On data-driven computation of information transfer for causal inference in discrete-time dynamical systems (Q2190698)

The authors give a new type of approach to capture the causal interaction in a dynamical system from time series data. The novelty lies in the concept of information transfer. The major purpose of this research work is to provide a data-driven approach for computing information transfer from time series data for inferring causal interaction in a dynamical system in a noisy environment. The authors consider the following dynamical system: \(z(t+1)=F(z(t))+\xi (t)\), where \(F=[F_{x}^{\top}\quad F_{x}^{\top}]^{\top}\), such that \(x(t+1)=F_{x}(x(t), y(t))+\xi_{x} (t)\) and \(y(t+1)=F_{y}(x(t), y(t))+\xi_{y} (t)\), where \(x\in \mathbb{R}^{|x|}\), \(y\in \mathbb{R}^{|y|}\), \(z=(x^{\top}, y^{\top})^{\top}\), and \(F_{x}: \mathbb{R}^{|x|+|y|}\rightarrow\mathbb{R}^{|x|}\), \(F_{y}: \mathbb{R}^{|x|+|y|}\rightarrow\mathbb{R}^{|y|}\) are assumed to be continuously differentiable. Here \(\xi (t) = (\xi_{x} (t)^{\top}, \xi_{y} (t)^{\top})^{\top}\) is an additive independent and identically distributed noise. Further, the dynamical system \(F_{\not{x}} (z(t))+\xi_{\not{x}} (t)\), where the dynamics in the \(x\) coordinate is frozen and given by \(x(t+1)=x(t)\) and \(y(t+1)=F_{y}(x(t), y(t))+\xi_{y} (t)\). To calculate the information transfer, it is needed to identify the evolution of conditional probability density function, \(\rho (z(t+1)|z(t))\), under the influence of the above dynamical systems. The authors apply a transfer operator theoretic framework involving transfer Perron-Frobenius and Koopman operators for the data-driven approximation of the system dynamics and the computation of information transfer. The transfer operators present a linear representation of the nonlinear systems by shifting focus from the state space to the space of functions and measures. The linear time-invariant dynamical system \(z(t+1)=A z(t)+\sigma\xi (t)\) is considered by the authors. Here \(z(t)\in \mathbb{R}^{n}\) and \(\xi (t)\) is vector-valued Gaussian random variable with zero mean and unit variance. Furthermore, the authors present a robust implementation of naturally structured dynamic mode decomposition algorithm for the finite-dimensional approximation of: (i) the Perron-Frobenius operator \(\mathbb{P}: M(Z)\rightarrow M(Z)\), defined by \[[\mathbb{P}\mu](A)=\int_{Z}\int_{W}\chi_{A} (T(x, v))d\vartheta (v) d\mu (z)=\int_{Z}p(z, A)d\mu (z),\] where \(\chi_{A}(z)\) is the indicator function for set \(A\) and \(p(z, A)\) is the transition probability function; (ii) the Koopman operator \(\mathbb{U}: F\rightarrow F\), which, for any given \(h\in \mathcal{ F}\), is defined as \[[\mathbb{U}h](z)=E_{\xi}[h(T(z, \xi))]=\int_{W}h(T(z, v))d\vartheta (z),\] where \(\mathcal{ F}\) is \(L_{\infty}(Z)\). The authors use the robust approximation of transfer operators for the computation of information transfer and causality inference for both linear and nonlinear dynamical systems. They begin the procedure of information transfer computation considering two-dimensional and two-subspace cases. After that the general case is studied. Moreover, to verify the efficiency of the developed algorithm and demonstrate their method the authors provide many examples.