JAMA Cholesky

[Cleve Moler <moler@mathworks.com>]

On Tue, 2 Apr 2002, Boyce Byerly wrote:

> 
> I sent out a message to that effect to the mailing list 
> associated with the group that maintains this.  Since
> the project *is* open source, I might just fix the problem
> myself and release my changes back to the group; that's the
> sort of way progress gets made for packages like that.   I
> haven't heard back from the group on this yet.  My experience
> has been that usually commercial software is no better than 
> open source (for example, StatBeans wanted a lot of money, and
> our opinion of that software is quite low.)

I didn't respond to your original message because I found it offensive.


> The symetry issue I'm less concerned with; memory is cheap,
> and it may well be that cutting that sort of corner would
> make Cholesky an odd case that doesn't play well with the 
> other features available in the matrix classes.

That's absolutely right.

> 
> -Boyce
> 
> 
> -----Original Message-----
> From: uemcre@unc.edu [mailto:uemcre@unc.edu]
> Sent: Monday, April 01, 2002 5:59 PM
> To: Multiple recipients of list
> Subject: http://math.nist.gov/javanumerics/jama/
> 
> 
> 
> I looked at the Cholesky program to see how it handled error conditions
> and found that it only tests that the matrix is symmetric. ...

You didn't look carefully enough.

> ... If the rest of
> the package is anything like this, it certainly is amateurish and not what
> I would expect from the Bureau.  

This remark is uncalled for.

> 1. There is no reason to store the whole matrix and thus no reason to
> check for symmetry.
> 
> 2. You MUST check for singularity and positive-definiteness

We do.

> The program fails on the matrix
>  0 1
>  1 1

Yes, this matrix is not positive definite. The JAMA CholeskyDecomposition
quickly discovers this fact.

Please read the documentation or the source code completely before
continuing this thread.  Pay particular attention to "isSPD".

  -- Cleve Moler
  moler@mathworks.com