On reversible cascades in scale-free and Erdős-Rényi random graphs

Let $G(V,E)$ be a simple directed graph, $\rho\in(\,0,1\,]$ and $S\subseteq V,$ where each vertex of $G$ can be in one of two states, active or inactive. The synchronous reversible cascade proceeds in rounds. In round zero, only the vertices in $S,$ called the seeds, are active. In round $i+1,$ a vertex with a positive indegree is activated (resp., deactivated) if at least (resp., less than) a $\rho$ fraction of its in-neighbors are active in round $i,$ where $i\in\mathbb{N}$. For $k\in\mathbb{N},$ define $\text{min-seed}^{(k)}(G,\rho)$ to be the minimum number of seeds needed so that all vertices will be active in or before round $k$.

The synchronous reversible cascade above is the same as the local interaction game with full rationality except that the latter is defined on infinite graphs with finite degrees [65, 51]. For local interaction games, Morris [65] studies conditions allowing finitely many seeds to activate each vertex sooner or later. Blume [13], Ellison [37], Young [85, 86] and Montanari and Saberi [61] study a variant where there is another variable specifying the vertices’ degree of rationality and the states are updated according to a Poisson clock. In their fully rational scenario, updating the state of a vertex $v$ means activating or deactivating $v,$ respectively, if at least or less than a certain fraction of $v$’s neighbors are active. They analyze the expected waiting time until all or most vertices enter the same state.

Consider the special cases of the synchronous reversible cascade with $\rho=1/2$ or $\rho=1/2+1/(2\,|\,V\,|)$. So a vertex is activated in a round if the simple or strict majority of its in-neighbors are active in the previous round; it is deactivated otherwise. Both special cases, among other similar dynamics, are suitable for modeling transient faults in majority-based fault-tolerant systems [74, 41, 42]. Call a set of seeds a $k$-round monopoly if it activates all vertices in or before round $k,$ where $k\in\mathbb{N}$. Any finite-round monopoly is called a dynamic monopoly. If a set of seeds does not lead to the deactivation of active vertices in any round, then it is said to be monotone. Peleg [73] shows an $\Omega(\sqrt{|\,V\,|}\,)$ lower bound on the minimum size of $2$-round monopolies in any simple undirected graph $G(V,E)$. Bounds are known on the minimum size of monotone dynamic monopolies in planar graphs [41], toroidal meshes, torus cordales, torus serpentini [42] and simple undirected graphs [73]. Berger [10] shows the existence of constant-size dynamic monopolies in an infinite family of simple undirected graphs. Optimal bounds on the minimum size of $1$-round monopolies are also derived for planar graphs, hypercubes, graphs with a given girth, graphs with diameter $2$ [56] and simple undirected graphs [56, 72]. Bermond et al. [11] derive bounds on the minimum size of $1$-round monopolies under the variant where a vertex $v$ is activated or deactivated, respectively, if the majority or minority of the vertices within distance $r$ from $v$ are active, $r\geq 2$. Peleg [74] surveys the above and many related results.

As an important variant, the irreversible cascade on a graph forbids the deactivation of vertices. In its most general form, a vertex $v$ is activated in a round if at least $\phi(v)$ of its in-neighbors are active in the previous round, where $\phi(v)\in\mathbb{N}$. It is sometimes defined as an asynchronous process because the order of activating the vertices does not affect the set of vertices that will be active. With various threshold functions $\phi(\cdot),$ irreversible cascades describe the propagation of permanent faults in majority-based systems [59, 40, 41, 42], spread of diseases [36], complex propagation [20, 44, 80], socio-economic cascades [47, 83] and cascading failures in infrastructure or organizational networks [83]. Assuming irreversible cascades, bounds are derived on the minimum number of seeds activating all vertices eventually for complete trees, rings, butterflies, wrapped butterflies, cube-connected cycles, shuffle-exchange graphs, DeBruijn graphs, hypercubes [59, 41], toroidal meshes [57, 42, 76, 55, 36], torus cordales, torus serpentini [57, 42, 36], chordal rings [40], multidimensional cubes [7], complete multipartite graphs, regular graphs [36], Erdős-Rényi random graphs [23, 22, 21], undirected connected graphs [24] and directed graphs with positive indegrees [24, 1]. More bounds are derived on the minimum number of seeds activating all vertices in one round for near-regular graphs, bounded-degree graphs [72] and, under several variants, simple undirected graphs [11].

A family of undirected graphs, $\{G_{n}(V_{n},E_{n})\mid|\,V_{n}\,|=n\}_{n=1}^{\infty},$ is scale-free if there is a constant $2<\gamma<3$ such that for sufficiently large $k\in\mathbb{Z}^{+}$ and as $n\to\infty,$ the fraction of vertices with degree $k$ in $G_{n}$ is proportional to $1/k^{\gamma}$. Directed scale-free graphs are defined similarly by using the in- or outdegrees of vertices instead. Many socio-economic, physical, biological and semantic networks are scale-free. Generative models of scale-free networks include preferential attachment models [5, 8, 6, 16, 68], copying models [18, 19, 52, 54], growth-deletion models [33, 29, 39], random-surfer models [12, 25], traffic-driven model [9], heuristically optimized trade-off model [38], hybrid models [75, 31, 71], semantic growth model [82] and random-graph models with given expected degree sequences [28, 30]. Many of these suitably describe the Webgraph [17, 35].

Let $G(V,E)$ be an undirected connected graph with diameter $\Delta,$ $\gamma>2$ be a constant and $\rho\in(\,0,1\,]$ such that the fraction of vertices with degree $k$ in $G$ is $O(1/k^{\gamma})$. This paper proves $\text{min-seed}^{(\Delta)}(G,\rho)=O(\lceil\rho^{\gamma-1}\,|\,V\,|\rceil)$. As scale-free graphs typically have small distances between vertices [70, 32, 15, 30, 31], activating all vertices within $\Delta$ rounds may be fast. Furthermore, the $O(\lceil\rho^{\gamma-1}\,|\,V\,|\rceil)$ bound continues to hold even if the synchronous reversible cascade is modified to proceed asynchronously instead.

For $n\in\mathbb{Z}^{+}$ and $p\in[\,0,1\,],$ the Erdős-Rényi random graph $G(n,p)$ is a simple undirected graph with vertices $1,2,\ldots,n$ where each of the $\binom{n}{2}$ possible edges appears independently with probability $p$ [14]. Assuming irreversible cascades, Chang and Lyuu [22, 21] consider the case where a non-isolated vertex of $G(n,p)$ is activated when at least a $\rho\in(\,0,1\,]$ fraction of its neighbors are active. They prove that for a sufficiently large constant $\beta>0,$ $n\in\mathbb{Z}^{+},$ $\delta\in\{1/n,2/n,\ldots,n/n\},$ $p\geq\beta(\ln(e/{\min\{\delta,\rho\}}))/(\rho n)$ and with probability $1-n^{-\Omega(1)}$ over $G(n,p),$ the minimum number of seeds needed to eventually activate at least $\delta n$ vertices is $\Theta(\min\{\delta,\rho\}\,n)$. The hidden constant in the $\Theta(\cdot)$ notation is independent of $p$. For a sufficiently large constant $\lambda>0,$ $n\in\mathbb{Z}^{+},$ $p\geq\lambda(\ln(e/\rho))/n$ and with probability $1-n^{-\Omega(1)}$ over $G(n,p),$ they also prove the existence of $O(\lceil\rho n\rceil)$ seeds that activate all vertices eventually.

This paper proves that for $n\in\mathbb{Z}^{+},$ $p=\Omega((\ln(e/\rho))/(\rho n))$ and with probability $1-\exp{(-n^{\Omega(1)})}$ over $G(n,p),$ $\text{min-seed}^{(1)}(G(n,p),\rho)=O(\rho n)$. Together with Chang and Lyuu’s $\Theta(\min\{\delta,\rho\}\,n)$ bound for irreversible cascades with an unbounded duration, our result shows that neither the reversibility of synchronous cascades nor the number of rounds played can change the asymptotically minimum number of seeds needed to activate all vertices for $p\geq\beta(\ln(e/{\min\{\delta,\rho\}}))/(\rho n)$ with a sufficiently large constant $\beta$ — it is always $\Theta(\rho n)$. Furthermore, our $O(\rho n)$ bound continues to hold even if the synchronous reversible cascade is modified to proceed asynchronously instead.

Other related topics include periodic behavior of synchronous reversible cascades [46, 45, 77, 78, 43, 62, 63, 64], inapproximability of minimum-seed problems [26, 24, 1], majority consensus computers [69], stable sets of active vertices [2, 4, 3, 48, 53], network decontamination [60, 58], maximization of social influence [34, 79, 49, 50, 66] and percolation theory [81], among others.

This paper is organized as follows. Sec. 2 presents the notations and the preliminaries. Sec. 3 derives general bounds on the minimum number of seeds needed to activate all vertices in a number of rounds. Secs. 4–5 investigate the cases of connected scale-free and Erdős-Rényi random graphs, respectively. Sec. 6 concludes the paper.

A directed graph $G(V,E)$ consists of a set $V$ of vertices and a set $E\subseteq V\times V$ of edges. An edge $(u,v)\in E$ goes from $u$ to $v$. An undirected graph is a directed one with each edge accompanied by the edge in the opposite direction. Unless otherwise specified, all graphs in this paper are simple, i.e., self-loops are not allowed [84]. For $v\in V,$ define

to be the open and closed in-neighborhoods of $v,$ respectively. The indegree of $v$ is $d^{\text{in}}(v)\equiv|\,N^{\text{in}}(v)\,|$. For $A\subseteq V,$ define $N^{\text{in}}(A)\equiv\cup_{a\in A}\,N^{\text{in}}(a)$ as the set of vertices incident on an edge coming into $A$. In case $G$ is undirected, write $N^{\text{in}}(\cdot),$ $N^{\text{in}}[\cdot]$ and $d^{\text{in}}(\cdot)$ simply as $N(\cdot),$ $N[\cdot]$ and $d(\cdot),$ respectively. For an undirected graph $G(V,E),$ $U\subseteq V$ and $i\geq 0,$ define $N^{i}[U]$ to be the set of vertices with distance less than or equal to $i$ from at least one vertex in $U$. That is, $N^{0}[U]=U$ and $N^{i+1}[U]=N^{i}[U]\cup N(N^{i}[U])$ for $i\geq 0$. Furthermore, each vertex can be in one of two states, active or inactive.

The synchronous reversible cascade on $G(V,E)$ with seed set $S\subseteq V$ and threshold $\rho\in(\,0,1\,]$ proceeds in rounds. In round zero, only the vertices in $S,$ called the seeds, are active. For each $i\in\mathbb{N},$ a vertex with a positive indegree is activated or deactivated in round $i+1,$ respectively, if at least or less than a $\rho$ fraction of its in-neighbors are active in round $i$. Vertices with indegree zero, instead, never change their states. For $k\in\mathbb{N},$ define $\text{\sf Active}^{(k)}(S,G,\rho)$ to be the set of active vertices in round $k$. Then define

which is the minimum number of seeds needed to activate all vertices in round $k$ (it is possible that a set of seeds activates all vertices in round $k$ by doing so before round $k$). Clearly, once all vertices are active, they will remain active forever. Therefore, $\text{min-seed}^{(k)}(\cdot,\cdot)$ monotonically decreases as $k$ increases.

Next, we describe an asynchronous reversible cascade. At any instant, one or more vertices may update their states. When a vertex with a positive indegree updates its state, it is activated or deactivated if at least or less than a $\rho$ fraction of its in-neighbors are active, respectively. Instead, vertices with indegree zero never update their states. The only requirement on the synchronization of updates is that, if the state of a vertex can change at time $t,$ then at least one vertex must change its state after time $t,$ where $t\geq 0$. Such a synchronization is said to be progressive. In other words, whenever changes of states are possible, an asynchronous reversible cascade cannot simply refuse them all forever. Define $\text{min-seed}^{\text{async}}(G,\rho)$ to be the minimum number of seeds that activate all vertices within a finite amount of time in every asynchronous reversible cascade with threshold $\rho$. So $\text{min-seed}^{\text{async}}(G,\rho)\leq s$ precisely when there exist $s$ seeds activating all vertices within a finite amount of time no matter how the updates of states are synchronized in a progressive way. Furthermore, define $\text{min-seed}^{\text{async}}_{\text{monotone}}(G,\rho)$ to be the minimum number of seeds meeting the following criteria in every asynchronous reversible cascade with threshold $\rho$:

• •

  All vertices are active after a finite amount of time;

• •

  No active vertices are ever deactivated.

So $\text{min-seed}^{\text{async}}_{\text{monotone}}(G,\rho)\leq s$ precisely when there exist $s$ seeds that, regardless of the (progressive) synchronizations of the reversible cascades, activate all vertices without ever deactivating any active one.

The following fact is folklore. See, e.g., [40, pp. 25, Sec. 2].

The following is Markov’s inequality.

We will use the following form of Chernoff’s bound [27].

For $n\in\mathbb{Z}^{+}$ and $p\in[\,0,1\,],$ the Erdős-Rényi random graph $G(n,p)$ is the simple undirected graph with vertices $1,2,\ldots,n$ where each of the possible $\binom{n}{2}$ edges appears independently with probability $p$ [14]. Below is an easy consequence of Chernoff’s bound.

The irreversible cascade is the modification of our synchronous reversible cascade that prohibits the deactivation of vertices. It can also be defined as an asynchronous process without affecting the set of vertices that will be active [40, pp. 25, Sec. 2]. Assuming irreversible cascades, Chang and Lyuu [22, 21] prove the following bound on activating vertices of Erdős-Rényi random graphs.

We begin with several lemmas that will be useful for analyzing reversible cascades on scale-free graphs.

The following lemma reduces the bound of Lemma 6 to $2$ if all degrees are less than or equal to $\rho$.

Next, we analyze the following picking of seeds in a directed graph $G(V,E)$: First, choose each vertex as a seed independently with some probability. Second, choose the vertices that cannot be activated in round $1$ also as seeds. The expected number of chosen seeds serves as an upper bound on $\text{min-seed}^{(1)}(G,\rho)$. Chang and Lyuu [23, Theorem 9] use a similar technique to analyze irreversible cascades. Their results are improved by Ackerman, Ben-Zwi and Wolfovitz [1, Sec. 3].