This dissertation deals with frequency domain
identification of linear systems in a deterministic set-up. A
robustly convergent algorithm for identification of
frequency response samples of a linear shift invariant plant is proposed.
An explicit bound on the identification error is
obtained based on suitable a priori assumptions about the plant
and the measurement noise. For a finite measurement duration, this algorithm yields (possibly) noisy point frequency
response samples of the plant and a worst case error bound.
Given such noisy
frequency response samples, two different
families of worst case identification algorithms are presented. Each
of these algorithms yields a model and a bound on the worst case
infinity norm of error between the plant and the model, based on a priori and (in some cases) a posteriori data. One of
the families of algorithms is robustly convergent and exhibits a
certain optimality for a fixed model order. Both the families of algorithms
are shown to be implementable as solutions to certain convex
optimisation problems. The ideas and numerical techniques
used for implementing these algorithms are further used to propose a method for
identification in the
nu-gap metric.