Dear Dr. Human (in Shark Research); Can you explain your problem with singular matrices ?
Perhaps someone can give a tutorial on generalized inverse matrices. Are your matrices over determined and inconsistent ? Have you considered singular value decomposition ? I have some old code for solving over determined inconsistent systems with SVD (Singular Value Composition). In SVD, you pick a threshold for excluding dimensions with a small percentage of the total 'hyper volume' and thus Your solution only uses the most salient eigenvectors (with significant contribution to the total hyper volume vector sum). This arises because noise always crates noise dimensions with little 'correlation'. You discard small eigenbasis vectors and only keep what you think are signal vectors. Of course you may be throwing out weak signals... IMHO FWIIW. Dr. D B Karron [EMAIL PROTECTED] 212 686 8748 (office/home) 212 448 0261 (fax) 347 886 9066 (dbk personal cell) -----Original Message----- From: morphmet [mailto:[EMAIL PROTECTED] Sent: Wednesday, March 16, 2005 9:27 AM To: morphmet Subject: generalised inverse matrices Hello All, I am having issues with singular matrices because the number of variables (linear morphometric measures) I have far exceeds the number of samples. I don't wish to introduce a bias into my analyses by selecting variables to include/exclude from my data matrix, hence I am wondering about the validity of using generalised inverse matrices. Is the use of generalised inverse matrices a valid/ accepted one? Are generalised inverse matrices statistically robust and will people believe my results? Has anyone published using generalised inverse matrices? How would I defend the use of generalised inverse matrices? Thanks in advance, Brett Human ************************ Brett Human, PhD Shark Researcher 27 Southern Ave West Beach SA 5024 Australia Ph: +61 8 8356 6891 email: [EMAIL PROTECTED] ************************ -- Replies will be sent to the list. For more information visit http://www.morphometrics.org
