Numerical Integration Over Simplexes and Cones

Author(s): P. C. Hammer, O. J. Marlowe and A. H. Stroud

Source: Mathematical Tables and Other Aids to Computation , Jul., 1956, Vol. 10, No. 55
(Jul., 1956), pp. 130-137
Published by: American Mathematical Society

Stable URL: https://www.jstor.org/stable/2002483

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

American Mathematical Society is collaborating with JSTOR to digitize, preserve and extend
access to Mathematical Tables and Other Aids to Computation

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
 This content downloaded from
103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
 NUMERICAL INTEGRATION OVER SIMPLEXES AND CONES 131

Particular formulas were developed by Hammer and Wymore for certain

symmetrical type regions including spheres and hyperspheres, cubes and hyper-
cubes. The methods they have used make it possible to obtain numerical integra-
tion formulas for a much larger class of regions than heretofore. However, it is
not practical to obtain specific integration formulas for all regions of interest.
Hence, in this paper we develop methods for obtaining numerical integration
formulas over simplexes which may be used, in principle, to approximate other
regions. Since one of the methods is applicable to cones in general if a formula is
given for the base, we include that in the development.
While we have calculated certain formulas in specific cases for n = 2, 3, and 4,
we use these as illustrations and hope to build a more complete table later.
Inductively we give integration formulas holding exactly for the k-th degree
polynomial in n variables over the n-simplex. Another method which may be
more valuable in some applications requires affine symmetry of the evaluation
points. Here the general theory is missing, but a general type of method is pro-
posed for which illustrations are provided.
2. Approximate integration formulas for cones. In the following a theorem of
Hammer and Wymore will be used which we state as follows:
THEOREM 1. If

E a,f(%) f (t)d V = (f),

then

E Waig (j) - g (rq)d V = WE (f)

j T~~~R

where T is an affine tran

W is the absolute value of t
in the domain of f and 1, *
This theorem permits us to consider convenient, specific regions R to develop
formulas for the class of all their affine transforms.
We want integration formulas of the form
* k

(1) JRfd V- ajf((,),

where the numbers aj are constants; 41, * kare points in the domain of f; R is
a bounded closure of an open set in E.. From this standpoint the most interesting
simple regions are the simplexes, which are special cones, since every polyhedron
is composed of a finite number of simplexes.
Let an n-dimensional region R be embedded in the hyperplane x = 1 in E,,+,
where we represent the points in E,+, by (t, x), where t is a point in En. Then the
set of all points xR, where 0 < x < 1, is a cone C with base R and vertex at the

~~~~j
origin in E,+,.
Let f(t, x) be a function defined over C and suppose that a suitable numerical
integration formula is given over the base R of C. If, for example,

(2) f fw l) d Vn-E ajf( %, I)

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
132 NUMERICAL INTEGRATION OVER SIMPLEXES AND CONES

then

(3) f f(%, x)dV = dx f f( , x)d VT7J xn E a,f(xtj, x)dx,

since the Jacobian of the affine transformation from R to xR is x-n. Define a

function
g (x) = ajf (xj, x)

and then we have

ffdv fxng (x)dx.

Now we ask for numerical integration formulas of the form

(4) JXng(x)dx- big(xi).

Since such formulas may certainly be found we then have

(5) f fdv E biajf (x j~, xi),

ci i

which is of the form required, that is, a weighted sum of integrand values.
THEOREM 2. If a formula (2) holds precisely for polynomials f of n variables of
at most degree m over a region R and if a formula (4) holds precisely for polynomials
g of at most degree m in x, then (5) holds over C for all polynomialfunctions f contain-
ing terms of at most degree m in its n + 1 variables.
We will forego a formal proof of Theorem 2 since it follows from the affine
invariance of the class of polynomials of at most a certain degree, so that

Xn E ajf (xtj, x) = f f(, x)d Vn.

xR

In order to develop specific formulas one may start with any of the numerous
formulas for a line segment integration and proceed to obtain a formula for a
triangle; using this obtain a formula for a tetrahedron and so on. This is the type
we consider now. In higher spaces, to keep the number of points smaller, the
integration formulas we choose for fol xng (x)dx are based on orthogonal poly-
nomials with weight function Xn so that with m values of xj the formula holds
precisely for a polynomial of degree at most 2m - 1 by formula (3). For n = 0
this gives the Gauss integration formula on the line with points of evaluation the
roots of the m-th Legendre polynomial. Then if the formula (2) for the base is
exact for polynomials of at most degree 2m - 1 in the first n variables, the final
formula holds for polynomials of degree 2m - 1 at most. Thus by this means
we obtain with mn points a formula valid for all (2m - 1)-degree polynomials over
a simplex in n-space. While we have discussed a range from 0 to 1, the affine
invariance permits us to give the formulas as valid for all simplexes, the necessary

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
 NUMERICAL INTEGRATION OVER SIMPLEXES AND CONES 133

adjustment of coefficients and points being given by the transformation as in

Theorem 1.
The orthogonal polynomials, qm(x), over the interval (0, 1) with weight func-
tions Xn are given by Christoffel's formula in terms of the Legendre polynomials,
Pm(x), on the initerval (0, 1):

Pm (x) Pm+l (x) ... Pm+n (x)

Pm (0) Pm+l (0) ... Pm+n (0)
(6) xnqm (X) = P'm (0) P m+ (0) ... P m+n (0)

p(nm- (O) p(n (o) ... p(n

In this formula and in the tables the value of n is one less than the dimension
of the space-for the line, n = 0; for the plane, n = 1; and so on. The resulting

TABLE I. n = 1

m j xi bi qm(x)
1 1 0.66666 66666 66666 667 0.50000 00000 00000 000 i14(2 - 3x)

2 1 0.35505 10257 21682 190 0.18195 86182 56022 831 16(3 - 12x + lOx2)
2 0.84494 89742 78317 810 0.31804 13817 43977 169

1 0.21234 05382 39152 944 0.06982 69799 01454 1224

3 2 0.59053 31355 59265 287 0.22924 11063 59586 248 -18(4 - 30x + 60x2 - 35x3)
3 0.91141 20404 87296 044 0.20093 19137 38959 648

1 0.13975 98643 43780 552 0.03118 09709 50008 0822

4 2 0.41640 95676 31083 175 0.12984 75476 08232 439 V110(5 - 60x + 210X2 - 280
3 0.72315 69863 61876 278 0.20346 45680 10271 322 + 126x4)
4 0.94289 58038 85482 299 0.13550 69134 31488 149

1 0.09853 50857 98826 4273 0.01574 79145 21692 2766

2 0.30453 57266 46363 885 0.07390 88700 72616 6584 f12 (6 - 105x + 560X2 - 1260x3
5 3 0.56202 51897 52613 862 0.14638 69870 84669 768 + 1260x4 - 462x6)
4 0.8019865821 26391 897 0.16717 46380 94369 395
5 0.96019 01429 48531 218 0.09678 15902 26651 7818

TABLE II. n = 2

m j xi bi qm (x)
1 1 0.75000 00000 00000 000 0.33333 33333 33333 333 '5 (4x-3)

2 1 0.45584 81559 88774 713 0.10078 58820 79825 431 l7 (15x2-20x+6)

2 0.87748 51773 44558 620 0.23254 74512 53507 905

1 0.29499 77901 11501 618 0.02995 07030 08580 6981

3 2 0.65299 62339 61648 121 0.14624 62692 59866 020 -l9(56x3-105x2+60x-10)
3 0.92700 59759 26850 269 0.15713 63610 64886 615

1 0.20414 85821 03227 136 0.01035 22407 49918 0652

2 0.48295 27048 95632 480 0.06863 38871 72923 0663 111 (21Ox4 -504x3
3 0.76139 92624 48137 593 0.14345 87897 99214 191 +420x2-140x+15)
4 0.95149 94505 53002 709 0.11088 84156 11277 894

1 0.14894 57870 52983 580 0.00411 38252 03099 00782

2 0.36566 65273 69113 217 0.03205 56007 22961 9169 -l13(792x5-2310x4
5 3 0.61011 36129 34480 701 0.08920 01612 21590 0168 +252Ox3-1260x2+280x-21)
4 0.82651 96792 28304 566 0.12619 89618 99911 440
5 0.96542 10600 81784 870 0.08176 47842 85770 9715

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
134 NUMERICAL INTEGRATION OVER SIMPLEXES AND CONES

TABLE III. n = 3

m j Xi bi qm(X)
1 1 0.80000 00000 00000 000 0.25000 00000 00000 000 A/6(4 - 5x)

2 21 0.89871
0.52985 79358 94884 910 0.06690 52498 06888 7467 A/8(10-30
34926 76543 662 0.18309 47501 93111 252

1 0.36326 46302 16511 947 0.01647 90592 82671 7230

3 2 0.69881 12691 63613 535 0.10459 98975 56806 681 ,10(20-105x+168X2-84X3)
3 0.93792 41006 19874 523 0.12892 10431 60521 608

1 0.26147 77888 30889 686 0.00465 83670 60069 48897

2 0.53584 64460 88250 229 0.04254 17241 42766 6674 A/12(35 -280x+756X2-840X3
3 0.79028 32299 69286 800 0.10900 43689 38641 000 +330x4)
4 0.95784 70805 66118 662 0.09379 55398 58522 9295

TABLE IV. Integration Formula over Triangle Exact for 7th Degree Polynomial
k yk ak

1 0.86113 63115 94052 580 0.34785 48451 37453 860

2 0.33998 10435 84856 264 0.65214 51548 62546 143
3 -0.33998 10435 84856 264 0.65214 51548 62546 143
4 -0.86113 63115 94052 580 0.34785 48451 37453 860

Points of evaluation (xi, XJYk)

j Xj XjYl = -XjY4 XjY2 = XjY3
1 0.13975 98643 43780 552 0.12035 22940 89888 328 0.04751 57045 30876 4547
2 0.41640 95676 31083 175 0.35858 53991 82305 152 0.14157 13593 61934 441
3 0.72315 69863 61876 278 0.62273 67399 39136 723 0.24585 96668 98990 366
4 0.94289 58038 85482 299 0.81196 18147 75453 378 0.32056 66993 96768 242

wik = biak = weights for points (xi, Xiyk)

j k = 1, 4 k = 2, 3
1 0.01084 64518 21050 5090 0.02033 45191 28957 5733
2 0.04516 80985 64739 8624 0.08467 94490 43492 5770
3 0.07077 61357 96171 8794 0.13268 84322 14099 443
4 0.04713 67363 86764 6765 0.08837 01770 44723 4729

qm(x) are orthogonal, but not necessarily normal. The roots are the valu
required. Normalization of the qm(x) gives the weights bj by:
rn-1

(7)=Z
(7) ~~~~~~bi =E E qi (Xj) ]2
i=O

where xj is a zero of qmn (x).

The orthogonal polynomials, qn(x), over the interval (0, 1) with weight func-
tion Xn are the Jacobi polynomials under the linear transformation x' = 2 (1 + x).
The standard definition of the Jacobi polynomials gives the weight function as
(1 - x)a(1 + x)f and the interval of orthogonality as (-1, 1). In this case,
a = 0, f3 = n and the linear transformation given above takes (1 + x)n into
and (-1, 1) onto (0, 1). From this fact, explicit representations of qm(x) are
available.
We consider an example of how to combine a formula for n = 0 (the Gauss
formula on the line) and one for n = 1, to obtain a formula for the triangle. In
the following discussion we consider a plane triangle with vertices (0, 0), (1, 1)

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
 NUMERICAL INTEGRATION OVER SIMPLEXES AND CONES 135

and (1,-1). Use of the formulas may be made for arbitrary triangles by applying
theorem 1. If we take, in each of the above cases, m = 4, a formula is obtained
which is exact for the general polynomial of degree 2m -1 = 7. For n = 1, xj
shall denote the roots of q4 (x), and bj the corresponding weight. The roots of the
Legendre polynomial, P4, shall be denoted by Yk, and the corresponding weights
by ak. Table IV gives the values of xj, bj, and Yk, ak; the points, (Xj, XjYk), at which
the integrand is to be evaluated, and the weights, Wjk, at these points.
The first three tables give: 1) the polynomials qm(x) obtained from equation
(6) for n = 1, 2, 3 respectively, and values of m indicated in each table, 2) the
roots xj of qm (x), and 3) the values of bj obtained from equation (7). The calcula-
tions were made at the University of Wisconsin Numerical Analysis Laboratory
using a CPC with Eugene Albright's 18-digit floating decimal board. The approxi-
mate error in the x,'s is no more than 1 in the seventeenth significant figure; the
approximate error in the bj's is no more than 1 in the sixteenth significant figure.
The sum of the bj's is 1/ (n + 1) where n is defined as in equation (6).
-3. Symmetrical formulas. While the foregoing development makes it possible,
in principle, to obtain numerical integration formulas for the n-simplex to hold
exactly for polynomials of at most degree k, we do not mean to suggest that such
formulas are the most desirable. One feature of these formulas is that they are
unsymmetrical-i.e., in a given simplex the particular points of evaluation are
not an invariant set under affine transformations taking the simplex onto itself.
In this section we give the preliminary results of our investigations resulting
from a requirement of affine symmetry. We give these results since we believe that
formulas for triangle and tetrahedron will be most useful and we have certain
specific formulas for these regions.
The requirement of affine symmetry we make is simply this: If an integration
formula involves calculation of the integrand at a certain point P to be multiplied
by a weight, w, then all images of P under all affine transformations of the region
onto itself will appear in the formula with the same weight, w. While such a re-
quirement would appear to increase the number of points, in dealing with poly-
nomials this is not the case; actually we have obtained fewer points than with
the other method.
Let us represent points in the space as vectors and write the vertices of a
3

triangle as V1, V2, V3, and the centroid as C = 3 E Vi. The first affine invariant

formula for the triangle is to use the centroid as the sole evaluation point with
weight equal to the area. This method works for all bounded regions in all finite
dimensional spaces to give a formula for the general linear function over the region.
For simplexes or for hypersquares, this is likely to be useful.
The quadratic function over tht! triangle we have shown to be integrated
exactly by evaluations at rVi + (1 - r)C, where i = 1, 2, 3. Since this is an
affinely symmetric set, we find that the weight (for each point) is one-third the
area, and r = t 2. For r = ? 2 the evaluation points are the distinct trisection
points of the median chords and for r = - the evaluation points are the mid-
points of the sides. These formulas we consider to offer prospects of extensive
usefulness, especially the latter. Since the general quadratic function has 6 terms,
we have used three fewer points than an arbitrary specification of evaluation
points would have permitted.

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
136 NUMERICAL INTEGRATION OVER SIMPLEXES AND CONES

The cubic polynomial is integrated over the triangle with r Vi + (1 - r) C

and C as evaluation points where r = 2; the weight associated with the centroid
is (--9A)A and the weight with each of the other three points is (25/48)A, where
A is the area of the triangle.
The quintic polynomial is integrated precisely with seven points in the triangle
using rVi + (1 - r)C, weight a; sVi + (1 - s)C, weight b; and C, weight c. We
1 + - - 1-5 155 - -{15155 + @1
find r= , =and a= 15-J A, b= (A?+
7 7 1200 / 1200 /
c = (9/40)A. Since the general quintic polynomial in two variables has 21 terms
this formula appears to be a type we call efficient noting that one might not hope
to accomplish a formula with fewer than 7 = 21/3 points. Here the "3" is the
number of degrees of freedom for each point due to coordinates and weight.
However, there are known hyperefficient formulas (cf. [1]), which use fewer poin
than indicated by this argument. While we will not reproduce the argument here,
we used a triangle with vertices (0, 0), (1, -1), (1, 1). Then the requirements of
the affine symmetry of the formula with the form of the region assured that all
monomials with odd powers of y could be omitted. This left 12 equations. We
chose five of these and solved them for a, b, c, r, and s, and verified that the
remaining 7 were satisfied.
Over the tetrahedron we have a formula for the general quadratic in three
variables involving rVi + (1- r)C, i = 1, 2, 3, 4, and weight a. We calculate
r-1/45 and the weight a = (Q)A, where A is the volume of the tetrahedron.
Another formula with points outside the tetrahedron results if r = -1/-f5. This
type of formula will generally be less used than one with points inside.
For the cubic polynomial over the tetrahedron we have r = I, a = (9/20)A,
and c = (-4-)A; the centroid C must now be inicluded and c is its weight. This
formula is efficient since it involves 5 points and the number of terms in the general
cubic is 20.
Generalization and extension of the affinely symmetric methods is now being
carried out. However, these specific results seem sufficiently useful to include now.
4. Conclusion. In this paper we have taken a step towards obtaining reason-
able numerical integration formulas over simplexes and cones. The cones which
are not simplexes have not been emphasized. However, the method devised for
cones permits obtaining a formula for integration over a cone (finite) provided
a formula is at hand for the base. Thus later on we hope to obtain other formulas
for the solid sphere which is a special cone based on its surface.
Error analysis has not been attempted. Experimental calculations on simple
regions indicated that for "reasonable" functions there is a decisive factor in favor
of the classical formulas here presented over Monte Carlo methods.
P. C. HAMMER
University of Wisconsin
Madison, Wisconsin
0. J. MARLOWE
Westinghouse Elec. Corp.
Atomic Power Div.
Pittsburgh, Pennsylvania
A. H. STROUD
Chemstrand Corp.
Alabama

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms
 NUMERICAL INTEGRATION OVER SIMPLEXES 137

This work is supported in part by a grant of Wisconsin Alumni Research Foundation funds
made by the Graduate Research Committee.
1. P. C. HAMMER & A. W. WYMORE, "Numerical integration over higher dimensional regions."
Unpublished manuscript.
2. G. SZEGO, Orthogonal Polynomials, Am. Math. Soc. Colloquium Publications, v. 23, New York,
1939.
3. G. W. TYLER, "Numerical integration of functions of several variables," Canadian J. Math.,
v. 5, 1953, p. 393-412.

Numerical Integration over Simplexes

1. Introduction. Integration formulae for numerical evaluation of integrals

over the simplex in n-space have been given inductively by Hammer, Marlowe,
and Stroud [1] so that it is possible in principle to determine a formula holding
exactly for the kth degree polynomial in n variables. In the same paper certain
affinely symmetric integration formulae are given for the triangle and tetrahedron.
Using the theory proposed by Hammer and Wymore [2], it is possible to extend
the usefulness of methods developed by transformations of the regions and by
use of Cartesian products.
In this paper we give two integration formulae of affinely symmetric type for
the simplex in n-space which respectively hold exactly for the quadratic poly-
nomial and the cubic polynomial in n variables. The method for establishing the
exact values of integrals needed we believe is new in that the "numerical" formulae
are used for the purpose.
2. The formula for cubic polynomials. Let the vertices of the n-simplex, Sn,
n
be Vo, * V*, In and then its centroid is given by C = E Vi/(n + 1). Let An be

the hypervolume of Sn.

THEOREM 1: An integration formula exact for the general cubic polynomial over
Snfor n > 1 is given by

r ~~~n
(1) f fdvn= an a f(Ui) + cnf(C)
Sn 0

where

an (n + 3)2 -(n + 1)2

4(n + 1)(n + 2) 4(n + 2)
and
2 n+ 1
Ui f+ Vi + + C i =O, ,n.
n +3 n +3

Proof: It may first be remarked that the points Ui are on the median lines of
Sn and that the statement of the theorem is in symmetric form. In particular,
under an affine transformation taking Sn onto itself, the set of points { Ui} is
invariant and the centroid C is fixed. Since there exists an affine transformation
mapping any simplex in En onto any other we may choose any particular simplex
Sn to carry out the proof. Our choice is specified by vertices as follows:

Vo = (, . . ,), V1 = (1, 0, * ,0),

V2 (1. 1. 0, ' ' '. 0.* ' Ven = (1, 0. ** 0. 1).

This content downloaded from

103.158.43.17 on Fri, 01 Sep 2023 08:44:57 +00:00
All use subject to https://about.jstor.org/terms