home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!mcsun!uknet!cam-cl!cam-cl!nmm
- From: nmm@cl.cam.ac.uk (Nick Maclaren)
- Newsgroups: comp.theory
- Subject: Re: Matrix Square Root
- Message-ID: <1992Sep10.075331.9873@cl.cam.ac.uk>
- Date: 10 Sep 92 07:53:31 GMT
- References: <Kevin.D.Brunson.1-080992181444@davmac27.oshaughnessy.lab.nd.edu> <dak.716054795@kaa>
- Sender: news@cl.cam.ac.uk (The news facility)
- Reply-To: nmm@cl.cam.ac.uk (Nick Maclaren)
- Organization: U of Cambridge Computer Lab, UK
- Lines: 41
-
- In article <dak.716054795@kaa>, dak@kaa.informatik.rwth-aachen.de (David
- Kastrup) writes:
- > Kevin.D.Brunson.1@nd.edu (Kevin D. Brunson) writes:
- > > ....
- > > The problem I have seems like a simple one: How do I calculate the
- > >square root of a symetric matrix? Theoretically, if a matrix A is positive
- > >semidefinite then there exists a matrix B that is the square root of A such
- > >that A = B*B. I have consulted Lancaster and Tismenetsky's "The Theory of
- > >Matrices" but a programmable algorithm is not obvious. ....
-
- > If you do an Eigenvector decomposition A=TDT^{-1}, where D is a diagonal
- > matrix containg the eigenvalues of the matrix, and T a matrix containing
- > the corresponding eigenvectors in its columns, TD^{0.5}T^{-1} will
- > have that property.
-
- I hit this problem when writing the multivariate normal random number
- generator for the NAG library, and there are three solutions:
-
- 1) Gaussian elimination on the original matrix plus a fudge constant
- on the diagonals (which I vaguely remember as being 5*maximum
- element*epsilon).
-
- 2) Cholesky decomposition ditto (but the fudge constant is larger,
- possible 13*...). This is the method that the NAG code uses.
-
- 3) A = T'DT decomposition, as described above, and replace negative
- elements of D by zero.
-
- The speed stakes are (2) > (1) > (3) and the accuracy stakes are
- exactly the converse. In all cases, the difference is a small constant
- factor, so there is nothing much in it. For the theory of (1) and (2),
- see Wilkinson and Reinsch - I don't know much about (3).
-
-
- Nick Maclaren
- University of Cambridge Computer Laboratory,
- New Museums Site, Pembroke Street,
- Cambridge CB2 3QG, England.
- Email: nmm@cl.cam.ac.uk
- Tel.: +44 223 334761
- Fax: +44 223 334679
-