[1] A possible alternative to the
marginal distribution would be to set the diagonal elements of Cmm (variances
of the missing dimensions) to some large value, while setting Cmp and Cpm
(covariances between missing and present information) to zero. This should
neutralize the contribution of the missing dimensions to probability
calculations. This is perhaps analogous to what Cooke et al. (1996) call
"variance weighting", a procedure that produced the best recognition rates of
all.