The integral in equation (4.1), including its first and second derivatives, has to be computed at least once in each iteration, usually several times, and therefore the choice of method and accuracy has to be made with care. There are several numerical integration method, e.g. Trapezoid, Simpson, Romberg, and Gaussian quadrature. I will take a closer look at the latter one.

The Gaussian quadrature is a generalization of the the Newton-Cotes formula,

where are the abscissas and the weights. Newton-Cotes is a
polynomial interpolation, where the abscissas are equally spaced in the interval (a,b).
Trapezoid and Simpson are both a variation of this interpolation. The idea of
the Gaussian quadrature is that the abscissas are no longer restricted to be equally
spaced, and that they can be chosen to give higher accuracy. Another attracting feature
is that the weights and abscissas can be arranged to make the integral exact when the
integrand is a product of a function *f*(*x*) and and a weight function *W*(*x*). The weight
function can be chosen to remove singularities. The rewritten Newton-Cotes formula is
then

and is exact if *f*(*x*) is a polynomial of degree . The abscissas and
the weights are determined by a set of polynomials that are orthogonal over the weight
function *W*(*x*). For quite a few ``classical'' orthogonal polynomials for a given weight
function, abscissas and weights are either tabulated or can be found by given algorithms
Press.

In this treatise, Gaussian quadrature with weight function has been
applied, and has proven to be very efficient. It is highly superior to other methods like the Trapezoid and the Simpson
method, both when it comes to speed and accuracy. To achieve satisfactory accuracy in
calculating equation (4.1) to, say, seven significant figures, 10
abscissas are needed, thus requiring the calculations of the integrand at only 10 points
in time. The argument of Gia applies here: If the number of abscissas used in
the Gaussian quadrature exceeds by one the number *N* in the truncated expansion in
equation (3.21), the error introduced by the Gaussian quadrature is totally
negligible compared with the error in the truncated expansion itself.

The abscissas and weights used can be easily deduced from the respective abscissas and weights for the Gauss-Legendre quadrature;

where *GL* denote Gauss-Legendre [page 276]NM.

The Gaussian quadrature method with the weight function chosen here seems to be comparable to the one used by Gia, and greatly superior to the integration methods used otherwise in connection with the AVP.

Mon Jul 5 02:59:28 MET DST 1999