D114 Adaptive Multidimensional Monte-Carlo Integration

RIWIAD is an adaptive multidimensional integration subroutine based on an original by G. Sheppey. It permits numerical integration of a large class of functions, in particular those that are irregular at the border of the integration region. The integral is always performed over the unit hypercube.

Structure:

SUBROUTINE subprogram
User Entry Names: RIWIAD
Files Referenced: Unit 6
External References: RNDM (V104) user-supplied FUNCTION subprogram
COMMON Block Names and Lengths: ANSWER2, INTERN7, OPTION3, PARAMS4,
RANDOM1, STORE77, STORE110001

Usage:

See Long Write-up for a description of all features. Here only the standard use is described.

The COMMON block PARAMS must always be set by the user: COMMON /PARAMS/ ACC,NDIM,NSUB,ITER

ACC

( REAL) Relative accuracy desired.

NDIM

( INTEGER) Number of dimension parameters.

NSUB

( INTEGER) Number of subvolumes allowed.

ITER

( INTEGER) Maximal number of iterations.

    FUNCTION EXAMPLE(Q)
    REAL EXAMPLE,Q
    DIMENSION Q(7)
    ...
    END

    EXTERNAL EXAMPLE
    COMMON /PARAMS/ ACC,NDIM,NSUB,ITER
    ACC=0.01
    NDIM=7
    NSUB=10000
    ITER=5
    CALL RIWIAD(EXAMPLE)
    ...

The program allows extensive user control via the COMMON blocks. See Long Write-up for details.

Method:

RIWIAD is iterative and in a given iteration it divides the unit hypercube into a certain number of subvolumes by means of a given set of intervals on each axis. Within each subvolume it estimates the mean value and variance of the integrand by random sampling, and then calculates the Riemann sum over the subvolumes. Using the variances found projected onto each axis it calculates a set of new interval divisions to be used in the next iteration. It returns when the desired accuracy is obtained or when the maximum number of iterations has been performed.

Restrictions:

There is, in principle, no limitations on the number of dimensions, although the present version only allows up to 9-dimensional integrals. The maximal dimensionality can easily be increased.

Notes:

1. The program is rather slow and should preferably be used only when other methods (for instance Gaussian quadrature) fail due to the irregular behaviour of the integrand. The time consumption is essentially proportional to the product of NSUB and ITER.
2. The non-CDC/Cray implementation of RIWIAD has IMPLICIT DOUBLE PRECISION(A-H,O-Z), i.e. all non- INTEGER variables are DOUBLE PRECISION, including the user-supplied external function.

D200

G.A. Erskine