Re: question about inverse_matrix() vs solve_linsys()

Hi Will

Both NCL functions are merely interfaces to LAPACK codes.
Both use the LAPACK DGETRF to perform the LU factorization.

The description for 'inverse_matrix' contains
" However, if efficiency is a concern, this is not the preferred method
(see solve_linsys)."

Hence, 'solve_linsys' looks like the way to go.

-----
solve_linsys: DGESV (DGETRF, DGETRS)

driver: http://www.netlib.org/lapack/double/dgesv.f
lu-fact: http://www.netlib.org/lapack/double/dgetrf.f
solver: http://www.netlib.org/lapack/double/dgetrs.f
-----

inverse_matrix: DGETRF, DGETRI

lu-fact: http://www.netlib.org/lapack/double/dgetrf.f
inverse: http://www.netlib.org/lapack/double/dgetri.f

D

On 4/25/13 6:56 PM, Will Hobbs wrote:
> Hi all
>
> The documentation for the inverse_matrix() function states that it uses LU factorisation to invert, but also states that for most solutions the function solve_linsys() is more efficient, but it also uses LU factorisation.
>
> Theoretically, that means that for an invertible matrix A,
>
>> Ainv = inverse_matrix(A)
>
> is the same as
>
>> I = conform(A, 0, -1) ;create identity matrix 'I'
>> do i = 0, nrow-1
>> I(i,i) = 1
>> end do
>> Ainv = solve_linsys(A, I)
>
> Mathematically they're the same, but is there any way in which the functions are coded that would make one more efficient than the other? Or is the 'inverse_matrix()' function just doing what I did in the second case?
>
> I'm dealing with large covariance matrices for a lot of model control runs, so even a modest difference in efficiency would be interesting to know.
>
> Many thanks
>
> Will
>
>
>
>
> _______________________________________________
> ncl-talk mailing list
> List instructions, subscriber options, unsubscribe:
> http://mailman.ucar.edu/mailman/listinfo/ncl-talk
>
_______________________________________________
ncl-talk mailing list
List instructions, subscriber options, unsubscribe:
http://mailman.ucar.edu/mailman/listinfo/ncl-talk
Received on Fri Apr 26 07:32:23 2013