Arxiv:2101.00414V1 [Physics.Soc-Ph] 2 Jan 2021 2

Dynamic Hidden-Variable Network Models

Harrison Hartle,1 Fragkiskos Papadopoulos,2 and Dmitri Krioukov1, 3 1Network Science Institute, Northeastern University, Boston, MA, USA 2Department of Electrical Engineering, Computer Engineering and Informatics, Cyprus University of Technology, 3036 Limassol, Cyprus 3Northeastern University, Departments of Physics, Mathematics, and Electrical&Computer Engineering, Boston, MA, USA Models of complex networks often incorporate node-intrinsic properties abstracted as hidden variables. The probability of connections in the network is then a function of these variables. Real- world networks evolve over time, and many exhibit dynamics of node characteristics as well as of linking structure. Here we introduce and study natural temporal extensions of static hidden-variable network models with stochastic dynamics of hidden variables and links. The rates of the hidden variable dynamics and link dynamics are controlled by two parameters, and snapshots of networks in the dynamic models may or may not be equivalent to a static model, depending on the location in the parameter phase diagram. We quantify deviations from static-like behavior, and examine the level of structural persistence in the considered models. We explore temporal versions of popular static models with community structure, latent geometry, and degree-heterogeneity. We do not attempt to directly model real networks, but comment on interesting qualitative resemblances, discussing possible extensions, generalizations, and applications.

I. INTRODUCTION without hidden variables. For example, hidden variables underly network models incorporating realistic fea- Networks are ubiquitous in nature [1–9], and their tures such as community structure (stochastic block mod- study relies heavily on the mathematical and compu- els [20]), latent geometry (random geometric graphs tational analysis of simple models [10, 11], typically in [21]), and degree-heterogeneity (soft configuration mod- the form of random networks built according to some els [22]). stochastic rules. In many models, nodes are assigned However, such models do not capture the dynamics characteristics (such as fitnesses [12, 13] or spatial co- of node-characteristics, nor the impact thereof on net- ordinates in a physical [14] or latent space [15–17]), work structure. The influence of dynamic node-states on which in turn affect the network’s structural forma- evolving link-structure has been investigated in the con- tion. Such models fall under the umbrella of hidden- text of adaptive networks [23–28], but in that case node- variables models [18], because they depend on internal states arise due to a highly complex feedback, interacting node-characteristics that are only implicitly expressed by with one another through co-evolving links. Such mod- the network structure, through effects on link-formation. els are more realistic and have interesting features, but Usually, hidden variables (HVs) are not externally spec- they do not directly explore the impact of dynamic node- ified as parameters – rather, their probability distribu- properties on dynamic network structure. tion is specified [12, 19], and they are sampled during There is a wide abundance of real-world examples of the network’s formation. Two sources of randomness un- dynamic node-properties influencing dynamics of net- derly such networks: the random HVs of nodes, and the work structure, such as: random formation of edges given those HVs. In general, a) changing habits, interests, jobs, and other at- hidden-variables models are defined by the following pro- tributes of people in social networks [29], cedure: 1. A random hidden-variable configuration H is b) changing geospatial coordinates of organisms dur- drawn with probability density ρ(H) from a set of ing formation of social ties, group-memberships, possible hidden-variable configurations H. and pathogenic contact networks [30–34], arXiv:2101.00414v1 [physics.soc-ph] 2 Jan 2021 2. Graph G is then drawn with conditional probability c) changing phenotypic traits of species as they bio- P(G|H) from a set of possible graphs G. logically evolve in ecological networks [35, 36], As a result, the overall probability of sampling any par- d) changing marketing and administrative strategies ticular graph G ∈ G is equal to of entities in economic networks [37, 38], Z e) changing demographic and infrastructural charac- P(G) = P(G|H)ρ(H)dH. (1) H teristics of cities in evolving highway and airport networks [39–41], Hidden-variables models, due to their capacity to en- code nodewise heterogeneity, are in many cases capa- f) changing gene-expression levels of neurons in devel- ble of exhibiting more structural realism than models oping connectomes [42, 43], 2

g) changing consumption-levels of residential nodes in graphs [21], soft configuration models [22], and hyper- evolving power grids [44, 45], bolic graphs [15]. These examples relate to, and partially encompass, several models of networks with dy- h) changing displayed content of websites on the namic node-properties that have been previously studied evolving world-wide web [46, 47]. – for instance dynamic latent space models [74–77], dy- These examples motivate the development of a simple namic random geometric graphs [78, 79], and dynamic modeling framework describing the impact of dynamic stochastic block models [72, 73]. The framework we study node-characteristics on dynamic link-structure. Such a is also widely generalizable to other contexts. framework would provide a temporal analogue of how Our study takes a step towards realistic modeling of node-properties influence network structure in hidden- dynamic networks with dynamic node properties. It in- variables models. In fact, it is standard practice to derive troduces a family of temporal network models that ex- temporal versions of static-network concepts [48–66], as tends static hidden-variables models to the temporal set- has been done for several models of static networks with ting, providing theoretical insight into the kinds of struc- hidden variables such as stochastic block models [67–73]. ture that can emerge as a consequence of the influence of Motivated by these considerations, here we study tem- hidden-variable dynamics on network-structure dynam- poral extensions of general static hidden-variables mod- ics. The framework can be used for studying real-world els, obtained by introducing dynamics of hidden variables temporal networks under the null hypothesis that physi- and of links. In these models each node has an evolv- cal or latent dynamic hidden variables drive the dynam- ing hidden variable, and each node-pair has a pairwise ics of network structure. Additionally, motivated by the affinity (equal to the connection probability in the static phenomenology emerging in these models, we speculate hidden-variables model), which is a function of the hid- that links in some real systems are out of equilibrium with den variables of both nodes. Pairwise affinities evolve respect to hidden-variables, partially explaining the pres- over time due to their dependence on a pair of evolving ence of long-ranged links in geometrically-embedded sys- hidden variables. The network itself evolves via node- tems and inter-group connectivity in modular systems. pairs being selected to re-evaluate their connections, re- In SectionII, we describe the properties that we use to sampling them with connection probability equal to the characterize the models we introduce. We then introduce pair’s affinity at the moment of re-evaluation. These sys- the static and temporal hidden-variables model families tems are governed by just two parameters beyond those in SectionIII, followed by various limiting regimes in Sec- of any static model: a rate of hidden-variable dynamics tionIV. SectionV provides several examples illustrating σ, and a rate of link-resampling ω. temporal hidden-variables models. We then consider a We find that these models have snapshots that are sta- variant of the family of models in SectionVI, incorpo- tistically equivalent to networks generated from the static rating an additional dynamic mechanism that enforces model in the following cases: static-model connection probabilities. The final sections are dedicated to descriptions of related work (Section a) if there is a sufficient timescale separation VII) and a discussion of our results and the implications (slow hidden-variable-dynamics relative to link- thereof (SectionVIII). Appendices provide the details of dynamics), several calculations and procedures left out of the main text. b) if connectivity is a deterministic function of hidden variables,

c) if hidden variables are held fixed, or II. DESIRED PROPERTIES OF DYNAMIC HIDDEN-VARIABLES MODELS d) if we add an additional dynamic mechanism whereby links actively respond to changes in hidden variables. This section outlines the properties that we use to characterize the family of dynamic hidden-variables models We also identify the conditions under which model net- that we introduce. Our goal is to construct natural tem- works evolve gradually, i.e., exhibit link-persistence, and poral versions of static networks with hidden variables, evaluate qualitative resemblances of snapshots to some and to understand the consequences of having introduced real networks which arise as deviations from static-model such dynamics. Our approach is via a Markov chain on behavior. We obtain analytical and numerical results graphs and hidden-variable configurations, with sources for effective connection probabilities (the probability of of randomness in the original static model being replaced a node-pair being connected given their current hidden- by random processes in the temporal model. variable values), directly quantifying deviations from Specifically, given a static hidden-variables model, i.e., static-model behavior in each case. a probability density on hidden-variable configurations The family of models we introduce is demonstrated H ∈ H and a conditional probability distribution on to have wide generality, as exemplified by temporal ex- graphs G ∈ G given H, the temporal extension yields tensions of four different static models with hidden vari- a probability distribution/density on temporal sequences ables: stochastic block models [20], random geometric of graphs and hidden-variable configurations, denoted 3

(t) T T (t) T T Given H, the graph sequence G is produced via a Markov G = G t=1 ∈ G and H = H t=1 ∈ H , respectively. We will evaluate the conditions under which chain with transition probability having auxiliary H- (t+1) (t) models within our framework satisfy the following prop- dependence, PG G G , H . Herein, we primarily erties: consider graph dynamics with H-dependence of the form (t+1) (t) (t+1) PG G G ,H , but also consider dynamics of (t+1) (t) (t+1) (t) a) Equilibrium Property: The marginal probability of the form PG G G ,H ,H in SectionVI. In a graph at any timestep is identical to its probabil- general, we could consider any choice of H-dependence – 0 ity in the static model; likewise for hidden variables. as long as G(t) is not influenced by H(t ) for any t0 > t, b) Persistence Property: The level of structural per- since that would entail graph-structure at time t being 0 sistence over time – quantified by, e.g., any graph dependent on HVs at future-times t > t. The ini- (1) similarity measure between graphs at adjacent tial graph G is sampled from the static-model con- (1) (1) timesteps – is high relative to the null expectation ditional probability P G H . The H-conditioned (of two i.i.d. static-model samples). temporally-joint graph probability distribution P (G|H) is then given by: c) Qualitative Realism: The graph-structure, HV- geometry (e.g., link-lengths), and/or dynamic be- T −1 haviors resemble observed characteristics of some (1) (1) Y (t+1) (t) P (G|H) = P G H PG G G , H . real-world systems at a qualitative level. t=1 (3) If the Equilibrium Property is satisfied, the tempo- Altogether, the temporally-joint graph probability distri- ral network in question is a strict extension of the static bution is given by model – individual snapshots are then indistinguishable from static-model realizations. If the Equilibrium Prop- Z erty is not satisfied, snapshots deviate from the static P (G) = P (G|H)p(H)dH, (4) T model, the resulting phenomenology of which we seek to H understand. The Persistence Property holding implies which is the temporal extension of Equation (1). a gradually evolving network, without sudden structural It is this strategy that underlies all temporal exten- transitions between networks at adjacent timesteps. In sions of static models that we consider. Static graphs most cases we have a parameter to tune the level of struc- without hyperparameters may also be included by disre- tural persistence, making the level of satisfaction of the garding H above, leaving only Equation (3), which be- Persistence Property fall along a continuum. To have comes a general Markov chain on graphs governed by Qualitative Realism simply means that the system ex- P G(t+1) G(t). Note that G can be seen as a multi- hibits some characteristics and behaviors that are anal- G plex network [82, 83] with layers representing timesteps. ogous to real-world systems – regardless of whether the detailed mechanisms are realistic or quantitatively ac- curate. In particular, we are interested in qualitative features relating to the dynamics of node-characteristics, A. Static Hidden-Variables Model and the effects of such dynamics on a network’s structural evolution. Here we describe the static hidden-variables model [18] (SHVM), which generates graphs by a two-step proce- dure. First, each node j (out of n total, labeled as III. MODELING FRAMEWORK {1, ..., n} = [n]) is assigned a hidden variable hj ∈ X , drawn independently with probability density ν(hj) This section provides an overview of our modeling ap- from set X . Thus the hidden-variable configuration is n n proach, and then defines static and temporal hidden- H = {hj}j=1 ∈ H = X and the joint hidden-variable variables models. We first describe our approach to con- Qn density is ρ(H) = j=1 ν(hj). Second, node-pairs ij structing temporal extensions of static models, which (1 ≤ i < j ≤ n) connect with pairwise probability produce length-T sequences of graphs G with a probabil- f (hi, hj), independently from one another. The condi- ity conditioned on a length-T sequence of hidden-variable tional probability P(G|H) of a graph G is thus given by configurations H. The latter arises from Markovian dynamics [80, 81] governed by conditional probability den- Y Aij 1−Aij (t+1) (t) (1) P(G|H) = (f (hi, hj)) (1 − f (hi, hj)) , sity PH H H . The initial configuration H is sampled from the static-model hidden-variable density 1≤iadjacency matrix probability density p(H) as a product: ij 1≤irandom graph. But since H is random, (G) is a prob- T −1 P (1) Y (t+1) (t) abilistic mixture of Equation5 over possible hidden- p(H) = ρ H PH H H . (2) n t=1 variable configurations H ∈ X via Equation1. 4

B. Temporal Hidden-Variables Model The probability of a graph-sequence G given H is the temporal product (3) of the following transition proba- We now describe a temporal version of the SHVM bilities,

(SectionIIIA), namely the temporal hidden-variables (t+1) A (t+1) (t) (t+1) (t) (t+1) Y ij 1−A P G G ,H = Y (1−Y ) ij , model (THVM). We denote by Aij the ij-th element G ij ij of G(t)’s adjacency matrix. The initial conditions 1≤i

Name Parameter Regime Equilibrium Property Tunable Persistence Single Static Graph ω = σ = 0 Yes No i.i.d. Graph Sequence ω = σ = 1 Yes No

Quasi-Static α2(σ, ω) ≈ 1 (Equation 12) Yes* Yes Complete link-resampling ω = 1, σ ∈ (0, 1) Yes Depends on f** Deterministic HV-to-graph ω = 1, f : X 2 → {0, 1} Yes Depends on f** Complete HV-resampling σ = 1, ω ∈ (0, 1) No Yes *** Fixed Hidden Variables σ = 0 Yes Yes Erd˝os-R´enyi-like σ/ω 1 No No Fixed graph structure ω = 0 Yes**** No****

TABLE I. Table of limiting cases of dynamics-parameters (σ, ω) for THVMs. The first and second columns provide a short-hand name and the associated parameter regime. The third column states whether the Equilibrium Property is satisfied, whereas the fourth column states whether the Persistence Property is satisfied (in a way that is tunable at any desired level, which for instance leaves out the case σ = ω = 0). *In the quasi-static regime, G(t) will have arisen from an HV-configuration closely resembling H(t), due to a timescale-separation. This implies approximate, rather than exact, satisfaction of the Equilibrium Property. ** When ω = 1 although the persistence property is in general lost due to each possible edge being resampled at every timestep, there is still some persistence present, tuned by σ and dependent upon the affinity function f. *** When σ = 1 the persistence property is tunably satisfied at the level of graph-structure, but not at all at the level of hidden variables, which are completely resampled every timestep. **** In the case of ω = 0, the initial graph remains fixed for all time, while HVs change. Since the initial condition is sampled from the static model, this regime technically satisfies the Equilibrium Property. It does so both at the level of graphs and at the level of hidden variables, but not at all at the joint level. Persistence is not tunable at the level of graphs, but is at the level of hidden variables.

A. Quasi-Static Regime (α2(σ, ω) ≈ 1) B. Complete link-resampling (ω = 1)

Here we consider the parameter regime quantified by Here we consider the case ω = 1 (top region of Figure the condition α2(σ, ω) ≈ 1 (upper-left region of Figure 1). This case resembles that of the quasi-static regime, 1), where but all links form based on current hidden-variable configurations, so there is no graph-encoded memory: ω P G(t+1) G(t), H = G(t+1) H(t+1). The result- α (σ, ω) = ∈ [0, 1], (12) G P 2 1 − (1 − ω)(1 − σ)2 ing Markov chain on H × G thus satisfies the Equilib- rium Property exactly, as opposed to approximately in in which networks have both random link-structure and the quasi-static regime (SubsectionIVA). Link-structure random hidden variables, and exhibit both the Persis- when ω = 1 is more correlated over time than two tence Property and the Equilibrium Property. The quan- i.i.d. samples from the SHVM (due to persistence in tity α2(σ, ω) is a naturally-arising function characterizing HV-configurations), but the specific level of persistence how effective connection probabilities differ from their depends on the form of the affinity function f(h, h0) and static-model counterparts (see AppendixA). The Equi- on σ. A variety of temporal network models have fully- librium Property is satisfied due to sufficient timescale resampled edges at each timestep [72, 85, 86]. separation: link-resampling happens quickly enough rel- As subset of the ω = 1 regime, consider THVMs with ative to hidden-variable motion for G(t) to remain caught binary affinity function f : X 2 → {0, 1}. In this case up with H(t). The dynamics can thus be considered all randomness comes from hidden variables, because f quasi-static, in the sense of quasi-static transformations deterministically maps HV-configurations to graphs. The in classical equilibrium thermodynamics [84]. Over time, static model’s conditional probability distribution in such the HV-configuration and link-structure both fully ex- cases is given by a product of indicator functions: plore their respective spaces, functioning as a tempo- Y ral network whose stationary distribution is the static P(G|H) = 1 {Aij = f (hi, hj)} , (13) hidden-variables model defined in SectionIIIA. Note 1≤i

(t) (t+1) (t+1) (t) Y Aij →Aij PG(G |G ) = pij , (14) Network structure fully resampled ω 1≤i

Here we consider the case for which all hidden variables C. Fixed Hidden Variables (σ = 0) are resampled at every timestep (σ = 1), so that no HV- driven structural persistence exists (right region of Fig- ure1). Note that walk-dynamics is parameterized by σ so Here we consider σ = 0 (left region of Figure1), that σ = 1 implies complete HV-randomization. If σ = 1, in which case all HVs are frozen in place, ensuring correlations among links (and non-links) do still exist due satisfaction of the Equilibrium Property. The ini- to simultaneous resampling; the set of node-pairs selected tial HV-configuration H(1) has the SHVM density ρ, for link-resampling at timestep t form links based upon but conditioning on some particular initial configura- the same underlying hidden-variable configuration H(t). tion H(1) yields fixed pairwise connection probabilities In this setting, ω quantifies the level of agreement among pij = f (hi, hj), resulting in temporal versions of edge- node-pairs as to what the HV-configuration is. For in- independent static networks [87–89]. Analytical expres- stance in spatial network models, if σ = 1, then ω directly sions for link-dynamics can be written straightforwardly controls the level of geometry-induced correlations. in terms of the set of values {pij}1≤i

(t) 0 (t−1) 0 q = q q = q = (1−σ)1{q = q }+σ% 0 , (19) V. TEMPORAL EXTENSIONS OF POPULAR P i i q STATIC NETWORK MODELS and (t) (t−1) (t) (t) (t−1) This section contains several examples of THVMs. In P Aij = 1 Aij , qi , qj = (1−ω)Aij +ωf (t) (t) . qi ,qj each subsection, we describe a static hidden-variables (20) model, its temporal extension according to the model- See Figure2 for visualized embeddings of network snap- ing framework of SectionIIIB, the effective connection shots from the stationary distribution of the example 0 probability that arises due to the dynamics, and offer m = 2, %1 = u = 1 − %2, fq,q0 = p1{q = q }. some additional discussion. We specifically consider temporal extensions of the following static network models: stochastic block models [20], random geometric graphs 3. Effective connection probabilities in hyper-SBMs [21], hypersoft configuration models [22], and hyperbolic graphs [15]. The block-dynamics of nodes in temporal hyper-SBMs introduces several novel features to the system. First, pairwise affinities change over time. Second, the set of A. Temporal Stochastic Block Models all existing links at time t need not have arisen from the group-assignments of time t. Temporal snapshots in This subsection considers temporal extensions of general thus deviate from the static model – the Equi- stochastic block models (SBMs), which are used to model librium Property does not necessarily hold. However, community structure in networks [20, 91, 102, 103]. even if snapshots do not resemble the static model, they do resemble a static model – an effective SBM. Consider two nodes, with current group-indices q, q0. Averaging 1. Static Hyperparametric SBMs over all past values of hidden variables, we obtain the ¯ effective connection probability fq,q0 for dynamic hyper- We consider a static network with conditionally SBMs. Since the SBM case is directly obtainable from Bernoulli-distributed edges amongst n nodes j ∈ [n], each discretization of the continuous model (see AppendixH) node having been randomly assigned to one of m groups we can use a discrete version of the general formula de- (a.k.a. communities, blocks, colors). Each node j in- rived in AppendixA, namely: dependently draws a group-index qj ∈ [m] = {1, ..., m} ¯ fq,q0 = α2fq,q0 from probability distribution % = {%q}q∈[m]. Each node- + (α1 − α2)(hfq,·i + hfq0,·i) (21) pair then connects with probability fqi,qj . In this defini- tion, the group-memberships {qj}j∈[n] are not externally + (1 − 2α1 + α2)hfi, 8

1.0 0.125

0.8 0.100

0.6 0.075

0.4 0.050

0.2 0.025

0.0 0 0.0 0.2 0.4 0.6 0.8 1.0

FIG. 2. Snapshots of a temporal stochastic block model: a modular network with dynamic group-assignments and link-resampling. The n = 100 nodes are partitioned into two groups with group-membership probabilities %1 = 0.4 = 0 1−%2, and group-memberships change in time by group-resampling with probability σ. The aﬃnity function is fq,q0 = p1{q = q } with p = 0.25, disallowing inter-group connections in the static model. Network snapshots are displayed via a spring-force layout algorithm [104], for various parameters (σ, ω) such that networks span a variety of structural outcomes. Node-coloration is by group-membership, and link-coloration is black for within-group links and green for between-group links. In the central panel, the eﬀective connection probability f¯1,2 between communities is plotted. Outside of the quasi-static regime, group- membership dynamics is fast enough for a substantial number of inter-group links to exist (f¯1,2 > 0), despite the inter-group connection formation probability being f1,2 = 0.

0 ¯ where coeﬃcients αb(σ, ω) for b ∈ {1, 2} are given by q }), terms in fq,q0 are evaluated as:

ω hf1,·i = up, αb = b , (22) 1 − (1 − ω)(1 − σ) hf2,·i = (1 − u)p, (24) hfi = p(u2 + (1 − u)2), and marginally-averaged aﬃnities are

from which the formula for f¯ 0 becomes X q,q hfq,·i = %q0 fq,q0 , 0 f¯ 0 = α p1{q = q } q0 q,q 2 (23)  X X 2up, q = q0 = 1 hfi = %qhfq,·i = %q%q0 fq,q0 .  0 q q,q0 + (α1 − α2) 2(1 − u)p, q = q = 2 (25)  p q 6= q0

Note that when σ = 1 we have α1(1, ω) = α2(1, ω) = ω 2 2 + (1 − 2α1 + α2)p(u + (1 − u) ). and Equation 21 reduces to the form of Equation 17. In the simple example case (m = 2,%1 = u, fq,q0 = p1{q = In particular, the between-group eﬀective connection 9 probability becomes 2. Temporal RGGs

¯ 2 2 f1,2 = p α1 − α2 + (1 − 2α1 + α2)(u + (1 − u) ) , To go from static RGGs to temporal RGGs, we incor- (26) porate coordinate-dynamics and link-resampling dynam- which is visualized in Figure2. In the extreme case ics. We consider here jump-dynamics, each node resam- of σ/ω 1 all links form between nodes with ef- pling its coordinate according to the static-model density fectively random group-assignments, making all pairs ν, with probability σ, each timestep t ∈ {2, ..., T } (the co- equally likely to connect, and reducing the system to a ordinate density follows Equation6, with ν(h) = 1 for the temporal Erd˝os-Rényi network of connection probability 2 2 uniform density on the unit interval). Link-resampling p(u + (1 − u) ). happens independently for each node-pair with probability ω each timestep. Since RGGs have deterministic connectivity, link-resampling of ij at time t guarantees 4. Temporal hyper-SBMs discussion (t) (t) (t) (t) that Aij = 1 if dX (hi , hj ) ≤ r and Aij = 0 otherwise. But if ij’s connectivity is not resampled at time t, links Interesting examples of Qualitative Realism arise in may fall out-of-equilibrium with respect to coordinates. temporal hyper-SBMs. For instance, group-dynamics of Note that we could also study temporal RGGs with walk- nodes yields inter-group connectivity, as is observed in dynamics, with either periodic or reflecting boundary real systems. If someone joins a different club, switches conditions; for simplicity, we study jump-dynamics here, political party, or emigrates to a new country, they at first leaving temporal RGGs with walk-dynamics for a future primarily carry ties to their original group – and thus study. upon changing group-membership, they suddenly have many inter-group links – not because of inter-group link- formation, but because of dynamic group-membership. 3. Effective connection probabilities in temporal RGGs Likewise, within-group connectivity can be lower than in the static model, as is the case in real systems due We now describe the effective connection probability to nodes having recently arrived from another group, or f¯(x) for RGGs between pairs of nodes for arbitrary (σ, ω). from neighbor-nodes having recently departed. These ef- The expression for f¯(x) in temporal RGGs is derived in fects arise outside the quasi-static regime, so we speculate AppendixB, and the result is provided here: that in some cases the non-equilibrium regime can bet- ter emulate real-world systems. We also note that we f¯(x) = α 1{x ≤ r} + 2r(1 − α ). (27) here considered group-resampling HV-dynamics (a dis- 2 2 crete version of jump-dynamics), but we could also con- The quantity α2 = α2(σ, ω), defined in Equation 22, di- sider a general Markov chain on group-assignments with rectly governs the level of locality in temporal RGGs. stationary distribution %. See Figure3 for a visualization of the function α2(σ, ω) and of network snapshots across a range of (σ, ω)-values. The effective connection probability f¯(x) has a step-like B. Temporal Random Geometric Graphs form, with connection probability α2 + 2r(1 − α2) for all x ≤ r and 2r(1 − α2) for all x > r. The above effective In this section we describe THVMs arising from static connection probability agrees perfectly with the results random geometric graphs (RGGs), which model the influ- of numerical simulations, see Figure4. ence of an underlying geometry on graph-structure [21].

4. Temporal RGGs discussion 1. Static Random Geometric Graphs

The naturally arising function α2(σ, ω) ∈ [0, 1] de- In random geometric graphs (RGGs), nodes are as- scribes the level of locality in network snapshots (see Fig- signed spatial coordinates as hidden variables, and node- ure3), and quantifies the Equilibrium Property. It inter- pairs are linked if their coordinates are closer than polates between the case of RGGs (α2(σ, ω) = 1) and ER some threshold distance r. Hence the affinity is binary, graphs (α2(σ, ω) = 0), resembling the structural transi- 2 f(hi, hj) = 1{dX (hi, hj) ≤ r}, with dX : X → [0, ∞) tion of the Watts-Strogatz model [108]. In this case, all denoting the geodesic distance in latent space X . Ex- links form locally, and it is dynamics of node positions amples of well-studied RGGs include Euclidean RGGs that induces the transition (alongside formation of lo- with periodic or nonperiodic boundary conditions [21], cal links at nodes’ new locations); a similar phenomenon spherical RGGs [107], and hyperbolic RGGs (the hyper- has been observed in contagion-dynamics among mobile bolic model with inverse-temperature parameter β = ∞ agents [109]. Also note, in dynamic RGGs, links can [15]). As a primary example we consider a simple one- exist that were not possible in the static model model: dimensional RGG with periodic boundary conditions: links of length greater than r, since the effective connec- ¯ X = [0, 1) and dX (hi, hj) = 1/2 − |1/2 − |hi − hj||. tion probability f(x) no longer goes completely to zero 10

1.0 1.00

0.8 0.75

0.6 0.50

0.4

0.25 0.2

0.0 0 0.0 0.2 0.4 0.6 0.8 1.0

FIG. 3. Snapshots of a temporal random geometric graphs: a geometrically-embedded network with dynamic node-coordinates and link-resampling. Coordinates of n = 100 nodes are sprinkled uniformly into a 1D ring of unit circumference, and change in time via jump-dynamics (coordinate-resampling with probability σ). The affinity as a function of distance is f(x) = 1{x ≤ r}, where r = 0.1 is the connection radius, disallowing long-ranged links in the static model. Snapshots are shown at various values of (σ, ω), with angular positions equal to 2π times spatial coordinates, and radial positions set to 1 with some added random noise. Link coloration is according to length: black links are of distances x ≤ r whereas green links are of distances x > r. In the central panel, the function α2(σ, ω) ∈ [0, 1] is visualized, which encodes the level of locality in temporal RGGs (see Equation 27). for x > r (see Equation 27). This is related to phe- Qualitative Realism are in conflict with the Equilibrium nomena observed in real-world networks: pairs of people Property. Note also that similar deviations of f¯(x) rel- may form friendships locally, but maintain those friend- ative to f(x) occur in THVMs arising from soft random ships after becoming geographically separated, resulting geometric graphs [110–113], for example the H2 model in the existence of long-ranged social ties that would not (see SectionVD). likely have formed at that distance. Likewise, the function f¯(x) is also less than one for distances x ≤ r, allowing for non-links that would be impossible in the static model. That phenomenon also appears in real-world sys- C. Temporal Hypersoft Configuration Model tems: instead of individuals knowing everyone in their local vicinity, non-links between closeby pairs may exist, In this section we consider a dynamic version of hy- due to them having only recently become proximate. As persoft configuration models (HSCMs), which model net- with the case of temporal hyper-SBMs, these examples of works with degree-heterogeneity [22]. 11 A) 1. Static Hypersoft Configuration Model

1.00 The static model we now consider is the hypersoft configuration model [22, 114] (HSCM), a hyperparametric version of a soft configuration model (SCM). SCMs come 0.75 in several varieties such as the Chung-Lu model [115], inhomogeneous random graphs [87], and the Norros-Reittu model [116]. Node-pairs connect with Aij-values being 0.50 independent (typically Bernoulli or Poisson distributed), such that on average, each node has a particular degree- 0.25 value. In hyperparametric SCMs, that degree-value is randomly assigned, according to some specified distribution of expected degrees. For example, one way to 0.00 obtain SCMs with a degree distribution that is Pareto- 0.1 0.2 0.3 0.4 0.5 mixed Poisson (with, say, power-law tail-exponent γ and expected degree hki), is for nodes j ∈ [n] to be assigned hidden variables hj ∈ [h−, ∞) drawn from a Pareto den- γ−1 −γ FIG. 4. The effective connection probability, in the- sity ν(h) = (γ − 1)h− h , with minimal HV-value ory and simulation, for 1D RGGs at various values h− = (γ − 2)hki/(γ − 1), and then for node-pairs to be of the dynamics-parameters (σ, ω). The static model connected with probability affinity-function f(x) is plotted with square markers. The solid lines are numerical estimates of the effective connec- 1 hihj ¯ f(hi, hj) = ≈ , (28) tion probability f(x) (with ω increasing as colors change 1 + nhki/hihj nhki from blue to yellow), whereas the dotted lines are the theoretical effective connection probability (Equation 27). the approximation holding when hihj/nhki 1. The expected degree of a node i in the static model is Z ∞ hki|hii = (n − 1) f(hi, h)ν(h)dh ≈ hi. (29) h− The actual degrees of nodes are sharply peaked around their expected degrees, and thus the above implies that 10 the degree distribution itself likewise has a power-law tail with exponent γ and mean hki.

8 2. Temporal HSCMs

6 Now we consider a temporal version of HSCMs. At each timestep, each node j, with probability σ, resamples (t) 4 its hidden variable hj from the static-model HV-density ν (jump-dynamics). Then, each node-pair ij (1 ≤ i < (t) 300 400 500 600 700 800 900 j ≤ n), with probability ω, has its indicator-variable Aij (t) (t) resampled from a Bernoulli of mean f(hi , hj ). In the static model, the HV-value hj alone determines the expected degree hk |h i. But in the temporal version, FIG. 5. Expected degree over time of a node in a j j (t) temporal hypersoft configuration model with jump- the quantity hi is time-evolving, and the expected de- dynamics of hidden variables. Each node’s expected degree dynamically trails behind the static-model expected gree (blue dotted curve) equilibrates towards its current degree, equilibrating at a geometric pace (See Figure5): static-model expected degree (green solid curve), as per Equation 30. In any realization, the actual degree over time (t) n (t−s)o fluctuates (purple curve), but its ensemble-average (orange E ki hi s≥0 solid curve) behaves as predicted. The average was ob- Z ∞ X (t−s) tained by simulating 1000 realizations with (n, hki, γ, ω, σ) = = (n − 1)ω (1 − ω)s f h , h ν(h)dh (t) T i (30) (200, 8, 2.8, 0.04, 0.01), keeping the HV-trajectory {hj }t=1 of s≥0 h− a single node j fixed across trials. D E X s (t−s) = ω (1 − ω) ki hi . s≥0 12

We can also average the above over all hidden-variable D. Temporal Hyperbolic Graphs values at timesteps earlier than t, to obtain an effective (t) expected degree that depends only on hj . To do this, In this section we consider a temporal extension of (t−s) (t) the hyperbolic model [15] (the 2 model, for short), a we use the probability density of hj given hj under H jump-dynamics: geometry-based network model simultaneously exhibiting sparsity, clustering, small-worldness [117, 118], degree (t) s s Ps x hj = (1 − σ) 1 (t) (x) + (1 − (1 − σ) ) ν(x), heterogeneity, community structure [119], and renormal- hj (31) izability [120]. Averaging Equation 30 over HVs at all timesteps t − s for s > 0, 2 1. Static H model Z ∞ h (t) (t) i X (t) k h = ω (1 − ω)s P x h hk |xidx E i i s i i 2 s≥0 h− The H model is parameterized by a number of nodes D E n, average degree hki, power-law exponent γ, and inverse- (t) = α1 ki hi + (1 − α1)hki, temperature β (which tunes the level of clustering). Hid- (32) den variables are polar coordinates, hj = (θj, rj), namely where α1(σ, ω) = ω /(1 − (1 − ω)(1 − σ)). In this a radial coordinate rj ∈ [0,R] encoding the popularity case α1 measures the level of equilibration of node- of node j and an angular coordinate θj ∈ [0, 2π), en- neighborhoods to their expected sizes. Having α1 ≈ 1 coding the similarity of node j to other nodes. These indicates the quasi-static regime whereas α1 ≈ 0 indi- coordinates are sampled according to separable density cates an averaged-out behavior so that the expected de- ν(θ, r) = νang(θ)νrad(r) where angles are distributed uni- gree of any given node is simply the expected average formly (νang(θ) = 1/2π) and radii have an exponentially degree hki of the network. growing density, γ − 1 sinh γ−1 r ν (r) = 2 , (33) rad 2 γ−1 3. Effective connection probabilities in temporal HSCMs cosh 2 R − 1 where R = R(n, hki, β, γ) is selected so that the mean We now discuss effective connection probabilities in degree is hki. The static-model affinity of node-pair ij is HSCMs. The formula derived in AppendixA applies, a Fermi-Dirac function [121] (a sigmoid) of the hyperbolic 0 but note that the affinity f(h, h ) (Equation 28) is a func- geodesic distance xij between i and j, 0 tion only of the product ψ = hh . Thus we can examine . (β/2)(xij −R) the effective connection probability as a function of ψ, f(hi, hj) = f(xij) = 1 1 + e , (34) denoted f¯(ψ). In order to calculate f¯(ψ) we first must compute the probability density of a product of hidden where xij = xij(hi, hj) is given by variables in past timesteps, given the value of the prod- cosh(xij) = cosh(ri) cosh(rj) uct at the current timestep. We then sum the expected (35) s − sinh(ri) sinh(rj) cos(θij), affinity given the product, weighted by ps = ω(1 − ω) , over all past timesteps s > 0. These calculations require with θij = π − |π − |θi − θj||. The connection probabil- a variety of intermediate steps, and are described in Ap- ity and coordinate-density in this model result in power- pendixC. law degree distributions (but could also give rise to other degree distributions if the radial coordinate-density was different), a similar feature to that exhibited by HSCMs 4. Temporal HSCMs discussion – but also, the geometry arising from inclusion of the angular coordinate yields a large clustering coefficient and Note that in HSCMs, non-equilibrium dynamics re- spatially localized link-structure, making this model also duces degree-heterogeneity; nodes with large HV-values similar to standard RGGs. Increasing the parameter β only transiently retain them. Equilibration, on the other yields more localized link-structure, approaching a step hand, allows for a full structural expression of the nodes’ function as β → ∞, leaving in that case an RGG (see internal heterogeneity. This implies that extremely het- SectionVB1) on the hyperbolic disk. As β → 0, typi- erogeneous real-world networks, if described by these cal link-lengths approach the system size and the model models, would typically be in the quasi-static regime. behaves similarly to the HSCM (see SectionVC1). We only considered jump-dynamics here (resampling of static-model expected degree-values), but we could al- 2 ternatively study walk-dynamics, where nodes’ HVs un- 2. Temporal H model dergo Brownian-like motion in a way that preserves ν. This could be achieved straightforwardly as described in To temporally extend the H2 model, we allow coor- D, alongside reflecting boundaries as studied in Appendix dinate dynamics so that each node j exhibits a trajec- (t) (t) (t) E. tory in the hyperbolic disk, hj = (θj , rj ) for t ∈ [T ]. 13 A) B) C)

D) E) F)

2 FIG. 6. Hidden-variable dynamics of nodes in a temporal H model, at increasing values of σ, with ﬁxed ω = 0.1. In each subplot, node-coordinates for 100 random nodes are shown at two adjacent timesteps, from a network with parameters (n, γ, β, R) = (500, 2.2, 5, 8). Each arrow points from the coordinate-location of a node at a given timestep (grey) to the coordinate-location of the same node at the next timestep (black). Subplots (A,B,C) depict jump-dynamics (coordinate-resampling with probability σ, otherwise remaining in place), whereas subplots (D,E,F) depict walk-dynamics (all nodes move to neighboring locations, with mean step-length parameterized by σ). Marker sizes are proportional to node degree-values. For small σ/ω (subplots A and D), nodes’ existing connections have arisen from approximately the present coordinates, making snapshots closely resemble the static hyperbolic model, as seen e.g. by the exhibited degree-heterogeneity. For larger σ/ω (subplots B and E), connections have arisen via mixtures of past and present coordinates, reducing degree- heterogeneity. For very large σ/ω (subplots C and F), the system behaves similarly to a temporal Erd˝os-R´enyi network.

For jump-dynamics, each node jumps to a random lo- the standard Fermi-Dirac form of f(x) (see Figure7). cation according to density ν(θ, r), with probability σ With decreasing ω/σ, the connection probability func- each timestep. For walk-dynamics, each node j steps tion smooths out and extends to a longer range due to (t+1) links being stretched more rapidly (for walk-dynamics), to a random location hj having angular and radial coordinates adjusted to relatively closeby values, with or more frequently (for jump-dynamics). This effect is increasingly large steps for larger σ-values; we describe more uniform and extends all the way out to long ranges the details of H2 walk-dynamics in AppendixF. Dynam- for jump-dynamics, whereas it is more localized for walk- ics of nodes on the hyperbolic disk is visualized in Fig- dynamics, for any given non-equilibrium value of (σ, ω). ure6, for both jump-dynamics and walk-dynamics. For Since the coordinates of 2 reflect popularity and simi- σ 1, nodes rarely resample their coordinates (in jump- H larity attributes, the effective connection probability and dynamics) and step to only very localized regions (in other non-equilibrium effects arising when outside of the walk-dynamics). On the other hand for σ ≈ 1, almost quasi-static regime have specific interpretations. The all nodes resample their coordinates at each timestep (in set of current links arose from nodes having been con- jump-dynamics) or move to a nearly-randomized loca- nected at past timesteps when their previous similarity tion (in walk-dynamics). We note that many other natu- attributes were compatible (small hyperbolic distance); ral and interesting choices for HV-dynamics exist, as we in real networks, such links may persist into the future discuss in SectionVIII and AppendixF. even if the similarity attributes change. For instance with social networks, consider friendships on Facebook, follow- ers on Twitter, or author collaborations: similarity be- 2 3. Effective connection probabilities in the temporal H tween connected pairs may decrease over time, but they model tend to remain connected. Likewise, it could take some time for two people that become more similar to discover In the temporal H2 model considered here, the effec- one another and to connect, in an online or traditional tive connection probability f¯(x) no longer remains in social network. 14

the space rather than being localized amongst neighboring groups of nodes. Degree-heterogeneity is also sup- pressed, as is the case for the temporal HSCM (see Sec- 0 A) 10 tionVC), because nodes accumulating large numbers of links due to being near the disk’s center do not stay near the disk’s center indefinitely. Clustering and het- - 1 2 10 erogeneity arise in the static H model due to the correlations in links from the underlying geometry. But in the static model, all links (and non-links) arise from the same - 2 10 underlying coordinate-configuration. When coordinates are dynamical, these correlations are weaker; nodes are linked with probabilities arising as a mixture of past and - 3 10 present coordinate-configurations.

3 6 9 12 15 VI. LINK-UPDATING IN RESPONSE TO HIDDEN-VARIABLE DYNAMICS

Finally, we describe an additional dynamical mecha- B) 0 nism that can be incorporated to achieve the Equilib- 10 rium Property exactly in temporal hidden-variables models, while retaining the Persistence Property, for all val- - 1 ues of σ and ω: links are updated directly in response 10 to changes in hidden variables, rather than only through link-resampling, to keep connection probabilities up-to- - 2 date (we refer to this mechanism as link-response). In 10 this model variant, G(t+1)’s probability distribution depends on each of G(t), H(t+1), and H(t), rather than on - 3 just the former two. We illustrate the mechanism at first 10 in the case of ω = 0. Suppose node-pair ij has a link 3 6 9 12 15 with probability pij = f(hi, hj), and that HVs (hi, hj) 0 0 are updated to become (hi, hj) in the next timestep. To ensure that the pair is then connected with probability 0 0 0 pij = f(hi, hj), we selectively delete now-less-likely edges FIG. 7. Effective connection probability function f¯(x) between connected pairs and selectively add now-more- in snapshots of a temporal hyperbolic model with likely edges between unconnected pairs. In particular: (n, γ, β, R) = (500, 2.2, 5, 8), for various values of ω. With slower link-resampling (smaller ω), links are increas- 0 0 a) If pij ≥ pij, then Aij = 1 ⇒ Aij = 1, and Aij = ingly allowed to dynamically stretch before being removed by + link-resampling, resulting in deviations from the static-model 0 ⇒ add link with probability qij, affinity f(x)(black dotted line). Coloration of the curve ¯ 0 0 f(x) is from yellow to blue as ω increases. The upper panel, b) If pij ≤ pij, then Aij = 0 ⇒ Aij = 0, and Aij = A), shows the case of jump-dynamics of coordinates. The − 1 ⇒ remove link with probability qij . lower panel, B), shows the case of walk-dynamics of coordinates. The choice of coordinate-dynamics is consequential The outcome needs to result in (A0 = 1|h0 , h0 ) = p0 . in the non-equilibrium regime, despite each having the same P ij i j ij Thus, stationary density. 0 a) If pij ≥ pij, the new connection probability satisfies 0 0 + + 1−pij 2 pij = pij + (1 − pij)qij. Hence, qij = 1 − 1−p . 4. Temporal H model discussion ij 0 b) If pij ≤ pij, the new connection probability satisfies (t) 0 Outside of the quasi-static regime, snapshots G do 0 − − pij 2 1 − p = pijq + (1 − pij). Hence, q = 1 − . not fully resemble the static H model – the Equilibrium ij ij ij pij Property is in general violated (despite the fact that each 0 + − link was formed via the static-model connection proba- Note that if pij = pij, then qij = qij = 0; no links will bility corresponding to the pairwise distance at the time form or break unless pairwise affinities change. Denoting (t) (t) (t) of that link’s formation). This phenomenon results in re- pij = f(hi , hj ) for t ∈ {1, ..., T }, the graph transition duced clustering because links become spread out across probability given H becomes: 15

inferred from real data and partially synthetically generated; the dynamics therein resembles THVMs with ω = 1 (t+1) (t) Y (t+1) (t) and σ = 0, but with varying average degree parameter PG G G , H = Yij Aij Aij , H , 1≤i

(t) eral dynamic latent space models [74–77] exist, as do dy- where Kij(Aij , H) incorporates the link-response dynamic random geometric graphs [78] (the latter being namics: continuous-time and inﬁnite-space, with nodes sprinkled as a Poisson process [137–139] and undergoing Brown- (t) n (t+1) (t)o + (t) (t) Kij Aij , H =1 pij ≥ pij qij 1 − Aij + Aij ian motion [140], with links remaining up-to-date as for

n (t+1) (t)o − (t) THVMs with ω = 1). A model with both dynamic HVs + 1 pij ≤ pij (1 − qij )Aij . and persistent links [141] was recently introduced, along- (38) side rigorous inference techniques and applications – but With the inclusion of link-response, arbitrary static not in reference to static network models. Other stud- hidden-variable networks can be extended to temporal ies investigated spreading on dynamic RGG-like graphs settings while satisfying the Equilibrium Property ex- [79, 109]. A few versions of dynamic SBMs are of partic- actly (See AppendixVI for a full derivation), and the Per- ular relevance; in one such paper [72], the model is a case sistence Property in a tunable fashion. Allowing ω > 0 of the temporal hyper-SBM studied in SectionVA with does not alter the Equilibrium Property’s exact validity, complete edge-resampling (ω = 1). Another study was of and it provides a more tunable level of structural persis- a temporal hyper-SBM with ω < 1 which thus exhibits tence. both link-persistence and group-assignment-persistence With G(t) indistinguishable from a static-model real- [73], influencing performance of community detection al- ization, all non-equilibrium phenomena of the types dis- gorithms and motivating the development of new ones. cussed inVA,VB,VC, andVD are prevented – this can Another area of relevant work is the rapidly emerging either enhance or hinder Qualitative Realism, depend- area of dynamic graph embeddings [75, 142–154], related ing on the context. If a single node’s HV is changed, it to the task of inference of hidden-variable trajectories will need to re-evaluate connections to all other nodes [155]. for which affinities have changed. This could be realis- We also note some additional works that are less- tic in some cases, since nodes themselves may be at the directly related to ours. Network-rewiring and MCMC most liberty to re-evaluate their connections. In other algorithms are widely used to sample static networks cases, more gradual structural transitions may be pre- [156–160]; in stationarity, these can be viewed as tem- ferred. This model-variant could thus serve well as a temporal networks satisfying the Equilibrium Property, with poral null model, especially for temporal networks with a level of persistence tunable via the number of iterations snapshots well-described by an SHVM. Despite structure between adjacent snapshots. Adaptive network mod- of THVMs with link-response being identical to that of els (for instance, SIS-dynamics [161] alongside contact- SHVMs, all dynamical features are open for study and switching [23, 24]), have dynamic node-properties that for comparison to real-world networks. evolve with time and guide network evolution, a com- monality with THVMs. Networks with node-growth and node-removal [162–165] have dynamic node-properties VII. RELATED WORK (degree-values as opposed to hidden variables) that influence link-formation. In the fitness model of growing We briefly review existing lines of research related to networks [12], static HVs and dynamic degrees both gov- our study. ern connection probabilities. Some static network models admit dual growing formulations [166] – analogously, if Several temporal network models are worth mention- the Equlibrium Property holds, THVM snapshots can be ing. Temporal analogs of specific static models have seen as dynamically produced static-model samples. been considered [67–73, 122, 123], many of which pre- serve the Equilibrium Property. Most such models have non-dynamic node properties, yielding models related to edge-Markovian networks [99, 100, 124–126] and dynamic VIII. DISCUSSION percolation [94–96]. The dynamic-S1 model [86] is a temporal extension of the static S1 model [15] consisting of In this work we have studied temporal network models a sequence of independent samples with HVs partially that are natural counterparts of static hidden-variables 16 models, obtained by inclusion of a dynamic mecha- can become more similar over time [177, 178] – or geo- nism for node-characteristics (jump-dynamics or walk- graphically move to closer-by coordinate locations. The dynamics) and dynamic mechanism for link-structure inclusion of interdependencies relating to dynamical pro- (link-resampling). Due to the wide generality of the cesses [179, 180] can allow for more interesting dynamics static hidden-variables framework, many popular static and realism, but at the cost of increased model complex- network models can be made temporal as THVMs. ity. With a single source of randomness in the static model, Real-world networks have dynamic node-properties which includes ω = 1 with deterministic connectivity that influence dynamics of link-structure. Examples of (SectionIVB) and σ = 0 with fixed initial HVs (Section such phenomena were set forth in SectionI, ranging IV C), the Equilibrium Property is exactly satisfied and across a wide variety of systems and scales. One di- the Persistence Property is controllable. If, however, the rect real-world application of THVMs could be to serve static model has two layers of randomness and links are as null models [65, 181] for evolving networks with dy- not completely refreshed each timestep (σ > 0 and ω < namic node-properties [75]. Dynamic embedding meth- 1), THVM snapshots are not in general distributed ac- ods [142–154], or generalizations of inference methods cording to the static model. Rather, numerous structural from dynamic SBMs [73], could potentially allow retrieval deviations arise, due to links falling out-of-equilibrium of H (and perhaps also σ, ω, and f) from an observed with respect to hidden variables – for instance, the ef- G. Links of real evolving networks may not in general fective connection probability f¯(h, h0) can substantially be fully equilibrated relative to the current set of node- differ from the affinity function f(h, h0) (see Figures4 characteristics, which is a dynamical behavior exhibited and7). Despite violating the Equilibrium Property, such by THVMs outside of the quasi-static regime. Hence in models arise naturally and exhibit Qualitative Realism in some cases, the Equilibrium Property and Qualitative interesting ways – for instance, the appearance of long- Realism may be in conflict, implying that caution should ranged links in temporal RGGs (SectionVA3) and inter- be used when applying static models to snapshots of group links in temporal hyper-SBMs (SectionVA3). An evolving networks. That said, static models do in many exception to the non-equilibrium dynamics arises in the cases accurately describe such snapshots; the internet, quasi-static regime (SectionIVA) in which case the Equi- for example, has exhibited a clear power-law degree-tail librium Property is approximately satisfied, due to all for decades [155, 182], evidently remaining in equilibrium A(t)-values arising from an HV-configuration closely re- from the perspective of THVMs (see the discussion in ij VC). sembling H(t). A second exception arises if we add a Overall, we expect that the present study will usefully third dynamical mechanism (SectionVI), namely link- inform general classifications of real-world networks ac- updating in direct response to HV-changes, which allows cording to the dynamics of node-properties and of how exact satisfaction of the Equilibrium Property (see Ap- those properties influence link-dynamics. pendixG) for all ( σ, ω). Both situations also lend themselves to tunable satisfaction of the Persistence Property, governed σ and ω. IX. ACKNOWLEDGEMENTS An assortment of possible modifications, improve- ments, and extensions are worth mentioning. Although We thank B. Klein, S. Redner, M. Shrestha, L. many questions are open within present framework, al- Torres, R. Van der Hofstad, and I. Voitalov for use- tered dynamics could also be considered. For HV- ful discussions and suggestions. This work was sup- dynamics, correlated motion akin to Langevin dynam- ported by ARO Grant Nos. W911NF-16-1-0391 and ics [167, 168] could provide insight into the forma- W911NF-17-1-0491, and by NSF Grant Nos. IIS- tion and persistence of communities. Altered link- 1741355 and DMS-1800738. F.P. acknowledges support structure and link-dynamics could be considered as well: by the TV-HGGs project (OPPORTUNITY/0916/ERC- some examples include directed and/or weighted links, CoG/0003), funded through the Cyprus Research and node-centric link-resampling dynamics [169], or pairwise- Innovation Foundation. individualized resampling rates. Continuous-time formulations of THVMs could allow some theoretical simplifi- cations; continuous time is used in studies of dynamical Appendix A: Effective connection probabilities percolation [94, 170, 171] and edge-Markovian networks [97–101], which could each be extended to a THVM- like framework by introducing hidden variables. Our re- Here we calculate effective connection probabilities for sults can also inform future studies of adaptive networks general THVMs with HVs evolving by jump-dynamics [172–176]; THVMs provide a simple setting in which dy- (HV-resampling with probability σ). We define the effec- ¯ 0 namic node-properties influence network-evolution. Un- tive connection probability f(h, h ) to be the probability (t) (t) (t) 0 derstanding such settings will provide a baseline for of Aij = 1 given hi = h and hj = h , in the limit as what to expect when coevolutionary feedbacks are also t → ∞. That is, present. An example of real-world links influencing node- f¯(h, h0) = lim A(t) = 1 h(t) = h, h(t) = h0 , (A1) properties is social influence, whereby acquainted pairs t→∞ P ij i j 17 where the limit t → ∞ is to wash out any initial condi- Appendix B: Temporal RGG effective connection tion. Due to the edge-resampling dynamics, the current probability (t) value of Aij arose from being last resampled at some time t − s, with s being a random nonnegative integer This section contains calculations of the effective con- s having distribution ps = ω(1 − ω) (where ω is the prob- nection probability for random geometric graphs on ability of link-resampling at any given timestep). The the unit interval with periodic boundaries and jump- effective connection probability is given by dynamics. This result could be obtained from Equation h i A6, but we show here an alternate derivation. The effec- ¯ 0 X (t−s) (t−s) (t) (t) 0 f(h, h ) = psE f hi , hj hi = h, hj = h . tive connection probability as a function of distances is s≥0 defined as the probability of two nodes being connected (A2) given that they are a distance d(t) = x apart, as t → ∞: To evaluate the above, we introduce a density P (x|h), ij s (t−s) (t) f¯(x) = lim A(t) = 1 d(t) = x . (B1) namely the density of hi (evaluated at x) given hi = P ij ij h. In our case, by jump-dynamics and conditioning on t→∞ (t) To calculate the above, we introduce the probability den- h = h, we have i sity on distances between node-pairs s timesteps prior to s s Ps(x|h) = (1 − σ) 1h(x) + (1 − (1 − σ) ) ν(x), (A3) when the distance-value is x, denoted Ps(y|x). We make use of the fact that d(t) can evolve in either of two ways: because h will have arisen from x after s timesteps via ij 2 either (a) zero jumps having occurred, that event hav- with probability (1 − σ) each timestep, neither i nor j ing probability (1 − σ)s, or via (b) at least one jump jumps, and thus the density is preserved. Otherwise, one having occurred, in which case the density is completely or both do jump, and their distance becomes completely randomized to ν(x). The expectation value appearing in randomized. The stationary density of distance x is the Equation A2 is equal to uniform on [0, 1/2], i.e., equal to 2 for all x ∈ [0, 1/2]. In a single time-advancement, jump-dynamics thus yields h (t−s) (t−s) (t) (t) i f h , h h = h, h = h0 2 2 E i j i j P1(y|x) = (1 − σ) 1x(y) + 2 1 − (1 − σ) . (B2) Z Z (A4) 0 0 0 0 Iterating the above logic, Ps(y|x) has two contributions: = f(x, x )Ps(x|h)Ps(x |h )dxdx , X X either neither node jumps at any time, or at least one node jumps at least once. Therefore, which, using Equation A3 and integrating over (x, x0), 2s 2s evaluates to: Ps(y|x) = (1 − σ) 1x(y) + 2 1 − (1 − σ) . (B3) (1 − σ)2sf(h, h0) We can compute f¯(x) via averaging the affinity s s 0 (t) + (1 − σ) (1 − (1 − σ) )(hf(·, h )i + hf(h, ·)i) (A5) f(hi, hj) = 1{dij ≤ r} over the distance-variable. That is, + (1 − (1 − σ)s)2hfi, h n o i ¯ X (t−s) (t) R f(x) = psE 1 dij ≤ r dij = x , (B4) where hf(·, h)i = hf(h, ·)i = X f(h, x)ν(x)dx and hfi = R 0 0 0 s≥0 X 2 f(x, x )ν(x)ν(x )dxdx . Finally, plugging Equation s A5 back into Equation A2, using ps = ω(1 − ω) and where the expectation term is s summing the geometric series that appear (P y = h n (t−s) o (t) i s≥0 1 d ≤ r d = x 1/(1 − y)), we obtain E ij ij Z 1/2 ¯ 0 0 f(h, h ) = α2f(h, h ) = Ps(y|x)1{y ≤ r}dy 0 0 (B5) + (α1 − α2)(hf(·, h )i + hf(h, ·)i) (A6) Z r 2s 2s + (1 − 2α1 + α2)hfi, = (1 − σ) 1x(y) + 2 1 − (1 − σ) dy 0 2s 2s where αb = αb(σ, ω) for b ∈ {1, 2} are given by = (1 − σ) 1{x ≤ r} + 2r 1 − (1 − σ) . ω Let s ∈ {0, 1, ...} be the delay since any given edge- αb(σ, ω) = . (A7) 1 − (1 − ω)(1 − σ)b indicator was last resampled. Recall that s has distri- s As an aside, we note that the average degree of the bution ps = ω(1 − ω) . Then, using the above, we find network is independent of (σ, ω). This can be seen that the effective connection probability for 1D RGGs by averaging Equation A2 over h and h0 and mak- with jump-dynamics is R X s 2s ing use of X Ps(x|h)ν(h)dh = ν(x) (which is true be- f¯(x) = ω (1 − ω) (1 − σ) 1{x ≤ r} cause Ps(x|h) describes the stationary distribution, re- s≥0 gardless of whether we consider walk-dynamics or jump- X s 2s (B6) dynamics). The result is hfi, regardless of σ and ω. This + ω (1 − ω) 2r 1 − (1 − σ) can be seen more directly in the case of jump-dynamics s≥0 0 by averaging Equation A6 over h and h . = α21{x ≤ r} + (1 − α2)2r, 18 with α2 = α2(σ, ω) arising from having evaluated sums where p1(φ|ψ) is the distribution of the product of a uni- P 2 s of geometric series of the form s≥0((1 − ω)(1 − σ) ) : form random variable and of one factor of a product, given that the value of that product is ψ. In the follow- ω ing, we walk through the remaining required calculations α2(σ, ω) = 2 . (B7) 1 − (1 − ω)(1 − σ) to obtain f1(ψ).

Appendix C: Effective connection probability in 1. Finding p(x|xy = ψ) terms of products of hidden variables Suppose that x and y are sampled uniformly on [0, 1]. This section describes effective connection probabili- Now condition on the fact that their product, xy, takes ties arising in temporal HSCMs, as studied in Section on the particular value xy = ψ. Then, what is the proba- VC. The static-model affinity f is a function of the prod- bility density of x alone? Note first that it must reside in uct of hidden variables, motivating study of the effective ¯ [ψ, 1], since ψ is the product of two numbers each in the connection probability f as a function of the product of range [0, 1], i.e., each reducing the value of the product. HVs as well. Within the acceptable range, the density is obtained as Consider one-dimensional hidden variables {hj}j∈[n] follows: each distributed uniformly on X = [0, 1]. This is appli- cable to HSCMs via the CDF-transform of arbitrary 1D Z 1 probability densities: if h has density ν, then u = F (h) = p(x|xy = ψ) ∝ 1ψ(xy)dy R h 0 0 0 ν(h )dh is distributed uniformly on [0, 1] (h− is the 1 h− 1 Z (C6) minimum value of h). Denote Ps(φ|ψ) as the probabil- ∝ 1ψ/x(y)dy x 0 ity density of φ = h(t−s)h(t−s) for some arbitrary pair ij i j = 1/x, (t) (t) given that hi hj = ψ. Then, where the ratio ψ/x is guaranteed to be in the range Z 1 ¯ X s [0, 1] since x ≥ ψ. Combining the above with the range f(ψ) = ω (1 − ω) Ps(φ|ψ)f(φ)dφ. (C1) 0 of acceptable values of x given xy = ψ, we have propor- s≥0 tionality For products of HVs each independently undergoing 1{x ∈ [ψ, 1]} jump-dynamics, we have p(x|xy = ψ) = c , (C7) x P (φ|ψ) = (1 − σ)2s1 (φ) s ψ and the normalizing coefficient c is determined by inte- s s + (1 − (1 − σ) ) (1 − σ) p1(φ|ψ) (C2) gration: + (1 − (1 − σ)s)2 µ(φ), Z 1 Z 1 dx 1 = p(x|xy = ψ)dx = c = c ln(1/ψ) with µ(φ) denoting the product density of hidden vari- x 0 ψ (C8) ables and p1(φ|ψ) the product HV-density conditioned 1 ⇒ c = . on a single jump. Then, ln(1/ψ) ¯ f(ψ) = αf(ψ) Therefore, Z 1 X s s + ω ((1 − σ)(1 − ω)) (1 − (1 − σ) ) f(φ)p1(φ|ψ)dφ 1{x ∈ [ψ, 1]} 0 p(x|xy = ψ) = , (C9) s≥0 x ln(1/ψ) Z 1 X s s 2 + ω (1 − ω) (1 − (1 − σ) ) f(φ)µ(φ)dφ. as is confirmed numerically in Figure8. s≥0 0 (C3) R 1 Note that 0 f(φ)µ(φ)dφ = hfi = hki/n. Then, evaluat- 2. Finding p1(φ|ψ) ing sums,

f¯(ψ) = α f(ψ)+(α −α )f (ψ)+(1−2α +α )hfi, (C4) Now suppose that one variable, say y, undergoes a ran- 2 1 2 1 1 2 dom jump (i.e., is resampled) and thus becomes a new b uniform variable on [0, 1]. The equality xy = ψ no longer with αb = ω/(1−(1−ω)(1−σ) ), and the quantity f1(ψ) being deﬁned as holds, but since it did hold prior to the jump, the variable x remains distributed according to p(x|xy = ψ). Z Therefore the new product’s value, which we denote by f1(ψ) = f(φ)p1(φ|ψ)dφ, (C5) φ = xy0 (where y0 is the post-jump version of y), has a 19

¯ Density of hidden variable x given product xy = 3. Calculating f1(ψ)

1{x [ , 1]} = 0.00 p(x|xy = ) = x ln(1/ ) = 0.25 We now average the aﬃnity over p1(φ|ψ), to get the ) = 0.50 contribution to the eﬀective connection probability com- 1 = 0.75 10 -- theoretical ing from one hidden variable jumping. This goes as = . empirical y

x 1

| Z

x 100 ¯ ( f1(ψ) = f(φ)p1(φ|ψ)dφ

p 0 Z 1 min(1/φ, 1/ψ) − 1 = f(φ) dφ 0.0 0.2 0.4 0.6 0.8 1.0 0 ln(1/ψ) ! x 1 1 Z ψ Z 1 f(φ) = f(φ)dφ + dφ − 1 . ln(1/ψ) ψ 0 ψ φ FIG. 8. The probability density of the value of one member (C12) x of a product xy conditioned on xy = ψ. In the absence of Using Equation C12, we can compute the eﬀective con- the conditionality, both x and y are distributed uniformly on nection probability f¯(ψ) for a given aﬃnity-function f(ψ) [0, 1]. via Equation C4.

Density of product = xy0 given product xy = Appendix D: Walk-Dynamics 101

100 Here we describe walk-dynamics in detail. Through- out this work, walk-dynamics in 1D is simulated by ﬁrst )

| 10 1 mapping random variables to the unit interval (by the ( 1

p inverse-CDF method [183]), doing a random walk on = 0.05 min(1/ , 1/ ) 1 2 p1( | ) = 10 = 0.30 ln(1/ ) [0, 1], then mapping back. For any one-dimensional prob- = 0.55 ability density ν(x) where x ∈ +, we define random -- theoretical R x 10 3 = 0.80 . empirical R variable u(x) = Fν (x), where Fν (x) = 0 ν(y)dy. The 0.0 0.2 0.4 0.6 0.8 1.0 probability density of u(x) is the uniform on [0, 1]. A random walk on [0, 1] is constructed via addition of uniform noise in the range [−2σ, 2σ], parameterized by σ ∈ [0, 1]. That is, after rescaling we have hidden-variable dynamics FIG. 9. The probability density of the value φ of the product φ = xy0, where y0 is uniformly sampled after having previ- 1 {|u0 − u| ≤ 2σ} P (u0|u) = . (D1) ously had random value y, and where xy was conditioned h 4σ to have value xy = ψ. Without any conditioning, all three x, y, y0 have marginal density uniform on [0, 1]. Note that the choice of [−2σ, 2σ] results in a mean jump- length parameterized by σ: Z density p (φ|ψ) of the following form: 0 0 0 0 1 h|u − u |i = |u − u |Ph(u |u)du Z 1 Z 2σ (D2) 0 φ 1 0 1 p1(φ|ψ) = 1{y ∈ [0, 1]}p 0 xy = ψ 0 dy = |x|dx = σ 0 y y 4σ −2σ Z 1 1{φ/y0 ∈ [ψ, 1]} dy0 where boundary conditions have been neglected in the = 0 0 0 (φ/y ) ln(1/ψ) y above case; when implementing boundary conditions, one Z 1 0 1 0 0 needs only to adjust the probability density Ph(u |u) ac- = 1{φ/y ∈ [ψ, 1]}dy . cording to the circumstance. See AppendixE for the case φ ln(1/ψ) 0 (C10) of reflecting boundaries. (1) (1) (1) Continuing with a change of variables, Drawing h from ν, we initialize u = Fν (h ) and iteratively time-advance as per the above to ob- (t) T 1 Z 1/φ tain {u }t=1. We then simply transform back via p (φ|ψ) = 1{y0/φ ∈ [1, 1/ψ]}d(y0/φ) h(t) = F −1(u(t)), to obtain one-dimensional dynamics 1 ln(1/ψ) ν 0 whose stationary distribution is ν. min(1/φ, 1/ψ) − 1 = . In dimensions greater than 1, walk-dynamics can be ln(1/ψ) simulated by first taking the multidimensional inverse- (C11) CDF transform, mapping the space X to a unit cube. The above is validated numerically in Figure9. Walk-dynamics can then be performed with whatever 20 custom boundary conditions are required on that unit or cube (boundary conditions that correspond to those of  xˆ + r, xˆ < r, X ), and the results can then be mapped back to the orig- 1  inal space X . For example in the 2 model (Appendix Pˆ(ˆx) = 2r xˆ ∈ [r, 1 − r], (E4) H 2r F), increments of change in the angular and radial coor-  1 + r − x,ˆ xˆ > 1 − r. dinates were chosen to be independent; this option was taken for simplicity, but non-independent cases would where thex ˆ-dependent coefficients of the first and third also be interesting to explore. Any transitional proba- terms arise from reflections of the formx ˆ − (−r) and bility density preserving the uniform on the unit cube 1−(ˆx−1). We seek a function Z :[−r, 1+r] → [0, 1] that would fall within the same framework. encodes the reflection properties of the walk-dynamics. The necessary Z is given by  Appendix E: Walk-dynamics with reflecting −x,ˆ xˆ < 0, boundary conditions  Z(ˆx) = x,ˆ xˆ ∈ [0, 1], (E5)  In this appendix we study walk-dynamics on X = [0, 1] 2 − x,ˆ xˆ > 1. with reflecting boundary conditions under uniform noise. The values ofx ˆ mapping to a given value of x0, namely In particular, we show that the stationary density is uni- those making up the inverse of Z, are given by form on [0, 1]. In turn, that implies that arbitrary 1D dynamics with density ν(h) can be made into a random  {−x0, x0}, x0 < r, walk of this type, by mapping initial h-values to [0, 1]  0 0 0 via the inverse-CDF transform, performing the reflect- {xˆ : Z(ˆx) = x } = {x }, x ∈ [r, 1 − r], (E6) ing random walk on [0, 1], then transforming the random  {x0, 2 − x0}, x0 > 1 − r. walk trajectories back to the original space (see appendix D). Let us compute the derivative of Z(ˆx), neglecting the Let X = [0, 1] be the HV-space and denote by measure-zero points of 0 and 1: x ←-U[0, 1] the value of a hidden variable. Then let  xˆ ∈ [−r, 1 + r] be an intermediate variable defined as −1, xˆ < 0, dZ(ˆx)  xˆ = x + u, where u ←-U[−r, r] is the uniform additive = 1, xˆ ∈ (0, 1), (E7) dxˆ noise which we use to simulate walk-dynamics. Lastly,  −1, xˆ > 1. let x0 = Z(ˆx) be the reflected variable, where the function Z encodes the reflecting boundary conditions. Note We now transform to find the density after one step of 0 that values of x in the ranges [0, r] and [1 − r, 1] are dynamics, as obtained from one of two different of values ofx ˆ: the case when reflected, and the case when not reflected. To −1 0 0 X dZ(ˆx) transform the density ofx ˆ into that of x0, we write x0 P (x ) = Pˆ(ˆx) dxˆ as a function ofz ˆ, as x0 = Z(ˆx) and use the general- xˆ:Z(ˆx)=x0  ized change-of-variables formula for probability densities Pˆ(−x0) + Pˆ(x0), x0 < r, [184]. We denote the densities of x, x,ˆ x0 as P (x), Pˆ(ˆx),  = ˆ 0 0 and P 0(x0), respectively. Then we have the following: P (x ), x ∈ [r, 1 − r],  Pˆ(x0) + Pˆ(2 − x0), x0 > 1 − r. (E8) dZ(ˆx) −1  0 X 2r, x0 < r, P (x ) = P (ˆx), (E1)  dxˆ 1 0 xˆ:Z(ˆx)=x0 = 2r, x ∈ [r, 1 − r], 2r  2r, x0 > 1 − r. by the rules of how probabilities densities transform [184]. The density ofx ˆ given x is = 1. 1{xˆ ∈ [x − r, x + r]} 1{x ∈ [ˆx − r, xˆ + r]} Therefore, P 0(x0) = 1 for all x0 ∈ [0, 1], and thus the Pˆ(ˆx|x) = = . 2r 2r uniform distribution is the stationary distribution of re- (E2) flecting walk-dynamics. Since x is uniform on [0, 1] the density ofx ˆ is then

1 Z Appendix F: Hyperbolic walk-dynamics Pˆ(ˆx) = Pˆ(ˆx|x)dx 0 Z 1 ˜ ˜ (1) 1 (E3) To sample h = (˜r, θ), and also to sample hj (a coordi- = 1{x ∈ [ˆx − r, xˆ + r]}dx 2 2r 0 nate from the initial timestep, i.e. the static H model), 1 we ﬁrst draw two independent random variables Ur and = |[0, 1] ∪ [ˆx − r, xˆ + r]| , 2r Uθ, each from the uniform distribution on [0, 1]. These 21 are then set equal to the cumulative density functions of ferent transition probabilities: ˜ νrad and νang, evaluated atr ˜ and θ, respectively: 1→1 0 − 0 Pij = 1 fij ≥ fij + (1 − q )1{fij < fij}, ij (G2) Z θ˜ ˜ 0→0 0 + 0 θ Pij = 1 fij < fij + (1 − qij)1{fij ≥ fij}, Uθ = νang(θ)dθ = , 0 2π (F1) − 0 + 0 Z r˜ γ−1 with qij = 1 − fij/fij, qij = 1 − (1 − fij)/(1 − fij) as cosh 2 r˜ − 1 Ur = νrad(r)dr = γ−1 . deﬁned in SectionVI. The remaining probabilities are 0 cosh 2 R − 1 obtained by normalization:

˜ 1→0 1→1 − 0 From the above, we can solve to obtain h in terms of Pij = 1 − Pij = q 1 fij < fij , ij (G3) (Ur,Uθ): 0→1 0→0 + 0 Pij = 1 − Pij = qij1 fij ≥ fij . θ˜ = 2πU , θ G→G0 Noting that P 0 and (G|H) are both separable into 2 γ − 1 H,H P r˜ = cosh−1 1 + cosh R − 1 U . a product over ij : 1 ≤ i < j ≤ n, we write γ − 1 2 r 0 A →A0 (F2) G→G Y Aij 1−Aij ij ij (G|H)P 0 = f (1 − f ) P . In the temporal setting, those initial variables are set P H,H ij ij ij (1) (1) 1≤i

Y 0 Y 0 0 0 Appendix G: Stationarity with Link-Response fij (1 − fij) = P(G |H ). (G7) 0 0 ij:Aij =1 ij:Aij =0 In this appendix, we show that the static-model graph probability distribution is preserved via the effect of link- The left-hand side of the above is exactly the static response as described in SectionVI. Specifically, we show model’s graph probability distribution given a hidden- that variable configuration (see Equation5 of the main text). Thus the static hidden-variables model is the stationary Z ! distribution of time-advancements with the link-response X G→G0 0 0 P(G|H)PH,H0 ρ(H)dH = P(G |H ), (G1) mechanism. H G∈G To show that these results hold even upon inclusion of the link-resampling mechanism (allowing ω > 0), con- G(t)→G(t+1) (t+1) (t) (t+1) (t) sider the following reasoning. Regardless of what the where PH(t),H(t+1) = PG(G |G ,H ,H ). We for now set ω = 0 and later argue that link-resampling link-response step yielded, each node-pair undergoing does not influence the results in question. First, we note link-resampling at rate ω will result in either (a) link- that the transition probability given (H,H0) is separable: ing according to the connection probability of the newly 0 0 A →A G→G Q ij ij updated hidden-variable configuration (with probability P 0 = P , with transition probabil- H,H 1≤i

[1] Heeralal Janwa, Steven Massey, Julian Velev, and [8] Giulio Cimini, Tiziano Squartini, Fabio Saracco, Diego Bud Bhubaneswar Mishra, “On the origin of biomolec- Garlaschelli, Andrea Gabrielli, and Guido Caldarelli, ular networks,” Frontiers in genetics 10, 240 (2019). “The statistical physics of real-world networks,” Nature [2] Niklas Boers, Bedartha Goswami, Aljoscha Rheinwalt, Reviews Physics 1, 58–71 (2019). Bodo Bookhagen, Brian Hoskins, and JürgenKurths, [9] Hyunju Kim, Harrison B Smith, Cole Mathis, Jason “Complex networks reveal global pattern of extreme- Raymond, and Sara I Walker, “Universal scaling across rainfall teleconnections,” Nature 566, 373 (2019). biochemical networks on earth,” Science advances 5, [3] Luciano da Fontoura Costa, Osvaldo N Oliveira Jr, eaau0149 (2019). Gonzalo Travieso, Francisco Aparecido Rodrigues, [10] Paul L Krapivsky and Sidney Redner, “A statistical Paulino Ribeiro Villas Boas, Lucas Antiqueira, physics perspective on web growth,” Computer Net- Matheus Palhares Viana, and Luis Enrique Cor- works 39, 261–276 (2002). rea Rocha, “Analyzing and modeling real-world phe- [11] Albert-LászlóBarabásiand Réka Albert, “Emergence nomena with complex networks: a survey of applica- of scaling in random networks,” science 286, 509–512 tions,” Advances in Physics 60, 329–412 (2011). (1999). [4] Luis EC Rocha, Fredrik Liljeros, and Petter Holme, [12] Ginestra Bianconi and A-L Barabási,“Competition and “Information dynamics shape the sexual networks of multiscaling in evolving networks,” EPL (Europhysics internet-mediated prostitution,” Proceedings of the Na- Letters) 54, 436 (2001). tional Academy of Sciences 107, 5706–5711 (2010). [13] Guido Caldarelli, Andrea Capocci, Paolo De Los Rios, [5] Geoffrey B West, James H Brown, and Brian J Enquist, and Miguel A Munoz, “Scale-free networks from vary- “A general model for the origin of allometric scaling laws ing vertex intrinsic fitness,” Physical review letters 89, in biology,” Science 276, 122–126 (1997). 258702 (2002). [6] Romualdo Pastor-Satorras and Alessandro Vespignani, [14] Marc Barthelemy, Morphogenesis of spatial networks Evolution and structure of the Internet: A statistical (Springer, 2018). physics approach (Cambridge University Press, 2004). [15] Dmitri Krioukov, Fragkiskos Papadopoulos, Maksim [7] Muhua Zheng, Antoine Allard, Patric Hagmann, Yasser Kitsak, Amin Vahdat, and MariánBoguná,“Hyper- Alemán-Gómez, and M Angeles´ Serrano, “Geomet- bolic geometry of complex networks,” Physical Review ric renormalization unravels self-similarity of the multi- E 82, 036106 (2010). scale human connectome,” Proceedings of the National [16]M Angeles´ Serrano, Marián Boguná, and Francesc Academy of Sciences 117, 20244–20253 (2020). Sagués, “Uncovering the hidden geometry behind 23

metabolic networks,” Molecular biosystems 8, 843–850 Journal of Statistical Mechanics: Theory and Experi- (2012). ment 2014, P09029 (2014). [17] Maksim Kitsak, Fragkiskos Papadopoulos, and Dmitri [37] Henk W Volberda and Arie Y Lewin, “Co-evolutionary Krioukov, “Latent geometry of bipartite networks,” dynamics within and between firms: From evolution to Physical Review E 95, 032309 (2017). co-evolution,” Journal of management studies 40, 2111– [18] MariánBogunáand Romualdo Pastor-Satorras, “Class 2136 (2003). of correlated random networks with hidden variables,” [38] Michael Stoica and Minet Schindehutte, “Understand- Physical Review E 68, 036112 (2003). ing adaptation in small firms: Links to culture and per- [19] Mark Newman, Networks (Oxford university press, formance,” Journal of Developmental Entrepreneurship 2018). 4, 1 (1999). [20] Brian Karrer and Mark EJ Newman, “Stochastic block- [39] Kenneth M Johnson, “Demographic trends in rural and models and community structure in networks,” Physical small town america,” (2006), 10.34051/p/2020.6. review E 83, 016107 (2011). [40] Dowell Myers, “Demographic dynamism and metropoli- [21] Mathew Penrose et al., Random geometric graphs, Vol. 5 tan change: Comparing los angeles, new york, chicago, (Oxford university press, 2003). and washington, dc,” Housing Policy Debate 10, 919– [22] Pim van der Hoorn, Gabor Lippner, and Dmitri Kri- 954 (1999). oukov, “Sparse maximum-entropy random graphs with [41] Juste Raimbault, “Modeling the co-evolution of a given power-law degree distribution,” Journal of Sta- cities and networks,” arXiv preprint arXiv:1804.09430 tistical Physics 173, 806–844 (2018). (2018). [23] SebastiánRisau-Gusmánand DamiánH Zanette, “Con- [42] Siobhan G Kuhar, Lei Feng, Susana Vidan, M Elizabeth tact switching as a control strategy for epidemic out- Ross, Mary E Hatten, and Nathaniel Heintz, “Changing breaks,” Journal of theoretical biology 257, 52–60 patterns of gene expression define four stages of cerebel- (2009). lar granule neuron differentiation,” Development 117, [24] Suwakan Piankoranee and Surachate Limkumnerd, “Ef- 97–104 (1993). fects of global and local rewiring on sis epidemic adap- [43] Matthew A McCormack, Kenneth M Rosen, Lydia tive networks,” in Journal of Physics: Conference Se- Villa-Komaroff, and George D Mower, “Changes in im- ries, Vol. 1144 (IOP Publishing, 2018) p. 012080. mediate early gene expression during postnatal devel- [25] CristiánHuepe, Gerd Zschaler, Anne-Ly Do, and Thilo opment of cat cortex and cerebellum,” Molecular brain Gross, “Adaptive-network models of swarm dynamics,” research 12, 215–223 (1992). New Journal of Physics 13, 073022 (2011). [44] Mohammad Moradi Dalvand, Seyed Bahram Zahir [26] Vincent Marceau, Pierre-AndréNoël,Laurent Hébert- Azami, and Hadi Tarimoradi, “Long-term load fore- Dufresne, Antoine Allard, and Louis J Dubé,“Adaptive casting of iranian power grid using fuzzy and artificial networks: Coevolution of disease and topology,” Physi- neural networks,” in 2008 43rd International Univer- cal Review E 82, 036116 (2010). sities Power Engineering Conference (IEEE, 2008) pp. [27] Guven Demirel, Federico Vazquez, GA Böhme, and 1–4. Thilo Gross, “Moment-closure approximations for dis- [45] Matteo De Felice, Andrea Alessandri, and Paolo M crete adaptive networks,” Physica D: Nonlinear Phe- Ruti, “Electricity demand forecasting over italy: Poten- nomena 267, 68–80 (2014). tial benefits using numerical weather prediction mod- [28] Thilo Gross and Hiroki Sayama, “Adaptive networks,” els,” Electric Power Systems Research 104, 71–79 in Adaptive networks (Springer, 2009) pp. 1–8. (2013). [29] I Barry Crabtree and Stuart J Soltysiak, “Identifying [46] Natalia Pasichnyk, Mykola Dyvak, and Roman Pasich- and tracking changing interests,” International Journal nyk, “Mathematical modeling of website quality charac- on Digital Libraries 2, 38–53 (1998). teristics in dynamics,” in Journal of Applied Computer [30] Ben B Chapman, Christer Brönmark,Jan-Ake˚ Nilsson, Science, Vol. 22 (Technical University Press, 2014) pp. and Lars-Anders Hansson, “The ecology and evolution 171–183. of partial migration,” Oikos 120, 1764–1775 (2011). [47] Sandra J Bean, “Emerging and continuing trends in vac- [31] Yannis Vardanis, Raymond HG Klaassen, Roine Strand- cine opposition website content,” Vaccine 29, 1874–1880 berg, and Thomas Alerstam, “Individuality in bird mi- (2011). gration: routes and timing,” Biology letters 7, 502–505 [48] Vincenzo Nicosia, John Tang, Cecilia Mascolo, Mirco (2011). Musolesi, Giovanni Russo, and Vito Latora, “Graph [32] Sonia Altizer, Rebecca Bartel, and Barbara A Han, metrics for temporal networks,” in Temporal networks “Animal migration and infectious disease risk,” science (Springer, 2013) pp. 15–40. 331, 296–302 (2011). [49] Elisenda Ortiz, Michele Starnini, and M Angeles´ Ser- [33] Martin Pfeffer and Gerhard Dobler, “Emergence of rano, “Navigability of temporal networks in hyperbolic zoonotic arboviruses by animal trade and migration,” space,” Scientific reports 7, 15054 (2017). Parasites & vectors 3, 35 (2010). [50] Dane Taylor, Sean A Myers, Aaron Clauset, Mason A [34] Andrew M Hein, Chen Hou, and James F Gillooly, “En- Porter, and Peter J Mucha, “Eigenvector-based central- ergetic and biomechanical constraints on animal migra- ity measures for temporal networks,” Multiscale Mod- tion distance,” Ecology letters 15, 104–110 (2012). eling & Simulation 15, 537–574 (2017). [35] Scott P Carroll, Andrew P Hendry, David N Reznick, [51] Hyoungshick Kim and Ross Anderson, “Temporal node and Charles W Fox, “Evolution on ecological time- centrality in complex networks,” Physical Review E 85, scales,” Functional Ecology 21, 387–393 (2007). 026107 (2012). [36] Torsten Held, Armita Nourmohammad, and Michael [52] Raj Kumar Pan and Jari Saramäki,“Path lengths, cor- Lässig,“Adaptive evolution of molecular phenotypes,” 24

relations, and centrality in temporal networks,” Physi- (2015) pp. 1079–1087. cal Review E 84, 016105 (2011). [70] Catherine Matias and Vincent Miele, “Statistical clus- [53] Aming Li, Sean P Cornelius, Y-Y Liu, Long Wang, and tering of temporal networks through a dynamic stochas- A-L Barabási,“The fundamental advantages of tempo- tic block model,” Journal of the Royal Statistical Soci- ral networks,” Science 358, 1042–1046 (2017). ety: Series B (Statistical Methodology) 79, 1119–1141 [54] Suyu Liu, Nicola Perra, MártonKarsai, and Alessandro (2017). Vespignani, “Controlling contagion processes in activity [71] Marianna Pensky, Teng Zhang, et al., “Spectral cluster- driven networks,” Physical review letters 112, 118702 ing in the dynamic stochastic block model,” Electronic (2014). Journal of Statistics 13, 678–709 (2019). [55] Nicola Perra, Andrea Baronchelli, Delia Mocanu, Bruno [72] Amir Ghasemian, Pan Zhang, Aaron Clauset, Cristo- Gon¸calves, Romualdo Pastor-Satorras, and Alessandro pher Moore, and Leto Peel, “Detectability thresholds Vespignani, “Random walks and search in time-varying and optimal algorithms for community structure in dy- networks,” Physical review letters 109, 238701 (2012). namic networks,” Physical Review X 6, 031005 (2016). [56] Ashwin Paranjape, Austin R Benson, and Jure [73] Paolo Barucca, Fabrizio Lillo, Piero Mazzarisi, and Leskovec, “Motifs in temporal networks,” in Proceedings Daniele Tantari, “Disentangling group and link persis- of the Tenth ACM International Conference on Web tence in dynamic stochastic block models,” Journal of Search and Data Mining (ACM, 2017) pp. 601–610. Statistical Mechanics: Theory and Experiment 2018, [57] Petter Holme, “Temporal network structures control- 123407 (2018). ling disease spreading,” Physical Review E 94, 022305 [74] Daniel K Sewell and Yuguo Chen, “Latent space mod- (2016). els for dynamic networks with weighted edges,” Social [58] Quan-Hui Liu, Xinyue Xiong, Qian Zhang, and Nicola Networks 44, 105–116 (2016). Perra, “Epidemic spreading on time-varying multiplex [75] Bomin Kim, Kevin H Lee, Lingzhou Xue, and Xiaoyue networks,” Physical Review E 98, 062303 (2018). Niu, “A review of dynamic network models with latent [59] Matthieu Nadini, Kaiyuan Sun, Enrico Ubaldi, Michele variables,” Statistics surveys 12, 105 (2018). Starnini, Alessandro Rizzo, and Nicola Perra, “Epi- [76] Purnamrita Sarkar, Sajid M Siddiqi, and Geogrey J demic spreading in modular time-varying networks,” Gordon, “A latent space approach to dynamic embed- Scientific reports 8, 2352 (2018). ding of co-occurrence data,” in Artificial Intelligence [60] Naoki Masuda and Petter Holme, Temporal network epi- and Statistics (2007) pp. 420–427. demiology (Springer, 2017). [77] Purnamrita Sarkar and Andrew W Moore, “Dynamic [61] Kaiyuan Sun, Andrea Baronchelli, and Nicola Perra, social network analysis using latent space models,” in “Contrasting effects of strong ties on sir and sis pro- Advances in Neural Information Processing Systems cesses in temporal networks,” The European Physical (2006) pp. 1145–1152. Journal B 88, 326 (2015). [78] Yuval Peres, Alistair Sinclair, Perla Sousi, and Alexan- [62] Dandan Li, Dun Han, Jing Ma, Mei Sun, Lixin Tian, dre Stauffer, “Mobile geometric graphs: Detection, cov- Timothy Khouw, and H Eugene Stanley, “Opinion dy- erage and percolation,” Probability Theory and Related namics in activity-driven networks,” EPL (Europhysics Fields 156, 273–305 (2013). Letters) 120, 28002 (2018). [79] Andrea Clementi and Riccardo Silvestri, “Parsimonious [63] Daniel M Dunlavy, Tamara G Kolda, and Evrim Acar, flooding in geometric random-walks,” Journal of Com- “Temporal link prediction using matrix and tensor fac- puter and System Sciences 81, 219–233 (2015). torizations,” ACM Transactions on Knowledge Discov- [80] James R Norris, Markov chains, 2 (Cambridge univer- ery from Data (TKDD) 5, 10 (2011). sity press, 1998). [64] Yugchhaya Dhote, Nishchol Mishra, and Sanjeev [81] Ehrhard Behrends, Introduction to Markov chains, Vol. Sharma, “Survey and analysis of temporal link pre- 228 (Springer, 2000). diction in online social networks,” in 2013 Interna- [82] Ginestra Bianconi, Multilayer networks: structure and tional Conference on Advances in Computing, Commu- function (Oxford university press, 2018). nications and Informatics (ICACCI) (IEEE, 2013) pp. [83] Manlio De Domenico, Albert Solé-Ribalta,Emanuele 1178–1183. Cozzo, Mikko Kivelä,Yamir Moreno, Mason A Porter, [65] Marta Sarzynska, Elizabeth A Leicht, Gerardo Chow- Sergio Gómez, and Alex Arenas, “Mathematical for- ell, and Mason A Porter, “Null models for community mulation of multilayer networks,” Physical Review X 3, detection in spatially embedded, temporal networks,” 041022 (2013). Journal of Complex Networks 4, 363–406 (2016). [84] Peter T Landsberg, “Foundations of thermodynamics,” [66] Laetitia Gauvin, AndréPanisson, and Ciro Cattuto, Reviews of modern physics 28, 363 (1956). “Detecting the community structure and activity pat- [85] Nicola Perra, Bruno Goncalves, Romualdo Pastor- terns of temporal networks: a non-negative tensor fac- Satorras, and Alessandro Vespignani, “Activity driven torization approach,” PloS one 9, e86028 (2014). modeling of time varying networks,” Scientific Reports [67] Tiago P Peixoto and Martin Rosvall, “Modelling se- 2, 469 (2012). quences and temporal networks with dynamic commu- [86] Fragkiskos Papadopoulos and Marco Antonio Rodr´ıguez nity structures,” Nature communications 8, 582 (2017). Flores, “Latent geometry and dynamics of proximity [68] Kevin S Xu and Alfred O Hero, “Dynamic stochastic networks,” Physical Review E 100, 052313 (2019). blockmodels for time-evolving social networks,” IEEE [87] Béla Bollobás, Svante Janson, and Oliver Rior- Journal of Selected Topics in Signal Processing 8, 552– dan, “The phase transition in inhomogeneous random 562 (2014). graphs,” Random Structures & Algorithms 31, 3–122 [69] Kevin Xu, “Stochastic block transition models for dy- (2007). namic networks,” in Artificial Intelligence and Statistics [88] Roberto Imbuzeiro Oliveira, “Concentration of the adja- 25

cency matrix and of the laplacian in random graphs with Vito Latora, “Disease spreading in populations of mov- independent edges,” arXiv preprint arXiv:0911.0600 ing agents,” EPL (Europhysics Letters) 82, 38002 (2009). (2008). [89] Linyuan Lu and Xing Peng, “Spectra of edge- [110] Mathew D Penrose et al., “Connectivity of soft random independent random graphs,” arXiv preprint geometric graphs,” The Annals of Applied Probability arXiv:1204.6207 (2012). 26, 986–1028 (2016). [90] Paul Erdos and Alfred Renyi, “On random graphs i,” [111] Michael Wilsher, Carl P Dettmann, and Ayalvadi Publ. math. debrecen 6, 18 (1959). Ganesh, “Connectivity in one-dimensional soft random [91] Carolyn J Anderson, Stanley Wasserman, and Kather- geometric graphs,” arXiv preprint arXiv:2007.06301 ine Faust, “Building stochastic blockmodels,” Social (2020). networks 14, 137–161 (1992). [112] Marcus Kaiser and Claus C Hilgetag, “Spatial growth [92] Harrison Hartle, Brennan Klein, Stefan McCabe, of real-world networks,” Physical Review E 69, 036103 Alexander Daniels, Guillaume St-Onge, Charles Mur- (2004). phy, and Laurent Hébert-Dufresne, “Network compari- [113] Carl P Dettmann and Orestis Georgiou, “Random geo- son and the within-ensemble graph distance,” Proceed- metric graphs with general connection functions,” Phys- ings of the Royal Society A 476, 20190744 (2020). ical Review E 93, 032313 (2016). [93] Richard W Hamming, “Error detecting and error cor- [114] Ivan Voitalov, Pim van der Hoorn, Maksim Kit- recting codes,” The Bell system technical journal 29, sak, Fragkiskos Papadopoulos, and Dmitri Krioukov, 147–160 (1950). “Weighted hypersoft configuration model,” Physical Re- [94] Jeffrey E Steif, “A survey of dynamical percolation,” view Research 2, 043157 (2020). in Fractal geometry and stochastics IV (Springer, 2009) [115] Fan Chung and Linyuan Lu, “Connected components in pp. 145–174. random graphs with given expected degree sequences,” [95] Yuval Peres and Jeffrey E Steif, “The number of infinite Annals of combinatorics 6, 125–145 (2002). clusters in dynamical percolation,” Probability theory [116] Ilkka Norros and Hannu Reittu, “On a conditionally and related fields 111, 141–165 (1998). poissonian graph process,” Advances in Applied Proba- [96] Davar Khoshnevisan, “Dynamical percolation on gen- bility 38, 59–75 (2006). eral trees,” Probability theory and related fields 140, [117] Karl Bringmann, Ralph Keusch, and Johannes Lengler, 169–193 (2008). “Average distance in a general class of scale-free [97] Matthew I Roberts, Batı S¸engül, et al., “Exceptional networks with underlying geometry,” arXiv preprint times of the critical dynamical erd˝os–rényi graph,” The arXiv:1602.05712 (2016). Annals of Applied Probability 28, 2275–2308 (2018). [118] Tobias Friedrich and Anton Krohmer, “On the diam- [98] RaphaëlRossignol, “Scaling limit of dynamical perco- eter of hyperbolic random graphs,” SIAM Journal on lation on critical erdös-rényi random graphs,” arXiv Discrete Mathematics 32, 1314–1334 (2018). preprint arXiv:1710.09101 (2020). [119] Ali Faqeeh, Saeed Osat, and Filippo Radicchi, “Char- [99] Andrea EF Clementi, Claudio Macci, Angelo Monti, acterizing the analogy between hyperbolic embedding Francesco Pasquale, and Riccardo Silvestri, “Flooding and community structure of complex networks,” Physi- time of edge-markovian evolving graphs,” SIAM journal cal review letters 121, 098301 (2018). on discrete mathematics 24, 1694–1712 (2010). [120] Guillermo Garc´ıa-Pérez, Marián Boguñá, and [100] John Whitbeck, Vania Conan, and Marcelo Dias M Angeles´ Serrano, “Multiscale unfolding of real net- de Amorim, “Performance of opportunistic epidemic works by geometric renormalization,” Nature Physics routing on edge-markovian dynamic graphs,” IEEE 14, 583–589 (2018). Transactions on communications 59, 1259–1263 (2011). [121] Mehran Kardar, Statistical physics of particles (Cam- [101] Aurelie Faure de Pebeyre, Fabien Tarissan, and Julien bridge University Press, 2007). Sopena, “On the relevance of the edge-markovian evolv- [122] Xiao Zhang, Cristopher Moore, and Mark EJ New- ing graph model for real mobile networks,” in 2013 IFIP man, “Random graph models for dynamic networks,” Wireless Days (WD) (IEEE, 2013) pp. 1–6. The European Physical Journal B 90, 200 (2017). [102] Tiago P Peixoto, “Entropy of stochastic blockmodel en- [123] Michel Mandjes, Nicos Starreveld, RenéBekker, and sembles,” Physical Review E 85, 056122 (2012). Peter Spreij, “Dynamic erd˝os-rényi graphs,” in Comput- [103] Tiago P Peixoto, “Nonparametric bayesian inference of ing and Software Science (Springer, 2019) pp. 123–140. the microcanonical stochastic block model,” Physical [124] Andrea EF Clementi, Francesco Pasquale, Angelo Review E 95, 012317 (2017). Monti, and Riccardo Silvestri, “Information spreading [104] Stephen G Kobourov, “Spring embedders and force in stationary markovian evolving graphs,” in 2009 IEEE directed graph drawing algorithms,” arXiv preprint International Symposium on Parallel & Distributed Pro- arXiv:1201.3011 (2012). cessing (IEEE, 2009) pp. 1–12. [105] Bo Söderberg, “Random graphs with hidden color,” [125] Ruijie Du, Hanxing Wang, and Yunbin Fu, Physical Review E 68, 015102 (2003). “Continuous-time independent edge-markovian random [106] Bo Söderberg, “General formalism for inhomogeneous graph process,” Chinese Annals of Mathematics, Series random graphs,” Physical review E 66, 066121 (2002). B 37, 73–82 (2016). [107] Alfonso Allen-Perkins, “Random spherical graphs,” [126] Ioannis Lamprou, Russell Martin, and Paul Spirakis, Physical Review E 98, 032310 (2018). “Cover time in edge-uniform stochastically-evolving [108] Duncan J Watts and Steven H Strogatz, “Collective dy- graphs,” Algorithms 11, 149 (2018). namics of ‘small-world’networks,” nature 393, 440–442 [127] Lorenzo Zino, Alessandro Rizzo, and Maurizio Por- (1998). firi, “Continuous-time discrete-distribution theory for [109] Arturo Buscarino, Luigi Fortuna, Mattia Frasca, and 26

activity-driven networks,” Physical review letters 117, International Conference on Information and Knowl- 228–302 (2016). edge Management (2019) pp. 469–478. [128] Iacopo Pozzana, Kaiyuan Sun, and Nicola Perra, “Epi- [145] Yu Xie, Chunyi Li, Bin Yu, Chen Zhang, and Zhouhua demic spreading on activity-driven networks with at- Tang, “A survey on dynamic network embedding,” tractiveness,” Physical Review E 96, 042310 (2017). arXiv preprint arXiv:2006.08093 (2020). [129] AngélicaSousa da Mata and Romualdo Pastor-Satorras, [146] Mounir Haddad, Cécile Bothorel, Philippe Lenca, “Slow relaxation dynamics and aging in random walks and Dominique Bedart, “Temporalnode2vec: Tempo- on activity driven temporal networks,” The European ral node embedding in temporal networks,” in Inter- Physical Journal B 88, 12 (2015). national Conference on Complex Networks and Their [130] Laura Alessandretti, Kaiyuan Sun, Andrea Baronchelli, Applications (Springer, 2019) pp. 891–902. and Nicola Perra, “Random walks on activity-driven [147] Di Jin, Sungchul Kim, Ryan A Rossi, and Danai networks with attractiveness,” Physical Review E 95, Koutra, “From static to dynamic node embeddings,” 052318 (2017). arXiv preprint arXiv:2009.10017 (2020). [131] Alessandro Rizzo and Maurizio Porfiri, “Innovation dif- [148] Simeon Spasov, Alessandro Di Stefano, Pietro Lio, and fusion on time-varying activity driven networks,” The Jian Tang, “Grade: Graph dynamic embedding,” arXiv European Physical Journal B 89, 20 (2016). preprint arXiv:2007.08060 (2020). [132] Michele Starnini and Romualdo Pastor-Satorras, “Tem- [149] Chuanchang Chen, Yubo Tao, and Hai Lin, “Dynamic poral percolation in activity-driven networks,” Physical network embeddings for network evolution analysis,” Review E 89, 032807 (2014). arXiv preprint arXiv:1906.09860 (2019). [133] Adrian Silvescu and Vasant Honavar, “Temporal [150] Pengyu Cheng, Yitong Li, Xinyuan Zhang, Liqun Chen, boolean network models of genetic networks and their David Carlson, and Lawrence Carin, “Dynamic em- inference from gene expression time series,” Complex bedding on textual networks via a gaussian process,” in systems 13, 61–78 (2001). Proceedings of the AAAI Conference on Artificial Intel- [134] Sophie Lebre, Jennifer Becq, Frederic Devaux, ligence, Vol. 34 (2020) pp. 7562–7569. Michael PH Stumpf, and Gaelle Lelandais, “Statis- [151] Linhong Zhu, Dong Guo, Junming Yin, Greg Ver Steeg, tical inference of the time-varying structure of gene- and Aram Galstyan, “Scalable temporal latent space in- regulation networks,” BMC systems biology 4, 130 ference for link prediction in dynamic social networks,” (2010). IEEE Transactions on Knowledge and Data Engineering [135] Steve Hanneke, Wenjie Fu, Eric P Xing, et al., “Discrete 28, 2765–2777 (2016). temporal models of social networks,” Electronic Journal [152] John Boaz Lee, Giang Nguyen, Ryan A Rossi, Nes- of Statistics 4, 585–605 (2010). reen K Ahmed, Eunyee Koh, and Sungchul Kim, [136] Andrew Mellor, “Event graphs: Advances and applica- “Dynamic node embeddings from edge streams,” IEEE tions of second-order time-unfolded temporal network Transactions on Emerging Topics in Computational In- models,” Advances in Complex Systems 22, 1950006 telligence (2020), 10.1109/tetci.2020.3011432. (2019). [153] Uriel Singer, Ido Guy, and Kira Radinsky, “Node [137] Roger E Miles, “On the homogeneous planar poisson embedding over temporal graphs,” arXiv preprint point process,” Mathematical Biosciences 6, 85–127 arXiv:1903.08889 (2019). (1970). [154] Srijan Kumar, Xikun Zhang, and Jure Leskovec, [138] Jesper Møller and Rasmus P Waagepetersen, “Mod- “Learning dynamic embeddings from temporal interac- ern statistics for spatial point processes,” Scandinavian tions,” arXiv preprint arXiv:1812.02289 (2018). Journal of Statistics 34, 643–684 (2007). [155] Fragkiskos Papadopoulos, Constantinos Psomas, and [139] Matthias Reitzner, “Poisson point processes: large de- Dmitri Krioukov, “Network mapping by replaying hy- viation inequalities for the convex distance,” (2013), perbolic growth,” IEEE/ACM Transactions on Net- 10.1214/ecp.v18-2851. working 23, 198–211 (2014). [140] SR Srinivasa Varadhan, Stochastic processes, Vol. 16 [156] Piet Van Mieghem, Huijuan Wang, Xin Ge, Siyu Tang, (American Mathematical Soc., 2007). and Fernando A Kuipers, “Influence of assortativity and [141] P Mazzarisi, P Barucca, F Lillo, and D Tantari, “A degree-preserving rewiring on the spectra of networks,” dynamic network model with persistent links and node- The European Physical Journal B 76, 643–652 (2010). specific latent variables, with an application to the in- [157] Jean-Gabriel Young, Giovanni Petri, Francesco Vac- terbank market,” European Journal of Operational Re- carino, and Alice Patania, “Construction of and effi- search 281, 50–65 (2020). cient sampling from the simplicial configuration model,” [142] Ranran Bian, Yun Sing Koh, Gillian Dobbie, and Physical Review E 96, 032312 (2017). Anna Divoli, “Network embedding and change mod- [158] Tom Bannink, Remco van der Hofstad, and Clara eling in dynamic heterogeneous networks,” in Proceed- Stegehuis, “Switch chain mixing times and triangle ings of the 42nd International ACM SIGIR Conference counts in simple random graphs with given degrees,” on Research and Development in Information Retrieval Journal of Complex Networks 7, 210–225 (2019). (2019) pp. 861–864. [159] Shankar Bhamidi, Guy Bresler, and Allan Sly, “Mix- [143] Palash Goyal, Sujit Rokka Chhetri, Ninareh Mehrabi, ing time of exponential random graphs,” in 2008 49th Emilio Ferrara, and Arquimedes Canedo, “Dynam- Annual IEEE Symposium on Foundations of Computer icgem: A library for dynamic graph embedding meth- Science (IEEE, 2008) pp. 803–812. ods,” arXiv preprint arXiv:1811.10734 (2018). [160] Ryan DeMuse, Terry Easlick, and Mei Yin, “Mixing [144] Yuanfu Lu, Xiao Wang, Chuan Shi, Philip S Yu, and time of vertex-weighted exponential random graphs,” Yanfang Ye, “Temporal network embedding with micro- Journal of Computational and Applied Mathematics and macro-dynamics,” in Proceedings of the 28th ACM 362, 443–459 (2019). 27

[161] Romualdo Pastor-Satorras, Claudio Castellano, Piet networks,” Computers & Mathematics with Applica- Van Mieghem, and Alessandro Vespignani, “Epidemic tions 65, 1645–1664 (2013). processes in complex networks,” Reviews of modern [174] Krzysztof Choromański,Micha l Matuszak, and Jacek physics 87, 925 (2015). Mi¸ekisz,“Scale-free graph with preferential attachment [162] Sergey N Dorogovtsev and Jose FF Mendes, “Evolution and evolving internal vertex structure,” Journal of Sta- of networks,” Advances in physics 51, 1079–1187 (2002). tistical Physics 151, 1175–1183 (2013). [163] Cristopher Moore, Gourab Ghoshal, and Mark EJ New- [175] Guido Caldarelli, Andrea Capocci, and Diego Gar- man, “Exact solutions for models of evolving networks laschelli, “A self-organized model for network evolu- with addition and deletion of nodes,” Physical Review tion,” The European Physical Journal B 64, 585–591 E 74, 036121 (2006). (2008). [164] Heiko Bauke, Cristopher Moore, Jean-Baptiste [176] Lia Papadopoulos, Jason Z Kim, JürgenKurths, and Rouquier, and David Sherrington, “Topological phase Danielle S Bassett, “Development of structural correla- transition in a network model with preferential at- tions and synchronization from adaptive rewiring in net- tachment and node removal,” The European Physical works of kuramoto oscillators,” Chaos: An Interdisci- Journal B 83, 519–524 (2011). plinary Journal of Nonlinear Science 27, 073115 (2017). [165] Luca Becchetti, Andrea Clementi, Francesco Pasquale, [177] RTAJ Leenders, “Longitudinal behavior of network Luca Trevisan, and Isabella Ziccardi, “Expansion and structure and actor attributes: modeling interdepen- flooding in dynamic random networks with node churn,” dence of contagion and selection,” Evolution of social arXiv preprint arXiv:2007.14681 (2020). networks 1, 165–184 (1997). [166] Dmitri Krioukov and Massimo Ostilli, “Duality between [178] Young-Ho Eom, Stefano Boccaletti, and Guido Cal- equilibrium and growing networks,” Physical Review E darelli, “Concurrent enhancement of percolation and 88, 022808 (2013). synchronization in adaptive networks,” Scientific re- [167] Tamar Schlick, Molecular modeling and simulation: ports 6, 1–7 (2016). an interdisciplinary guide: an interdisciplinary guide, [179] Marco Mancastroppa, Raffaella Burioni, Vittoria Col- Vol. 21 (Springer Science & Business Media, 2010). izza, and Alessandro Vezzani, “Active and inactive [168] Marco Antonio Rodr´ıguez Flores and Fragkiskos Pa- quarantine in epidemic spreading on adaptive activity- padopoulos, “Similarity forces and recurrent compo- driven networks,” arXiv preprint arXiv:2004.07902 nents in human face-to-face interaction networks,” (2020), 10.1103/physreve.102.020301. Physical review letters 121, 258301 (2018). [180] Genki Ichinose, Yoshiki Satotani, Hiroki Sayama, and [169] Emmanuel Jacob and Peter Mörters,“The contact pro- Takashi Nagatani, “Reduced mobility of infected agents cess on scale-free networks evolving by vertex updat- suppresses but lengthens disease in biased random ing,” Royal Society open science 4, 170081 (2017). walk,” arXiv preprint arXiv:1807.01195 (2018). [170] HäggströmOlle, Peres Yuval, and E Steif Jeffrey, “Dy- [181] Nicholas J Gotelli, “Research frontiers in null model namical percolation,” in Annales de l’Institut Henri analysis,” Global ecology and biogeography 10, 337–343 Poincare (B) Probability and Statistics, Vol. 33 (Else- (2001). vier, 1997) pp. 497–528. [182] Georgos Siganos, Michalis Faloutsos, Petros Faloutsos, [171] Christophe Garban, Gábor Pete, and Oded Schramm, and Christos Faloutsos, “Power laws and the as-level in- “The scaling limits of near-critical and dynamical perco- ternet topology,” IEEE/ACM Transactions on network- lation,” Journal of the European Mathematical Society ing 11, 514–524 (2003). 20, 1195–1268 (2018). [183] Luc Devroye, “Nonuniform random variate generation,” [172] Kevin E Bassler, Erwin Frey, and RKP Zia, “Coevolu- Handbooks in operations research and management sci- tion of nodes and links: Diversity-driven coexistence in ence 13, 83–121 (2006). cyclic competition of three species,” Physical Review E [184] Patrick Billingsley, Probability and measure (John Wi- 99, 022309 (2019). ley & Sons, 2008). [173] Hiroki Sayama, Irene Pestov, Jeffrey Schmidt, Ben- [185] Fragkiskos Papadopoulos, Maksim Kitsak, M Angeles´ jamin James Bush, Chun Wong, Junichi Yamanoi, and Serrano, MariánBoguná, and Dmitri Krioukov, “Pop- Thilo Gross, “Modeling complex systems with adaptive ularity versus similarity in growing networks,” Nature 489, 537–540 (2012).