Finite impulse response models: a non-asymptotic analysis of the least squares estimator (Q2040046)

The authors study finite impulse response (FIR) systems which are in some sense the simplest discrete-time dynamical systems. Of special interest is the linear regressor case which leads to a linear regression model with ``time-shifted'' random covariates. The authors derive non-asymptotic near-optimal estimation and prediction bounds for the least squares estimator (LSE) of the parameters. Especially they present high-probability bounds on the estimation error and the prediction risk of the LSE based on two concentration inequalities on the norm of sums of dependent covariate vectors and on singular values of their covariance operator. It follows a discussion on the optimality of the obtained results based on the Cramer-Rao risk lower bound in the Gaussian case.