Re: multidimensional integration

["Robert G. Brown" <rgb@phy.duke.edu>]

[-- Attachment #1: Type: text/plain, Size: 5753 bytes --]

Manoj Warrier writes:

> Is not Monte-Carlo integration the recommended method for
> multi-dimensional integration?
> The GSL specific routines for this are described at:
> http://www.gnu.org/software/gsl/manual/gsl-ref_23.html#SEC371

I >>think<< that this is in an edge case of sorts.  MC is definitely
right for very high dimensionalities (say, 100 dimensions).  In those
cases it is the only game in down since even bisecting each dimension a
single time leaves one with 2^100 cells to sum over, which is "bad".  In
1 dimension, though, it is very slow to converge compared to more
analytic and adaptive methods, especially for smooth integrands with
support in only certain parts of the integration region.

In between there is clearly a crossover.  In (say) five dimensions
bisection introduces 2^5 or 32 cells -- covering any one of those cells
more finely in an adaptive mesh isn't THAT expensive in its scaling if
the function has reasonable support and smoothness.  Adaptive methods in
this case give you the benefit of good control over errors provided only
that significant "features" if the integrand are large enough to be
picked up by the initial meshes.

> I think your point (b) below hits the nail on the head.
> I have not used these techniques, and would like to hear other,
> more expert opinion.

The person who wants to do this problem has already indicated that MC
methods are very slow and not horribly accurate for this problem,
suggesting that it is on the multidimensional mesh side of the
crossover.  In any event, even in 2 dimensions it is likely desirable to
have a real multidimensional integration routine instead of just calling
a 1 d routine twice, once inside the other.  In 2d the refinement mesh
would have to cover the regions where the integrand is supported in
stripes that reach all the way across the long dimensions, resulting in
lots of calls at points where the integrand has little support -- this
gets to be LOTS of calls in signficant subvolumes in d = 3 or 4 or 5 as
one has to refine the mesh across all dimensions where the support has a
signficant projection, even if the support itself has relatively little
volume.

I'm guessing that HALF does something like mesh refinement by halving
the scale (doubling the mesh) successively in subvolumes only to the
extent required to get each successive subvolume to converge, so that a
coarse mesh is used over all regions where the integrand is smooth and
small but a fine one covers regions where it varies rapidly and/or is
large.  How smart it is with respect to interpolating functions or the
like, I don't know.

It sounds like there is nothing like this in the GSL, though, for the
intermediate/low dimensions where it would solidly beat MC.

    rgb

> 
> Best Regards.
> Manoj
> 
> On Wed, 29 Jun 2005, Robert G. Brown wrote:
> 
>> Hi fellow GSL'ers.
>> 
>> We have a postdoc in our department who is preparing to integrate
>> something. In his previous position at another place, he used NAG to do
>> this, and has the requisite code already in place.  He requested that we
>> buy and install a single copy of NAG just for him and a student to be
>> able to use this one routine to do this one integral on just one
>> computer, at a cost of many hundreds of dollars.
>> 
>> I suggested that he look into using the GSL instead, since it is a very
>> high-quality library to my own direct experience and of course is both
>> free and universally installed in our department.  GSL and NAG both use
>> QUADPACK as the basis for their 1D integrals (and have nearly identical
>> call structure) so I figured that the transition would actually be
>> painless.
>> 
>> However, the integrand he has to integrate is actually defined and
>> integrated over somewhere between 5 to 7 dimensions (with rectangular
>> limits).  The routine he used from NAG was actually d01fcc, which is NOT
>> from QUADPACK but rather implements the multidimensional adaptive
>> routine HALF with a custom interval rule.  When I looked at GSL's online
>> manual (version 1.6 as of this last December) I didn't see a
>> multidimensional integration routine equivalent to d01fcc.
>> 
>> SO, questions:
>> 
>>   a) Is a multidimensional integration routine equivalent to d01fcc
>> implemented or under development, and if so, where is it and/or how do I
>> get a version that has it?  I looked at the CVS tree and didn't
>> immediately see one.  In principle I could probably use e.g. a
>> multidimentional ODE solver but I'd think that having a d01fcc
>> equivalent would be much more efficient.
>> 
>>   b) If not, does anybody have any suggestions on the "best" way to
>> attack this sort of integral using existing tools?  At five dimensions I
>> suspect that just calling 1 dim integrations five levels deep would
>> result in an awful lot of wasted energy and time.  Framing it as an ODE
>> set also seems like it would work but likely not be terribly efficient
>> or terribly easy to control error-wise.
>> 
>>   c) On a related note, has anybody done a head-to-head performance
>> comparison of GSL with NAG -- either time/efficiency performance or
>> numerical accuracy type performance?  This isn't a significant issue on
>> this particular project but is an issue that I expect to see come up in
>> the future.
>> 
>>    rgb
>> 
> 
> -- 
> Manoj Warrier (manoj.warrier@ipp.mpg.de)
> 
> Stellaratortheorie, Max-Planck Institut Fur Plasmaphysik
> TeilInstitut Greifswald Wendelsteinstrasse 1
> D-17491 Greifswald Germany Tel: +49-3834-882434
> 
> --------- History of Computing 10-11-3003 ---------------
> Then there used to be this great user friendly OS which
> overwrote your MBR whenever you installed it.
> ---------------------------------------------------------

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]