Graphons and the $H$-property

Structural system theory deals with the problem of understanding when a given network topology can sustain a prescribed system property. Typical such properties are, e.g, controllability and stability. In more detail, consider a network of $n$ mobile agents $x_{1},\ldots,x_{n}$, whose communication topology is described by a directed graph (digraph) $G=(V,E)$, with the nodes $v_{1},\ldots,v_{n}$ representing the agents and directed edges $v_{i}v_{j}$ indicating the information flow (with the convention that a directed edge $v_{i}v_{j}$ indicates that agent $x_{j}$ can access state information from $x_{i}$). Given the digraph $G$, a system dynamics $\dot{x}(t)=f(x(t))$ is said to be adapted to $G$ if the dynamics of $x_{i}(t)$ depend only on its incoming neighbors in $G$: $\frac{\partial f_{i}}{\partial x_{j}}\neq 0\Rightarrow v_{j}v_{i}\in E$.

We denote by $\Sigma_{G}$ the set of differentiable dynamics adapted to $G$. Next, given a desired system property $\mathcal{S}$ (e.g., asymptotic stability at a given equilibrium), we say that the digraph $G$ sustains $\mathcal{S}$ if there exists a dynamics $f\in\Sigma_{G}$ satisfying the property $\mathcal{S}$. This research line was initiated by C.-T. Lin in his seminal paper [3] and led to several extensions, among which we mention the work of the authors on structural stability [4] and structural ensemble controllability [5] as they are the basis of the present work.

Graphons have recently been introduced in [6, 7] to study large graphs. A graphon can be seen as both the limit object of a convergent sequence (where convergence is in the cut-norm [8]) of graphs of increasing size, and as a statistical model from which to sample random graphs (we elaborate on this below). Mathematically, a graphon is a symmetric, measurable function $W:[0,1]^{2}\to[0,1]$, with $W(x,y)=W(y,x)$ for all $x,y\in[0,1]^{2}$. We note that a graphon should be viewed as an equivalence class of such functions, where $W_{1}\sim W_{2}$ if there exists a measure-preserving map $\phi:[0,1]\to[0,1]$ such that $W_{1}(x,y)=W_{2}(\phi(x),\phi(y))$. This is the continuous equivalent of saying that graphs that are equal up to relabeling of their nodes are in the same equivalence class. We will overlook this distinction for the sake of clarity, but all our statements lift to the equivalence class.

We next describe how to sample a graph $G_{n}$ on $n$ nodes from a graphon $W$.

In this abstract, we characterize the graphons $W$ with the property that network topologies sampled from $W$ are structurally stable [4] or structurally ensemble controllable [5] with probability one. Precisely, we have the following definition [2]:

The problem we address can now be formulated as characterizing the $H$-property in graphons.

As a stepping-stone to the study of the general case of graphons, we have studied in our recent work [2] the class of step-graphons (see Definition 2 below), and provided essentially necessary and sufficient conditions for the $H$-property to hold in that context. We summarize below the key objects introduced to solve the problem as well as the main statement, as they are a blueprint for the study of the general case.

We start with the following definition:

It should be clear that for a given step-graphon $W$, there exists infinitely many partition sequences (obtained, e.g., by sub-dividing intervals where the step-graphon is constant).

Step-graphons are related to stochastic block models [9], with the difference that nodes in stochastic block models are deterministically placed in intervals (called communities in that context) whereas, in the graphon case, the nodes are randomly placed.

Through our extant work, we have highlighted three key objects associated with a step-graphon $W$ as being of particular importance. We introduce them below:

It should be clear from the sampling procedure given above that the concentration vector describes the expected proportion of sampled nodes in each interval. The support of a step-graphon can be described by a graph, called skeleton graph, which is defined as

The last object we introduced is derived from the skeleton graph. Let $F=\{f_{1},\ldots,f_{r}\}$, we let the edge-incidence matrix $B\in\mathbb{R}^{q\times r}$ of a graph $S=(U,F)$ be given as

Owing to the factor $\frac{1}{2}$ in (1), all columns of $B$ are probability vectors, i.e., all entries are nonnegative and sum to one. The edge polytope of $S$, introduced in [10], is

We illustrate the above three key objects in Figure 2.

To state the result, we introduce the following three conditions:

Condition $A$:

The skeleton graph $S$ has an odd cycle.

Condition $B$:

The concentration vector $x^{*}$ belongs to the relative interior of the edge polytope $\mathcal{X}(S)$, i.e., $x^{*}\in\operatorname{int}\mathcal{X}(S)$.

Condition $B^{\prime}$:

$x^{*}\in\mathcal{X}(S)$.

The above conditions can be shown to be independent of the choice of a partition sequence $\sigma$ [2]. The main results proved in [1, 2] are reproduced below

For example, the step-graphon in Figure 2 satisfies Conditions A and B and, hence, has the $H$-property.

We now discuss the general case of graphons. One first observes that while some objects introduced above, such as the concentration vector, admit a relatively simple translation to the infinite-dimensional case, the skeleton graph does not admit a simple infinite-dimensional equivalent. Since the edge-polytope was defined from the edge-incidence matrix of the skeleton graph, naive limits $n\to\infty$ do not yield useful results.

To circumvent this difficulty, we are forced to take a different approach to describing the edge-polytope of a graphon, that by-passes the use of the skeleton graph. Relying on some technical results in [2]—precisely, what we refer to as the $A$-matrix of a step-graphon—we arrive at the following characterizations:

The following result, which characterizes the map $\Phi(W)$ when $W$ is a step-graphon, highlights its importance to the analysis of the $H$-property.

We thus introduce the following condition as an extension of Condition $A$:

Condition $A_{\rm ext}$:

$\Phi_{\overline{W}}$ is surjective.

For $W$ a step-graphon with skeleton graph $S$, we define

We state, without a proof, that $\mathrm{L}(W)$ does not depend on the choice of partition sequence $\sigma$ (and hence, of $S$) for $W$.

The space $\mathrm{L}(W)$ can be thought of as a functional equivalent of the edge-polytope of a step-graphon, as the next proposition shows:

Even though $\mathrm{L}(W)$ is a functional space, it still relies on $S$ for its definition and thus cannot be applied directly to general $W$. To resolve the issue, we define

Note that the requirement $\|\Phi_{\overline{W}}(c)\|_{\mathrm{L}^{1}}=1$ is equivalent to $\int_{\operatorname{supp}W}c(s,t)\mathrm{d}s\mathrm{d}t=1$. It should be clear that $\mathcal{X}(W)$ is well defined for all graphons.

We have the following result:

This proposition states that the $\mathcal{X}(W)$ introduced above, which did not rely on the skeleton graph of $W$, equals the space $\mathrm{L}(W)$ and together with Proposition 3 shows that it can indeed serve as an equivalent of the edge-polytope applicable to the case of general graphons. This leads to the following extension of Condition B:

Condition $B_{\rm ext}$:

$\mathbf{1}\in\operatorname{int}\mathcal{X}(W)$.

We are now in a position to state the main result of this abstract:

The proofs of the above results will appear in an upcoming paper.