Most performance measures of pilot-assisted multiple-input multiple-output
(MIMO) systems are functions that depend on both the linear precoding filter
and the pilot sequence. A framework for the optimization of these two
parameters is proposed, based on a matrix-valued generalization of the concept
of effective signal-to-noise ratio (SNR) introduced in a famous work by Hassibi
and Hochwald. The framework applies to a wide class of utility functions of
said effective SNR matrix, most notably a well-known mutual information
expression for Gaussian inputs, an upper bound on the minimum mean-square error
(MMSE), as well as approximations thereof. The approach consists in decomposing
the joint optimization problem into three subproblems: first, we describe how
to reformulate the optimization of the linear precoder subject to a fixed pilot
sequence as a convex problem. Second, we do likewise for the optimization of
the pilot sequence subject to a fixed precoder. Third, we describe how to
generate pairs of precoders and pilot sequences that are Pareto optimal in the
sense that they attain the Pareto boundary of the set of feasible effective SNR
matrices. By combining these three optimization problems into an iteration, we
obtain an algorithm which allows to compute jointly optimal pairs of precoders
and pilot sequences with respect to some generic utility function of the
effective SNR.Comment: 32 pages, 9 figures, submitted to IEEE Transactions on Information
Theor