suave {cubature}  R Documentation 
Integration with SUbregionAdaptive Vegas Algorithm
Description
Suave uses vegas()
like importance sampling combined with a
globally adaptive subdivision strategy: Until the requested accuracy is
reached, the region with the largest error at the time is bisected in the
dimension in which the fluctuations of the integrand are reduced most. The
number of new samples in each half is prorated for the fluctuation in that
half.
Usage
suave(
f,
nComp = 1L,
lowerLimit,
upperLimit,
...,
relTol = 1e05,
absTol = 1e12,
minEval = 0L,
maxEval = 10^6,
flags = list(verbose = 0L, final = 1L, smooth = 0L, keep_state = 0L, level = 0L),
rngSeed = 0L,
nVec = 1L,
nNew = 1000L,
nMin = 50L,
flatness = 50,
stateFile = NULL
)
Arguments
f 
The function (integrand) to be integrated as in

nComp 
The number of components of f, default 1, bears no relation to the dimension of the hypercube over which integration is performed. 
lowerLimit 
The lower limit of integration, a vector for hypercubes. 
upperLimit 
The upper limit of integration, a vector for hypercubes. 
... 
All other arguments passed to the function f. 
relTol 
The maximum tolerance, default 1e5. 
absTol 
the absolute tolerance, default 1e12. 
minEval 
the minimum number of function evaluations required 
maxEval 
The maximum number of function evaluations needed, default 10^6. Note that the actual number of function evaluations performed is only approximately guaranteed not to exceed this number. 
flags 
flags governing the integration. The list here is exhaustive to keep the documentation and invocation uniform, but not all flags may be used for a particular method as noted below. List components:

rngSeed 
seed, default 0, for the random number
generator. Note the articulation with 
nVec 
the number of vectorization points, default 1, but can be set to an integer > 1 for vectorization, for example, 1024 and the function f above needs to handle the vector of points appropriately. See vignette examples. 
nNew 
the number of new integrand evaluations in each subdivision. 
nMin 
the minimum number of samples a former pass must
contribute to a subregion to be considered in that region's
compound integral value. Increasing nmin may reduce jumps in
the 
flatness 
the parameter p, or the type of norm used to compute the fluctuation of a sample. This determines how prominently "outliers," i.e. individual samples with a large fluctuation, figure in the total fluctuation, which in turn determines how a region is split up. As suggested by its name, flatness should be chosen large for "flat" integrands and small for "volatile" integrands with high peaks. Note that since flatness appears in the exponent, one should not use too large values (say, no more than a few hundred) lest terms be truncated internally to prevent overflow. 
stateFile 
the name of an external file. Vegas can store its entire internal state (i.e. all the information to resume an interrupted integration) in an external file. The state file is updated after every iteration. If, on a subsequent invocation, Vegas finds a file of the specified name, it loads the internal state and continues from the point it left off. Needless to say, using an existing state file with a different integrand generally leads to wrong results. Once the integration finishes successfully, i.e. the prescribed accuracy is attained, the state file is removed. This feature is useful mainly to define ‘checkpoints’ in longrunning integrations from which the calculation can be restarted. 
Details
See details in the documentation.
Value
A list with components:
 nregions
the actual number of subregions needed
 neval
the actual number of integrand evaluations needed
 returnCode
if zero, the desired accuracy was reached, if 1, dimension out of range, if 1, the accuracy goal was not met within the allowed maximum number of integrand evaluations.
 integral
vector of length
nComp
; the integral ofintegrand
over the hypercube error
vector of length
nComp
; the presumed absolute error ofintegral
 prob
vector of length
nComp
; the\chi^2
probability (not the\chi^2
value itself!) thaterror
is not a reliable estimate of the true integration error.
References
T. Hahn (2005) CUBAa library for multidimensional numerical integration. Computer Physics Communications, 168, 7895.
See Also
Examples
integrand < function(arg) {
x < arg[1]
y < arg[2]
z < arg[3]
ff < sin(x)*cos(y)*exp(z);
return(ff)
} # end integrand
suave(integrand, lowerLimit = rep(0, 3), upperLimit = rep(1, 3),
relTol=1e3, absTol=1e12,
flags=list(verbose=2, final=0))