[eigen-commits] commit/eigen2: bjacob: Fix 4x4 matrix inversion. Applying Euler's trick is more tricky than what "high performance matrix inversion" websites would have you believe. Our 4x4 matrix inversion wasn't numerically stable, because in applying the Euler trick we didn't take the 2x2 block of biggest determinant. As a result, with float, we got relative errors above 1% every 1000 random matrices, and we got completely wrong results every 10000 matrices.

[commits-noreply@xxxxxxxxxxxxx]

1 new changeset in eigen2:

http://www.bitbucket.org/eigen/eigen2/changeset/6d247b06ffe9/
changeset:   r1750:6d247b06ffe9
user:        bjacob
date:        2009-10-27 12:35:25
summary:     Fix 4x4 matrix inversion. Applying Euler's trick is more tricky than what "high performance matrix inversion" websites would have you believe. Our 4x4 matrix inversion wasn't numerically stable, because in applying the Euler trick we didn't take the 2x2 block of biggest determinant. As a result, with float, we got relative errors above 1% every 1000 random matrices, and we got completely wrong results every 10000 matrices.

Note that this decreases the performance, but we're still significantly faster than the brutal cofactors approach.
affected #:  1 file (292 bytes)

Repository URL: http://bitbucket.org/eigen/eigen2/

-- 
This is a commit notification from bitbucket.org.
You are receiving this because you have the service 
enabled, addressing the recipient of this email.