible range of \(r\) depends on constraints. For binary metadata, attainable values are shaped not only by class counts but also by the degree sequence, graph topology, and whether metadata are fixed to specific vertices [1908.04588]. That work studies three ensembles—metadata-graph space, graph space, and metadata space—and shows that empirical assortativity should be compared to the feasible range under the relevant ensemble rather than blindly to \([-1,1]\) [1908.04588].

Higher-order organization can also invalidate interpretation based on first-neighbor degree correlations alone. In highly assortative networks, two graph-generation procedures can produce similar Pearson \(r\) but very different percolation behavior because they induce different **correlations above the first shell** [1109.0560]. To expose this, the paper defines a generalized Pearson coefficient at chemical distance \(\ell\),
\[
r_{\ell}= \frac{\langle k k'\rangle_{\ell} - \langle (k + k')/2\rangle^{2}_{\ell}}{\langle (k^2 + k'^2)/2\rangle_{\ell}- \langle (k + k')/2\rangle^{2}_{\ell}},
\]
with ordinary assortativity recovered as \(r_1=r\) [1109.0560]. This suggests that nearest-neighbor Pearson assortativity is sometimes too coarse to characterize long-range degree organization.

Related extensions replace degree by a richer local summary. The **two-walks degree assortativity** coefficient applies the same Pearson form to
\[
\tilde{k}_{i}=\sum_{j\in N(i)}k_j,
\]
thereby incorporating second-neighbor information [1704.03943]. That work reports the existence of graphs and real networks that are degree-disassortative but two-walks-assortative, while no assortative-disassortative examples were found in its exhaustive computational study; the latter is explicitly stated as a conjectural nonexistence claim rather than a theorem [1704.03943].

## 5. Heavy tails, asymptotics, and failure modes

The most sustained critique of the Pearson assortativity coefficient concerns **heavy-tailed degree distributions**. In large scale-free networks, especially disassortative ones, the magnitude of the coefficient can decrease with network size, making comparisons across differently sized graphs unreliable [1204.0266; 1202.3071]. One standard expression is
\[
\rho_n= \frac{\sum_{ij\in E} d_i d_j - \frac{1}{|E|}\left(\sum_{i\in V} d_i^2\right)^2}{\sum_{i\in V} d_i^3 - \frac{1}{|E|}\left(\sum_{i\in V} d_i^2\right)^2},
\]
which makes explicit the dependence on second and third degree moments [1204.0266].

For scale-free degree exponents \(\gamma\in(1,3)\), one analysis shows that in disassortative graphs the lower bound
\[
\rho_n \ge \rho_n^-
\]
satisfies \(\rho_n^-\to 0\), so even strong disassortativity can become invisible as the graph grows [1204.0266]. A related paper proves that when the degree distribution has an infinite third moment, any limit point of the Pearson assortativity coefficient is **non-negative** under broad regularity conditions, despite genuine negative dependence being possible in the graph structure [1202.3071]. These results are extended to directed heavy-tailed networks, where Pearson coefficients \(r_\alpha^\beta\) can converge to non-negative limits across large regions of the \((\gamma_+,\gamma_-)\)-plane [1310.6528].

A still stronger critique is given for sufficiently heavy-tailed scale-free networks with exponent
\[
2<\tau<\frac{7}{3},
\]
maximum degree \(\Delta=\Theta(n^{1/(\tau-1)})\), and tail \(\Pr(\deg(v)\ge k)=\Theta(k^{1-\tau})\). Under these assumptions,
\[
r(G)=\Theta\!\big(-n^{-(\tau-2)/(\tau-1)}\big),
\]
so the coefficient is forced to be negative and tends to \(0^{-}\) regardless of the actual wiring preference [2508.04608]. That paper argues that the extreme sparsity of the upper-right corner of the degree-degree joint distribution is unavoidable in this regime, mechanically pushing the covariance negative [2508.04608].

Because of these failures, several papers advocate **rank-based alternatives** such as Spearman’s rho and Kendall’s tau for heavy-tailed networks [1204.0266; 1202.3071; 1310.6528]. The common rationale is that rank correlations are not dominated by large raw degree magnitudes, are more stable across graph size, and better capture monotone dependence when high moments diverge [1310.6528]. This suggests that Pearson assortativity is most informative when second and third moments are well behaved and degree heterogeneity is not itself the dominant asymptotic effect.

## 6. Model-specific behavior and links to motifs, clustering, and dynamics

Despite its limitations, the Pearson assortativity coefficient remains analytically tractable in several network models. In **sparse random intersection graphs** with non-vanishing clustering, it is explicitly tied to the degree moments of the asymptotic distribution [1209.4675]. In the active model \(G_s(n,m,P)\), under sparsity assumptions, fixed \(s\), and finite third moment \(\mathbb E Z^3<\infty\), the paper gives asymptotic formulas for \(r\) and states that the limit is determined by the first three moments of the asymptotic degree distribution [1209.4675]. A central conclusion is that assortativity is asymptotically **non-negative**, and positive when clustering does not vanish and the third moment is finite; the positivity argument uses Hölder’s inequality to ensure \(a_1a_3\ge a_2\) [1209.4675].

That same paper connects assortativity to local triangle structure through two auxiliary quantities: the expected number of common neighbors of adjacent nodes,
\[
h(G)=\mathbb E^* d(v_1,v_2),
\]
and the expected degree of a neighbor of a degree-\(k\) vertex,
\[
b_k(G)=\mathbb E_k^* d(v_1).
\]
Its formulas show that the increase of \(b_k\) with \(k\) is driven by the triangle-closing term, linking degree-degree correlation directly to clustering [1209.4675]. This supports the broader interpretation that in random intersection graphs, assortativity emerges naturally from the shared-attribute mechanism.

A different tractable setting is **Bernoulli random graph superpositions**, where the limiting adjacent-degree distribution is represented as
\[
\bar f_{2}=\delta_{(1,1)} \ast (\bar f_1 \otimes \bar f_1) \ast f'_{2},
\]
and the model assortativity has the closed form
\[
\Cor(D_1^*, D_2^*) = \frac{P_{21}( P_{43} + P_{33}) - P_{32}^2}{P_{21}(P_{43} + P_{32}) - P_{32}^2 + \mu P_{21}^2 ( P_{21} + P_{32})}
\]
[2002.11809]. The paper proves a moment inequality implying that this limiting Pearson assortativity is always **nonnegative** under the stated assumptions [2002.11809].

Assortativity can also emerge from **component selection** rather than from the substrate network. For the giant component formed by site percolation on an uncorrelated random network, a generating-function calculation yields a general formula for \(r\), and for finite third moment the paper proves
\[
r\le 0 \qquad (p \ge p_{\rm c}),
\]
so the giant component is generically **disassortative** in the percolating phase [1807.08164]. Near the threshold, the average degree of neighbors of degree-\(k\) nodes scales as \(k^{-1}\), reinforcing the disassortative interpretation [1807.08164].

Several works relate Pearson assortativity to local motif structure. One graphon-based study rewrites assortativity in terms of homomorphism densities of paths, stars, and triangles, with a central formula
\[
r_W = \frac{ (n-3)t(P_3,W) + \frac{3n\,t(C_3,W)}{(n-1)(n-2)} - \frac{(n-2)t(P_2,W)^2}{t(P_1,W)} }{ (n-3)3t(S_3,W) + t(P_2,W) - \frac{(n-2)t(P_2,W)^2}{t(P_1,W)} }
\]
[2503.03061]. That paper uses Archimedean copula graphons to generate networks to target assortativity without rewiring, thereby making the coefficient a controllable function of subgraph frequencies [2503.03061].

Finally, Pearson assortativity is often consequential for dynamical systems on networks. In a directed network of theta neurons, four directed assortativity coefficients \(r(\alpha,\beta)\) are used as control parameters in constructing effective connectivity matrices, and the paper reports that \((\mathrm{in},\mathrm{out})\) and especially \((\mathrm{in},\mathrm{in})\) can significantly alter bifurcation structure, whereas \((\mathrm{out},\mathrm{in})\) and \((\mathrm{out},\mathrm{out})\) have no visible dynamical effect in that model [2004.00206]. In partitioned SIS dynamics, boundary dominance implies a strictly negative \(B\to I\) assortativity component, showing that interface-resolved assortativity can encode nonlinear flow geometry that is invisible to the scalar coefficient [2601.19422].

## 7. Methodological uses, inference, and contemporary extensions

The Pearson assortativity coefficient remains a standard descriptive statistic, but recent work treats it increasingly as an object for decomposition, control, and inference rather than as a self-sufficient summary. Localization methods define node-level or multiscale assortativity distributions, which in empirical networks are often **skewed, overdispersed, and multimodal** rather than concentrated around the global value [1708.01236]. Edge-based decompositions separate globally disassortative networks into mixtures of many assortative edges and fewer, stronger disassortative ones [1212.6456]. Partition-based profiles further split the coefficient into interface-specific components, yielding exact decomposition theorems for scalar assortativity [2601.19422].

For generative modeling, copula graphons provide one route to targeting Pearson assortativity through motif control [2503.03061]. Another recent line proposes an extension of geometric inhomogeneous random graphs with tunable assortativity after arguing that single-valued coefficients, including Pearson, do not sufficiently capture degree-dependent wiring preferences in heavy-tailed latent-space models [2508.04608]. This suggests an ongoing shift from scalar estimation toward fine-grained joint and conditional endpoint distributions.

Privacy has also become an explicit concern. A recent differential privacy framework for network assortativity defines the classical coefficient
\[
r=\frac{M^{-1}\Sigma_{e_{ij} \in E}{d_id_j}-\left[ M^{-1}\Sigma_{e_{ij} \in E}{\frac{1}{2}\left( d_i+d_j \right)} \right] ^2}{M^{-1}\Sigma_{e_{ij} \in E}{\frac{1}{2}\left( d_{i}^{2}+d_{j}^{2} \right)}-\left[ M^{-1}\Sigma_{e_{ij} \in E}{\frac{1}{2}\left( d_i+d_j \right)} \right] ^2}
\]
and focuses on unbiased private estimation of the numerator, termed the **assortativity factor** \(r_u\) [2505.03639]. It proposes three algorithms—\(\mathbf{Local_{ru}}\), \(\mathbf{Shuffle_{ru}}\), and \(\mathbf{Decentral_{ru}}\)—and proves unbiasedness of the corresponding estimators, with \(\mathbf{Shuffle_{ru}}\) reported as the most accurate under its privacy model [2505.03639].

Taken together, these developments show that the Pearson assortativity coefficient remains foundational because of its analytic simplicity, broad applicability, and direct interpretation as an edge-endpoint correlation. At the same time, the literature now treats it as a statistic whose meaning depends strongly on graph class, moment conditions, metadata constraints, partition structure, and the scale at which mixing is examined. This suggests that, in contemporary network analysis, the coefficient is best viewed not as a complete description of assortative structure but as one carefully contextualized observable within a broader methodological toolkit.

Source: https://www.emergentmind.com/topics/pearson-assortativity-coefficient