Thread Subject: Large Overdetermined System

I have a very large overdetermined linear system of
equations that I would like to solve. It is a system of
form A*x = b, where A is M x N with M > N, and A and b are
both known. All entries are real.

When the product of M*N is sufficiently small,
MATLAB's "lscov" works well. However, when M x N becomes
something like 11000 x 9000 (which is the kind of problem I
am solving), MATLAB gives an out of memory error message
that I cannot get around by just closing other programs or
using pack, clearing workspace variables, etc.

Might there be a way to load part of my matrix into memory
at a time and solve the problem in steps?

"J RP" <philNOSPAMsonj@hotSPAMMENOTmail.com> wrote in message
<g3p1ag$8u9$1@fred.mathworks.com>...
> I have a very large overdetermined linear system of
> equations that I would like to solve. It is a system of
> form A*x = b, where A is M x N with M > N, and A and b are
> both known. All entries are real.
>
> When the product of M*N is sufficiently small,
> MATLAB's "lscov" works well. However, when M x N becomes
> something like 11000 x 9000 (which is the kind of problem I
> am solving), MATLAB gives an out of memory error message
> that I cannot get around by just closing other programs or
> using pack, clearing workspace variables, etc.
>
> Might there be a way to load part of my matrix into memory
> at a time and solve the problem in steps?

Not really. Better is to use a sparse form
to do this. Often when you have that large
of a matrix, it has many zeros, i.e., it is
sparse.

So since an 11000x9000 martrix will take
nearly a gigabyte of ram to store, you will
gain if there are zeros.

Next, try using an iterative solver to solve
the problem, not lscov.

John

"J RP" <philNOSPAMsonj@hotSPAMMENOTmail.com> wrote in message
<g3p1ag$8u9$1@fred.mathworks.com>...
> I have a very large overdetermined linear system of
> equations that I would like to solve. It is a system of
> form A*x = b, where A is M x N with M > N, and A and b are
> both known. All entries are real.
>
> When the product of M*N is sufficiently small,
> MATLAB's "lscov" works well. However, when M x N becomes
> something like 11000 x 9000 (which is the kind of problem I
> am solving), MATLAB gives an out of memory error message
> that I cannot get around by just closing other programs or
> using pack, clearing workspace variables, etc.
>
> Might there be a way to load part of my matrix into memory
> at a time and solve the problem in steps?

Not really. Better is to use a sparse form
to do this. Often when you have that large
of a matrix, it has many zeros, i.e., it is
sparse.

So since an 11000x9000 martrix will take
nearly a gigabyte of ram to store, you will
gain if there are zeros.

Next, try using an iterative solver to solve
the problem, not lscov.

John

You might want to give TLSQR a shot: http://tomopt.com/scripts/register.php

If you use the save command before the solver call you can send me the
input data. Then you need not waste any time.

Best wishes, Marcus
Tomlab Optimization Inc.
http://tomopt.com/
http://tomdyn.com/