Core Concepts
The core message of this paper is that incorporating prior structural information about the system matrix A* in the form of a convex constraint set K can lead to significantly improved finite-time identification performance compared to the unconstrained ordinary least squares (OLS) estimator, especially when the intrinsic dimension of A* is much smaller than the ambient dimension n.
Abstract
The paper considers the problem of finite-time identification of a linear dynamical system (LDS) of the form xt+1 = Axt + ηt+1, where A is the unknown system matrix to be estimated from T samples of a single trajectory (xt)Tt=0. The authors assume that A* is strictly stable, i.e., its spectral radius ρ(A*) < 1.
The key contributions are as follows:
The authors derive non-asymptotic error bounds in the Frobenius norm for the constrained least squares estimator of A*, where the constraint set K captures prior structural information about A*. The bounds depend on the local size of K at A*, as measured by Talagrand's γ1 and γ2 functionals.
The authors instantiate their general results for three specific examples of K:
K is a d-dimensional subspace of Rn×n
K is a suitably scaled ℓ1 ball, capturing sparsity of A*
K consists of matrices formed by sampling a bivariate convex function on a uniform grid, capturing convex structure of A*
For these examples, the authors show that the sample complexity T required to accurately estimate A* can be significantly smaller than the unconstrained case, especially when the intrinsic dimension of A* (d, sparsity level k, or complexity of the underlying convex function) is much smaller than the ambient dimension n.
The proofs rely on novel concentration inequalities for suprema of second-order subgaussian chaos processes, which are of independent interest.
Overall, the paper demonstrates the benefits of leveraging structural constraints in the finite-time identification of linear dynamical systems.
Stats
The paper does not contain any explicit numerical data or statistics. The key quantities used in the analysis are:
ρ(A*): Spectral radius of the system matrix A*
J(A*): A quantity defined in (2.1) that captures the stability of A*
γ1, γ2: Talagrand's gamma functionals that measure the local complexity of the constraint set K
Quotes
"We assume prior structural information on A* is available, which can be captured in the form of a convex set K containing A*. For the solution of the ensuing constrained least squares estimator, we derive non-asymptotic error bounds in the Frobenius norm that depend on the local size of K at A*."
"Upon instantiating our bounds for the aforementioned choices of K, we obtain the following corollaries."