Universal Algebraic Controllers and System Identification

Abstract

In this document, some structured operator approximation theoretical methods for system identification of nearly eventually periodic systems, are presented. Let $\mathbb{C}^{n\times m}$ denote the algebra of $n\times m$ complex matrices. Given $\varepsilon>0$, an arbitrary discrete-time dynamical system $(\Sigma,\mathcal{T})$ with state-space $\Sigma$ contained in the finite dimensional Hilbert space $\mathbb{C}^n$, whose state-transition map $\mathcal{T}:\Sigma\times ([0,\infty)\cap \mathbb{Z})\to \Sigma$ is unknown or partially known, and needs to be determined based on some sampled data in a finite set $\hat{\Sigma}=\{x_t\}_{1\leq t\leq m}\subset \Sigma$ according to the rule $\mathcal{T}(x_t,1)=x_{t+1}$ for each $1\leq t\leq m-1$, and given $x\in \hat{\Sigma}$. We study the solvability of the existence problems for two triples $(p,A,\varphi)$ and $(p,A_\eta,\Phi)$ determined by a polynomial $p\in \mathbb{C}[z]$ with $\deg(p)\leq m$, a matrix root $A\in\mathbb{C}^{m\times m}$ and an approximate matrix root $A_\eta\in\mathbb{C}^{r\times r}$ of $p(z)=0$ with $r\leq m$, two completely positive linear multiplicative maps $\varphi:\mathbb{C}^{m\times m}\to \mathbb{C}^{n\times n}$ and $\Phi:\mathbb{C}^{r\times r}\to \mathbb{C}^{n\times n}$, such that $\|\mathcal{T}(x,t)-\varphi(A^t)x\|\leq\varepsilon$ and $\|\Phi(A_\eta^t)x-\varphi(A^t)x\|\leq\varepsilon$, for each integer $t\geq 1$ such that $\|\mathcal{T}(x,t)-y\|\leq \varepsilon$ for some $y\in \hat{\Sigma}$. Some numerical implementations of these techniques for the reduced-order predictive simulation of dynamical systems in continuum and quantum mechanics, are outlined.