home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky sci.math.stat:2657 sci.math.num-analysis:3648
- Path: sparky!uunet!psinntp!curly.appmag.com!verano!pa
- From: pa@verano.sba.ca.us (Pierre Asselin)
- Newsgroups: sci.math.stat,sci.math.num-analysis
- Subject: Re: Double decorrelation
- Message-ID: <1209@verano.sba.ca.us>
- Date: 19 Dec 92 18:10:00 GMT
- References: <1gq7rgINNcqk@agate.berkeley.edu>
- Followup-To: sci.math.stat
- Organization: None. Santa Barbara, CA
- Lines: 28
-
- In <1gq7rgINNcqk@agate.berkeley.edu>
- shein@nima.berkeley.edu (Soren Hein) writes:
-
- >Given a matrix X (M by N) of stochastic variables, find two matrices
- >A (M by M) and B (N by N) such that for the transformed matrix
- > Y = A X B,
- >we have
- > E[Y Y^T] = I_M
- > E[Y^T Y] = I_N
-
- I'm not sure it can be done. If you can settle for diagonal matrices
- instead of identities, a singular value decomposition (essentially)
- will do the job. I assume you know the covariance matrix of your
- Xij's.
-
- 1) Restrict A and B to orthogonal matrices. Then,
- Y Y^T = A X X^T A^T independent of B, and
- Y^T Y = B^T X^T X B independent of A.
-
- 2) Form the deterministic matrix E[X X^T] and diagonalize.
- The eigenvectors are the columns of A.
-
- 3) Form the deterministic matrix E[X^T X] and diagonalize.
- The eigenvectors are the rows of B.
- --
-
- --Pierre Asselin, Santa Barbara, California
- pa@verano.sba.ca.us
-