On Wed, 11 Jun 2003 18:39:00 +0000 (UTC), M Law <[EMAIL PROTECTED]> wrote:
> In linear discriminant analysis, we first pool the within class > covariance matrix to form S_w, and then consider inv(S_w) * (m_1 - m_2) > for the direction for projection. > > The use of pooled covariance matrix essentially assumes that the > covariance matrices of the two classes are similar. What happens > if this assumption is significantly violated? Is there any > special name of this problem? Has this been considered in the > literature? The statistical packages that perform linear discriminant function typically do allow for the option of "separate covariances." How are you getting your introduction to DF, that you don't know this? Are you just guessing ahead of the class? One way to look at the result is like this: Unless you have REALLY big Ns, you do not get very good results when you have to estimate all those extra parameters. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html "Taxes are the price we pay for civilization." Justice Holmes. . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
