Re: Mahalanobis distance & covariance matrices

From: Ray Koopman (koopman_at_sfu.ca)
Date: 09/26/04 

Date: 25 Sep 2004 21:52:06 -0700

Gottfried Helms <helms@uni-kassel.de> wrote in message
news:<cj4plb$qa9$05$1@news.t-online.com>...
> [...]
> An alternative could then be, that you estimate a correlation
> for the population and keep that correlation-matrix constant
> for all new incoming cases, and use the mahalanobis-distances
> according to that population correlation-matrix.
>
> I've done not much with distance-based statistics, so I don't
> know for sure, but may be in cluster-analysis or discriminance-
> analysis this aspect is handled implicitely, since it seems
> quite obvious, and classifying of unknown objects is directly
> in the main focus of these procedures.

Traditional discriminant analysis uses Mahalanobis distances based
on a such fixed covariance matrix, but it is the pooled within-group
covariance matrix, not the overall covariance matrix.