--- Begin Message ---
- From: Brian Gough <bjg at network-theory dot co dot uk>
- To: Andrew Jason Penner <ajpenner at Physics dot UManitoba dot CA>
- Date: Sun, 11 Aug 2002 22:53:42 +0100 (BST)
- Subject: Re: Bug in gsl_linalg_LU_decomp
- References: <15698.55864.207582.291866@debian><Pine.OSF.4.21.0208090843330.28930-200000@aurora.physics.umanitoba.ca>
Andrew Jason Penner writes:
> Okay, attached is a program that will compute the LU decomposition
> of a Hermitian matrix
>
> [1 2 3]
> [2 3 4]
> [3 4 5]
>
> It should fail, however it does produce values in the end. The
> output is sent to the screen.
The LU decomposition is valid (i.e. if you multiply p'*l*u it gives
the original matrix).
The solver uses the decomposition and gets an answer, it's inaccurate
but there is no division by zero, due to finite precision.
What's needed is an estimate of the error or condition number, which
we currently don't have a function for. The underlying blas
triangular solvers don't do it.
--- End Message ---