JAM Journal of Applied Mathematics 1687-0042 1110-757X Hindawi Publishing Corporation 10.1155/2014/276585 276585 Research Article Numerical Solution of High Order Bernoulli Boundary Value Problems Costabile F. Napoli A. Abbasbandy Saeid 1 Department of Mathematics University of Calabria 87036 Rende Italy unical.it 2014 1072014 2014 31 10 2013 28 05 2014 22 06 2014 10 7 2014 2014 Copyright © 2014 F. Costabile and A. Napoli. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

For the numerical solution of high order boundary value problems with special boundary conditions a general procedure to determine collocation methods is derived and studied. Computation of the integrals which appear in the coefficients is generated by a recurrence formula and no integrals are involved in the calculation. Several numerical examples are presented to demonstrate the practical usefulness of the proposed method.

1. Introduction

Higher order differential equations arise in a variety of different areas of science, engineering, and technology (see [1, 2]) since they model a wide spectrum of phenomena.

Particularly, the solutions of fifth-order BVPs model viscoelastic flows  and the seventh-order BVPs model induction motors with two rotor circuits [4, 5]. Ordinary differential equations of sixth and eighth order arise in modeling instability when an infinite horizontal layer of fluid is heated from below and is subject to the action of rotation . Moreover, high order boundary value problems arise in hydrodynamic, hydromagnetic stability , and other branches of applied sciences.

In  the authors presented a class of collocation methods for the numerical solution of high order boundary value problems: (1)y(n)(x)=f(x,y(x)),xI=[a,b],(2)B[x,y]=g,xI, where n>1, y(x)=(y(x),y(x),,y(q)(x)), 0q<n, and B is a linear operator on the boundary I, gRn.

The idea in  is the following: the differential problem (1)-(2) is written in the following equivalent integral form: (3)y(x)=Pn-1[y,x]+abGn-1(x,t)f(t,y(t))dt, where Pn-1[y,x] is the unique polynomial which satisfies the boundary conditions (4)B[x,Pn-1]=g and Gn-1(x,t) is a kernel (Green) function. Gn-1(x,t) is such that B[x,Gn-1]=0 and it is differentiable under the integral sign such that (3) satisfies (1).

Thus, from (3) and (4) we obtain a collocation polynomial which approximates the solution of problem (1)-(2).

In the present work the authors use this technique to derive collocation methods for the numerical solution of (1) with the particular boundary conditions (5)y(a)=β0,y(k)(b)-y(k)(a)=βk+1,k=0,,n-2 with βk, k=0,,n-1, being real constants.

Conditions (5) are called the Bernoulli boundary conditions, since they are related to the Bernoulli interpolation problem . They have physical and engineering interpretation , but to the authors’ knowledge, they have not been considered previously in the literature.

In [9, 10] the BVP (1)–(5) is considered: in  a nonconstructive proof of the existence and uniqueness of solution is given, and in  Picard’s method is applied in connection with Newton’s method for the numerical solution of the problem.

The present paper is organized as follows: in Section 2 we summarize some theoretical results on the existence and uniqueness of solution for problem (1)–(5). Then, in Section 3, we present the method for the numerical solution of this type of problems, which produces smooth, global approximations in the form of polynomial functions. In Section 4 we give an a priori estimation of error and, in Section 5, we present some particular cases. In Section 6 we propose an algorithm to compute the numerical solution of (1)–(5) in the nodal points and then, in Section 7, we present some numerical examples of both linear and nonlinear BVPs which confirm the theoretical results.

2. Preliminaries

Let Bk(x) be the Bernoulli polynomial of degree k  and let us set (6)Sk(t)=Bk(t)-Bk(0). Moreover, let (7)h=b-a,Δfa(k)=f(k)(b)-f(k)(a). The following theorems hold.

Theorem 1 (see [<xref ref-type="bibr" rid="B3">12</xref>]).

Let fC(n+1)[a,b]. Then (8)f(x)=f(a)+k=1nSk(x-ah)hk-1k!Δfa(k-1)+Rn[f,x], where Rn[f,x] is the remainder term (9)Rn[f,x]=abGn(x,t)f(n+1)(t)dt with Gn(x,t) being Peano’s kernel: (10)Gn(x,t)=1n![(x-t)+n-k=1nSk(x-ah)hk-1k×(nk-1)(b-t)n-k+1k=1nSk(x-ah)hk-1k].

Theorem 2 (see [<xref ref-type="bibr" rid="B3">12</xref>]).

If fC(n-1)[a,b], then the polynomial (11)Pn[f,x]=f(a)+k=1nSk(x-ah)hk-1k!Δfa(k-1) satisfies the Bernoulli interpolation problem (12)Pn[f,a]=f(a),ΔPn(k)=Pn(k)[f,b]-Pn(k)[f,a]=f(k)(b)-f(k)(a)=Δfa(k),kkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkkk=0,,n-1.

The proof of the existence and uniqueness of solution of (1)–(5) is based on (3) , under the hypothesis that the function f(x,y) satisfies the Lipschitz condition (13)|f(x,y1(x))-f(x,y2(x))|k=0qLk|y1(k)(x)-y2(k)(x)| in a certain domain interval of [a,b]×Rq+1.

3. The Collocation Method

Let y(x) be the solution of (1)–(5). If xi, i=1,,m, are m distinct points in [a,b] and y(x)C(n+m)[a,b], using Lagrange interpolation, we get (14)y(n)(x)=i=1mli(x)y(n)(xi)+R¯m(y,x), where (15)R¯m(y,x)=(x-x1)(x-xm)m!y(n+m)(ξx),ξx(a,b) and li(t) are the fundamental Lagrange polynomials on the m points xi.

Inserting (14) into (3), in view of (1), we obtain (16)y(x)=Pn-1[y,x]+i=1mf(xi,y(xi))abGn-1(x,t)li(t)dt+abGn-1(x,t)R¯m(y,t)dt. Hence the following identity holds: (17)y(x)=Pn-1[y,x]+i=1mpn,i(x)f(xi,y(xi))+Tn,m(y,x), where (18)pn,i(x)=abGn-1(x,t)li(t)dt,i=1,,m,(19)Tn,m(y,x)=abGn-1(x,t)R¯m(y,t)dt.

This suggests defining the polynomials (20)yn,m(x)=Pn-1[y,x]+i=1mpn,i(x)f(xi,yn,m(xi)), where yn,m(x)=(yn,m(x),yn,m(x),,yn,m(q)(x)), 0qn-1.

Theorem 3.

The polynomial of degree n+m implicitly defined by (20) satisfies the relations (21)yn,m(a)=y(a),(22)yn,m(k)(b)-yn,m(k)(a)=βk+1,k=0,,n-2,(23)yn,m(n)(xi)=f(xi,yn,m(xi)),i=1,,m; that is, yn,m(x) is a collocation polynomial for (1)–(5) on the nodes xi, i=1,,m.

Proof.

From (18), pn,i(a)=pn,i(b)=0, i=0,,m, and thus relations (21) follow from direct computation. To prove (22) we derive Gn-1(x,t)  k times, k=1,,n-2, with respect to x, and using the well-known relation  Bs(x)=sBs-1(x), s>0, we get (24)kxkGn-1(x,t)={g1(x,t)=(x-t)n-k(n-k)!-j=knBj-k(x-ah)hj-k-1(b-t)n-j-1(j-k)!(n-j-1)!xtg2(x,t)=-j=knBj-k(x-ah)hj-k-1(b-t)n-j-1(j-k)!(n-j-1)!x<t. From the property of Bernoulli polynomials Bs(1)=(-1)sBs(0), we have g1(a,t)=g2(b,t); thus (25)pn,i(k)(a)=abg1(a,t)li(t)dt=abg2(b,t)li(t)dt=pn,i(k)(b). Hence (26)yn,m(k)(b)-yn,m(k)(a)=y(k)(b)-y(k)(a)+i=1m(pn,i(k)(b)-pn,i(k)(a))f(xi,yn,m(xi))=y(k)(b)-y(k)(a). From this, (22) follows. Next, by deriving yn,m(x)  n times, we obtain (27)yn,m(n)(x)=Pn-1(n)[y,x]+k=1n-1pn,k(n)(x)f(xk,yn,m(xk))=k=1n-1lk(x)f(xk,yn,m(xk)) and this implies (23).

4. The Error

In what follows for all yC(q)[a,b] we define the norm y=max0sq{maxatb|y(s)(t)|}  and the constants (28)L=k=0qLk,R=maxaxb|R¯m(y,x)|. Further, we define (29)Qm=max0sq{maxaxbi=1m|pni(s)(x)|}Dn,s=maxaxbk=sn-1|Bk-s((x-a)/h)|(k-s)!(n-k)!,Δ=max0sq{hn-s+1(n-s+1)!+hn-s-1Dn,s}.

An a priori estimation of the global error is possible.

Theorem 4.

With the previous notations, suppose that LQm<1. Then (30)y-yn,mRΔ1-LQm.

Proof.

By deriving (17) and (20) we get (31)y(s)(x)-yn,m(s)(x)=i=1n-1pni(s)(x)[f(xi,y(xi))-f(xi,yn,m(xi))]+sxsabGn-1(x,t)R¯n,m(y,t)dt. Now, since (32)sxsabGn-1(x,t)R¯n,m(y,t)dt=1(n-s)!ax(x-t)n-sR¯n,m(y,t)dt-k=sn-1Bk-s(x-ah)hk-s-1(k-s)!(n-k+1)!×ab(b-t)n-k+1R¯n,m(y,t)dt, we have (33)|sxsabGn-1(x,t)R¯n(y,t)dt|hn-s+1(n-s+1)!R+hn-s-1RDn,s. Thus (34)|y(s)(x)-yn,m(s)(x)|i=1n-1|pni(s)(x)|k=0qLk|y(s)(xi)-yn(s)(xi)|+hn-s+1(n-s+1)!R+hn-s-1RDnsLy-ynQm+RΔ and inequality (30) follows.

5. Particular Cases

Now we present explicitly the formulas for some values of n.

For the computation of pn,i(x) we need axtkli(t)dt and xbtkli(t)dt. Letting (35)Fi1(x)=axli(t)dt,Mi1(x)=xbli(t)dt,Fik(x)=axFi,k-1(t)dt,Mik(x)=xbMi,k-1(t)dt,k2, and integrating by parts k times, we obtain (36)axtkli(t)dt=j=0k(-1)jk!(k-j)!xk-jFi,j+1(x),xbtkli(t)dt=j=0kk!(k-j)!xk-jMi,j+1(x).

5.1. The Fifth-Order Case

Now we consider the case of the fifth-order BVP (37)y(v)(x)=f(x,y(x)),x[0,1],y(0)=β0,y(k)(1)-y(k)(0)=βk+1,k=0,,3. In this case Green’s function is (38)G4(x,t)={14![t4(1-x)+t3(2x2-2x)+t2(-2x3+3x2-x)+t(x4-2x3+x2)]tx14![-t4x+t3(2x2+2x)-t2(2x3+3x2+x)+t(x4+2x3+x2)-x4]xt,4!p5,i(x)=(x4-2x3+x2)[Fi2(x)-Mi2(x)]-2x(2x2-3x+1)[Fi3(x)+Mi3(x)]+12x(x-1)[Fi4(x)-Mi4(x)]+24(1-x)Fi5(x)-24xMi5(x). Hence (39)y5,m(x)=P4[y,x]+i=1mp5,i(x)f(xi,y5,m(xi)). By deriving (44) we get (40)y5,m(s)(x)=P4(s)[y,x]+i=1mp5,i(s)(x)f(xi,y5,m(xi))s=1,,5, where p5,i(s)(x) can be easily computed using the same technique as for p5,i(x).

5.2. The Seventh-Order Case

Consider (41)y(vii)(x)=f(x,y(x)),x[0,1],y(0)=β0,y(k)(1)-y(k)(0)=βk+1,k=0,,5.

In this case Green’s function is (42)G6(x,t)=16!{t6(1-x)+3t5(x2-x)+5t4(-x3+32x2-x2)+5t3(5x4-2x3+x2)+t2(-3x5+152x4-5x3+x2)+t(x6-3x5+52x4-x22)tx-t6x+3t5(x2+x)+5t4(-x3-32x2-x2)+5t3(5x4+2x3+x2)+t2(-3x5+152x4-5x3+x2)+t(x6-3x5+52x4-x22)-x6xt. Hence (43)6!p7,i(x)=(x6-3x5+52x4-x22)[Fi2(x)-Mi2(x)]-x(6x4-15x3+10x2-1)[Fi3(x)+Mi3(x)]+30(x-1)2[Fi4(x)-Mi4(x)]-60x(2x2-3x+1)[Fi5(x)+Mi5(x)]+360x(x-1)[Fi6(x)-Mi6(x)]+6!(1-x)Fi7(x)-6!xMi7(x),(44)y7,m(x)=P6[y,x]+i=1mp7,i(x)f(xi,y7,m(xi)).

5.3. Order <inline-formula><mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M97"><mml:mi>n</mml:mi><mml:mo>=</mml:mo><mml:mn>8</mml:mn><mml:mo>,</mml:mo><mml:mn>9</mml:mn><mml:mo>,</mml:mo><mml:mn>10</mml:mn></mml:math></inline-formula>

For n=8, we have (45)7!p8,i(x)=(x84-x7+76x6-712x4+x26)[Fi1(x)+Mi1(x)]+(x7-72x6+72x5-76x3+x6)[Fi2(x)-Mi2(x)]-7x(x5-3x4+52x3-x2)[Fi3(x)+Mi3(x)]-7x(6x4-15x3+10x2-1)[Fi4(x)-Mi4(x)]+210x2(x-1)2[Fi5(x)+Mi5(x)]-420x(2x2-3x+1)[Fi6(x)-Mi6(x)]+2520x(x-1)[Fi7(x)+Mi7(x)]+7!(1-x)Fi8(x)+7!xMi8(x).

For n=9, we get (46)8!p9,i(x)=(x8-4x7+143x6-73x4+23x2)[Fi2(x)-Mi2(x)]+4x(-2x6+7x5-7x4+73x2-x3)[Fi3(x)+Mi3(x)]+28x2(2x4-6x3+5x4-1)[Fi4(x)-Mi4(x)]+56x(-5x4+15x3-10x2+1)[Fi5(x)+Mi5(x)]+1680x2(x-1)2[Fi6(x)-Mi6(x)]+3360x(-2x2+3x-1)[Fi7(x)+Mi7(x)]+20160x(x-1)[Fi8(x)-Mi8(x)]+8!(1-x)Fi9(x)-8!xMi9(x).

For n=10, we obtain (47)9!p10,i(x)=(x105-x9+32x8-75x6+x4-310x3)×[Fi1(x)+Mi1(x)]-x(x8-92x7+6x6-215x4+2x2-310)×[Fi2(x)-Mi2(x)]+3x2(3x6-12x5+14x4-7x2+2)[Fi3(x)+Mi3(x)]+12x(-6x6+21x5-21x4+7x2-1)[Fi4(x)-Mi4(x)]+252x2(2x4-6x3+5x2-1)[Fi5(x)+Mi5(x)]+504x(-6x4+15x3-10x2+1)[Fi6(x)-Mi6(x)]+15120x2(x-1)2[Fi7(x)+Mi7(x)]+30240x(-2x2+3x-1)[Fi8(x)-Mi8(x)]+181440x(x-1)[Fi9(x)+Mi9(x)]+9!(1-x)Fi10(x)+9!xMi10(x).

6. Algorithms

To calculate the approximate solution of problem (1)–(5) by (20) at x[a,b], we need the values yj(k)=yn,m(k)(xj), j=1,,m, and k=0,,q. These values can be calculated by solving the following system: (48)yi(k)=Pn-1(k)[y,xi]+j=1mpnj(k)(xi)f(xj,yj)i=1,,m,k=0,,q with yj=(yj,yj,,yj(q)), 0qn-1.

To solve it, if we put (49)Y=(Y0,,Yq)T,Yk=(y1(k),,ym(k)),F(Y)=(Fm,,Fm)TRq+1,Fm=(f1,,fm),fi=f(xi,yi),C=(B0,,Bq)T,Bk=(Pm(k)[y,x1],,Pn-1(k)[y,xm]),A=(A0000000Aq),Ak=(p1,1(k)p1,m(k)pm,1(k)pm,m(k)) with pi,j(k)=pn,j(k)(xi), k=0,,q, we write (48) as (50)Y-AF(Y)=C or, equivalently, Y=G(Y), where (51)G(Z)=AF(Z)+C.

For the existence and the uniqueness of the solution of (50) the following result holds.

Proposition 5.

Let L be defined as in (28). If T=LA<1, the system (50) has a unique solution which can be calculated by an iterative method (52)Y(ν+1)=G(Y(ν)),ν=1,2, with a fixed Y(0)Rs, s=m(q+1), and G defined as in (51). Moreover, if Y is the exact solution of the system, (53)Y(ν)-YTν1-TY(1)-Y(0).

Proof.

If V=(V0,,Vq)T, Vk=(v1(k),,vm(k)) and W=(W0,,Wq)T, Wk=(w1(k),,wm(k)), then G(V)-G(W)ALV-W; hence G is contractive. Thus the result follows from the well-known contraction mapping theorem.

To calculate the elements A0,,Aq of the matrix A we need the values Fis(xj) and Mis(xj), s=1,,n, i,j=1,m, where Fis(x) and Mis(x) are defined in (35).

Since li(t)=k=1,kim((t-xk)/(xi-xk)), it suffices to compute (54)cxj=tkctk-1ct1rm,i(t)dtdt1dtk-1, where c=a or c=b, r0,0(t)=1, and (55)rm,i(t)=(t-x1)(t-xi-1)(t-xi+1)(t-xm)i=1,2,,m. Let us define (56)g0,1,c(i)(x)=x-c and, for s=1,,m-1, (57)gs,j,c(i)(x)=cx=tjctj-1ct1(t-z1(i))(t-z2(i))(t-zs(i))dtdt1dtj-1, where (58)zj(i)={xjifj<ixj+1ifjij=1,,m-1. We can easily compute (59)g0,j,c(i)(x)=(x-c)jj!. For the computation of (57) we use the recursive algorithm  (60)gs,j,c(i)(x)=(x-zs(i))gs-1,j,c(i)(x)-jgs-1,j+1,c(i)(x). Thus, if Wi=k=1,kim(xi-xk), we get (61)Fik(xj)=gm-1,k,a(i)(xj)Wi,Mik(xj)=(-1)kgm-1,k,b(i)(xj)Wi.

7. Numerical Examples

Now we present some numerical results obtained by applying method (20) to find numerical approximations of the solutions of some test problems. As the true solutions are known, we considered the error function E(x)=|y(x)-yn,m(x)|. To solve the nonlinear system (48) we used the so-called modified Newton method  (the same Jacobian matrix is used for more than one iteration) and algorithm (60) for the computation of the entries of the matrix. Equidistant points are used as nodal points. Analogous results are obtained in the considered examples by using as nodes the zeros of Chebyshev polynomials of first and second kind.

Example 1.

Consider the following (62)y(v)(x)=y(x)-(15+10x)ex,x[0,1],y(0)=0,y(1)=0,y(1)-y(0)=-(e+1),y′′(1)-y′′(0)=-4e,y′′′(1)-y′′′(0)=3-9e, with solution y(x)=x(1-x)ex. Figure 1 shows the graph of the error function E(x) for two different values of m.

Error function of problem (62) for m=4 (a) and for m=6 (b).

Example 2.

Consider the following (63)y(v)(x)=-24e-5y+481+x5,x[0,1],y(0)=0,y(1)=log2,y(1)-y(0)=-12,y′′(1)-y′′(0)=34,y′′′(1)-y′′′(0)=-74, with solution y(x)=log(x+1). The graph of E(x), for two different numbers of nodes, is plotted in Figure 2.

Error function of problem (63) for m=4 (a) and for m=7 (b).

Example 3.

Consider (64)y(vii)(x)=-y-ex(2x2+12x+35),x[0,1],y(0)=y(1)=0,y(1)-y(0)=-(e+1),y′′(1)-y′′(0)=-4e,y′′′(1)-y′′′(0)=3(1-3e),y(iv)(1)-y(iv)(0)=8(1-2e),y(v)(1)-y(v)(0)=5(3-5e), with solution y(x)=x(1-x)ex. Figure 3 shows the graph of E(x).

Error function of problem (64) for m=4 (a) and for m=8 (b).

Note that the equation in (64) is the same as that in Example 2 of , but the boundary conditions are different.

Example 4.

Consider (65)y(ix)(x)=y(x)-9ex,x[0,1],y(0)=1,y(1)=0,y(1)-y(0)=-e,y′′(1)-y′′(0)=1-2e,y′′′(1)-y′′′(0)=2-3e,y(iv)(1)-y(iv)(0)=3-4e,y(v)(1)-y(v)(0)=4-5e,y(vi)(1)-y(vi)(0)=5-6e,y(vii)(1)-y(vii)(0)=6-7e, with solution y(x)=(1-x)ex. Figure 4 shows the graph of E(x).

Error function of problem (65) for m=4 (a) and for m=8 (b).

The equation in (65) is the same as that in [1, 2], but the boundary conditions are different.

Example 5.

Consider (66)y(x)(x)=e-xy2(x),x[0,1],y(0)=1,y(k)(1)-y(k)(0)=e-1,k=0,,8 with solution y(x)=ex. Figure 5 shows the graph of E(x).

Error function of problem (66) for m=3 (a) and for m=6 (b).

Note that the equation in (66) is the same as that in , but the boundary conditions are different. The conditions in  are the so-called Lidstone-type conditions. Problems of this type have been analyzed in  using a similar technique.

8. Conclusions

This paper presents a class of collocation methods for nth order differential equations with Bernoulli boundary conditions. For two positive integers n,m a polynomial of degree n+m approximating the exact solution is given explicitly. Numerical experiments support theoretical results. Further developments can be done, concerning particularly numerical estimates of the error.

Conflict of Interests

The authors declare that there is no conflict of interests regarding the publication of this paper.

Hossain B. Islam S. A novel numerical approach for odd higher order boundary value problems Mathematical Theory and Modeling 2014 4 5 1 11 Wazwaz A. Approximate solutions to boundary value problems of higher order by the modified decomposition method Computers & Mathematics with Applications 2000 40 6-7 679 691 10.1016/S0898-1221(00)00187-5 MR1776685 2-s2.0-0034276401 Karageorghis A. Phillips T. N. Davies A. R. Spectral collocation methods for the primary two-point boundary value problem in modelling viscoelastic flows International Journal for Numerical Methods in Engineering 1988 26 4 805 813 10.1002/nme.1620260404 2-s2.0-0023998211 Richards G. Sarma P. R. R. Reduced order models for induction motors with two rotor circuits IEEE Transactions on Energy Conversion 1994 9 4 673 678 10.1109/60.368342 2-s2.0-0028742405 Siddiqi S. S. Ghazala A. Muzammal I. Solution of seventh order boundary value problem by differential transformation method World Applied Sciences Journal 2012 16 11 1521 1526 Siddiqi S. S. Twizell E. H. Spline solutions of linear twelfth-order boundary-value problems Journal of Computational and Applied Mathematics 1997 78 2 371 390 10.1016/S0377-0427(96)00164-1 MR1433517 ZBL0865.65059 Chandrasekhar S. Hydrodynamic and Hydromagnetic Stability 1981 New York, NY, USA Dover Costabile F. Napoli A. A class of collocation methods for high order boundary value problems submitted Costabile F. Serpe A. On Bernoulli boundary value problems Le Matematiche 2007 62 2 163 173 Costabile F. Serpe A. Bruzio A. Gelman L. Hukins D. W. L. Hunter A. Korsunsky A. No classic boundary conditions Proceedings of the World Congress on Engineering 2007 London, UK 918 921 Jordan C. Calculus of Finite Differences 1965 3rd New York, NY, USA Chelsea MR0183987 Costabile F. Expansion of real functions in Bernoulli polynomials and applications 273 Conferences Seminars Mathematics 1999 Univesty of Bari 1 13 Costabile F. Napoli A. Collocation for high-order differential equations with Lidstone boundary conditions Journal of Applied Mathematics 2012 2012 20 120792 MR2956524 10.1155/2012/120792 Agarwal R. P. Boundary Value Problems for Higher Order Differential Equations 1986 Singapore World Scientific 10.1142/0266 MR1021979 Costabile F. Napoli A. A class of collocation methods for numerical integration of initial value problems Computers & Mathematics with Applications 2011 62 8 3221 3235 10.1016/j.camwa.2011.08.036 MR2837755 2-s2.0-80053345310 Quarteroni A. Sacco R. Saleri F. Numerical Mathematics 2000 New York, NY, USA Springer MR1751750