Effi ci ent eval uat i on of i ntegral s of order l / r l / r 2, l / r 3
usi ng Gauss quadrat ure
LIU JUN, * G. BE E R a n d J. L. ME E K Department of Civil Engineering, University of Queensland, Australia An efficient met hod is presented for the numerical evaluation of integrals as t hey are encountered in the two- and three-dimensional i mpl ement at i on of the Boundary Element Method where the inte- grands are of 0(1/r), 0(1/ r 2) and 0(1/ra). Key Words: Boundary Element Method, ~ntegration INTRODUCTION The two- and three-dimensional i mpl ement at i on of t he Boundary Element Method (BEM) requires the evaluation of integrals where the integrand is 0( l / r ) , 0(1/ r 2) and 0(1/ra), where r is a distance from a source point. For a general i mpl ement at i on using isoparametric boundary ele- ment s these integrands have to be evaluated numerically. The Gauss-Quadrature (G-Q) formul a ~ has been used almost exclusively in the pa s t ) - s For the efficient i mpl ement at i on of the BEM it is very i mport ant to select the mi ni mum number of Gaussian points required to give a sufficient precision of integration. This may include the subdivision of the element into a number of sub-elements. This selec- t i on of t he integration order has been discussed in previous paper s ) ' a In this paper, we examine these met hods in more detail and present a new met hod for one- and two-dimen- sional integration. Some examples are presented and com- parison is made wi t h the met hods given in reference 2. ONE-DIMENSIONAL GAUSS-QUADRATURE The one-dimensional formula for G-Q can be written as: 1 /(x) d x = y . A . / ( x . ) ( 1 ) n = l - - 1 In the above N is the t ot al number of integration (Gaussian) points, x n their co-ordinates and A n the weight coefficients. The number of points needed depends on the required precision of the integration and the nature of the integrand. Stroud and Secrest I give an estimate of the error as: 1 f ( x ) dx -- ~ Anf (Xn) <e H (2) n = l - - 1 where H~l d2Nf [dxZN[ and e ~ 4[(22N(2N)!) *Visiting scholar, Institute of Geology, Academia Simica, Bering, China. Paper presented at the 7th Boundary Element Forum in Japan sponsored by JASCOME. Discussion closes November 1985. For the simple one-dimensional integration shown in Fig. 1, the value of H can be determined in the following way (here the Iacobi an is constant): 1 L f =- - r =- - ( 1 - - ~ ) +R r 2 ( 2N) ! L 2N H> 22Nr2N+l (3) He 4 e = - - ~ (4) / (4,)2N/L2~v T w o e x t r e m e values c a n b e substituted for r: r = R gives maxi mum error r = R + L gives mi ni mum error I f we choose a value of r = R the error will be over- estimated and the estimate for the number of integration points will probabl y be conservative. This is shown in Fig. 2 where the predicted (curve 1) and actual (curve 2) number of integration points required for a given precision as a funct i on of R]L is shown. Obviously, we must work wi t h a value of r/ L = r*/L where R/ L <r*[L <R[ L + 1. Here we have: r* R a . . . . ~ - ( 5 ) L L L In equation (5) al l is a translation of the curve towards the N-axis as shown in Fig. 2 (curve 3). Substituting (5) into equation (4) and taking eo = 4(L/4R) 2N, the equation (4) for the function f l = 1/r can be written as: e(a x) = eo (6) Figure 1. One-dimensional element 0264-682Xl85/030118-06 $02.00 1 1 8 E n g i n e e r i n g A n a l y s i s , 1 9 8 5 , V o l . 2, N o . 3 1 9 8 5 C M L P u b U c a t i o n s Ef f i ci ent evaluation o f integrals using Guass quadrature: Li u Jun, G. Beer and J. L. Meek N 1 0 6 2 I " . C U R v z 3 1.0 2 0 " 30 Figure 2. Predicted error curve 1 (r*/L = R/L)~ curve 3 (r*]L = R [L + O. 3 5) - and actual error (curve 2) Using expressions ( 6) - ( 8) we pl ot diagrams in Figs. 3, 4 and 5 whi ch give t he Gaussian poi nt s needed f or t he inte- grat i on o f these f unct i ons versus t he closeness o f t he source poi nt ( R/ L) f or an allowable error e~ n) o f 10 - a and l 0 -4 respectively. I n all o f t he expressions, a l l is t aken t o be 0. 35. For a given ratio R/ L t he number , N, o f Gaussian poi nt s depends on bot h t he chosen allowable error and t he nat ure o f t he integrand. To check t he equat i ons ( 6) - ( 8) , as an exampl e, Fig. 4 shows a compar i son curve o f t he predi ct ed and act ual number o f i nt egrat i on poi nt s needed t o achieve a cert ai n accur acy (Ca(2)= 10 -a) f or f unct i on f 2 = 1/ta. We can see t hat bot h curves are very close t o each ot her. Therefore using t he above equat i on in t he numeri cal i nt egrat i on, we can obt ai n a result whi ch meet s t he accur acy requi red and is economi cal in comput er resources. Tables 1-3 show t he limiting values o f R[L f or part i cu- lar number s o f i nt egrat i on poi nt s and t he accur acy o f inte- N 6 2 Figure 3. Predicted number o f integration points, N f or f unct i on 1/r E~U a-" 1 0 - lO-,a 10-2 ZO " R / L N 10 6 10 - 4 lO-Z_ 10-z : Figure 4. Predicted and actual number o f integration poi nt s f or f unct i on 1/r 2 Similarly, t he expressi on f or t he f unct i on 1"2 = 1/ r 2 can be wri t t en as: e (2) = ( 2 N + 1) eo ( 7) and f or t he f unc t i onf a = 1/ra: e (3) = ( N + I ) ( 2 N + 1) eo ( 8 ) Where e 0), e (2) and e 0 ) are allowable errors. N 1 0 6 F ~ r e S. tion 1/r a 1 0 - 4 lO-Z Predicted number o f integration points f or func- Table 2. e = lO-3 f = 1/r jr= 1#2 f = 1#3 2 1.9882 (1.6382) 2.9730 (2.6230) 3.9127 (3.5627) 3 0.9961 (0.6461) 1.3776 (1.0276) 1.7357 (1.3857) 4 0.7050 (0.3550) 0.9279 (0.5779) 1.1346 (0.7846) 5 0.5730 (0.2230) 0.7283 (0.3783) 0.8712 (0.5212) 6 0.4990 (0.1490) 0.6179 (0.2679) 0.7267 (0.3767) 7 0.4521 (0.1021) 0.5486 (0.1986) 0.6364 (0.2864) 8 0.4198 (0.0698) 0.5012 (0.1512) 0.5749 (0.2249) 9 0.3963 (0.0463) 0.4668 (0.1168) 0.5305 (0;1805) 10 0.3785 (0.0285) 0.4407 (0.0907) 0.4968 (0.1468) Table 1. Number of integration points needed for precision,, e = 10- 3 f = 1/r f = l/r 2 f = 1/r 3 2 1.1180 (0.7680) 1.6719 (1.3219) 2.2003 (1.8503) 3 0.6786 (0.3286) 0.9386 (0.5886) 1.1825 (0.8325) 4 0.5287 (0.1787) 0.6958 (0.3458) 0.8508 (0.5008) 5 0.4551 (0.1051) 0.5785 (0.2285) 0.6920 (0.3420) 6 0.4119 (0.0619) 0.5100 (0.1600) 0.5998 (0.2498) 7 0.3835 (0.0335) 0.4654 (0.1154) 0.5399 (0.1899) 8 0.3636 (0.0136) 0.4340 (0.0840) 0.4979 (0.1479) 9 0.3487 (0.0000) 0.4107 (0.0607) 0.4668 (0.1168) 10 0.3374 0.3928 (0.0428) 0.4428 (0.0928) Ef f i ci ent evaluation o f integrals using Guass quadrature: Li u Jun, G. Beer and J. L. Meek Table 3. e = 10-" N ~ R/L f = 1/r f = 1/r 2 f = 1# 3 2 3. 5355 (3.1855) 5.2869 (4.9369) 6.9479 (6.6079) 3 1. 4620 (1.1120) 2. 0221 (1.6721) 2. 5477 (2.1977) 4 0.9402 (0.5902) 1. 2373 (0.8873) 1. 5130 (1.1630) 5 0. 7213 (0.3713) 0.9168 (0.5668) 1. 0967 (0.7467) 6 0. 6046 (0.2546) 0. 7486 (0.3986) 0.8804 (0.5304) 7 0.5329 (0.1829) 0.6466 (0.2966) 0.7502 (0.4002) 8 0. 4848 (0.1348) 0.5787 (0.2287) 0.6639 (0.3139) 9 0.4504 (0.1004) 0. 5305 (0.1805) 0.6028 (0.2528) 10 0. 4247 (0.0747) 0. 4945 (0.1445) 0. 5575 (0.2075) r - - . _ R o L ~ L ~ ~i" L ; ~ L L Lo - I F ~ r e 6. An element and its sub-elements gration. The values of the R / L shown in the brackets are obt ai ned according to curve 3, r*/L = R[ L + 0.35. Other- wise t hey are according to curve 1, (r*[L = R[ L) , i.e. the value used in reference 2. For the case where the source points is very near to the integration region and the number of the integration points which can be used in a program is limited, the region has to be divided into some sub-elements. This has been suggested by Watson. 2,a However, the scheme presented here is different from Watson' s, as we subdivide this element un- equally instead of equally. This is explained for the simple example in Fig. 1. Figure 6 shows an element which is subdivided into four sub-elements. Each sub-element has its own local co-ordinate which varies from - - 1 to + 1. According to Fig. 6, t he general formulae can be obt ai ned as follows: +1 M+I N( / ) I = f / ( ~ ) d ~ = E E Hnq)f(~On D) (9) d ] = 1 n = l - - 1 where M + 1 is t ot al number of t he sub-elements. For all sub-elements except the last, we have: Lo 3 ~ = 2-/~(D = - ~ ( 1 - - 2 / ) R , _ ( ~ o ) (10) ~/ 1 + 2 / for the last sub-element: 1 (11) LM+I \ Lo] I f t he source point is very near to the element, the t ot al number of the integration points needed is more than the limiting value listed in Tables 1-3. Using the expression (11) and Tables 1-3, the number M is obtained. Then the number N q ) for each sub-element is found using (10), (11) and (6), (7), (8). The numerical integration is carried out using equat i on (9). Two examples are analysed below. The first example is: 1 I = (1.1 - - x ) a - - 1 wi t h the allowable error ea (a) = 10 -4. The different results are listed in Table 4: We can see t hat if the integration region is not sub- divided (M = 0), the results is obtained according to t he limiting value ( = 10) in the program, so t hat the comput a- t i on error is greater by far t han e~a)(= 10-a). On the con- trary, i f M = 3 the result is very good. It should be poi nt ed out t hat the comput er t i me depends on the limitation of integration points used in the program as the integration region has to be subdivided. Let us consider a second example: 1 I - - (1.1 - - x) 2 - 1 The result is listed in Table 5: In Table 5, M, N and e* are the same as t he above example. N L is a limitation of the integration points which can be used in the program. We can find t hat if the largest NL(=1 0 ) is used, the M , N are smaller when e * < e [ 2). That means comput er t i me is shortened. Table 4 M N e* 0 10 10 -2 3 21 1.5 X 10 -6 He r e i n , M + 1 i s t o t a l n u m b e r o f s ub- e l e - m e n t s , N i s t o t a l n u m b e r of i n t e g r a t i o n p o i n t s , a n d e* i s c o m p u t a t i o n e r r o r Table 5. e~ ~)' ~ = 10-2 N L ~ M + 1 N e* (%) l 2 6 18 0.04 3 5 16 0.05 4 4 14 0.I0 5 4 14 0.10 6 4 12 0.I0 7 3 12 0 . 4 0 8 3 12 0 . 4 0 9 3 12 0 . 4 0 10 2 12 0 . 4 0 T W O - D I M E N S I O N A L G A U S S - Q U A D R A T U R E The two-dimensional G-Q formula can be written as: 1 1 f f f ( ~ ( 1 ) ' ~ ( 2 ) ) d ~ ( l ) d ~ ( 2 ) - 1 - 1 Ef f i ci ent evaluation o f integrals using Guass quadrature: Li u Jun, G. Beer and J. L. Me e k N" t N 2 = E E A 0 ) A ( 2 ) C t e 0 ) g 2 ) ~ ( 1 2 ) l = 1 ] = 1 In the above, N. r (7 = 1, 2) is the number of integration points in ~('t) (7 = 1, 2) direction. The number N~ depends on t he required precision of integration. The order (N1, N2) of the formula to be used is determined from the upper bound for the error given by Stroud and Secrest: 2 IA~I < 2 ~ e,r/-L r (13) 3, =1 where and <H-r 4 e v - 22N, r (2N, r)! The norrnalised error is ea = A~f f . The function 1/r 2 is taken to be representative of actual integrands and the case considered is a rectangular region of integration of dimensions L1 xL2 (Fig. 7). Proceeding as before, it is found that N~ should be chosen such that: e (2) = (2N, r + 1)eo (14) For the case x a qt Sb, where is a quadrilateral element, N~ is chosen by R/ L ~ in terms of (14) as for the one- dimensional problem. Being similar to the one.dimensional problem, when N~ exceeds the limiting value of the number of Gaussian points which can be used in the program, the region has to be divided into some sub-regions (Fig. 7). Then the integration formula must be written as: 1 1 f f - - 1 - - 1 f(/j(1),/j(2)) d/j0) d/j(2) M, + I M2 + I N z N 2 a ( 1 ) ~ ( 2 ) { ~ ( 1 ) ~ ( 2 h = E E E E - , - j , , . , , ' , i - !=1 k=l i=1/=1 for all of the sub-regions except the last one, we have: Sj~ = 2-(/' k)~r + ~ (1 -- 2 g ~ ) R g ~) L (ct, k) k ) ( R' r ) = 1 + 2 ( k-~-~,/ (15) (16) [ . 2 _ i I t ! L1 Figure 7. A n el ement and sub-regions a 1 I " [ - " " 2 J J I D ] Figure 8. I t ,, ,O 2 . , Distribution o f Gaussian poi nt s - exampl e 1 for last one: 2 \ 2 ~ / (17) R(M'v +1) = 2M 7 __R~ L~. y+l ) L. v Some examples are now given for the two-dimensional problem. The following examples can be solved exactly in order to check whether the method is correct or not. Firstly, the integral: 1 1 - 1 - - 1 is computed. The integration region and the source point x a are shown in Fig. 8. The allowable error e~ 3) is 10 -2. The exact solution is 2.327344790. In the process of numerical integration, the region is divided into four sub-regions. Fifty-eight integration points are used which are distri- buted in Fig. 8. The result is 2.32729505, and computation error e* is 0.21 x 10 -3. Obviously, the result is sufficiently accurate. In the second example, the integral is: +1 1 - a t dr/ r - i - i the integration region is shown in Fig. 9. The exact solu- tion is 3.343673397. We compute this integral according to two different cases: (1) the allowable error is 10 -2, the result is 3.34523876 without any subdivision. The number of the integration points is 81 (Fig. 9a), the computation error e* is equal to 0.49 x 10- 3. (2) The allowable error is taken to be 10- 4. In the computation process, the region is divided into ten sub-regions, the total number of the inte- gration points is 187 distribution of which is shown in Fig. 9b the result is 3.34369565 e* = 0.68 x 10- 6. The above results show that this met hod given by authors is very efficient for the functions 0(I/r), 0( I / r 2) Efficient evaluation of integrals using Guass quadrature." Liu Jun, G. Beer and J. L. Meek 1 ] t _ L 1 0 2 ( a ) + I 2 I 9 I 1 6 I 1 9 1 1 1 6 I 1 6 I - - - - I - I 1 6 1 9 q 1 6 - ( b ) Figure 9. Distribution o f integration point -- example 2 and 0( 1/ r a) with different singularity and for different allowable error. COMPARISON Lachat and Watson 2'a give the following formula for the integral of (1/r2): (2Np + 1)Otp2NP <K (18) where ap = (ds/dr/p) = Lp[4R and K is a constant which is chosen in the light of experience. It should be pointed out t hat the physical meaning of K is explicit. The value is believed t o be dependent on the required integration error and a value of K= 10 -s is suggested in reference 5. The pl ot of expression (18) comput ed wi t h this value is shown in Fig. 10. Comparing this wi t h the curves if Fig. 5 it is found t hat the humber of integration points given by equation (18) is always larger t han by equation (7) when (R[L) is specified. This means t hat expression (18) is conservative. We comput e the integral: +1 +1 dx f f ( 1 . 4 - x ) a - - 1 --1 wi t h the met hod given by the authors and in reference 2. The exact solution is 0. 9447190. The results are listed in Table 6. The distribution of integration points are shown in Fig. 11. N Figure 10. ~ ~ ~ 1 ~ 1 ) = i o-~ IO-L 1.o 26 : R / L . Table 6 Methods's~'-...~ M N I e* % Author's 4 84 0.9441790 0.64 X 10 -s Reference 2 9 188 0.9441792 0.20 X 10 -6 1 6 1 6 _ 1 6 1 6 1 1 6 I . 4 X a ( a ) ( b ) Figure 11. Distribution of Gaussian points: (a) author's method; (b) method o f ref 2 SUMMARY AND CONCLUSIONS The efficient evaluation of integrals where the integrand varies rapidly and is of order ( l / r ) , (1/r2), (1/ r a) has been investigated in detail. Integrands of this t ype occur in the numerical i mpl ement at i on of the BEM met hod in two- and three-dimensional elastostatics, where r is the distance to a source point. \ X o Ef f i c i e nt eval uat i on o f integrals usi ng Guass quadrat ure: Li u Jun, G. Beer and J. L. Me e k I t has been f ound t hat in part i cul ar f or t hree-di mensi onal Boundar y El ement analysis t he comput at i onal effort and t her ef or e comput i ng time, needed t o i nt egrat e t he Ker nel - Shape f unct i on pr oduct s is significant and in some cases may be mor e significant t han t he ot her stages o f comput a- t i on i ncl udi ng t he sol ut i on o f t he syst em o f equat i ons. The me t hod present ed in this paper allows t o select, f r om a table, t he number o f i nt egrat i on poi nt s needed f or t he numeri cal i nt egrat i on t o achieve a prescri bed precision. This number is dependent on t he order o f t he i nt egrand and t he pr oxi mi t y o f t he source poi nt t o t he i nt egrat i on region. The values o f t he tables are checked by simple exampl es where t he integral can be eval uat ed anal yt i cal l y and good agreement was f ound bet ween predi ct ed and act ual preci- sion o f t he i nt egrat i on. When t he source poi nt is very close t o t he i nt egrat i on regi on a subdivision i nt o subregions has been suggested as t he number o f i nt egrat i on poi nt s needed may exceed t he number available in t he comput er program. I n t he present paper t he aut hor s are present i ng a new and mor e effi ci ent met hod o f less i nt egrat i on poi nt s f or a mor e accurat e result. The met hods o f i nt egrat i on present ed here are expect ed t o make t he eval uat i on o f Boundar y El ement coeffi ci ent mat ri ces mor e effi ci ent and save a significant amount o f comput at i onal ef f or t especially f or t hree-di mensi onal analysis. They have been i mpl ement ed in t he comput er program BEFE whi ch was devel oped b y Beer. s Recent exampl es on t hree-di mensi onal stress analysis pr obl ems have given excel l ent results wi t h regard t o accur acy and effi ci ency. There is a possi bi l i t y t hat t he i nt egrat i on can be made even mor e effi ci ent by using coor di nat e t r ansf or mat i ons near t he singular poi nt . This has been pr oposed r ecent l y by Higashimachi et al. 9 and Mustoe. I These met hods have been investigated in mor e detail by t he aut hor s and are discussed in a sequel t o t hi s paper. ACKNOWLEDGEMENTS The aut hor s woul d like t o t hank t he Aust ral i a/ Chi na Counci l f or provi di ng suppor t f or Mr Liu Jun. REFERENCES 1 Stroud, A. H. and Secrest, D. Gaussian Quadrature Formulae, Prentice-Hall, Engelwood Cliffs, NJ, 1966 2 Lachat, J. C. and Watson, J. O. Efffective numerical treatment of boundary integral equation, J. N u n Meth. Eng. 1976, 10, 981 3 Lachat, J. C. and Watson, J. O. In Boundary Integral Equation Method: Computational Application in Applied Mechanics, eds T. A. Cruse and F. J. Rizzo, AMD 11, ASME, New York, 1975 4 Chadouet, A. and Afzal, M. CA, ST, or 3D: Three-dimensional boundary element analysis computer code, in Boundary Ele- ments, Prec. 5th Int. Conf., Hiroshima, eds. C. A. Brebbia, T. Futagami and M. Tanaka, Springer-Vedag, Japan, 1983, pp. 787-809 5 Higashimachi, T., Ezawa, Y., Okamoto, Y. and Kamimiya, Y. Some investigations concerning precision o f three-dimensional analysis by the Boundary Element Method, 60th General Meeting o f Japan Soc. Mech. Engrs, No. 830-1, April, 1983, pp. 7-9 (in Japanese) 6 Cristessu, M. and Loubignac, G. Gaussian quadrature formulae functions with singularities in 1/R over triangles and quad- ranges, in Recent Advances in Boundary Element Methods, ed. C. A. Brebbia, Pentech Press, London, 1978, pp. 375-390 7 Pina, H. L. G. and Fernandes, J. L. M. Some numerical integra- tion formulae over triangles and squares with a 1]R singularity, Appl. Math. Modelling 1981, 5, 209 8 Beer, G. BEFE- a combined boundary element-f'mite ele- ment computer program. Adv. in Eng. Software 1983, 6, 103 9 Higashimachi, T., Okamoto, N., Ezawa Y.I Aizawa, T. and Ire, A. Interactive structural analysis system using the advanced boundary element method, Prec. 5th Int. Conf., Hiroshima, eds C. A. Brebbia, T. Futagami and M. Tanaka, Springer- Verlag, Japan, 1983, pp. 847-856 10 Mustoe, G. G. W. Advanced integration schemes over boun- dary elements and volume cells for two- and three-dimensional non-linear analysis, in Developments in Boundary Element Met hods - 3, eds I. K. Banerjee and S. Mukherjee, Applied Science, London, 1984, pp. 213--269