We present and discuss the structure and design of optimum multivariable de
cision feedback equalizers (DFEs), The equalizers are derived under the con
straint of realizability, that is, causal and stable filters and finite dec
ision delay. The design is based on a known dispersive discrete-time multiv
ariable channel model with infinite impulse response. The additive noise is
described by a multivariate ARMA model. Equations for obtaining minimum me
an square error (MMSE) and zero-forcing DFEs are derived under the assumpti
on of correct past decisions. Design of a realizable MMSE DFE requires the
solution of a linear system of equations in the model parameters, No spectr
al factorization is required.
We derive novel necessary and sufficient conditions for the existence of ze
ro-forcing DFEs and point out the significance of these conditions for the
design of multiuser detectors.
An optimal MMSE DFE will contain IIR filters in general, Simulations indica
te that the performance improvement obtained with an IIR DFE is reduced mor
e than for a (suboptimal) FIR DFE when error propagation is taken into acco
unt since the use of IIR feedback filters tends to worsen the error propaga
tion.