# The rich are loopy\: a spectral framework for loop structure in the rich club { :label: manuscript } :author: { :name: Leo Torres :email: [email protected] :affiliation: Nora, Leipzig, Germany } :: :abstract: The rich club coefficient $\phi(k)$ quantifies whether high-degree nodes interconnect more densely than expected, but it is blind to the structure of those interconnections and offers no principled way to select the degree threshold $k$. We introduce the *cycle core* $H_k$, the maximal subgraph of vertices with degree at least $k$ that forms cycles among themselves. The nested sequence $H_1 \supseteq H_2 \supseteq \cdots$ reveals how the cycle core evolves as the degree threshold increases, capturing structure that $\phi(k)$ cannot see. We develop two spectral tools for this sequence. The first is structural\: the *$k$-blocking operator* $\mathbf{M}_k$ generalizes the non-backtracking matrix by blocking $k - 1$ directions at each node. We prove that its spectrum is equivalent to that of a modified non-backtracking matrix weighted with a specific function of the nodes' degrees. Thus, $\mathbf{M}_k$ couples cycle structure with the degree sequence of the graph. We show the spectral radius of $\mathbf{M}_k$ peaks at an intrinsic threshold $k^*$, yielding the *optimal rich club* $H_{k^*}$ without null models or parameter tuning. The second tool is dynamical\: the $k$-blocking walk defines a sub-stochastic Markov chain whose leading eigenvalue $\lambda_1(k)$ measures how long a process can persist on the cycle core before being expelled to the periphery. We evaluate the framework on fifteen networks spanning different domains, wherein $H_{k^*}$ closely recovers the published rich clubs. :: :toc: ## Introduction { :label: sec-intro } The rich club coefficient $\phi(k)$ of a network :cite:colizza2006,zhou2004:: measures whether high-degree nodes interconnect more densely than expected. It is defined as $\phi(k) = 2E_k / (N_k(N_k - 1))$, where $N_k$ and $E_k$ are the vertices and edges among vertices with degree at least $k$. It has found wide use in neuroscience :cite:vandenheuvel2011:: and beyond :cite:opsahl2008::, but it captures only density: a star and a cycle of high-degree nodes can yield the same $\phi(k)$ despite having fundamentally different structure. A further difficulty is the choice of threshold\: $\phi(k)$ is a function of $k$, and there is no principled way to select a single value. In this work, we propose an alternative. For each positive integer $k$, let $G_k$ be the graph induced by nodes of degree at least $k$. Define the *cycle core* $H_k = \operatorname{2\text{-}core}(G_k)$ as the maximal subgraph of $G_k$ that forms loops or cycles among themselves. The nested sequence $H_1 \supseteq H_2 \supseteq H_3 \supseteq \cdots$ progressively strips away nodes that do not participate in cyclic interconnections at increasing degree thresholds (:ref:sec-cycle-core::). Unlike $\phi(k)$, which sees only density, the cycle core sequence sees *structure*\: whether high-degree nodes form cycles, how those cycles nest as the degree threshold rises, and at what threshold the cyclic core collapses. On top of $H_k$ we build two new spectral tools. *A structural invariant.* We define an operator $\mathbf{M}_k$ that generalizes the non-backtracking matrix :cite:krzakala2013:: by blocking $k - 1$ directions at each node of $V(H_k)$, while the non-backtracking matrix blocks only one direction. The spectrum of $\mathbf{M}_k$ captures the interaction between the cycle structure of $H_k$ and the degree heterogeneity of nodes in $V(H_k)$ relative to the full network. We show that this interaction is not reducible to the degree sequence alone (:ref:sec-operator::). The spectral radius of $\mathbf{M}_k$ peaks at an intrinsic threshold $k^*$, giving the *optimal rich club* $H_{k^*}$, a principled answer to the threshold selection problem (:ref:sec-entrenchment-profile::). *A dynamical process.* We use $\mathbf{M}_k$ to define a random walker on the oriented edges of $H_k$. This yields a sub-stochastic Markov chain $P_k$ in which the walker either remains in the cycle core or is expelled to the periphery. The leading eigenvalue $\lambda_1(P_k)$ measures the per-step survival probability of a process trying to persist on $H_k$. This walk is distinct from the standard non-backtracking walk on the same nodes\: the blocking constraints reweight transition probabilities by an analytically derived factor that depends on cycle structure within $H_k$ rather than only on degree. $\lambda_1(P_k)$ as a function of $k$ traces how the walk's expected persistence time on $H_k$ falls off as the threshold tightens (:ref:sec-dynamics::), a profile not derivable from the structural spectrum of $\mathbf{M}_k$. We compute $H_{k^*}$ on fifteen networks that have previously been identified as having either a rich club or a general core-periphery structure, comparing against published ground truths and three coreness-ranking baselines (the $k$-core decomposition :cite:seidman1983::, Della Rossa et al.'s $\alpha$-coreness :cite:dellarossa2013::, and a Cucuringu et al.'s spectral baseline :cite:cucuringu2016::). $H_{k^*}$ recovers the published rich clubs without null models or parameter tuning. Further, we show the blocking walk is qualitatively and quantitatively different from a degree-proportional non-backtracking random walk. The dynamical quantity $\lambda_1$ also tracks the structural evolution of the core of the Internet's autonomous systems across two years of daily snapshots :cite:leskovec2005:: (:ref:sec-empirical::). ## The cycle core { :label: sec-cycle-core } Let $G$ be a simple connected graph on $n$ vertices and $m$ edges. Throughout, $d(v)$ denotes the degree of vertex $v$ in $G$. For each integer $k \geq 1$, let $V_k = \{v \in V(G) : d(v) \geq k\}$ denote the set of vertices with degree at least $k$. Recall the $2$-core of a graph is the maximal subgraph in which every vertex has degree at least $2$; equivalently, it is the subgraph that remains after iteratively removing all vertices of degree less than $2$. A graph has a non-empty $2$-core if and only if it contains a cycle. :definition: { :label: def-cycle-core :title: Cycle core } For a graph $G$ and an integer $k \geq 1$, the *cycle core* at threshold $k$ is $$ H_k = \operatorname{2\text{-}core}(G_k), $$ where $G_k$ is the subgraph of $G$ induced on vertices with degree at least $k$. :: The cycle core $H_k$ captures vertices that are both high-degree *and* participate in cycles among other high-degree vertices. Taking the $2$-core removes all tree-like appendages from the high-degree subgraph, retaining only the cyclic part. :proposition: { :label: prop-filtration :title: Filtration property } The cycle cores form a nested sequence $H_1 \supseteq H_2 \supseteq H_3 \supseteq \cdots$ that eventually becomes empty. :: :proof: Since $G_{k+1} \subseteq G_k$, the subgraph $H_{k+1}$ is a subgraph of $G_k$ with minimum degree at least $2$. The $2$-core is the *maximal* such subgraph, so $H_{k+1} \subseteq H_k$. For $k$ larger than the maximum degree in $G$, no vertex has degree at least $k$, so $G_k$ and $H_k$ are both empty. :: :remark: { :label: rem-kcore } The cycle core filtration is distinct from the standard $k$-core decomposition :cite:seidman1983,batagelj2003::, which has found wide use in network analysis :cite:dorogovtsev2006::. The $k$-core $K_k$ is the maximal subgraph with minimum degree $k$; the cycle core $H_k$ is the $2$-core of the subgraph induced by vertices with *global* degree at least $k$. For $k = 2$ they coincide\: $H_2 = \operatorname{2\text{-}core}(G)$. For $k \geq 3$ the two filtrations generally differ. :: :proposition: { :label: prop-kcore-inclusion :title: Cycle core contains k-core } For every $k \geq 2$, $K_k \subseteq H_k$. :: :proof: Every $v \in K_k$ has $\deg_{K_k}(v) \geq k$, hence $d(v) \geq k$ and $v \in V_k$. The induced subgraph $K_k = G[V(K_k)]$ is a subgraph of $G_k$ with minimum degree at least $k \geq 2$, so $K_k \subseteq \operatorname{2\text{-}core}(G_k) = H_k$. :: The inclusion is generically strict, and the gap is often large. :ref:fig-Hk-vs-Kk:: shows $|H_k|$, $|K_k|$, and $|H_k \setminus K_k|$ as functions of $k$ on the ten rich-club and $k$-core networks listed in :ref:sec-empirical:: (the five small Borgatti-Everett networks have too narrow a $k$-support to display the pattern); on every one, $K_k$ collapses to the empty set at a much lower $k$ than $H_k$ does. The reason is structural\: $K_k$ demands that every member have at least $k$ neighbours *that themselves survive the $k$-core peeling*, a self-referential condition that propagates the peeling globally. $H_k$ demands only that a vertex have global degree at least $k$ and lie on a cycle among other $V_k$-vertices, a strictly weaker condition that does not require the neighbours to be deep in any hierarchy. The cycle filtration is therefore longer-lived than the $k$-core filtration. :html: { :path: Hk_vs_Kk.html :static: Hk_vs_Kk.png :label: fig-Hk-vs-Kk :alt: |H_k|, |K_k|, |H_k \ K_k| versus k on the ten rich-club and k-core paper networks } :caption: Cycle core persists past the standard $k$-core. Vertex counts $|H_k|$ (cycle core), $|K_k|$ ($k$-core), and the gap $|H_k \setminus K_k|$ versus $k$ on the ten rich-club and $k$-core networks of :ref:sec-empirical::. $K_k$ vanishes well before $H_k$ on every example. :: :html: { :path: cycle_core_widget.html :label: fig-cycle-core } :caption: Cycle cores of Zachary's Karate Club network. Use the slider to vary the degree threshold $k$. :: :ref:fig-cycle-core:: illustrates the cycle core on a small network. At $k = 2$, $H_k$ contains 33 of 34 nodes. At $k = 4$, it shrinks to 16, retaining only those with degree at least 4 that participate in cycles among themselves. At $k = 9$, only a triangle of three high-degree nodes survives. Two further nodes have degree at least 9 but are not in $H_9$\: each has at most one neighbour of degree at least 9 within $V_9$, so neither can close a cycle at that threshold. The cycle core $H_k$ is a subgraph of $G_k$, and $\phi(k)$ measures the edge density of $G_k$. Two networks with the same $\phi(k)$ can have very different cycle cores. One may have a large, richly connected $H_k$ while the other has an acyclic $G_k$ with $H_k = \emptyset$. The cycle core captures the topological structure that density misses. Computing $H_k$ requires thresholding by degree and extracting the $2$-core, both of which run in time linear in the number of edges. The non-backtracking matrix is the algebraic object that organises cycle counts on a graph; we will work with its restriction to the cycle core throughout the rest of the paper. :definition: { :label: def-nb-matrix :title: Non-backtracking matrix } Let $G = (V, E)$ be a simple undirected graph and write $\vec{E}(G) = \{(u, v) : \{u, v\} \in E\}$ for its set of directed edges (each undirected edge contributes two oriented copies). The *non-backtracking matrix* $\mathbf{B}(G)$ of $G$ is the $|\vec{E}(G)| \times |\vec{E}(G)|$ matrix with entries $$\mathbf{B}(G)\bigl[(u, v),\,(x, y)\bigr] \;=\; \begin{cases} 1 & \text{if } v = x \text{ and } y \neq u, \\ 0 & \text{otherwise.} \end{cases}$$ For each integer $k \geq 2$ we write $\mathbf{B}_k := \mathbf{B}(H_k)$ for the non-backtracking matrix of the cycle core; equivalently, $\mathbf{B}_k$ is the principal submatrix of $\mathbf{B}(G)$ on directed edges with both endpoints in $V(H_k)$. :: ## The blocking operator { :label: sec-operator } We now define an operator whose eigenvalue spectrum is closely tied to the structure of the cycle core. :definition: { :label: def-blocking-operator :title: Blocking operator } For a graph $G$ and an integer $k \geq 2$, the *$k$-blocking operator* $\mathbf{M}_k$ is a matrix indexed by states $(u, v, W)$ where\: :itemize: :-: $v \in V(H_k)$ is the *current node*, :-: $u \in N(v) \cap V(H_k)$ is the *arrival node*, :-: $W \subset N(v) \setminus \{u\}$ with $|W| = k - 2$ is the set of *blocked nodes*. :: A transition $(u, v, W) \to (v, x, W')$ occurs when $x \in N(v) \cap V(H_k)$, $x \neq u$, $x \notin W$, and $W' \subset N(x) \setminus \{v\}$ with $|W'| = k - 2$. An entry of $\mathbf{M}_k$ is equal to $1$ if and only if the transition between the row-state and the column-state is valid, and zero otherwise. :: The operator $\mathbf{M}_k$ can be interpreted as the non-normalized transition matrix of the *$k$-blocking walker*\: at node $v$ it is forbidden from returning to the arrival node $u$ and from using any of the $k - 2$ blocked nodes in $W$. In total, $k - 1$ directions are blocked. For $k = 2$, the blocked set is empty and the operator reduces to the non-backtracking matrix. :theorem: { :label: thm-trace :title: Trace formula } For all $n \geq 1$, $$ {:label: eq-trace} \operatorname{tr} \left( \mathbf{M}_k^n \right) = \sum_{\mathbf{w}} \; \prod_{i=1}^{n} \binom{d(v_i) - 2}{k - 2}, $$ where the sum ranges over all non-backtracking closed walks $\mathbf{w} = (v_1, \ldots, v_n)$ of length $n$ among vertices in $V(H_k)$. The walk is cyclically non-backtracking, i.e., indices on $\mathbf{w}$ are interpreted modulo $n$ so $v_{i+1} \neq v_{i-1}$ holds for every $i$ including the wraparound at $i = 1$ and $i = n$. :: :proof: By definition, $\operatorname{tr}(\mathbf{M}_k^n)$ counts closed sequences of states compatible with the transition rule of :ref:def-blocking-operator::. Each such sequence projects to a non-backtracking closed walk $\mathbf{w} = (v_1, \ldots, v_n)$ among nodes in $V(H_k)$ by taking the current-vertex component $v_i$ and ignoring $u_i$ and $W_i$ of each state. The cyclic-NB constraint is automatic\: the transition out of state $s_n$ back into $s_1$ forces $v_2 \neq v_n$, and the transition into $s_1$ from $s_n$ sets $u_1 = v_n$, which the next transition out of $s_1$ then forbids as the new destination, giving $v_{n-1} \neq v_1$. Furthermore, given $\mathbf{w}$, the number of compatible state sequences is $\prod_i \binom{d(v_i) - 2}{k-2}$. Indeed, at vertex $v_i$, the blocked set must be a $(k-2)$-subset of $N(v_i) \setminus \{v_{i-1}, v_{i+1}\}$ (indices mod $n$), and these choices are independent across vertices. :: The weight $\binom{d(v_i) - 2}{k-2}$ uses the *global* degree, not the degree within $H_k$. Blocked directions can point to any neighbor, including low-degree vertices outside $H_k$. The trace formula therefore couples the cycle structure of $H_k$ with how each node is embedded in the broader network. :remark: { :label: rem-peripheral } In networks with heterogeneous degree distributions, high-degree nodes are rare and, when $k$ is large, most of a node's neighbors have low degree. The blocked set $W$ therefore consists almost entirely of low-degree neighbors, and blocking them does not restrict $H_k$-internal transitions. This is consistent with an interpretation where the blocked directions represent connections to the periphery that the node *sacrifices* to maintain its position in the core. In other words, a $k$-blocking walk may be interpreted as aiming to remain entrenched in the high-degree areas of the network. :: To turn the trace formula into a statement about $\mathbf{M}_k$'s spectrum we need a companion matrix carrying the same trace data. :definition: { :label: def-DB :title: Degree-weighted non-backtracking matrix } Let $\mathbf{B}_k$ denote the non-backtracking matrix of $H_k$, indexed by directed edges $(u, v)$ among nodes in $V(H_k)$. Let $\mathbf{D}_k$ be the diagonal matrix with entries $\mathbf{D}_k[(u,v),(u,v)] = \binom{d(v) - 2}{k-2}$ (the global degree of $v$ in $G$). The *degree-weighted non-backtracking matrix* is $\mathbf{D}_k \mathbf{B}_k$. Because $H_k$ has minimum degree at least $2$, $\mathbf{B}_k$ has trivial kernel; combined with $\mathbf{D}_k$ invertible (every binomial entry is $\geq 1$), $\mathbf{D}_k \mathbf{B}_k$ is invertible whenever $H_k$ is non-empty. :: :theorem: { :label: thm-spectral :title: Spectrum of the blocking operator } The non-zero eigenvalues of $\mathbf{M}_k$ coincide, with multiplicity, with the non-zero eigenvalues of $\mathbf{D}_k \mathbf{B}_k$. As immediate consequences\: :enumerate: :-: $\mathbf{M}_k$ has a non-zero eigenvalue if and only if $H_k$ is non-empty; :-: $\operatorname{rank}(\mathbf{M}_k) = 2|E(H_k)|$, so $\operatorname{nullity}(\mathbf{M}_k) = |\text{states}(\mathbf{M}_k)| - 2|E(H_k)|$. :: :: :proof: Expanding $\operatorname{tr}((\mathbf{D}_k \mathbf{B}_k)^n)$ over closed non-backtracking walks, each walk $\mathbf{w} = (v_1, \ldots, v_n)$ contributes $\prod_i \binom{d(v_i) - 2}{k-2}$\: $\mathbf{B}_k$ enforces the non-backtracking constraint, and $\mathbf{D}_k$ supplies the binomial weight at each visited vertex. Comparing with :ref:thm-trace::, $\operatorname{tr}((\mathbf{D}_k \mathbf{B}_k)^n) = \operatorname{tr}(\mathbf{M}_k^n)$ for all $n \geq 1$. The traces of all powers of a matrix determine the multiset of its non-zero eigenvalues, by the Newton-Girard identities. This proves the main statement. For (1)\: if $H_k$ is empty there are no NB closed walks among $V(H_k)$, so $\mathbf{D}_k \mathbf{B}_k$ is the empty operator (the $0\times 0$ matrix). Conversely, a cycle $v_1, \ldots, v_g$ in $H_k$ contributes $g$ distinct length-$g$ closed walks to $\operatorname{tr}(\mathbf{M}_k^g)$, each with the same binomial product, giving $\operatorname{tr}(\mathbf{M}_k^g) \geq g \prod_i \binom{d(v_i) - 2}{k-2} > 0$. For (2)\: equality of non-zero spectra implies equal rank; since $\mathbf{D}_k$ and $\mathbf{B}_k$ are both invertible, $\operatorname{rank}(\mathbf{M}_k) = 2|E(H_k)|$, and rank-nullity on $\mathbf{M}_k$ yields $\operatorname{nullity}(\mathbf{M}_k) = |\text{states}(\mathbf{M}_k)| - 2|E(H_k)|$. :: If all nodes in $V(H_k)$ have the same degree $d$ in $G$, then the eigenvalues of $\mathbf{M}_k$ are $\binom{d - 2}{k - 2}$ times those of $\mathbf{B}_k$, and $\mathbf{D}_k \mathbf{B}_k$ carries no information beyond the non-backtracking matrix. Thus, the operator is most informative when degrees of nodes in $V(H_k)$ vary. The spectrum of $\mathbf{D}_k \mathbf{B}_k$ is not determined by the degree sequence alone\: degree-preserving rewiring changes $\mathbf{B}_k$ while preserving $\mathbf{D}_k$, so two graphs with identical degree sequences can yield distinct spectra. Nor is it determined by the cycle structure alone\: attaching pendants to nodes in $V(H_k)$ raises their global degrees and so changes $\mathbf{D}_k$, while $H_k$ and $\phi(k)$ remain untouched. The operator therefore sees structure that neither the degree sequence nor $\phi(k)$ can detect on its own. The state space of $\mathbf{M}_k$ is much larger than that of $\mathbf{D}_k \mathbf{B}_k$\: each directed edge of $H_k$ contributes $\binom{d(v) - 1}{k - 2}$ states to $\mathbf{M}_k$, one per choice of blocked set $W$, and this multiplicity grows quickly with $k$ and with the degrees of $V(H_k)$. :ref:thm-spectral:: collapses the blow-up\: the non-zero spectrum of $\mathbf{M}_k$ is recovered from the $2|E(H_k)|$-dimensional companion $\mathbf{D}_k \mathbf{B}_k$ alone, which is the matrix we work with in practice. The non-backtracking factor $\mathbf{B}_k$ is itself a principal submatrix of the non-backtracking matrix of $G$, obtained by restricting rows and columns to directed edges with both endpoints in $V(H_k)$. The next section follows its leading eigenvalue as a function of $k$ and reads off what the peak says about $H_k$. ## The entrenchment profile { :label: sec-entrenchment-profile } This section studies how the spectral radius of $\mathbf{D}_k \mathbf{B}_k$ depends on the degree threshold $k$. The behaviour is non-monotone, and the location of its maximum picks out a distinguished threshold and an associated subgraph. We call this curve the *entrenchment profile*\; the name will be motivated in :ref:sec-dynamics::, where $\rho(k)$ is shown to be a degree-weighted spectral measure of cycle entrenchment, closely related to but spectrally distinct from the per-step survival rate $\lambda_1(P_k)$ of the $k$-blocking walk. :definition: { :label: def-profile :title: Entrenchment profile, optimal threshold, optimal rich club } The *entrenchment profile* of a graph $G$ is the function $$\rho(k) \;=\; \rho(\mathbf{D}_k \mathbf{B}_k), \qquad k = 2, 3, \ldots, \max_v d(v).$$ The *optimal threshold* is $k^* = \operatorname{argmax}_k \rho(k)$, and the *optimal rich club* is $H_{k^*}$. :: The entrenchment profile encodes both the cycle structure within $H_k$ (via $\mathbf{B}_k$) and the degree heterogeneity of those vertices (via the binomial weights in $\mathbf{D}_k$). :html: { :path: rho_curves.html :static: rho_curves.png :label: fig-rho-curves :alt: Entrenchment profiles across network families } :caption: Entrenchment profile $\rho(k)$ (A), cycle-core decay $|H_k|/|H_3|$ (B), and rich-club coefficient $\phi(k)$ (C) on six representative networks spanning the rich-club, $k$-core, and Borgatti-Everett frameworks. :: The profile rises as the $\mathbf{D}_k$ weights grow, peaks at $k^*$, and crashes to zero when $H_k$ becomes empty (:ref:fig-rho-curves::A). The shape reflects two opposing forces visible in the trace formula :ref:eq-trace::\: the binomial weight $\binom{d - 2}{k - 2}$ rises with $k$ (up to $k = \lfloor d/2 \rfloor + 1$), while the support $H_k$ contracts; by :ref:thm-spectral::, $\rho > 0$ iff $H_k$ is non-empty, pinning the crash to where $H_k$ empties (:ref:fig-rho-curves::B). The location of $k^*$ varies widely ($k^* = 4$ for Krackhardt, $k^* = 205$ for Facebook), tracking the upper tail of the degree distribution\: $k^*$ is bounded by $\max_v d(v)$, and within that bound, networks with more vertices at high degree sustain cycles up to a higher threshold. Panel C shows the rich-club coefficient $\phi(k)$ on the same networks for reference. At small $k$ the binomial weights collapse to $\approx 1$ and $\rho \approx \rho(\mathbf{B})$, the ordinary non-backtracking spectral radius\: large, but uninformative about the core. Empirically, on every network we tested (:ref:sec-empirical::) the maximum is unique, interior (so $k^* > 2$), and well-separated from its neighbours, so $k^*$ is stable to small perturbations of the degree sequence. :remark: { :label: rem-regular } For a $d$-regular graph the maximum admits a closed form\: $H_k = G$ for every $k \leq d$, so $\rho(\mathbf{D}_k\mathbf{B}_k) = \binom{d-2}{k-2}\rho(\mathbf{B})$ and the binomial prefactor places $k^* = \lfloor d/2 \rfloor + 1$. No analogous closed form is known in the non-regular case. :: Computing $k^*$ requires evaluating $\rho(\mathbf{D}_k \mathbf{B}_k)$ across $k$ via power iteration on the sparse matrix $\mathbf{D}_k \mathbf{B}_k$ (size $2|E(H_k)|$, cost $O(|E(H_k)|)$ per iteration). On the sparse networks of :ref:sec-empirical:: a single $k^*$ completes in well under a minute on a laptop. ### Deviation against the configuration-model null { :label: sec-deviation } The leading-order shape of $\log \rho_G(k)$ is explained by the degree sequence; the genuine rich-club coupling, beyond what degree-only statistics can detect (cf. the Kojaku-Masuda :cite:kojaku2018:: configuration-null impossibility), concentrates in the *residual* obtained by subtracting a degree-preserving null. Concretely, for a graph $G$ with degree sequence $\pi_G$, draw an ensemble $\{G^{(s)}\}_{s=1}^{S}$ of configuration-model graphs with the same degree sequence and average their log-profiles\: $$ \overline{\log \rho_{\mathrm{CM}}}(k) \;=\; \frac{1}{S}\sum_{s=1}^{S} \log \rho\bigl(\mathbf{D}_k \mathbf{B}_k(G^{(s)})\bigr), $$ and define the *deviation profile* of $G$ as $$ \delta_G(k) \;=\; \log \rho_G(k) \;-\; \overline{\log \rho_{\mathrm{CM}}}(k). $$ By construction $\delta_G(k) = 0$ when $G$ is itself a typical configuration-model sample, and $\delta_G(k) > 0$ when $G$ packs cycles onto its high-degree vertices more densely than degree-preserving randomization would predict. Averaging over $S = 1000$ samples per graph (:ref:fig-deviation::), the null behaves correctly\: regular-graph and bimodal-CM controls sit at $\delta \approx 0$ across the whole support. Real networks separate cleanly\: C. elegans and macaque cortex climb to $\delta_G \approx 30\text{ to }40$ log-units of $\rho$ at high $k$, orders of magnitude above what the degree sequence alone predicts. :html: { :path: deviation_profiles.html :static: deviation_profiles.png :label: fig-deviation :alt: Deviation profiles against the configuration-model null } :caption: Real networks (green) and synthetic controls (other colors) compared against degree-preserving configuration-model nulls. Top\: raw entrenchment profile $\rho_G(k)$ per graph. Bottom\: deviation $\delta_G(k) = \log \rho_G(k) - \langle \log \rho_{\mathrm{CM}} \rangle(k)$ with $\pm 1\sigma$ band from $S = 1000$ CM samples; regular graphs and bimodal-CM controls sit at $\delta = 0$, while real networks (C. elegans, macaque cortex) climb to $\delta \approx 30\text{ to }40$ log-units of $\rho$ at high $k$, signalling rich-club coupling beyond what the degree sequence alone can explain. :: ## The blocking walk as a Markov chain { :label: sec-dynamics } The $k$-blocking walk defines a natural stochastic process on the cycle core. At each step, the walker at state $(u, v, W)$ transitions to a uniformly random valid state $(v, x, W')$, choosing a fresh blocked set at the target. The blocked set $W$ may block all exits to $V(H_k)$, forcing the walker to the periphery. The walk is *sub-stochastic* on the cycle core, with a sink representing expulsion from the core. ### The edge-projected chain { :label: sec-edge-chain } A key structural observation is that the blocked set has *no memory*\: at each step, $W$ is drawn fresh from the distribution of valid $(k-2)$-subsets of $N(x) \setminus \{v\}$, independently of the previous blocked set. The full chain on states $(u, v, W)$ therefore projects to a Markov chain on directed edges $(u, v)$ among nodes in $V(H_k)$. :theorem: { :label: thm-edge-chain :title: Edge-projected blocking walk } The $k$-blocking walk projected onto directed edges of $H_k$ is the sub-stochastic Markov chain $P_k$ on $\vec{E}(H_k)$ with transition probabilities $$ P_k\bigl((u,v) \to (v,x)\bigr) \;=\; \frac{1}{d(v) - 1}\;\mathbf{1}\!\bigl[x \in N(v) \cap V(H_k),\; x \neq u\bigr]. $$ Equivalently $P_k = \mathbf{\Delta}^{-1} \mathbf{B}_k$, where $\mathbf{\Delta}$ is the diagonal matrix on $\vec{E}(H_k)$ with $\mathbf{\Delta}[(u, v),(u,v)] = d(v) - 1$. This chain has the same size as $\mathbf{B}_k$ and the same non-trivial eigenvalues as the row-normalized full $k$-blocking walk on states. :: :proof: At state $(u, v, W)$ the walker picks the next vertex $x$ uniformly from $N(v) \setminus (W \cup \{u\})$, a set of size $d(v) - 1 - (k - 2)$. Marginalizing over $W$, distributed uniformly on $\binom{N(v) \setminus \{u\}}{k - 2}$, the marginal probability that the next vertex is a specific $x \in N(v) \setminus \{u\}$ is $$ \mathbb{E}_W\!\left[\frac{\mathbf{1}[x \notin W]}{d(v) - 1 - (k - 2)}\right] \;=\; \frac{1}{d(v) - 1 - (k - 2)} \cdot \frac{\binom{d(v) - 2}{k - 2}}{\binom{d(v) - 1}{k - 2}} \;=\; \frac{1}{d(v) - 1}, $$ using $\binom{d - 2}{k - 2} / \binom{d - 1}{k - 2} = (d - 1 - k + 2)/(d - 1)$. The walk transitions to $(v, x)$ iff $x \in V(H_k)$; otherwise the walker is expelled. Since $W$ is drawn fresh at each step, conditional on the current edge $(u, v)$ the marginal distribution of $W$ is uniform on $\binom{N(v) \setminus \{u\}}{k - 2}$, so the projection is Markovian (an exact lumping of the full chain over the blocked-set coordinate $W$), and the spectrum of $P_k$ coincides with the non-zero spectrum of the row-normalized full chain. :: The transition probability $1/(d(v) - 1)$ is independent of $k$\: the marginalization erases the role of the blocked set entirely, leaving only the standard non-backtracking random walk on $G$ restricted to $\vec{E}(H_k)$. Concretely, what $P_k$ drops is the per-step identity of the blocked set, an entropy of $\log\binom{d(v) - 1}{k - 2}$ bits per visit to $v$, but these modes mix to uniform in a single step under the fresh-redraw rule and so contribute nothing to the spectrum. All $k$-dependence in $P_k$ enters through the support truncation $V(H_k)$, i.e., through which edges are kept after the threshold, not through any per-step weighting. The row sum at edge $(u, v)$ is $$ \sum_x P_k\bigl((u, v) \to (v, x)\bigr) \;=\; \frac{m_v - 1}{d(v) - 1}, \qquad m_v = |N(v) \cap V(H_k)|, $$ and the deficit $1 - (m_v - 1)/(d(v) - 1) = (d(v) - m_v)/(d(v) - 1)$ is the per-step *expulsion probability*, the fraction of $v$'s neighbors that lie in the periphery $V \setminus V(H_k)$. When all peripheral neighbors of $v$ vanish (the regular case, or any vertex with no peripheral exits), the row sum is $1$ and the chain is locally stochastic at $v$. ### Persistence { :label: sec-persistence } The leading eigenvalue $\lambda_1(P_k) \leq 1$ of the sub-stochastic chain measures the per-step *survival rate* of the walk on the cycle core. The expected persistence time, the number of steps before expulsion, scales as $1 / (1 - \lambda_1)$. :itemize: :-: $\lambda_1 = 1$\: the walk is never expelled. At low $k$, the blocked set blocks few directions and exits to $V(H_k)$ are always available. :-: $0 < \lambda_1 < 1$\: the walk persists for a finite expected time. Higher $k$ blocks more directions, increasing the expulsion probability. :-: $\lambda_1 = 0$\: $H_k$ is empty and the walk dies immediately. :: Empirically, $\lambda_1(P_k)$ starts close to $1$ at low $k$ and decays monotonically toward $0$ as $k$ approaches the maximum degree (:ref:fig-rho-lambda::, solid orange curves). On C. elegans, $\lambda_1 \approx 0.998$ at $k = 3$ and $\lambda_1 = 0.12$ at $k^* = 44$; equivalently, the expected persistence time $1/(1 - \lambda_1)$ drops from roughly $460$ steps to $\approx 1.1$. At the optimal threshold the walk is ejected at essentially the next hop, even though $H_{k^*}$ remains non-empty. The functions $\rho(k)$ and $\lambda_1(k)$ are *dynamical counterparts* in the sense that both parametrize the cycle core by $k$ and both measure aspects of its capacity, but they read off different aspects. $\rho(k)$ tracks the degree-weighted cycle structure of $H_k$ through the binomial diagonal $\mathbf{D}_k$, while $\lambda_1(k)$ tracks how long a stochastic process can resist expulsion. Their shapes differ accordingly (:ref:fig-rho-lambda::), with $\rho$ rising, peaking at $k^*$, and crashing when $H_k$ empties, whereas $\lambda_1$ starts near $1$ and decays monotonically across the whole $k$-support, with no analogue of the peak. The mechanism behind that asymmetry is the explicit-$k$ binomial weight $\binom{d-2}{k-2}$ in $\mathbf{D}_k$, which lets degree concentration outrun support contraction at the sweet spot $k^*$; $P_k$ carries no such weighting, so raising $k$ can only shrink the cycle core and make the walker easier to expel. The optimal threshold is therefore a structural feature identified by $\rho$, not a dynamical optimum that the walker would independently pick out; on $H_{k^*}$ the walker barely persists because that is where the structural sweet spot sits, not because the dynamics endorse it. :html: { :path: rho_lambda_overlay.html :static: rho_lambda_overlay.png :label: fig-rho-lambda :alt: rho(k) and lambda_1(P_k) overlaid on shared k-axis for three networks } :caption: Survival rate $\lambda_1(P_k)$ (solid orange, right linear-axis) and entrenchment profile $\rho(k)$ (dashed navy, left log-axis) on three rich-club networks, plotted on the shared $k$-axis. Grey dotted line marks $k^*$. $\lambda_1(P_k)$ decays monotonically with no peak; $\rho(k)$ rises and peaks at $k^*$. :: :remark: { :label: rem-dirichlet :title: Dirichlet-Laplacian framing } $P_k = \mathbf{\Delta}^{-1} \mathbf{B}_k$ is the standard non-backtracking transition matrix on $G$ restricted to $\vec{E}(H_k)$, the *Dirichlet variant* of the non-backtracking random-walk Laplacian of Jost, Mulas, and Torres :cite:jost2023::, with the periphery $V \setminus V(H_k)$ as absorbing boundary. Its principal eigenvalue $\mu_1(\mathrm{Id} - P_k) = 1 - \lambda_1(P_k)$ is the corresponding Dirichlet spectral gap. :: ### Comparison with a degree-biased baseline { :label: sec-degree-baseline } A natural concern is that any walk biased toward high-degree neighbors should have a similar survival profile, since both blocking and the eventual expulsion past the boundary of $V(H_k)$ are sensitive to degree. This concern is sharpened by the analytical result of Kojaku and Masuda :cite:kojaku2018:: that a single core/periphery pair under the configuration-model null is impossible with only two blocks, so any core/periphery signal beyond degree heterogeneity must come from non-degree structure (see :ref:sec-position:: for the full conversation). We control for this with a baseline\: the *degree-proportional* non-backtracking walk that on the same state space (directed edges of $H_k$) makes the transition $$ {:label: eq-degree-baseline} P^{\mathrm{deg}}_k \!\big((u, v) \to (v, x)\big) = \mathbf{1}[x \in V(H_k)]\,\frac{d(x)}{\sum_{w \in N(v) \setminus \{u\}} d(w)}, $$ with the remaining mass leaking to an absorbing state. This baseline encodes pure degree bias\: it favors moving toward the high-degree core but ignores cycle structure entirely. :ref:fig-lambda1:: overlays $\lambda_1(P^{\mathrm{deg}}_k)$ (dashed) on the blocking-walk profile. On heterogeneous networks the two profiles separate substantially in the rich-club regime, for example, on C. elegans at $k^* = 44$ the unbiased blocking walk has $\lambda_1 = 0.12$ versus $\lambda_1^{\mathrm{deg}} = 0.35$ for the degree-biased baseline (gap $-0.23$). The degree-proportional walk persists longer because biasing transitions toward high-degree (interior) neighbours suppresses peripheral expulsion; the blocking walk picks uniformly and is thus more exposed to expulsion. The size of the gap $|\lambda_1 - \lambda_1^{\mathrm{deg}}|$, which ranges 0.05 to 0.32 across the fifteen networks, measures how strongly degree heterogeneity inflates persistence relative to an unbiased non-backtracking baseline; it is large precisely on the heterogeneous networks where degree heterogeneity is itself large. More fundamentally, the two walks differ in what they encode about $G$. The degree-proportional walk's transition probabilities are a fixed function of neighbour degrees, so $\lambda_1^{\mathrm{deg}}$ reports back only the degree sequence we put in; the blocking walk picks uniformly on valid exits, biasing the dynamics in no direction, so $\lambda_1$ is sensitive to the joint structure of degree and cycle topology as they appear in $G$. :html: { :path: lambda1_profiles.html :static: lambda1_profiles.png :label: fig-lambda1 :alt: Survival rate profiles on four rich-club networks } :caption: Survival rate $\lambda_1(P_k)$ vs $k$ on four rich-club ground-truth networks; solid: blocking walk, dashed: degree-proportional baseline of :ref:eq-degree-baseline::. :: ## Empirical validation { :label: sec-empirical } We evaluate $k^*$ and $H_{k^*}$ on fifteen networks for which the literature has identified a core or core-periphery structure under one of three frameworks\: the *rich club* (degree-thresholded density assessed against degree-preserving randomizations), the *$k$-core decomposition* (iterative degree peeling), and the *Borgatti-Everett block model* (idealized core/periphery fit to the adjacency matrix). Five networks per framework, with published claims, are compared against $H_{k^*}$ under that framework's lens; full dataset sources and sizes are listed in :ref:sec-appendix-datasets::. For each network we compute $\rho(k)$ in log-space, take $k^* = \operatorname{argmax}_k \rho(k)$, extract $H_{k^*}$ and the survival rate $\lambda_1(P_{k^*})$, and contrast $H_{k^*}$ against three coreness-ranking baselines\: $k$-core coreness :cite:seidman1983::, Della Rossa et al. $\alpha$-coreness :cite:dellarossa2013::, and the magnitude of the bottom Fiedler eigenvector of the random-walk Laplacian (Cucuringu et al. Lap-Core, :cite:cucuringu2016::). Della Rossa $\alpha$-coreness is skipped for $|V| > 10{,}000$ due to its $O(nm \log n)$ cost. Per-network values for every metric across all fifteen networks appear in :ref:fig-empirical::; the three framework-specific subsections below walk through what those numbers mean for each ground-truth claim. :html: { :path: empirical_scatter.html :static: empirical_scatter.png :label: fig-empirical :alt: per-network metric values across all fifteen networks } :caption: Per-network cycle-core analysis on the fifteen empirical networks. Each panel shows one metric on the $x$-axis-indexed network roster. Rich-club (blue): 1 C. elegans, 2 Macaque, 3 Internet AS, 4 Airport, 5 ca-AstroPh. $k$-core (red): 6 Facebook, 7 Yeast PPI, 8 Email Enron, 9 cit-HepTh, 10 Gnutella. Borgatti-Everett (green): 11 Karate, 12 Krackhardt, 13 Davis SW, 14 Pol. books, 15 NetScience. Click any framework marker in the legend at the top to toggle that framework's points on or off, useful for isolating a single group across all panels. $k^*$ and $|H_{k^*}|$ use log scale; the remaining panels are linear. Two $k$-core overlap measures are shown: "k-shell top-$|H|$ ov" is the size-matched overlap between $H_{k^*}$ and the top $|H_{k^*}|$ vertices by coreness, while "k-shell max-shell ov" is the overlap between $H_{k^*}$ and the entire maximum $k$-shell (which is often larger than $H_{k^*}$). "DR ov" and "Lap ov" are size-matched overlaps with the top $|H_{k^*}|$ vertices ranked by Della Rossa $\alpha$-coreness and Lap-Core, respectively (missing dots: Della Rossa skipped for $|V| > 10{,}000$). :: ### Results: rich-club ground truth { :label: sec-results-richclub :nonum: } *C. elegans.* $k^* = 44$ falls within the Towlson rich-club regime $[35, 73]$. $H_{k^*}$ contains exactly the eleven ground-truth command interneurons\: AVAL, AVAR, AVBL, AVBR, AVDL, AVDR, AVEL, AVER, DVA, PVCL, PVCR. No null model or parameter tuning was used. The maximum $k$-shell ranking contains all eleven (the max-shell is a superset, with 82 nodes), and Della Rossa $\alpha$-coreness picks ten of the same eleven (0.91); two completely different mechanisms (sub-stochastic blocking walk vs greedy persistence-probability ranking) converge on nearly identical answers. *Macaque cortex.* $H_{k^* = 41}$ contains 26 areas, including all 12 to 15 Harriger et al. identified. Della Rossa $\alpha$-coreness agrees on 25 of the 26 (96%); the 25 are a subset of a maximum $k$-shell of comparable size. $|\Delta\lambda_1| = 0.28$. *Internet AS 2000.* $H_{k^* = 389}$ is exactly four ASes, all canonical tier-1 providers\: AS701 (UUNET/MCI), AS1239 (Sprint), AS3561 (MCI Internet), AS7018 (AT\&T). $\phi(k^*) = 1.00$\: $H_{k^*}$ is the complete four-vertex subgraph, the sharpest possible rich-club signal. Della Rossa picks the same four as its top four (1.00); the maximum $k$-shell contains all four. *Airport.* $H_{k^* = 123}$ has 45 airports; the top ten by degree (AMS, FRA, CDG, IST, ATL, PEK, ORD, MUC, DME, DXB) are all canonical mega-hubs. The published "20 to 30" is below 45, but Della Rossa independently selects 40 of the same 45 (0.89) and $|\Delta\lambda_1| = 0.27$, so the broader core is structural rather than spurious. Multi-core interpretation discussed in :ref:sec-discussion::. *ca-AstroPh.* $H_{k^* = 191}$ has 69 authors, within the 50 to 100 range McAuley et al. predicted, with $|\Delta\lambda_1| = 0.20$ comparable to the rest of the panel; Della Rossa $\alpha$-coreness is not computable at this network size. The maximum $k$-shell contains seven of the 69 (0.10)\: degree-based $k$-shell ranking and $H_{k^*}$ select substantively different authors. ### Results: k-core ground truth { :label: sec-results-kcore :nonum: } The cycle core may be tighter or broader than the maximum $k$-shell, but it is chosen by spectral peak rather than peeling depth, and the two select substantively different sets. On the Internet AS the four tier-1 providers in $H_{k^*}$ form a strict subset of the 21-node max-shell. On yeast PPI, $H_{k^*}$ contains only 15 nodes against a 130-node max-shell, sharing just one member. On cit-HepTh the 16-node $H_{k^*}$ shares no nodes at all with the 52-node max-shell, yet contains the famous AdS/CFT and string-theory landmarks (Maldacena's hep-th/9711200 leads with 2468 citations). P2P Gnutella reverses the pattern\: a 1176-node $H_{k^*}$ broader than the 1004-node max-shell, with the two overlapping on 44% of $H_{k^*}$. Iterative peeling selects topically tight clusters; cycle-core selects the cyclically interconnected high-citation papers regardless of topic. ### Results: Borgatti-Everett ground truth { :label: sec-results-be :nonum: } On the small canonical networks the cycle core agrees with classical block-model analysis\: karate's $H_{k^*}$ is the seven-node inner clique anchored by the instructor (vertex $0$) and the president ($33$), with both the $k$-shell and Della Rossa rankings recovering five of seven members; Krackhardt's $H_{k^*}$ overlaps the Borgatti-Everett advice core $\{2, 14, 18, 1, 5, 21\}$ on five of eight members. Political books $H_{k^*}$ is a tight three-book bridging core, with one of three in the maximum $k$-shell (0.33), and Della Rossa selecting two of three (0.67). Network science is the outlier where the baselines largely disagree with $H_{k^*}$\: top-$|H|$ overlaps with $k$-shell and Lap-Core are both $0.00$, and Della Rossa selects one of four (0.25). The four nodes of $H_{k^*}$ form a small cyclic component among modest-degree vertices that all three baselines, ranking by degree or by Laplacian centrality, simply ignore; this is consistent with the tree-like global structure Holme :cite:holme2005:: noted. ### Baseline comparison summary { :label: sec-baseline-summary } :html: { :path: baselines_summary.html :static: baselines_summary.png :label: fig-baselines :alt: Δλ_1 and baseline overlap across all 15 networks } :caption: (A) $|\Delta\lambda_1| = |\lambda_1(P_k) - \lambda_1(P_k^{\mathrm{deg}})|$ at $k^*$ across all fifteen networks, coloured by framework. (B) Overlap with $H_{k^*}$ for three baseline coreness rankings ($k$-shell, Della Rossa $\alpha$-coreness, Lap-Core). :: Three patterns are visible in :ref:fig-baselines::. First, $|\Delta\lambda_1|$ ranges from $0.06$ on the large networks with the most heterogeneous degree distributions (Facebook ego, yeast PPI, cit-HepTh) up to $0.28$ on rich-club connectomes (macaque cortex)\: the gap is largest where a few high-degree vertices dominate the cycle structure, and the unbiased non-backtracking walk is most exposed there to peripheral expulsion. Second, Della Rossa $\alpha$-coreness is the most aligned baseline, selecting $0.89$ to $1.00$ of $H_{k^*}$ on every rich-club network where it is computable; two mechanisms, sub-stochastic blocking walks and greedy persistence-probability rankings, converge on nearly the same nodes, indicating $H_{k^*}$ identifies a structurally meaningful set rather than an artifact of the operator construction. Third, Lap-Core overlaps zero with $H_{k^*}$ on every large network with a heterogeneous degree distribution (Airport, ca-AstroPh, Internet AS, Facebook ego, yeast PPI, Email Enron, cit-HepTh, polbooks, NetScience), and is nontrivial only on the small dense Borgatti-Everett graphs (Davis 0.69, Krackhardt 0.38, Karate 0.29) and the moderate-size connectomes (macaque 0.38, C. elegans 0.27)\: the Fiedler vector tracks graph cuts and community structure rather than core-periphery, aligning with $H_{k^*}$ only when the core-periphery split is also a community split, consistent with :cite:cucuringu2016::. The maximum $k$-shell is a superset of $H_{k^*}$ on rich-club networks (overlap $1.00$ on C. elegans, AS, macaque) but $1.5\times$ to $7.5\times$ larger\: cycle core is the more selective characterization. ### Temporal evolution of the internet core { :label: sec-temporal } The SNAP AS dataset provides 733 daily snapshots of the Internet topology from November 1997 to December 1999, a period of explosive growth (3000 to 6000 ASes) that includes the MCI WorldCom merger (September 1998), the largest telecom merger of the era. We sample biweekly and apply the cycle-core pipeline at a range of degree thresholds rather than only at $k^*$, which lets us probe the shell structure surrounding the rich club. *Layered shell structure.* The AS topology is not "small rich club + flat periphery". On the 2000-01-02 snapshot the cycle-core profile reveals a graded hierarchy\: $|H_{k=500}| = 3$ (a complete triangle, $\phi=1$), $|H_{k=200}| = 8$ (still a near-clique), $|H_{k=100}| = 20$ (denser-than-random core of the second tier), $|H_{k=50}| = 45$, and $|H_{k=5}| \approx 600$, with the remaining $\approx 5800$ ASes forming the loose periphery. Each shell is denser than the next outward (the rich-club coefficient $\phi(k)$ drops from $1.0$ at the centre to $\approx 0.013$ at the boundary), but the structure is graded, not bimodal. The same shell hierarchy appears qualitatively on every snapshot we sampled (:ref:fig-as-shells::). :html: { :path: as_temporal_shells.html :static: as_temporal_shells.png :label: fig-as-shells :alt: cycle-core shell sizes at four k thresholds across the 1997-2000 AS time series } :caption: Cycle-core sizes $|H_k|$ at four thresholds $k \in \{50, 100, 200, 300\}$ on biweekly Internet AS snapshots, November 1997 to November 1999. The inner core ($k=300$, dark) is identity-stable at three ASes for almost the entire record; the intermediate shells ($k=200, 100, 50$) grow approximately monotonically as the network doubles in size. :: *Stability and accretive growth.* The inner core is identity-stable: at $k=300$, $|H_k|$ is exactly the three-AS triangle $\{$AS701 (UUNET), AS3561 (MCI), AS1239 (Sprint)$\}$, the original tier-1 backbone providers, for almost the entire two-year record, expanding to five only in the last few months (adding AS1 and AS7018, the latter being AT&T's IP backbone). The intermediate shells grow but *accretively*\: the 1997 $H_{k=100}$ set of seven ASes is a strict subset of the 2000 $H_{k=100}$ set of twenty. No founder is displaced. While $|V|$ doubles (3015 to 6474), $|H_{k=100}|$ nearly triples (7 to 20)\: the rich-club expansion is real, just gated behind the inner three. The threshold $k^*$ drifts smoothly upward from $\approx 234$ in November 1997 to $\approx 395$ in November 1999 as the tier-1 providers accumulate degree. *Structural vs. temporal nestedness.* On any single snapshot the family $\{H_k\}_k$ is monotonically nested by construction\: $H_k \subseteq H_{k-1}$ for all $k$, because raising the threshold can only remove vertices. That nestedness is automatic. What is *not* automatic, and what the AS data shows empirically, is a second kind of nestedness across time at fixed $k$\: $H_k(G_{t_1}) \subseteq H_k(G_{t_2})$ for $t_1 < t_2$, on every threshold we tested. Founders are retained, new members are accreted. This is a structural property of the underlying social/economic process rather than a corollary of the definition. A counterfactual makes the distinction concrete. Consider a co-authorship network spanning a paradigm shift, or an online social platform where accounts churn through bans and retirements. In such a network the high-degree set rotates\: yesterday's central authors stop publishing, today's central accounts did not exist last year. For each snapshot the cycle-core filtration still satisfies $H_k \subseteq H_{k-1}$, but $H_k(G_{t_1}) \not\subseteq H_k(G_{t_2})$, and the temporal subset relation fails. The Internet AS topology happens to be a domain where the high-degree set is anchored by long-lived infrastructure (physical fiber backbones, regulatory inertia), so temporal nestedness holds; in other domains it would not. *Comparison with the baselines.* Only the cycle core delivers temporal nestedness on the AS data (:ref:fig-as-accretion::). For each method, computing its top-$N$ set in 1997 and again in 2000, the fraction of the 1997 set that survives in 2000 is\: cycle core $100\%$ at top-3 and top-7, $90\%$ at top-20; $k$-shell coreness $100\%$ at top-3 but $86\%$ at top-7 and $60\%$ at top-20; Della Rossa $\alpha$-coreness $100/71/45\%$; Lap-Core $0\%$ across all $N$ (the Fiedler vector picks completely different nodes as the network grows, presumably because it tracks community boundaries that shift with the periphery). The ranking-based baselines lack structural nestedness even within a single snapshot, because top-$N$ by a scalar ranking is not a nested family as $N$ varies\: a node ranked slightly outside the top-$N$ can climb in at the next snapshot and displace a previous member. The cycle core inherits temporal stability *because* its filtration is structurally nested. :html: { :path: as_accretion.html :static: as_accretion.png :label: fig-as-accretion :alt: per-method temporal stability of the top-N core on the AS time series } :caption: 1997 top-$N$ set survival in 2000, by method, for $N \in \{3, 7, 20\}$ on the Internet AS time series. Cycle core preserves $\geq 90\%$ at every $N$ because its filtration is structurally nested; $k$-shell coreness and Della Rossa $\alpha$-coreness lose identity as $N$ grows (top-$N$ by a scalar ranking is not a nested family); Lap-Core fails entirely (the Fiedler vector reorders under network growth). :: ## Discussion { :label: sec-discussion } We introduced the $k$-blocking operator $\mathbf{M}_k$ and two objects derived from it\: the structural invariant $\mathbf{D}_k \mathbf{B}_k$, which captures the interaction between cycle structure and degree heterogeneity via the trace formula, and the edge-projected $k$-blocking walk chain $P_k$, a sub-stochastic Markov chain whose leading eigenvalue $\lambda_1$ measures persistence on the cycle core. The optimal threshold $k^*$ gives an intrinsic characterization of the rich club, and $\lambda_1(k)$ gives a dynamical measure of entrenchment, i.e., how long a process can resist expulsion from the core. The temporal analysis of the Internet core demonstrates that these tools can track structural evolution in ways invisible to density-based measures. ### Position within the core-periphery literature { :label: sec-position } The literature on core-periphery structure has converged from several distinct directions, and the cycle core sits in identifiable conversation with each. Three foundational frameworks frame essentially every modern analysis. Borgatti and Everett :cite:borgatti2000:: introduced the discrete block-model formulation in which the core forms a clique, the periphery is empty, and core-periphery edges are dense; a network is then assessed by how closely its adjacency matrix matches this idealized pattern. The rich-club coefficient :cite:colizza2006,zhou2004,opsahl2008:: focuses on density among the highest-degree subgraph, with significance assessed against degree-preserving randomizations. The $k$-core decomposition :cite:seidman1983,batagelj2003,dorogovtsev2006:: instead peels the network from the outside in, producing a layered structure of nested shells. The cycle core is degree-thresholded like rich-club, layered like $k$-core, but instead of measuring density on $V_k$ it asks which vertices in $V_k$ participate in cycles among themselves. The intrinsic threshold $k^*$ collapses the layered filtration into a single binary cut at the spectral peak, recovering a Borgatti-Everett-style core/periphery split without an external null model or block fit. The closest dynamical relative is the work of Della Rossa, Dercole and Piccardi, who define a *persistence probability* $\alpha_S$, the chance that a standard random walker currently in $S$ remains in $S$ at the next step, and grow $S$ greedily from the least-connected node to obtain a non-decreasing core-periphery profile :cite:dellarossa2013::. The blocking walk shares this random-walk epistemology but differs in three ways\: the chain is *sub-stochastic*, with mass leaving to the periphery rather than being conserved over a growing set; we measure the *asymptotic* survival rate $\lambda_1(P_k)$ rather than the one-step persistence; and the walk is non-backtracking and degree-aware via the blocking constraint, so it sees cycle structure that a uniform walker cannot. A second lineage applies spectral methods to the adjacency or Laplacian matrix. Cucuringu, Rombach, Lee and Porter :cite:cucuringu2016:: introduce a family of detectors, Path-Core (an edge-leave-one-out betweenness), LowRank-Core (a low-rank approximation of $\mathbf{A}$), and Lap-Core (the bottom eigenvector of the random-walk Laplacian), together with an objective for choosing the core/periphery cut. Our spectral object is the *non-backtracking* matrix family $\mathbf{D}_k \mathbf{B}_k$ rather than $\mathbf{A}$ or its Laplacian, and where their methods produce a coreness score from a single matrix, ours produces a parametric family indexed by $k$ whose spectral radius peaks at the intrinsic threshold; the family itself is the diagnostic. A third lineage casts core-periphery as model fitting and carries an important warning for any method built on top of high-degree vertices. Zhang, Martin and Newman :cite:zhang2015:: fit a two-block stochastic block model with $p_{11} > p_{12} > p_{22}$ via expectation-maximization with belief propagation, and show that in the weak-structure limit the optimal classifier reduces to a degree-based split. Kojaku and Masuda :cite:kojaku2018:: sharpen this conclusion under the configuration-model null\: a single core/periphery pair is *mathematically impossible* with only two blocks, because excess intra-core edges relative to the configuration model force equal excess intra-periphery edges. Real core-periphery beyond degree heterogeneity therefore requires at least three blocks, a core, a periphery, and a third structure such as a community or another core/periphery pair. This analytical result is what the degree-proportional baseline in :ref:fig-lambda1:: is designed to address\: $P_k^{\mathrm{deg}}$ shares its state space with $P_k$, so the gap $\lambda_1(P_k) - \lambda_1(P_k^{\mathrm{deg}})$ measures structure beyond degree by construction, and the substantial gap on networks with heterogeneous degree distributions (e.g., $|\Delta\lambda_1| = 0.23$ at $k^* = 44$ on the C. elegans connectome) is the empirical counterpart of the inference-theoretic conclusion. Two further studies clarify what kind of object one should expect to find. Kojaku and Masuda :cite:kojaku2017:: argue that real networks often contain *multiple* non-overlapping core-periphery pairs and detect them by maximizing a Borgatti-Everett- style quality function across a partition; the worldwide airport network alone yields ten geographically clustered pairs in their analysis, suggesting that the broadness of $H_{k^*}$ on the airport network may reflect aggregation of multiple regional cores rather than a single overgrown one. Gallagher, Young, and Foucault Welles :cite:gallagher2020:: distinguish between two qualitative types of core-periphery\: *hub-and-spoke* (a single binary core/periphery, in the Borgatti-Everett style) and *layered* (a nested shell decomposition, in the $k$-core style), and show via Bayesian model selection that real networks are distributed across both types, with neither universal. The cycle core sits naturally inside this typology\: the filtration $H_1 \supseteq H_2 \supseteq \cdots$ is a layered structure indexed by global degree threshold, while the cut at $k^*$ delivers a hub-and-spoke core $H_{k^*}$ on demand, so a single object yields either characterization depending on whether one queries the filtration or the spectral peak. For an extended taxonomy of core-periphery flavors, including the bow-tie structure of directed networks, the onion topology of attack-resistant networks, and the nestedness measures developed in ecology, we refer to the survey by Csermely, London, Wu, and Uzzi :cite:csermely2013::. ### Variants and refinements of the rich-club coefficient { :label: sec-rc-variants } The original rich-club coefficient :cite:colizza2006,zhou2004:: has accumulated a substantial refinement literature in the two decades since, along five broad axes. *Weighted* variants replace edge counts with weights and quantify how concentrated the high-strength subgraph is (Opsahl et al. :cite:opsahl2008::, Serrano :cite:serrano2008::). *Significance refinements* sharpen the configuration-model null, either analytically (Jiang and Zhou :cite:jiang2008::), via generative samplers (Mondragón and Zhou :cite:mondragon2012::; Mondragón :cite:mondragon2014::), or as joint hypothesis tests over a $k$-range (Muscoloni and Cannistraci :cite:muscoloni2017::). *Alternative quantities at the same threshold* change what counts as "rich", for instance by demanding connectedness (Ma and Mondragón :cite:mamondragon2015::) or generalising beyond degree to arbitrary node attributes (Cinelli :cite:cinelli2019generalized,cinelli2017dyadic::). *Hierarchical and temporal extensions* probe rich-club ordering across coarse-grained scales (McAuley et al. :cite:mcauley2007rich::) and along time-respecting paths (Pedreschi et al. :cite:pedreschi2022::). *Higher-order generalisations* lift the analysis from pairwise edges to hyperedges (Nakajima et al. :cite:nakajima2023::). The cycle core occupies a different position within this lineage. Rather than modifying the *quantity* measured at threshold $k$, it modifies the *condition* under which a vertex is retained at threshold $k$\: vertices must not only have degree at least $k$ but also participate in cycles among other vertices satisfying the same condition. The intrinsic threshold $k^*$ from the spectral peak of $\mathbf{M}_k$ further dispenses with the threshold-selection problem common to all $\phi(k)$ variants. These two structural moves are independent of the extensions above and could in principle be combined with any of them. ### Open directions { :label: sec-open } *Multi-core extension.* The cycle core delivers a single $H_{k^*}$. Networks with genuinely multiple regional cores, such as the worldwide airport network or political blog networks :cite:kojaku2017::, may be better described by a partition into several local cycle cores, each with its own degree threshold. An extension of $k^*$ that detects and ranks multiple peaks in $\rho(\mathbf{D}_k \mathbf{B}_k)$ restricted to subgraphs is a natural next step. *Analytical null model.* The deviation profile $\delta_G(k)$ uses an empirical configuration-model average; an analytical expression for $\mathbb{E}[\rho(\mathbf{D}_k \mathbf{B}_k)]$ over the configuration-model ensemble, accessible in principle via the trace formula and known asymptotic results for non-backtracking walk counts on configuration-model graphs, would replace the Monte-Carlo baseline by a closed form and yield a principled normalization for cross-network comparison. *Phase transitions.* The entrenchment profile's rise-peak-crash pattern should exhibit a sharp transition for random graph ensembles, connecting to the $k$-core phase transition literature. *Weighted and directed networks.* The $k$-blocking walk generalizes naturally to weighted graphs (where edge weights modulate transition probabilities) and directed graphs (where the non-backtracking constraint becomes asymmetric). Rich-club analysis in brain connectomes often involves weighted and directed synaptic data. ## Appendix\: Datasets { :label: sec-appendix-datasets } Fifteen networks across three frameworks; each entry gives sizes, source citation, and the published claim against which we compare $H_{k^*}$. *Rich-club ground truth (5 networks).* :itemize: :-: *C. elegans connectome*, 279 neurons, 2287 synaptic connections :cite:varshney2011::. Towlson et al. :cite:towlson2013:: identified eleven command interneurons via $\phi(k)/\phi_{\text{rand}}(k) \geq 3$ over the regime $35 \leq k \leq 73$. :-: *Macaque cerebral cortex*, 91 areas, 1401 connections (Network Repository compilation of tracer data). Harriger et al. :cite:harriger2012:: identified 12 to 15 cortical areas (prefrontal, parietal, premotor, cingulate). :-: *Internet autonomous-system topology*, 6474 ASes, 12572 peering links (January 2000 snapshot, :cite:leskovec2005::). Zhou and Mondragón :cite:zhou2004:: identified $\sim$12 tier-1 providers via $\phi(k)$ analysis and BGP classification. :-: *Global airport network*, 3397 airports, 19230 routes (OpenFlights). Colizza et al. :cite:colizza2006:: and Guimerà et al. :cite:guimera2005:: identified the top 20 to 30 airports. :-: *Astrophysics coauthorship*, 17903 authors, 196972 collaborations :cite:leskovec2007evolution::. McAuley et al. :cite:mcauley2007rich:: report a pronounced rich-club among 50 to 100 highly-cited authors. :: *$k$-core ground truth (6 networks; Internet AS appears in both groups).* :itemize: :-: *Internet autonomous-system topology*, as above; Carmi et al. :cite:carmi2007:: applied $k$-shell decomposition to characterize the Internet's "medusa" three-layer structure. :-: *Facebook ego networks*, 4039 users, 88234 edges (McAuley-Leskovec). :-: *Human protein-protein interactions*, 21521 proteins, 338625 interactions :cite:agrawal2018::, in the spirit of Wuchty and Almaas :cite:wuchty2005:: who peeled the yeast PPI. :-: *Enron email network*, 33696 accounts, 180811 messages. Kitsak et al. :cite:kitsak2010:: identified the innermost $k$-shell as the most influential spreaders. :-: *High-energy-physics theory citations*, 27400 papers, 352021 citations. Eom and Fortunato :cite:eom2011:: showed highly-cited papers concentrate in the deepest $k$-shells. :-: *Gnutella P2P snapshot*, 62561 nodes, 147878 edges, where Carmi et al. :cite:carmi2007:: identified a "medusa" three-layer $k$-shell decomposition. :: *Borgatti-Everett ground truth (5 networks).* :itemize: :-: *Zachary's karate club*, 34 nodes, 78 edges; the canonical small social-network test case :cite:zachary1977::. :-: *Krackhardt's high-tech managers* advice network, 21 nodes :cite:borgatti1999::. :-: *Davis Southern Women* event-attendance bipartite, 32 nodes :cite:borgatti2000::. :-: *Political books* network of Krebs, 81 nodes, 91 edges (largest connected component); core-periphery analysis by Holme :cite:holme2005::, community structure by Newman :cite:newman2006finding::. :-: *Network-science coauthorship* of Newman, 915 nodes, 952 edges (largest connected component); analyzed in :cite:holme2005::. :: :references: @article{zachary1977, title={An information flow model for conflict and fission in small groups}, author={Zachary, W. W.}, year={1977}, journal={Journal of Anthropological Research}, volume={33}, number={4}, pages={452--473}, } @article{colizza2006, title={Detecting rich-club ordering in complex networks}, author={Colizza, V. and Flammini, A. and Serrano, M. A. and Vespignani, A.}, year={2006}, journal={Nature Physics}, volume={2}, doi={https://doi.org/10.1038/nphys209}, } @article{krzakala2013, title={Spectral redemption in clustering sparse networks}, author={Krzakala, F. and Moore, C. and Mossel, E. and Neeman, J. and Sly, A. and Zdeborová, L. and Zhang, P.}, year={2013}, journal={Proceedings of the National Academy of Sciences}, volume={110}, doi={https://doi.org/10.1073/pnas.1312486110}, } @article{vandenheuvel2011, title={Rich-club organization of the human connectome}, author={van den Heuvel, M. P. and Sporns, O.}, year={2011}, journal={Journal of Neuroscience}, volume={31}, doi={https://doi.org/10.1523/JNEUROSCI.3539-11.2011}, } @article{varshney2011, title={Structural properties of the Caenorhabditis elegans neuronal network}, author={Varshney, L. R. and Chen, B. L. and Paniagua, E. and Hall, D. H. and Chklovskii, D. B.}, year={2011}, journal={PLoS Computational Biology}, volume={7}, doi={https://doi.org/10.1371/journal.pcbi.1001066}, } @article{towlson2013, title={The rich club of the C. elegans neuronal connectome}, author={Towlson, E. K. and Vertes, P. E. and Ahnert, S. E. and Schafer, W. R. and Bullmore, E. T.}, year={2013}, journal={Journal of Neuroscience}, volume={33}, doi={https://doi.org/10.1523/JNEUROSCI.3784-12.2013}, } @article{zhou2004, title={The rich-club phenomenon in the Internet topology}, author={Zhou, S. and Mondragón, R. J.}, year={2004}, journal={IEEE Communications Letters}, volume={8}, doi={https://doi.org/10.1109/LCOMM.2004.823426}, } @article{leskovec2005, title={Graphs over time\: densification laws, shrinking diameters and possible explanations}, author={Leskovec, J. and Kleinberg, J. and Faloutsos, C.}, year={2005}, journal={ACM SIGKDD}, doi={https://doi.org/10.1145/1081870.1081893}, } @article{guimera2005, title={The worldwide air transportation network\: anomalous centrality, community structure, and cities' global roles}, author={Guimerà, R. and Mossa, S. and Turtschi, A. and Amaral, L. A. N.}, year={2005}, journal={Proceedings of the National Academy of Sciences}, volume={102}, doi={https://doi.org/10.1073/pnas.0407994102}, } @article{opsahl2008, title={Prominence and control\: the weighted rich-club effect}, author={Opsahl, T. and Colizza, V. and Panzarasa, P. and Ramasco, J. J.}, year={2008}, journal={Physical Review Letters}, volume={101}, doi={https://doi.org/10.1103/PhysRevLett.101.168702}, } @article{seidman1983, title={Network structure and minimum degree}, author={Seidman, S. B.}, year={1983}, journal={Social Networks}, volume={5}, doi={https://doi.org/10.1016/0378-8733(83)90028-X}, } @article{batagelj2003, title={An O(m) algorithm for cores decomposition of networks}, author={Batagelj, V. and Zaversnik, M.}, year={2003}, journal={arXiv preprint cs/0310049}, doi={https://doi.org/10.48550/arXiv.cs/0310049}, } @article{dorogovtsev2006, title={k-core organization of complex networks}, author={Dorogovtsev, S. N. and Goltsev, A. V. and Mendes, J. F. F.}, year={2006}, journal={Physical Review Letters}, volume={96}, doi={https://doi.org/10.1103/PhysRevLett.96.040601}, } @article{borgatti2000, title={Models of core/periphery structures}, author={Borgatti, S. P. and Everett, M. G.}, year={2000}, journal={Social Networks}, volume={21}, doi={https://doi.org/10.1016/S0378-8733(99)00019-2}, } @article{dellarossa2013, title={Profiling core-periphery network structure by random walkers}, author={Della Rossa, F. and Dercole, F. and Piccardi, C.}, year={2013}, journal={Scientific Reports}, volume={3}, doi={https://doi.org/10.1038/srep01467}, } @article{zhang2015, title={Identification of core-periphery structure in networks}, author={Zhang, X. and Martin, T. and Newman, M. E. J.}, year={2015}, journal={Physical Review E}, volume={91}, doi={https://doi.org/10.1103/PhysRevE.91.032803}, } @article{gallagher2020, title={A clarified typology of core-periphery structure in networks}, author={Gallagher, R. J. and Young, J.-G. and Foucault Welles, B.}, year={2021}, journal={Science Advances}, volume={7}, doi={https://doi.org/10.1126/sciadv.abc9800}, } @article{csermely2013, title={Structure and dynamics of core/periphery networks}, author={Csermely, P. and London, A. and Wu, L.-Y. and Uzzi, B.}, year={2013}, journal={Journal of Complex Networks}, volume={1}, doi={https://doi.org/10.1093/comnet/cnt016}, } @article{cucuringu2016, title={Detection of core-periphery structure in networks using spectral methods and geodesic paths}, author={Cucuringu, M. and Rombach, P. and Lee, S. H. and Porter, M. A.}, year={2016}, journal={European Journal of Applied Mathematics}, volume={27}, doi={https://doi.org/10.1017/S095679251600022X}, } @article{kojaku2017, title={Finding multiple core-periphery pairs in networks}, author={Kojaku, S. and Masuda, N.}, year={2017}, journal={Physical Review E}, volume={96}, doi={https://doi.org/10.1103/PhysRevE.96.052313}, } @article{kojaku2018, title={Core-periphery structure requires something else in the network}, author={Kojaku, S. and Masuda, N.}, year={2018}, journal={New Journal of Physics}, volume={20}, doi={https://doi.org/10.1088/1367-2630/aab547}, } @article{harriger2012, title={Rich club organization of macaque cerebral cortex and its role in network communication}, author={Harriger, L. and van den Heuvel, M. P. and Sporns, O.}, year={2012}, journal={PLoS One}, volume={7}, doi={https://doi.org/10.1371/journal.pone.0046497}, } @article{leskovec2007evolution, title={Graph evolution: densification and shrinking diameters}, author={Leskovec, J. and Kleinberg, J. and Faloutsos, C.}, year={2007}, journal={ACM Transactions on Knowledge Discovery from Data}, volume={1}, doi={https://doi.org/10.1145/1217299.1217301}, } @article{mcauley2007rich, title={Rich-club phenomenon across complex network hierarchies}, author={McAuley, J. J. and da Fontoura Costa, L. and Caetano, T. S.}, year={2007}, journal={Applied Physics Letters}, volume={91}, doi={https://doi.org/10.1063/1.2773951}, } @article{carmi2007, title={A model of Internet topology using k-shell decomposition}, author={Carmi, S. and Havlin, S. and Kirkpatrick, S. and Shavitt, Y. and Shir, E.}, year={2007}, journal={Proceedings of the National Academy of Sciences}, volume={104}, doi={https://doi.org/10.1073/pnas.0701175104}, } @article{wuchty2005, title={Peeling the yeast protein network}, author={Wuchty, S. and Almaas, E.}, year={2005}, journal={Proteomics}, volume={5}, doi={https://doi.org/10.1002/pmic.200400962}, } @inproceedings{agrawal2018, title={Large-scale analysis of disease pathways in the human interactome}, author={Agrawal, M. and Zitnik, M. and Leskovec, J.}, year={2018}, booktitle={Pacific Symposium on Biocomputing}, doi={https://doi.org/10.1142/9789813235533_0011}, } @article{kitsak2010, title={Identification of influential spreaders in complex networks}, author={Kitsak, M. and Gallos, L. K. and Havlin, S. and Liljeros, F. and Muchnik, L. and Stanley, H. E. and Makse, H. A.}, year={2010}, journal={Nature Physics}, volume={6}, doi={https://doi.org/10.1038/nphys1746}, } @article{eom2011, title={Characterizing and modeling citation dynamics}, author={Eom, Y.-H. and Fortunato, S.}, year={2011}, journal={PLoS One}, volume={6}, doi={https://doi.org/10.1371/journal.pone.0024926}, } @article{borgatti1999, title={Models of core/periphery structures}, author={Borgatti, S. P. and Everett, M. G.}, year={1999}, journal={Connections}, volume={21}, } @article{holme2005, title={Core-periphery organization of complex networks}, author={Holme, P.}, year={2005}, journal={Physical Review E}, volume={72}, doi={https://doi.org/10.1103/PhysRevE.72.046111}, } @article{newman2006finding, title={Finding community structure in networks using the eigenvectors of matrices}, author={Newman, M. E. J.}, year={2006}, journal={Physical Review E}, volume={74}, doi={https://doi.org/10.1103/PhysRevE.74.036104}, } @article{jost2023, title={Spectral theory of the non-backtracking Laplacian for graphs}, author={Jost, J. and Mulas, R. and Torres, L.}, year={2023}, journal={Discrete Mathematics}, volume={346}, number={10}, pages={113536}, doi={https://doi.org/10.1016/j.disc.2023.113536}, } @article{serrano2008, title={Rich-club vs rich-multipolarization phenomena in weighted networks}, author={Serrano, M. A.}, year={2008}, journal={Physical Review E}, volume={78}, doi={https://doi.org/10.1103/PhysRevE.78.026101}, } @article{jiang2008, title={Statistical significance of the rich-club phenomenon in complex networks}, author={Jiang, Z.-Q. and Zhou, W.-X.}, year={2008}, journal={New Journal of Physics}, volume={10}, doi={https://doi.org/10.1088/1367-2630/10/4/043002}, } @article{mondragon2012, title={Random networks with given rich-club coefficient}, author={Mondragón, R. J. and Zhou, S.}, year={2012}, journal={European Physical Journal B}, volume={85}, doi={https://doi.org/10.1140/epjb/e2012-21026-3}, } @article{mondragon2014, title={Network null-model based on maximal entropy and the rich-club}, author={Mondragón, R. J.}, year={2014}, journal={Journal of Complex Networks}, volume={2}, doi={https://doi.org/10.1093/comnet/cnu006}, } @article{muscoloni2017, title={Rich-clubness test\: how to determine whether a complex network has or doesn't have a rich-club}, author={Muscoloni, A. and Cannistraci, C. V.}, year={2017}, journal={arXiv preprint}, doi={https://doi.org/10.48550/arXiv.1704.03526}, } @article{mamondragon2015, title={Rich-cores in networks}, author={Ma, A. and Mondragón, R. J.}, year={2015}, journal={PLoS ONE}, volume={10}, doi={https://doi.org/10.1371/journal.pone.0119678}, } @article{cinelli2019generalized, title={Generalized rich-club ordering in networks}, author={Cinelli, M.}, year={2019}, journal={Journal of Complex Networks}, volume={7}, doi={https://doi.org/10.1093/comnet/cnz002}, } @article{cinelli2017dyadic, title={Rich-club ordering and the dyadic effect\: two interrelated phenomena}, author={Cinelli, M. and Ferraro, G. and Iovanella, A.}, year={2017}, journal={Physica A}, volume={490}, doi={https://doi.org/10.1016/j.physa.2017.08.122}, } @article{pedreschi2022, title={The temporal rich club phenomenon}, author={Pedreschi, N. and Battaglia, D. and Barrat, A.}, year={2022}, journal={Nature Physics}, volume={18}, doi={https://doi.org/10.1038/s41567-022-01634-8}, } @article{nakajima2023, title={Higher-order rich-club phenomenon in collaborative research grant networks}, author={Nakajima, K. and Shudo, K. and Masuda, N.}, year={2023}, journal={Scientometrics}, volume={128}, doi={https://doi.org/10.1007/s11192-022-04621-1}, } ::