Maximum likelihood estimators and worst case optimal algorithms for system identification (Q1105558)

Suppose X and Z are finite-dimensional linear spaces over the real field R and Y is a normed space over R. Consider two linear operators, S and N, called respectively a solution operator and an information operator, where \(S: X\to Z\) and \(N: X\to Y\). The following identification problem is considered: estimate Sx for every \(x\in X\) using the perturbed information y given by \(y=Nx+\rho\) where \(\rho\) describes the noise contaminating Nx. Any approximation z to Sx is constructed by an algorithm (or estimator) A using the available data y, i.e., \(z=A(y)\) where \(A: Y\to Z\). The authors suggest different algorithms for the solution of this general problem. In particular, they describe the situations when the so-called robust-interpolatory algorithm and the projection algorithm coincide. But the main attention is paid to the maximum likelihood estimators and their relationship with the worst case optimal algorithms. The authors present some general results which are illustrated by considering three interesting particular cases when the noise \(\rho\) has a uniform distribution, normal distribution and Laplace distribution.