Thread Subject: SVD output

Matlab's SVD() function yields a result very close to, but
different than the result when using LAPACK's sgesvd()
method in C++. Any idea why? By default, svd uses double-
precision, but I'm using the single precision version of
svd, which, according to Matlab's documentation, utilizes
sgesvd in its computation. Since they both use sgesvd,
shouldn't the results be equal?

I'm using ACML v4.0.1 and Matlab 7.4.0

Thanks,
Bryan

On May 6, 9:22=A0am, "Bryan " <bdg146.removeT...@gmail.com> wrote:
> Matlab's SVD() function yields a result very close to, but
> different than the result when using LAPACK's sgesvd()
> method in C++. =A0Any idea why? =A0By default, svd uses double-
> precision, but I'm using the single precision version of
> svd, which, according to Matlab's documentation, utilizes
> sgesvd in its computation. =A0Since they both use sgesvd,
> shouldn't the results be equal?
>
> I'm using ACML v4.0.1 and Matlab 7.4.0

How are the results different?

Greg

Greg Heath <heath@alumni.brown.edu> wrote in message
> How are the results different?
>
> Greg


Given a <168 x 2> input of single-precision values between
0 and 1:

The max difference between the V output matrices is
1.7881e-7 and the max difference between the U output
matrices is 3.7625e-7. The sigma values are the same.

Computing X(input) - U*S*V' shows that the C++ sgesvd()
output is better than the Matlab SVD() output. However,
that may be pure coincidence and a different input may
yield different results.

"Bryan " <bdg146.removeThis@gmail.com> wrote in message
<fvpm1r$bps$1@fred.mathworks.com>...
> Matlab's SVD() function yields a result very close to, but
> different than the result when using LAPACK's sgesvd()
> method in C++. Any idea why? By default, svd uses double-
> precision, but I'm using the single precision version of
> svd, which, according to Matlab's documentation, utilizes
> sgesvd in its computation. Since they both use sgesvd,
> shouldn't the results be equal?
>
> I'm using ACML v4.0.1 and Matlab 7.4.0
>
> Thanks,
> Bryan
The uniqueness of SVD decomposition is a recurrent topic.
Google for "svd decomposition not unique".
Anyway, I agree that same routines should (on principle)
agree in their results... unless they resort to other
routines (like rand()) which might differ from machine to
machine and explain the issue.
To be convinced you should check whether or not your present
results agree with the definition (i.e. check that the
product equals the original matrix).
Hope this helps.
Regards

"carlos lopez" <clv2clv_00000000_@adinet.com.uy> wrote in
message <fvpulg$rdh$1@fred.mathworks.com>...
> The uniqueness of SVD decomposition is a recurrent topic.
> Google for "svd decomposition not unique".
> Anyway, I agree that same routines should (on principle)
> agree in their results... unless they resort to other
> routines (like rand()) which might differ from machine to
> machine and explain the issue.
> To be convinced you should check whether or not your
present
> results agree with the definition (i.e. check that the
> product equals the original matrix).
> Hope this helps.
> Regards

Thanks for the input.

I understood that the SVD is not unique, but I assumed
that, since the underlying computations are both being
conducted with LAPACK, their solution should be unique. I
understand your point though regarding relying on other,
non-uniform routines though.

I have checked the solution via X-U*S*V' and comparing
that result to 0. However, there seems to be no definite
conclusion as to which gives the better result. For some
data, the C++ output is closer to 0, while for others,
Matlab is closer.

So I suppose getting the C++ svd output to equal the
Matlab svd output isn't really possible?