A Virtual Space Embedding for the Analysis of Dynamical Networks Giacomo Innocenti and Paolo Paoletti Abstract— Networks of dynamical systems are widespread both in nature and in technology applications, but the interplay between the node dynamics, their interconnections and the macroscopic behavior is still not completely understood. In fact, the intrinsically large dimension of the phase space of such networks represents the main obstacle for their analysis and control. In this paper we propose a technique that is based on embedding the network in a (potentially virtual) spatial coordinate system to obtain a simplified description of the overall behavior. We derive conditions under which both the network dynamics and its stability properties can be successfully analyzed using such approach. A network of Fitzhugh-Nagumo systems is used as an illustrative example to show the effectiveness of the proposed technique. I. I NTRODUCTION The analysis and control of networks composed by the interconnection of nonlinear dynamical systems represent intriguing fields of investigation due to their wide application to the understanding of natural systems and to the control of artificial ones. For example, in the last decades networks of dynamical systems, such as cellular neural networks or analog processor arrays, have been introduced in engineering applications as a paradigm for analog computation, image analysis and pattern generation [1], [2]. Similarly, dynamical networks play crucial roles in biological systems for gener- ating spatio-temporal patterns in nervous systems of various organisms, including humans, see e.g. [3], [4]. Recently, dynamical networks has also been exploited to understand phenomena of opinion dynamics [5] and infectious diseases spreading [6]. However, beside its relevance for a wide variety of sci- entific fields, the interplay between the individual node dynamics and interconnections, and the behavior exhibited by the network itself is still not completely understood. With this respect, the main obstacle is represented by the intrinsically large dimension of the network phase space that scales linearly with the number of nodes and that therefore can easily reach the point where traditional analysis and control techniques developed for low-dimensional systems may not be easily applied anymore. On the other hand, a dramatic reduction of the model complexity can be achieved by assuming that the dynamics of each node is actually a sampling of the dynamics of a continuous system governed by an underlying partial differential equation (PDE). The evolution of each node is G. Innocenti is with the Dipartimento di Ingegneria dell’Informazione, Universit` a di Firenze, I-50139 Firenze, Italy, giacomo.innocenti@unifi.it P. Paoletti is with the Centre for Engineering Dynamics, School of Engineering, University of Liverpool, L69 3GH Liverpool, UK, P.Paoletti@liverpool.ac.uk required to be close (in some norm) to the one observed at the point where such node is collocated, see [7] for an example for the linear case and [8] for some preliminary results for chains of nonlinear systems. We stress that the continuous spatial coordinate system used to collocate the nodes of the network needs not to represent a physical space. It is also worth pointing out that, although it is known that there exists phenomena, such as propagation failure, that may occur on networks but not on PDEs [9], such degenerate situations seem to play a negligible role for all the applications mentioned before and this justifies the need for further investigations on the relationships between networks and behavior of PDEs. In this paper we propose a technique for analyzing the be- havior of networks where individual nodes are governed by a nonlinear ordinary differential equation (ODE). In section II such network is embedded in a continuous spatial coordinate system to define a PDE whose dynamics interpolates the behavior of each node. Conditions for the PDE to correctly represent the network dynamics are stated in section III. Here we focus mainly on propagating phenomena where the PDE can be rewritten in a moving coordinate frame to effectively obtain an ordinary differential equation whose solution provides the profile of the propagating wave. A procedure for determining the stability of such profile is also discussed in the same section. We then introduce an illustrative example in section IV to show how the proposed approach can be successfully used to analyze the behavior of a non trivial nonlinear dynamical network. Some final remarks and future outlooks are discussed in section V. Notation and preliminaries. Throughout the paper we will denote the sets of real and complex numbers as, re- spectively, R and C, whereas N 0 will represents the set of natural number including 0, while Z will denote the integer set. Given a function f (ξ ): R w R n we will refer to its derivative as the linear operator df (ξ) = ∂f 1 ∂ξ 1 ... ∂f 1 ∂ξw . . . ∂fn ∂ξ 1 ... ∂f 1 ∂ξw . We will also make use of the multi–index notation, i.e., given the vectors h = [h 1 ,...,h w ] T N w 0 and x = [x 1 ,...,x w ] T R w we will refer to the following quantities as follows h!= h 1 ! ...h w !, x h = x h1 1 ...x hw w h x = h1 ∂x h1 1 ... hw ∂x hw w = |h| ∂x h1 1 ...∂x hw w . 52nd IEEE Conference on Decision and Control December 10-13, 2013. Florence, Italy 978-1-4673-5717-3/13/$31.00 ©2013 IEEE 1331