Monte Carlo Methods Appl. 19 (2013), 281 – 330 DOI 10.1515 / mcma-2013-0013 © de Gruyter 2013 Worst case error for integro-differential equations by a lattice-Nyström method Davoud Rostamy, Mohammad Jabbari and Mahshid Gadirian Abstract. In this paper, we make an offer of the lattice approximate method for solving a class of multi-dimensional integro-differential equations with the initial conditions. Also, we analyze the worst case error measured in weighted Korobov spaces for these equations. Finally, numerical examples complete this work. Keywords. QMC-Nyström, lattice quadrature, worst case error, multi-dimensional integral equation, multi-dimensional integro-differential equations. 2010 Mathematics Subject Classification. 65N60, 65N15. 1 Introduction The paper aims to extend the integral equation to the following system of Fredholm integro-differential equations given by Z Z ! p.u/ D g.x/ C 1 1 .x; y/u.y/d y C 2 2 .x; y/ ru.y/d y; (1.1) D D with the initial conditions 8 ˆ < u.x/ D b0 in x 2 @D1 ; or ˆ ! : ru.x/ D b2 in x 2 @D2 ; (1.2) where @D1 ; @D2 @D and !.x/ ru.x/ p.u/ D ˛0 .x/u.x/ C ˛ 1 (1.3) ! are given for x 2 D D Œ0; 1d , 1 ; 2 2 R and d 2 N. Also, 1 , ! 2 , g, ˛0 and ˛ 1 functions from a weighted Korobov space. This problem describes several interesting physical and financial phenomena (see [7–9, 13, 14, 22]). On the other hand, the existence and uniqueness of the above problem were proved by many authors (see [3, 11, 12, 17, 18, 24]). The paper proposes a convergence analysis of discretization schemes for these equations. The convergence Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 282 D. Rostamy, M. Jabbari and M. Gadirian analysis is cast into the general setting of information based complexity (IBC). Numerical solution by collocation discretization is proposed, with collocation points chosen as lattice points of a QMC integration rule. 1.1 Preliminary (See [4, 5, 10].) The weighted Korobov spaces are characterized by a smoothness parameter ˛ > 1 and weights 1 1 2 > 0 where j moderates the behavior of the function with respect to the j th variable; a small j means d .D/ denote that the function depends weakly on the j th variable. Let H D H; ˛ a weighted Korobov space, where D .j /j 1 is a sequence of positive weights and ˛ > 1 is a smoothness parameter. For any Z X 2 i h:x u.x/ D u.h/e O with u.h/ O D u.x/e 2 ih:x d x; D h2Zd the norm of u in H is given by kukH D X 1=2 ju.h/j O ra .; h/ ; 2 h2Zd therefore, if we consider X @u D u.h/2 O ihj e 2 i h:x @xj d for j D 1; : : : ; d; (1.4) h2Z then we have kru.x/kH;1 D X 1=2 ku.h/2 O ihk21 r˛ .; h/ ; h2Zd where r˛ .; h/ D d Y r˛ .j ; hj /; (1.5) j D1 and ´ 1 r˛ .j ; hj / D j 1 jhj j˛ if hj D 0; otherwise. Thus, we have r˛ .; h/ 1 for all h 2 Zd . If we use the Cauchy–Schwarz inequality, it shows that for all u 2 H the following inequality is made: X 1=2 X 1=2 X 1 2 kuksup ju.h/j O ju.h/j O r˛ .; h/ : r˛ .; h/ d d d h2Z h2Z h2Z Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 Worst case error for integro-differential equations 283 Therefore we have d Y kuksup kukH 1 C 2.˛/j 1=2 ; (1.6) j D1 where .x/ WD 1 X j x j D1 denotes the Riemann zeta function. Furthermore the inequality in (1.6) becomes equality when u is a multiple of the function X e 2 ih:x : r˛ .; h/ d h2Z The rest of this paper is organized as follows. The paper investigates two cases of (1.1) in Sections 2 and 3. In these sections, we will obtain that the worst case ˛ error achieves the optimal rate of convergence O.n 2 Cı /, ı > 0, in weighted Korobov spaces, for a sufficiently large n. We assume that t1 ; : : : ; tn 2 D form a set of rank-1 lattice points. On the other hand, tractability in the absolute sense means that the minimal value of n is needed in the Quasi Monte Carlo Nyström (QMCNyström) method to reduce the worst case error to " 2 .0; 1/ and it is a bounded polynomial in d and " 1 . Also, the tractability and strong tractability of the QMCNyström method in the absolute or normalized sense are investigated. Of course, we know that strong tractability means that the bound is independent of d . Also, we will show that strong QMC-Nyström tractability in the absolute sense holds iff 1 X j < 1; (1.7) j D1 and QMC-Nyström tractability in the absolute sense holds iff Pd j D1 j lim sup < 1: d !1 log.d C 1/ (1.8) Moreover, strong tractability in the normalized sense is defined in term of the normalized error with respect to the initial error. The conditions (1.7) and (1.8) are also sufficient conditions for strong QMC-Nyström tractability in the normalized sense [23,25]. Finally, some numerical results for two cases show that the proposed method has merited. Also, some important propositions and results are proved in appendices. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 284 2 D. Rostamy, M. Jabbari and M. Gadirian First case study of (1.1) In this section, we will consider 1 .x; y/ D .x; y/a1 .y/, ! 2 .x; y/ D .x; y/! a2 .y/, ! ! D 1 D 2 , ˛0 .x/ D 1 and ˛1 .x/ D 0 in (1.1). Therefore we have the following equation: Z Z ! u.x/ D g.x/ C .x; y/a1 .y/u.y/d y C .x; y/a2 .y/ ru.y/d y ; (2.1) D D with the initial conditions (1.2), where the kernel is assumed to be of the form .x; y/ WD k.x y/, with k.x/ having period one in each component of x. Further, we assume that g, k belong to a weighted Korobov space H. We approximate u in (2.1) by using the Nyström method based on QMC rules, that is, equal-weight integration rules. We assume that t1 ; : : : ; tn 2 D and the approximation of u is given by n X un .x/ WD g.x/ C k.x ti /a1 .ti /un .ti / C k.x ti /! a2 .ti / run .ti / ; (2.2) n i D1 by differentiating from (2.2), we have n rx un .x/ WD rx g.x/ C X rx k.x n ti / a1 .ti /un .ti / C ! a2 :.ti /run .ti / ; (2.3) i D1 where the function values un .t1 /; : : : ; un .tn / are obtained by solving the following linear system from (2.2) and (2.3): 8 n X ˆ ˆ ˆ k.tj ti /a1 .ti /un .ti / u .t / WD g.t / C n j j ˆ ˆ n ˆ ˆ i D1 ˆ ˆ < C k.tj ti /! a2 .ti / run .ti / ; (2.4) n ˆ X ˆ ˆ ˆ rx k.tj ti / a1 .ti /un .ti / rx un .tj / WD rx g.tj / C ˆ ˆ ˆ n ˆ i D1 ˆ : C! a2 .ti / run .ti / ; for j D 1; 2; : : : ; n. We recall that the system of (2.4) is as a reduced linear system by required conditions given in equations (1.2). Therefore, in this section we investigate the worst case error of the QMCNyström method which essentially the worst possible error u un , measured in sup norm. Also, we assume that ! k1 .x y/ WD k.x y/a1 .y/; k2 .x y/ WD k.x y/! a2 .y/: Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 Worst case error for integro-differential equations 285 Hence we make a good set of points t1 ; : : : ; tn which leads to a as small worst case error as possible. There are alternative methods for tending to this goal, so we select a rank-1 lattice rule as a QMC-rule with points given by ti D ¹ inz º with i D 1; 2; : : : ; n (see [6,16]). In particular, we need a class of lattice-Nyström methods such that z is introduced by the generating vector which is an integer vector having no factor in common with n, and the braces around a vector indicate that each component of the vector is to be replaced by its fractional part. 2.1 Lower and upper bounds on the worst case error for (2.1) Let C 1;v .D/ be a set of continuous functions u W D ! R which are one time continuously differentiable in D and such that for all t 2 Dn@D, the following estimate holds: 8 ˆ if v < 0; <1 kruk c.u/ 1 C jlog .t/j if v D 0; ˆ : v .t / if v > 0; where 1 < v < 1, c.u/ is a positive constant and .t/ D min0<t<1 ¹t; 1 tº is the distance from t 2 .0; 1/ to the boundary of the interval .0; 1/. We equip the space of bounded linear operators from C 1;v .D/ to C 1;v .D/ with the usual operator norm kKk D sup kKuksup ; kuksup 1 for a given kernel, we are interested in the integral operator K W C 1;v .D/ ! C 1;v .D/ given by Z .x; y/ a1 .y/u.y/ C ! a2 .y/:ru.y/ d y D Z Z D 1 .x; y/u.y/d y C 3 .x; y/d y; Ku D D D where 3 .x; y/ D ! 2 .x; y/ ru.y/ and ! 2 .x; y/ D 2 .x; y/! a2 .y/ with Z Z kKk D max j1 .x; y/jd y C max j3 .x; y/jd y; x2D x2D D D and the corresponding discrete operator Kn W C 1;v .D/ ! Pn Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 286 D. Rostamy, M. Jabbari and M. Gadirian given by n Kn u D 1X k.x n ti / a1 .ti /u.ti / C ! a2 .ti / ru.ti / i D1 n 1X D k1 .x n n ti /u.ti / C i D1 with 1 X! k2 .x n n kKn k D max x2D ti / ru.ti /; i D1 1X jk1 .x n n ti /j C max x2D i D1 1X jk3 .x n ti /j; i D1 where t1 ; : : : ; tn 2 D. Thus Z Z jk3 .y/jd y kk1 ksup C kk3 ksup jk1 .y/jd y C kKk D D D and n kKn k D max x2D 1X jk1 .x n n ti /j C max x2D i D1 1X jk3 .x n ti /j kk1 ksup C kk3 ksup ; i D1 where the inequalities become equalities when k1 and k3 are constant functions. Here, we assume that the operator K is a compact and bounded operator (see Appendix 1). ! If g, k1 , k2 2 H are given, then we study the solution of (2.1) to the re-solvent formula ! S.g; k1 ; k2 / WD u; which we express as u D g C Ku or as .I K/u D g, where I W C 1;v .D/ ! C 1;v .D/ denotes the identity operator Iu D u also, we assume that the operator .I K/ 1 exists. Thus, by using the Fredholm alternative, we have k.I K/ 1 k < 1 and u D .I K/ 1 g. Therefore we have ! X 2 i h:x O .Ku/.x/ D u.h/e O k1 .h/ C kO2 .h/:2 ih h2Zd implying ! u.h/ O D g.h/ O C kO1 .h/ C kO2 .h/ 2 i h u.h/; O Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 Worst case error for integro-differential equations 287 hence we conclude the following results: g.h/ O ; ! O O k1 .h/ C k2 .h/ 2 i h u.h/ O D 1 (2.5) and kukH Xˇ ˇ ˇ D ˇ h2Zd k.I ˇ2 1=2 ˇ g.h/ O ˇ r˛ .; h/ ! ˇ kO1 .h/ C kO2 .h/ 2 ih 1 1 K/ (2.6) kC kgkH ; ! where the inequality becomes equality when g, k1 and k2 are constant functions. We use QMC-Nyström method by the algorithm ! An .g; k1 ; k2 / WD un or un D g C Kn un : Suppose that n WD k.I Then the operator .I 1 Kn / k.I Kn / 1 K/ kk.K Kn /Kn k < 1: exists and 1 1 C k.I k K/ 1 kkKn k : 1 n Then un is well defined and we have un D .I 1 Kn / g: Therefore we observe that n < 1 is essentially as a related condition on the value of n and the equality of the points t1 ; : : : ; tn . We assume that ˇ > 0 and > 1 are fixed. Therefore we recall that ! S.g; k1 ; k2 / D .I and K/ ! An D An .g; k1 ; k2 / D .I 1 g Kn / 1 g: Hence we define the worst case error of a QMC-Nyström method by en;d .An / WD sup !! k1 ;k2 ;k4 ;g2Æ ! kS.g; k1 ; k2 / ! An .g; k1 ; k2 /ksup ; Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 288 D. Rostamy, M. Jabbari and M. Gadirian where ® ! ! Æ D k1 ; k2 ; g W kgkH 1; kk1 kH ˇ; kk2 kH;1 ˇ; krki kH;1 ˇ for i D 1; 2; k.I and ! k4 .x ! y/ D r k2 .x K/ 1 ¯ k ; y/:ru.tj /: Also, we write the following inequality: ! kS.g; k1 ; k2 / ! An .g; k1 ; k2 /ksup en;d .An / kgkH : Note that the constants ˇ and in Æ are mutually independent. We define the initial error associated with the zero algorithm A0 0 as e0;d .A0 / WD sup !! k1 ;k2 ;k4 ;g2Æ ! kS.g; k1 ; k2 /ksup : For " 2 .0; 1/, we are interested in finding the smallest value of n for which en;d .An / "; corresponding to tractability in the absolute sense, or en;d .An / "e0;d .A0 /; corresponding to tractability in the normalized sense. For " 2 .0; 1/ and d 1, we define the following set (see [13, 14]): nabs ."; d / WD min¹n W 9QMC-Nyström method An with en;d .An / "º: The integral equation in this section is said to be QMC-Nyström tractable in the absolute sense iff there exist nonnegative constants C , p and q independent of " and d such that (see [4]) nabs ."; d / C " p d q; for all " 2 .0; 1/; d 1; and the problem is said to be strongly QMC-Nyström tractable in the absolute sense iff the above condition holds with q D 0. Furthermore, tractability and strong tractability in the normalized sense are defined in a similar way, with nabs ."; d / replaced by nnor ."; d / WD min¹n W 9QMC-Nyström method An with en;d .An / "e0;d º: Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 289 Worst case error for integro-differential equations On the other hand, we obtain the initial error, from (1.6) and (2.6), ! kS.g; k1 ; k2 /ksup k.I K/ 1 kkgkH d Y 1 C 2.˛/j 1=2 j D1 d Y 1 C 2.˛/j 1=2 ; j D1 that is an upper bound on the initial error e0;d and does not depend on ˇ. Now, we obtain a lower bound on initial error. We assume that L 1 1 .x; y/ D c WD min ˇ; ; where Z LD D ! k2 rgd y ! ! ! and k2 is a constant vector such that kk2 kH;1 ˇ and kr k2 kH;1 ˇ. We define g such that 1 g.h/ O 1 D ; ! Gr ˛ .; h/ O O .k1 .h/ C k2 .h/:2 ih/ where G WD d Y 1 C 2.˛/j 1=2 j D1 and kgkH 1. On the other hand, if we put ai .x/ D 1, bi .y/ D c, and ki .x y/ D ai .x/bi .y/, i D 1; 2, in the equation Z Z ! u.x/ D g.x/ C k1 .x y/u.y/d y C k2 .x y/ ru.y/d y ; (2.7) D D then we have u.x/ D g.x/ C c1 C c2 ; (2.8) R ! where c1 D D cu.y/d y and c2 D D k2 :ru.y/d y. By differentiating from (2.8), we have ru.x/ D rg.x/: (2.9) R Substituting (2.8) and (2.9) in (2.7), we write Z Z ! 1 c1 D c g.y/d y C k2 rgd y C c2 .c 1 c D D 1/ : Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 290 D. Rostamy, M. Jabbari and M. Gadirian Therefore c2 is a free parameter and c1 is obtained in terms of c2 ; then we have 1 C L k.I K/ 1 k D kuk 1 C c C L C c2 .c 1/ C c2 D ; 1 c 1 c L 1 such that c . By the above assumptions, (1.6) and (2.5), we know d Y 1 X e 2 i h:x 1=2 ! D kS.g; k1 ; k2 /ksup D 1 C 2.˛/j ; G r˛ .; h/ sup d j D1 h2Z ! and we know that for proving the above inequality, S.g; k1 ; k2 / is a factor of X e 2 ih:x : r˛ .; h/ d h2Z Then we have d Y 1=2 ! kuksup D kS.g; k1 ; k2 /k D kukH 1 C 2.˛/j : j D1 On the other hand, we have kukH D 1, because X 1=2 2 kukH D r˛ .; h/ ju.h/j O h2Zd D X h2Zd D 1=2 1 r˛ .; h/ jGr˛ .; h/j2 1=2 1 X 1 G2 r˛ .; h/ d h2Z X 1 1 1 C D G r˛ .; 0/ r˛ .; h/ 1=2 h¤0 d XY 1 1 D C j jhj Qd G j D1 1 h¤0 j D1 d d Y X 1 Y D 1C j 2 jhj G j D1 j D1 ˛ 1=2 ˛ 1=2 D 1: h¤0 In the above, we assume that g D k1 D c and we show another lower bound for k.I K/ 1 k D 1 cc . Thus, we have the following proposition. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 291 Worst case error for integro-differential equations ! ! Proposition 2.1. If k1 ; k2 ; k3 ; k4 2 Æ, then ! d d Y Y 1=2 1=2 c max ; e0;d 1 C 2.˛/j 1 C 2.˛/j : 1 c j D1 j D1 Hence we have a lower bound and upper bound on the initial error with the same dependence on d ; in other words we know exactly how the initial error increases with d . Proposition 2.2. If c WD min.ˇ; Nyström method (2.4) satisfies L 1 /, then the worst case error for the QMC- d c 2.˛/1 1 Y ; en;d .An / max .1 C 2.˛/w˛ j / 1 c n˛ n !1=2 1 ; j D1 where w˛ 1 is a constant independent of n and d . R ! L 1 ! Proof. We consider k1 D c WD min.ˇ; /, k2 and L D D k2 rgd y such that ! ! kk2 kH;1 < ˇ and kr k2 k ˇ; hence we write u " Z un D c g.y/d y 1 c D !# n 1X g.ti / : n i D1 Therefore we have the inequality ! sup kS.g; k1 ; k2 / en;d .An / ! An .g; k1 ; k2 /ksup kgkH 1 1 c D 1 sup kgkH 1 ˇ! n ˇ 1X ˇ g.ti /ˇ ˇ n ˇZ ˇ ˇ c ˇ g.y/d y ˇ D i D1 c wor-int e .t1 ; : : : ; tn /: c n;d Hence we conclude wor-int en;d .t1 ; : : : ; tn / wor-int en;1 1 2 n 1 0; ; ; : : : ; n n n D 2.˛/1 n˛ 1=2 ; (2.10) wor-int where en;d .t1 ; : : : ; tn / denotes the worst case integration error in H using quadrature points t1 ; : : : ; tn . If we consider Sharygin’s lower bound [4], then this rate of convergence of O.n ˛=2 / is optimal for the integration problem in weighted Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 292 D. Rostamy, M. Jabbari and M. Gadirian Korobov spaces. In fact, it was proved in [4, 5] that a generating vector z for a rank-1 lattice rule constructs component-by-component to achieve the rate of convergence O.n ˛=2Cı /; ı > 0. Also, some authors have proved (see [20, 21]) that wor-int en;d .t1 ; : : : ; tn / d 1 Y .1 C 2.˛/w˛ j / n !1=2 1 ; (2.11) j D1 where w˛ WD min.1; 1=.21 jmin j// 1, with the minimum of the function 1 < min < 1C2 ˛ denoting 1 X cos.2hx/ .x/ D h˛ hD1 (also, see [15, 16]). Moreover, it was proved in [25,26] that the integration problem in weighted Korobov spaces is strongly QMC tractable iff (1.7) holds, and QMC tractable iff (1.8) holds. In the following we obtain an upper bound for the worse case error. We recall that .I Kn /un D g and .I Kn /u D .I K/u C .K Kn /u D g C .K Kn /u: Therefore we obtain u un D .I Kn / 1 .K Kn /u: Thus, we have the following inequality: ! kS.g; k1 ; k2 / ! An .g; k1 ; k2 /ksup D ku un ksup k.I Kn / 1 kk.K Kn /uksup : Therefore we have k.I Kn / 1 k 1 C k.I K/ 1 kkKn k ; 1 n where n WD k.I K/ 1 kk.K Kn /Kn k < 1: Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 (2.12) 293 Worst case error for integro-differential equations Therefore we conclude kKn k kk1 ksup C kk3 ksup .kk1 kH C kk3 kH / d Y 1 C 2.˛/j 1=2 : (2.13) j D1 Hence we write ku un ksup 1 C k.I 1 k.kk K/ 1 kH k.I 1 k.K C kk3 kH / K/ Qd 1 kk.K j D1 1 C 2.˛/j 1=2 Kn /Kn k Kn /uksup : The term k.K Kn /Kn k controls whether or not n < 1 while k.K Kn /uksup determines the rate of convergence. It remains to obtain bounds on these two terms. Let t1 ; : : : ; tn be rank-1 lattice points generated by z, that is, ti D ¹ inz º where ¹xº D x Œx. We have Z Z ! k2 .x y/ ru.y/d y k1 .x y/u.y/d y C ..K Kn /u/.x/ D D D n 1X k1 .x n n 1 X! k2 .x n ti /u.ti / i D1 X D ti / ru.ti / i D1 X UO 1x .h/ h2Zd ¹0º hz0 .mod n/ UO 2x .h/; h2Zd ¹0º hz0 .mod n/ where U1x .y/ D k1 .x y/u.y/; ! U2x .y/ D k2 .x y/ ru.y/; Z O U1x .h/ D k1 .x y/u.y/e 2 ih:y d y D Z X D kO1 .l/e 2 i l:.x y/ u.y/e D D X l2Zd D X 2 i h:y dy l2Zd kO1 .l/e 2 i l:x Z u.y/e 2 i.hCl/:y dy D kO1 .l/e 2 i l:x u.h O C l/; l2Zd Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 294 D. Rostamy, M. Jabbari and M. Gadirian and UO 2x .h/ D ! k2 .x Z y/ ru.y/e D Z X ! D kO2 .l/e 2 i l:.x D 2 i h:y y/ dy :ru.y/e 2 i h:y dy l2Zd Z X! 2 il:x O D k2 .l/e : ru.y/e 2 i.hCl/:y dy D l2Zd X! kO2 .l/e 2 il:x u.h O C l/2 i:.h C l/: D l2Zd Hence we have X Kn /u D .K X kO1 .l/u.h O C l/e 2 i l:x h2Zd ¹0º l2Zd hz0 .mod n/ X X! kO2 .l/u.h O C l/e 2 il:x 2 i:.h C l/: h2Zd ¹0º l2Zd hz0 .mod n/ Proposition 2.3. Suppose there exists an integer vector z for Sn;d .z/ defined by Sn;d .z/ D X l2Zd X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; h C l/r˛ .; h/ 1=2 (2.14) such that Sn;d .z/ < 1 : 4ˇ 2 Then the worst case error for the lattice-Nyström method satisfies en;d .An / .1 C 2ˇ/ 1 Qd j D1 1 C 2.˛/j 4ˇ 2 Sn;d .z/ 1=2 ˇ. C 1/Sn;d .z/: Proof. See Appendix 2. In [4, 8, 9, 15, 16, 19], we observe the following CBC algorithm for constructing a generating vector z. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 295 Worst case error for integro-differential equations Algorithm 2.4. If n is a prime number, then: Step 1: Set z1 D 1. Step 2: For t D 2; 3; : : : ; d , with z1 ; : : : ; z t 1 already chosen and fixed, find a z t 2 ¹1; 2; : : : ; n 1º to minimize Sn;t .z1 ; : : : ; z t 1 ; z t /. In this case, the components of the generating vector z can be restricted to the set ¹1; 2; : : : ; n 1º. It causes to the optimal rate of convergence O.n ˛=2Cı / for ı > 0. Proposition 2.5. If we assume that 1 X 1=.˛ 2ı/ j < 1; j D1 and n is a prime number such that n .8ˇ 2 /2 26˛ d Y 1 C 2.1 C 2 3˛ 1=2 / .˛/j 2 ; (2.15) j D1 then the generating vector z constructed by Algorithm 2.4 completes the optimal rate of convergence, with en;d.An / CQ d;ı n ˛=2Cı en;d .An / Cd;ı n ˛=2Cı e0;d for all ı 2 .0; min.2 n D n.ı; d /. 3˛ ; .˛ 1/=2//, where Cd;ı and CQ d;ı are independent of Proof. According to [9], we suppose that z 2 ¹1; 2; : : : ; n 1ºd is constructed by Algorithm 2.4, such that n is a prime number. Then we have Sn;d .z / 1 ın1=.2p/ d Y p 1=p 1 C 2.1 C ı p /1=2 .˛p/j ; j D1 for all p 2 .1=˛; 1 and ı 2 .0; 2 3˛ . We now obtain a sufficient condition on n to ensure that 1 Sn;d .z/ < : 4ˇ 2 It is enough to choose n such that the upper bound in the above inequality with 1 p D 1 and ı D 2 3˛ is not greater than 8ˇ 2 . In other words, if we write (2.15), then 1 Sn;d .z/ : 4ˇ 2 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 296 D. Rostamy, M. Jabbari and M. Gadirian If we consider Sn;d .z/ d Y 1 ın1=.2p/ p 1=p 1 C 2.1 C ı p /1=2 .˛p/j j D1 1 ; 8ˇ 2 then we have d Y ın1=.2p/ 8ˇ 2 p 1=p 1 C 2.1 C ı p /1=2 .˛p/j : j D1 Now, we put p D 1 and ı D 2 n .8ˇ 2 /2 26˛ 3˛ ; then d Y 1 C 2.1 C 2 3˛ 1=2 / .˛/j 2 ; j D1 and we conclude from the above propositions that en;d .An / 2.1 C 2ˇ/ˇ. C 1/ ın1=.2p/ d Y p 1=p 1 C 2.1 C ı p /1=2 .˛p/j .1 C 2.˛/j /1=2 ; j D1 for all p 2 .1=˛; 1 and ı 2 .0; 2 d Y 3˛ . On the other hand, we know from [4] .1 C xj / .d C 1/ xj j D1 log.d C1/ Pd ; (2.16) j D1 for all xj > 0, we see that the requirement (2.15) on n does not grow with d if (1.7) holds, and it grows only polynomially with d when (1.8) holds. The conditions (1.7) and/or (1.8) are also sufficient to ensure that en;d .An / does not grow faster than polynomially with d . If we assume that p D 1=.˛ 2ı/ and ı min.2 3˛ ; .˛ 1/=2/, then we have en;d .An / D O.n 1=.2p/ ˛=2Cı / .˛ 2ı/=2 because in the above formula n Dn D n ˛=2Cı . However, we will need to assume stronger conditions on the weights if we have the optimal rate of convergence at the same time. On the other hand, we analyze tractability in the normalized sense. For given " 2 .0; 1/, we find the smallest n for which en;d .An / ". We observe that it is sufficient to insist that 1 Sn;d .z/ ; (2.17) 1=2 Q " 1 .1 C 2ˇ/ˇ. C 1/ jdD1 1 C 2.˛/j C 4ˇ 2 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 297 Worst case error for integro-differential equations the right-hand side of which is less than 1=.4ˇ 2 /. Using Algorithm 2.4, we generate a vector z satisfying (2.17) such that " d 1 Y p 2 n pr min 1 C 2.1 C ı p /1=2 .˛p/j 2p p2.1=˛;1 and ı2.0;2 3˛ ı j D1 " 1 .1 C 2ˇ/ˇ. C 1/ d Y !2p #! .1 C 2.˛/j /1=2 C 4ˇ 2 ; (2.18) j D1 where pr.x/ denotes the smallest prime number greater than or equal to x. Hence we conclude that nnor.";d // is less than or equal to the right-hand side of (2.18). Hence, for tractability in the normalized sense we obtain " d 1 Y p 2 nor n ."; d / pr min 1 C 2.1 C ı p /1=2 .˛p/j 2p 3˛ ı 2.1=˛;1 and ı2.0;2 j D1 #! 1 2 2p " .1 C 2ˇ/ˇ. C 1/ C 4ˇ : 2.2 Numerical experiments for the first case study In this section, we present some numerical results for the proposed scheme (2.4) by using the CBC algorithm. We carry out (2.4) by using an AMD Opteron computer with 15 Gigabytes RAM memory with 2.2 GHz CPU for these experiments. In (2.4), we assume that Qd 2 d Y e xi e xi k.x/ D i D1 ; p.x/ D p ; p.x/ 2 2 i D1 D 1, @D1 D @D; D D Œ0; 1d and u.x/ D d Y .xi e xi2 1/: i D1 Hence we obtain g.x/ by the following cases. Therefore we will compare exact solution with approximation solution. The evolution of the absolute error of this method, ku un k1 , for d D 10; 20 and n D 107; 523; 1009 are given in Tables 1–4 based on CPU times. In this case, we will consider two different examples in (2.1). Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 298 D. Rostamy, M. Jabbari and M. Gadirian Example 2.6. We have 8 a .y/ D 1; y 2 D; ˆ ˆ < 1 b0 D 0; ˆ ˆ ! :! a 2 .y/ D 0 ; y 2 D: Example 2.7. We have 8 a1 .y/ D 1; y 2 D; ˆ ˆ < ! ! b2 D 0; ˆ ˆ ! :! a 2 .y/ D 1 ; y 2 D: n CPU time(s) 107 523 1009 10.515 139.036 479.259 ku un k 0.251e-9 0.103e-8 0.631e-7 Table 1. d D 10 for Example 2.6. n CPU time(s) 107 523 1009 23.675 305.823 1123.862 ku un k 0.121e-8 0.3213e-8 0.321e-7 Table 2. d D 20 for Example 2.6. n CPU time(s) 107 523 1009 12.764 175.523 387.246 ku un k 0.745e-8 0.194e-8 0.971e-7 Table 3. d D 10 for Example 2.7. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 299 Worst case error for integro-differential equations n CPU time(s) 107 523 1009 52.876 426.985 1432.765 ku un k 0.642e-7 0.544-8 0.398e-7 Table 4. d D 20 for Example 2.7. 3 Second case study of (1.1) or integro-differential equation with convection ! as a vector in (1.1); therefore we study In this case, we consider ˛0 D 0, and ˛ 1 the equation Z p.u/ D g.x/ C 1 D Z 1 .x; y/u.y/d y C 2 D ! 2 .x; y/ ru.y/d y; (3.1) with the initial conditions (1.2) and (1.3). Therefore we have Z Z ! ! a1 ru.x/ D g.x/ C 1 1 .x; y/u.y/d y C 2 2 .x; y/ ru.y/d y: (3.2) D D We approximate u using the Nyström method based on quasi-Mont-Carlo rules. Let t1 ; : : : ; tn be selected points in D that we use to approximate u: n n 1 X 2 X ! ! 2 .x; ti / run .ti / a1 run .x/ D g.x/ C 1 .x; ti /un .ti / C n n i D1 i D1 by integrating the above equation and we assume that there is a suitable vector ! a1 such that Z u.x/ D ! a :ru.x/d xI 1 therefore we have Z un .x/ D g.x/d x C n Z 1 X 1 .x; ti /un .ti /d x n i D1 C 2 n n Z X ! 2 .x; ti / run .ti /d x: i D1 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 300 D. Rostamy, M. Jabbari and M. Gadirian Here, we obtain un .t1 /; : : : ; un .tn /, run .t1 /; : : : ; run .tn / by solving the following linear system: 8 n ˆ 1 X ˆ ! ˆ 1 .tj ; ti /un .ti / a ru .t / D g.t / C ˆ 1 n j j ˆ ˆ n ˆ ˆ i D1 ˆ ˆ ˆ n ˆ ˆ 2 X ! ˆ ˆ ˆ C 2 .tj ; ti / run .ti /; ˆ ˆ n < i D1 ! (3.3) Z n Z ˆ 1 X ˆ ˆ ˆ un .tj / D 1 .x; ti /un .ti /dx g.x/d x C ˆ ˆ n ˆ tj ˆ i D1 tj ˆ ˆ ˆ ! ˆ Z n ˆ ˆ 2 X ˆ ˆ C 2 .x; ti / run .ti /d x ; ˆ ˆ : n i D1 tj where j D 1; : : : ; n, and therefore we have 2n equations and 2n unknowns. We e W C 1;v .D/ ! C 1;v .D/ by define the integral operator K Z Z ! e D 1 2 ru.y/d y; (3.4) Ku 1 .x; y/u.y/d y C 2 .x; y/ D D and we consider Z Z e D 1 max kKk x2D D jk1 jd y C 2 max x2D D jk3 jd y; (3.5) ! such that k1 .x y/ D 1 .x; y/, k2 .x y/ D ! 2 .x; y/ and k3 D ! 2 ru.y/. The 1;v e n W C .D/ ! C 1;v .D/ is given by corresponding discrete operator K n X e n u D 1 k1 .x K n ti /un .ti / C i D1 with e n k D max kK x2D n 2 X ! k2 .x n ti / run .ti /; (3.6) i D1 n 1 X jk1 .x n i D1 ti /j C max x2D n 2 X jk3 .x n ti /j; (3.7) i D1 where t1 ; : : : ; tn 2 D. Thus the above kernels in (3.7) are convolution kernels with periodic form. Therefore we define Z Z e kKk D 1 jk1 .y/jd y C 2 jk3 .y/jdy 1 kk1 ksup C 2 kk3 ksup ; (3.8) D D Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 301 Worst case error for integro-differential equations and we write n e n k D 1 max kK x2D 1X jk1 .x n n ti /j C 2 max x2D i D1 1X jk3 .x n ti /j: i D1 Therefore we conclude that fn k 1 kk1 ksup C 2 kk3 ksup : kK (3.9) 3.1 Lower and upper bounds on the worst case error for (3.1) ! We assume that g; k1 ; k2 2 H are given functions and we study the solutions of ! S.g; k1 ; k2 / D u from (3.2) as ! e a1 ru D g C Ku; e D! where Ku a1 ru g. Hence we have e 1 .! a ru uDK (3.10) g/; 1 (3.11) since .x; y/ D k.x y/. On the other hand, we write Z Z ! e Ku.x/ D 1 k1 .x y/u.y/d y C 2 k2 .x y/ ru.y/d y D D X Z X 2 ih:.x y/ 2 i h:y O D 1 k1 .h/e u.h/e O dy D h2Zd h2Zd Z X ! C 2 kO2 .h/e 2 i h:.x D D 1 X X h2Zd 2 ih:x kO1 .h/u.h/e O C 2 h2Zd y/ u.h/2 O ih:e 2 i h:y dy h2Zd X ! kO2 .h/u.h/:2 O ihe 2 i h:x : h2Zd Therefore we have e Ku.x/ D X ! 2 i h:x u.h/e O 1 kO1 .h/ C 2 kO2 .h/ 2 i h : h2Zd Hence, if we put v.h/ D ! a1 ru.h/; then we have ! v.h/ O D g.h/ O C u.h/ O 1 kO1 .h/ C 2 kO2 .h/ 2 ih ; Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 (3.12) 302 D. Rostamy, M. Jabbari and M. Gadirian or we write u.h/ O D v.h/ O g.h/ O : ! 1 kO1 .h/ C 2 kO2 .h/ 2 i h (3.13) Thus we conclude that kukH ˇ2 !1=2 ˇ g.h/ O ˇ D ˇ r˛ .; h/ ! ˇ h2Z 1 kO1 .h/ C 2 kO2 .h/ 2 ih e 1 k k! kK a ru gk : ˇ X ˇˇ ˇ ˇ d v.h/ O 1 (3.14) H ! We use QMC-Nyström method by the algorithm An .g; k1 ; k2 / WD un or e n 1 .! un D K a1 ru g/: (3.15) e K e n k < 1; K/ (3.16) Also, if we assume that e n WD kK e 4 1 en kk.K e n 1 exists and we have then the operator K e kK 1 k e 1 kkK e nk 1 C kK : en 1 (3.17) Now, we consider ˇ > 0, > 1, ! e u D S.g; k1 ; k2 / D K and 1 .! a1 ru ! e n 1 .! un D An .g; k1 ; k2 / D K a1 ru g/ g/: Also, we assume that the worst case error of the QMC-Nyström method is introduced by en;d .An / WD sup ! k1 ;k2 ;g2Æ ! kS.g; k1 ; k2 / ! An .g; k1 ; k2 /ksup ; (3.18) where ® ! e D g; k1 ; k2 W k! Æ a1 ru ! e gkH 1; kk1 kH ; kk2 kH;1 ˇ; kK 1 ¯ k : ! Thus, if g 2 H is a linear function, then for all k1 ; k2 we have ! kS.g; k1 ; k2 / ! An .g; k1 ; k2 /ksup en;d .An /kgkH : Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 (3.19) 303 Worst case error for integro-differential equations Moreover, we define the initial error associated with the zero algorithm A0 0 as follows: e0;d WD sup kS.g; k; k1 /ksup : (3.20) ! k1 ;k2 ;g2Æ Therefore, the initial error follows from (1.6) and (3.15) is e kS.g; k1 ; k2 /ksup kK 1 kk! a1 ru gkH d Y 1 C 2.˛/j 1=2 ; j D1 thus we write ! 8g; k1 ; k2 ; 2 Æ; kS.g; k1 ; k2 /ksup d Y 1 C 2.˛/j 1=2 : (3.21) j D1 Now, we obtain a lower bound on the initial error. Let y/ D k1 ; k1 .x ! k2 .x ! y/ D ˛ 1 !k where k1 is a constant so that kk1 kH D k1 ˇ and we assume k˛ 1 H;1 ˇ and ! 1 e k . Now g.x/ is defined such that ka1 ru gkH 1 and kK u.h/ O D ! 1 a1 ru.h/ g.h/ O D ; ! Gr .; h/ ˛ O O 1 k1 .h/ C 2 k2 .h/ 2 i h (3.22) ! thus for this choice of g, k1 and k2 , we have d Y 1 X e 2 i h:x 1=2 ! D 1 C 2.˛/j : kS.g; k1 ; k2 /ksup D G r˛ .; h/ sup d (3.23) j D1 h2Z Hence we conclude the following proposition for the initial error based on the before section. Proposition 3.1. Let k1 .x kk1 kH ˇ; ! y/ D k1 and k2 .x ! where y/ D ˛ 1 !k D k! k˛ k 2 kH ˇ 1 H e and kK 1 k : Then d Y 1 C 2.˛/j j D1 1=2 e0;d d Y 1 C 2.˛/j 1=2 : j D1 Proof. See Appendix 3. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 304 D. Rostamy, M. Jabbari and M. Gadirian Proposition 3.2. Suppose there exists an integer vector z for which Sn;d .z/ is defined in (2.14) and we assume that Sn;d .z/ < 1 : .1 C 2 /2 ˇ 2 Then the worst case error for the lattice-Nyström method satisfies Q .1 C 2 /ˇ 1 C ˇ.1 C 2 / jdD1 1 C 2.˛/j en;d .An / Sn;d .z/; 1 ˇ 2 .1 C 2 /2 Sn;d .1 C 2 ˇ/ ˛ 2 Cı and we conclude en;d .An / D O.n /. Proof. We obtain a sufficient condition on n to ensure that Sn;d .z/ < 1 : C 2 / 2 ˇ 2 .1 It is enough to choose n such that the upper bound in Section 2 with p D 1 and ı D 2 3˛ is not greater than 2ˇ 2 .1 C /2 . In other words, if 1 2 n 2.1 C 2 /2 ˇ 2 26˛ 2 d Y 3˛ 1=2 1 C 2.1 C 2 / .˛/j 2 ; (3.24) j D1 then Sn;d .z/ 1 2ˇ 2 . 1 C 2 /2 because Sn;d .z/ d Y 1 ın1=.2p/ p 1=p 1 C 2.1 C ı p /1=2 .˛p/j j D1 1 : 2ˇ 2 .1 C 2 /2 Therefore we have the following inequality: ın 1=.2p/ 2 2 2ˇ .1 C 2 / d Y p 1=p 1 C 2.1 C ı p /1=2 .˛p/j : j D1 Moreover, by Propositions 2.5 and 3.2, we conclude en;d .An / d 2.1 C ˇ.1 C 2 // Y .1 C 2.˛/j /.1 C 2 /ˇ ın1=.2p/ j D1 d Y p 1=p 1 C 2.1 C ı p /1=2 .˛p/j : j D1 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 305 Worst case error for integro-differential equations On the other hand, if we put Sn;d .z/ < 1 ; 2.1 C 2 /2 ˇ 2 then ˇ 2 .1 C 2 /2 1 D : 2 2 2ˇ .1 C 2 / 2 Therefore we have the following inequality: ˇ 2 .1 C 2 /2 Sn;d 1 ˇ 2 .1 1 2 C 2 /2 Sn;d .z/ Also, if we put p D 1=.˛ 1 for all p 2 .˛ 2ı/ with ı min.2 en;d .An / D O.n because in the above formula we have n 3˛ ; .˛ ˛=2Cı 1=.2p/ Dn ; 1 and ı 2 .0; 2 3˛ : 1/=2/, then we obtain / .˛ 2ı/=2 Dn ˛=2Cı . Using the property d Y xj j D1 log.d C1/ Pd .1 C xj / .d C 1/ for all xj > 0; j D1 we see that the requirement (3.24) on n does not grow with d if holds, and it grows only polynomially with d when Pd j D1 j lim sup <1 log.d C 1/ d !1 P1 j D1 j <1 holds. Proposition 3.3. Suppose n is a prime number satisfying (3.24). Then the generating vector z is constructed by Algorithm 2.4, so it achieves the optimal rate of convergence, with en;d.An / en;d .An / Cd;ı n ˛=2Cı and CQ d;ı n ˛=2Cı ; e0;d for all ı 2 .0; min.2 3˛ ; .˛ 1/=2//, where Cd;ı and CQ d;ı are independent of n but depend on ı and d additionally if we write 1 X 1=.˛ 2ı/ j < 1; j D1 then the numbers Cd;ı , CQ d;ı and the requirement (3.24) on n are bounded independently of d . Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 306 D. Rostamy, M. Jabbari and M. Gadirian On the other hand, for tractability in the absolute sense, we find the smallest n for en;d .An / ". From the before proposition, we see that it is sufficient to insist that 1 Sn;d .z/ 1 " .1 C 2 /ˇ.1 C ˇ.1 C 2 // (3.25) 1 ; Qd 2 2 j D1 .1 C 2.˛/j /ˇ .1 C 2 / the right-hand side of which is less than 1=ˇ 2 .1 C 2 /2 . Using Proposition 2.5, we observe that Algorithm 2.4 will generate a vector z satisfying (3.25) if we demand that " d 1 Y p 2 n pr min 1 C 2.1 C ı p /1=2 .˛p/j (3.26) 2p 3˛ p2.1=˛;1 and ı2.0;2 ı j D1 " 1 .1 C 2 /ˇ.1 C ˇ.1 C 2 // d Y !2p #! 2 .1 C 2.˛/j / C ˇ .1 C 2 / 2 ; (3.27) j D1 where pr.x/ denotes the smallest prime number greater than or equal to x. Hence we conclude that nabs ."; d / is less than or equal to the right-hand side of (3.27). Similarly, for tractability in the normalized sense we obtain " d 1 Y p 2 nor 1 C 2.1 C ı p /1=2 .˛p/j n ."; d / pr min 2 p2.1=˛;1 and ı2.0;2 3˛ ı p j D1 #! 1 2 2 2p " .1 C 2 /ˇ.1 C ˇ.1 C 2 // C ˇ .1 C 2 / : 3.2 Numerical experiments for the second case study In this section, we present some numerical results for the proposed scheme (3.3) by using the CBC algorithm. We carry out (3.3) by using an AMD Opteron computer with 15 Gigabytes RAM memory with 2:2 GHz CPU for these experiments. In (3.3), we assume that Qd 2 d xi Y e xi i D1 e ; p.x/ D 1 .x/ D 2 .x/ D p ; p.x/ 2 2 i D1 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 Worst case error for integro-differential equations 307 1 D 2 D 1, @D1 D @D; D D Œ0; 1d and u.x/ D d Y .xi e xi2 1/: i D1 Hence we obtain g.x/ by the following cases. Therefore we will compare exact solution with approximation solution. The evolution of the absolute error of this method, ku un k1 , for d D 10; 20 and n D 107; 523; 1009 are given in Tables 5–8 based on CPU times. In this case, we will consider two different examples in (3.3). Example 3.4. We have ´ ! ! a 1 .y/ D 1 ; y 2 D; b0 D 0; Example 3.5. We have 8 ! <! a 1 .y/ D 1 ; y 2 D; ! : ! b2 D 0: n CPU time(s) 107 523 1009 238.515 743.036 1654.259 ku un k 0.643e-9 0.543e-8 0.687e-6 Table 5. d D 10 for Example 3.4. n CPU time(s) 107 523 1009 197.695 875.985 1487.092 ku un k 0.121e-8 0.459e-9 0.043e-7 Table 6. d D 20 for Example 3.4. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 308 D. Rostamy, M. Jabbari and M. Gadirian n CPU time(s) 107 523 1009 226.987 654.841 1304.765 ku un k 0.942e-7 0.044e-8 0.654e-6 Table 7. d D 10 for Example 3.5. n CPU time(s) 107 523 1009 152.876 426.985 1752.765 ku un k 0.642e-7 0.544-8 0.398e-7 Table 8. d D 20 for Example 3.5. A Appendix 1 Consider the following sequence of linear operators: Kn W C 1;v .D/ ! Pn given by n Kn u D 1X k1 .x n n ti /u.ti / C i D1 1 X! k2 .x n ti / ru.ti /; i D1 where Pn is the space of polynomials of degree less than or equal to n in D. Therefore the dimension of this space is finite. On the other hand, we assume that the kernel functions satisfy the following conditions: max x2D and max x2D n X jk1 .x ti /j D B1;n < 1 i D1 n X ! kk2 .x ti /k D B2;n < 1: i D1 Also, we consider the following linear operator: K W C 1;v .D/ ! C 1;v .D/ Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 Worst case error for integro-differential equations 309 given by Z Ku D Z k1 .x D t/u.t/dt C ! k2 .x D t/ ru.t/dt: Theorem A.1. Each operator Kn W C 1;v .D/ ! Pn has the following properties: (1) it is a bounded operator, (2) dim Kn .Pn / < 1, (3) the operator Kn is compact. Proof. It is clear that we can prove that Kn is a sequence of bounded operators (see [1, Chapter 12]) and according to the theorem of finite dimensional domain or rang (see [2, Chapter 8]), the proof is completed. Theorem A.2 (Sequence of compact linear operators). Let ¹Kn º be a sequence of compact linear operators from the normed space C 1;v .D/ into the Banach space Pn . If ¹Kn º is uniformly operator convergent, say, kKn Kk ! 0, then the limit operator K is compact. Proof. See [2]. Theorem A.3 (Continuity). Let X and Y be normed spaces. Then every compact linear operator K W X ! Y is bounded, hence continuous. Proof. See [2]. B Appendix 2. Proof of Proposition 2.3 It follows from the Cauchy–Schwarz inequality that ˇ X X ˇ k.K Kn /uksup D sup ˇˇ UO 1x .h/ C x2D h2Zd ¹0º hz0 .mod n/ ˇ ˇ D sup ˇˇ x2D X h2Zd ¹0º hz0 .mod n/ X ˇ ˇ UO 2x .h/ˇˇ kO1 .l/u.h O C l/e 2 i l:x h2Zd ¹0º l2Zd hz0 .mod n/ C X ˇ X! ˇ 2 i l:x O k2 .l/u.h O C l/e 2 i.h C l/ˇˇ h2Zd ¹0º l2Zd hz0 .mod n/ Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 310 D. Rostamy, M. Jabbari and M. Gadirian DI ‚ ƒ X…„ O jk1 .l/jju.h O C l/j X h2Zd ¹0º l2Zd hz0 .mod n/ DII …„ ƒ X ˇ! ˇ ˇkO2 .l/u.h O C l/2 i.h C l/ˇ : ‚ X C h2Zd ¹0º l2Z d hz0 .mod n/ We have X X I D jkO1 .l/jju.h O C l/j h2Zd ¹0º l2Zd hz0 .mod n/ " X jkO1 .l/j l2Zd X ju.h O C l/j2 r˛ .; h C l/ h2Zd ¹0º hz0 .mod n/ X h2Zd ¹0º hz0 .mod n/ by the Cauchy–Schwarz inequality X X jkO1 .l/j2 l2Zd X l2Zd X 1=2 # ; 1=2 ju.h O C l/j r˛ .; h C l/ 2 X h2Zd ¹0º hz0 .mod n/ jkO1 .l/jkukH l2Zd D kukH 1 r˛ .; h C l/ h2Zd ¹0º hz0 .mod n/ D 1=2 X l2Zd X 1 r˛ .; h C l/ X h2Zd ¹0º hz0 .mod n/ jkO1 .l/j2 r˛ .; h/1=2 1=2 1 r˛ .; h C l/ 1=2 l2Zd X l2Zd X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; h C l/r˛ .; h/ 1=2 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 311 Worst case error for integro-differential equations D kukH kk1 kH Sn;d .z/ kk1 kH k.I K/ 1 kkgkH Sn;d .z/ ˇSn;d .z/; and X II D Xˇ ˇ ˇkO2 .l/ u.h O C l/2 i.h C l/ˇ h2Zd ¹0º l2Zd hz0 .mod n/ X ! kkO2 .l/k21 l2Zd X ku.h O C l/2 i.h C l/k21 r˛ .; h 1=2 C l/ h2Zd ¹0º hz0 .mod n/ X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; h C l/ 1=2 by the Cauchy–Schwarz inequality 1=2 X ! X 2 2 O ku.h O C l/2 i.h C l/k1 r˛ .; h C l/ kk2 .l/k1 h2Zd ¹0º hz0 .mod n/ l2Zd X l2Zd D X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; h C l/ X ! 1=2 X kkO2 .l/k21 krukH;1 l2Zd l2Zd 1=2 X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; h C l/ 1=2 X ! 1=2 2 O D krukH;1 kk2 .l/k1 r˛ .; l/ l2Zd X l2Zd X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; l/r˛ .; h C l/ 1=2 ! D krukH;1 kk2 kH Sn;d .z/ ˇSn;d .z/: Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 312 D. Rostamy, M. Jabbari and M. Gadirian In the above equation, we assume that krukH;1 1. Thus, we obtain k.K Kn /uksup . C 1/Sn;d .z/: Similarly, we write .K Kn /Kn D KKn Kn Kn Z n 1X D k1 .x y/ k1 .y n D tj /u.tj / j D1 n 1 X! k2 .y C n ! tj / ru.tj / dy j D1 ! k2 .x Z C 1 n y/ D n X rk1 .y tj /u.tj / j D1 n 1X ! C r k2 .y n ! tj /ru.tj / d y j D1 n 1X k1 .x n ti / i D1 n 1X k1 .ti n tj /u.tj / j D1 n 1 X! C k2 .ti n ! tj /ru.tj / j D1 n 1 X! k2 .x n i D1 n 1X ti / rk1 .ti n tj /u.tj / j D1 n 1X ! C r k2 .ti n ! tj /ru.tj / j D1 D 1 n n X Z k1 .x y/k1 .y tj /u.tj /d y j D1 n 1X k1 .x n ! ti /k1 .ti tj /u.tj / i D1 Z n 1X C k1 .x n ! y/k2 .y tj /ru.tj /d y j D1 n 1X k1 .x n ! ti /k2 .ti ! tj /ru.tj / i D1 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 313 Worst case error for integro-differential equations Z n ! 1X k2 .x C n y/rk1 .y tj /u.tj /d y j D1 n ! 1 X! k2 .x n ti /rk1 .ti tj /u.tj / i D1 C 1 n n X Z ! k2 .x ! y/r k2 .y tj /ru.tj /d y j D1 n 1 X! k2 .x n ! ti /r k2 .ti ! tj /ru.tj / i D1 DW I C II C III C IV: Therefore we have ˇ n Z 1 X ˇˇ kI C IIk D sup ˇ k1 .x ˇ D x2D n y/k1 .y tj /d y j D1 n 1X k1 .x n ˇ ˇ ˇ tj /ˇ ˇ ti /k1 .ti i D1 ˇ Z n 1 X ˇˇ k .x C sup ˇ ˇ D 1 x2D n ! y/k2 .y tj /d y j D1 n ! ti /k2 .ti 1X k1 .x n i D1 DW sup x2D 1 n n X ji j C sup x2D j D1 1 n n X ˇ ˇ ˇ tj / :ru.tj /ˇ ˇ ! jiij j D1 and kIII C IVk D sup x2D n ˇZ 1 X ˇˇ ! ˇ k2 .x n D y/:rk1 .y tj /d y j D1 n 1 X! k2 .x n ti /:rk1 .ti i D1 ˇ Z n ! 1 X ˇˇ C sup k2 .x ˇ ˇ D x2D n ! y/:r k2 .y ˇ ˇ tj /ˇˇ tj /d y j D1 n 1 X! k2 .x n i D1 ! ti /:r k2 .ti ˇ ˇ ˇ tj / :ru.tj /ˇ ˇ ! Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 314 D. Rostamy, M. Jabbari and M. Gadirian n n 1X 1X DW sup jiiij C sup jivj; x2D n x2D n j D1 j D1 where Z X iD k1 .x h2Zd ¹0º hz0 .mod n/ Z X X D D h2Zd ¹0º hz0 .mod n/ X D y/k1 .y X h2Zd ¹0º hz0 .mod n/ kO1 .l/e 2 i l:x ii D h2Zd dy 2 i.hCl/:y dy 2 i.hCl/:tj ; ! y/k2 .y tj /:ru.tj /e 2 i h:y dy D X ¹0º hz0 .mod n/ l2Zd X X D tj /e 2 i h:y tj /e l2Zd h2Zd ¹0º hz0 .mod n/ D k1 .y Z kO1 .l/e 2 i l:x kO1 .h C l/e k1 .x X k1 .y D Z X dy l2Zd X X y/ kO1 .l/e 2 il:.x h2Zd ¹0º l2Zd hz0 .mod n/ D 2 i h:y tj /e D kO1 .l/e 2 i l:x ! k2 .y Z tj /:ru.tj /e 2 i.hCl/:y dy D kO1 .l/e 2 i l:x kO3 .h C l/e 2 i.hCl/:tj ; h2Zd ¹0º l2Zd hz0 .mod n/ ! where we define k3 .y tj / D k2 .y tj / ru.tj /, Z X ! iii D k2 .x y/:rk1 .y tj /e 2 i h:y d y h2Zd ¹0º hz0 .mod n/ D X h2Zd ¹0º hz0 .mod n/ D X D Z X ! kO2 .l/e 2 i l:.x D y/ :rk1 .y 2 i h:y dy l2Zd Z X! 2 il:x O k2 .l/e : rk1 .y h2Zd ¹0º l2Zd hz0 .mod n/ tj /e tj /e 2 i.hCl/:y dy D Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 315 Worst case error for integro-differential equations X! O 1 .h C l/e kO2 .l/e 2 i l:x :rk X D h2Zd ¹0º hz0 .mod n/ ¹0º hz0 .mod n/ D D ¹0º hz0 .mod n/ h2Zd ¹0º hz0 .mod n/ y/ ! :r k2 .y h2Zd ¹0º hz0 .mod n/ ! ! where k4 D r k2 .y 2 i h:y dy tj /:ru.tj /e 2 i h:y tj /:ru.tj /e 2 i.hCl/:y dy D l2Zd 2 i.hCl/:tj ; l2Zd tj /ru.tj /. Therefore we write k.K j D1 X j D1 n ˇ X ˇ 1 ˇ C sup ˇ n x2D j D1 n ˇ X ˇ 1 ˇ C sup ˇ n x2D j D1 X X kO1 .l/kO1 .h C l/e 2 i l:x e ˇ ˇ ˇ 2 i.hCl/:tj ˇ h2Zd ¹0º l2Zd hz0 .mod n/ n ˇ X ˇ 1 ˇ C sup ˇ n x2D X h2Zd ¹0º hz0 .mod n/ X X kO1 .l/kO3 .h C l/e 2 i l:x e ˇ ˇ ˇ 2 i.hCl/:tj ˇ l2Zd X! O 1 .h C l/e 2 i l:x e kO2 .l/:rk ˇ ˇ ˇ 2 i.hCl/:tj ˇ h2Zd ¹0º l2Zd hz0 .mod n/ X X! ! kO2 .l/:kO4 .h C l/e 2 i l:x e ˇ ˇ ˇ 2 i.hCl/:tj ˇ h2Zd ¹0º l2Zd hz0 .mod n/ ! X jkO1 .l/jjkO1 .h C l/j C jkO1 .l/jjkO4 .h C l/j h2Zd ¹0º l2Zd hz0 .mod n/ dy l2Zd ! X! kO2 .l/e 2 il:x :kO4 .h C l/e X Kn /Kn k n ˇ X ˇ 1 ˇ sup ˇ n x2D tj /:ru.tj /e Z X! ! kO2 .l/e 2 i l:x : r k2 .y X D ! y/:r k2 .y Z X ! kO2 .l/e 2 i l:.x h2Zd D ! k2 .x D h2Zd X ; l2Zd Z X iv D 2 i.hCl/:tj ! ! O 1 .h C l/j C jkO2 .l/ ! C jkO2 .l/ rk k4 .h C l/j Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 316 D. Rostamy, M. Jabbari and M. Gadirian ! ! ! kk1 k2H C kk1 kH kk3 k C kk2 kH;1 krk1 kH;1 C jk2 kH;1 kk4 kH;1 Sn;d .z/ ! ! kk1 k2H C kk1 kH kk2 kH krukH;1 C kk2 kH;1 krk1 kH;1 ! ! C kk2 kH;1 kr k2 kH;1 krukH;1 Sn;d .z/: ! If k1 ; k2 2 Æ and krukH;1 1, then we have Kn /Kn k .ˇ 2 C ˇ 2 C ˇ 2 C ˇ 2 /Sn;d .z/ D 4ˇ 2 Sn;d .z/: k.K On the other hand, we write n D k.I 1 K/ kk.K Kn /Kn k 4ˇ 2 Sn;d .z/; so we have n < 1. Hence we write Sn;d .z/ < 1 : 4ˇ 2 Also, we observe that k.I Kn / 1 k 1 C k.I Kn / 1 C 2ˇ Qd 1 k.kk k C kk k / 1 H 3 H 1 1 1 j D1 1 C 2.˛/j Qd j D1 1 C 2.˛/j 1=2 n 1=2 4ˇ 2 Sn;d .z/ d Y 1=2 1 C 2ˇ 1 C 2.˛/j : 2 4ˇ Sn;d .z/ (B.1) j D1 On the other hand, we conclude that ku un ksup k.I Kn / 1 kk.K Kn /uksup d 1=2 .1 C 2ˇ/ Y 1 C 2.˛/ k.K j 1 4ˇ 2 Sn;d .z/ Kn /uksup j D1 d 1=2 .1 C 2ˇ/ Y 1 C 2.˛/j 2 1 4ˇ Sn;d .z/ j D1 k.I K/ 1 kkgkH kk1 kH C ˇ Sn;d .z/ 1=2 Q .1 C 2ˇ/ jdD1 1 C 2.˛/j .ˇ C ˇ/Sn;d .z/: 1 4ˇ 2 Sn;d .z/ Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 317 Worst case error for integro-differential equations C Appendix 3. Proof of Proposition 3.1 y/ D k1 as a constant function such that Again we consider k1 .x e 1 k ; kk1 kH D k1 ˇ and kK ! also, we assume that k1 .x y/ and k2 .x y/ are separable kernels. Then, we have Z Z ! ! a1 ru.x/ D g.x/ C 1 k1 .x y/u.y/d y C 2 k2 .x y/ ru.y/d y; D D or we write ! a1 ru.x/ D g.x/ C 1 ˛1 .x/ Z D Z b1 .y/u.y/d y C 2 ˛2 .x/ D b2 .y/ ru.y/d y; D g.x/ C 1 c1 C 2 c2 : We find c1 and c2 by integration of the above equation, thus we have Z u.x/ D g.x/d x C .1 c1 C 2 c2 /x C d1 : Hence we have Z Z g.x/ C 1 c1 C 2 c2 D g.x/ C 1 k1 D Z C 2 D g.y/d y C .1 c1 C 2 c2 /y C d1 d y g.y/ C 1 c1 C 2 c2 d y; therefore c2 D Z Z 1 c1 C d1 1 k1 g.y/d yd y C 2 1 2 K21 22 D Z C 2 g.y/d y C 1 c1 C d2 c1 1 c1 1 2 D D l.c1 /: Then we have Z u.y/ D g.y/d y C 1 c1 y C 1 2 Z Z 1 k1 y 1 c1 C d1 2 1 c1 ; g.y/d yd y C 1 2 K21 22 D Z C 2 g.y/d y C 1 c1 C d2 c1 D Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 318 D. Rostamy, M. Jabbari and M. Gadirian and Z un .y/ D g.y/d y C 1 c1 y C 2 n Z 1X 1 k1 y g.y/d y n ti " 1 1 2 K21 22 i D1 C 1 c1 C d1 2 n 2 X C g.ti / C 2 1 c1 C d2 n # ! c1 1 c1 i D1 and so u un D Z Z 1 1 2 K21 2 22 1 2 k1 y g.y/d yd y D n Z C 2 n Z 1X g.y/d y n ti i D1 ! g.y/d y D 1X g.ti / n : i D1 Hence we have en;d .An / sup kS.g; k1 ; k2 / An .g; k1 ; k2 /ksup kgkH 1 D 1 k1 1 k 1 1 2 Z Z g.y/d yd y D n Z 1X g.y/d y n ti i D1 ! wor-int C en;d .t1 ; : : : ; tn / ; so we have DF en;d .An / ‚ Z Z 1 k1 1 k 1 1 2 1 D …„ ƒ n Z 1X g.y/d yd y g.y/d y n ti i D1 ! wor-int C en;d .t1 ; : : : ; tn / : Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 319 Worst case error for integro-differential equations On the other hand, by using (2.10) and (2.11), we have en;d .An / 1 k1 k 1 1 2 1 1 max F C 2ı.˛/1 n˛ 12 ; d 1 Y .1 C 2ı.˛/w˛ j / n FC ! 12 ! 1 : j D1 e n un D ! We know that if K a1 ru g, then we have e n 1 .! un D K a1 ru fn u D ! Also, we have K a1 ru g/: g and then fn uDK 1 e n u D .K en Moreover, if we write K .! a1 ru g/: e C Ku e or K/u e n 1 .K en uDK e CK e n 1 Ku; e K/u then we have un e n 1 Ku e uDK e n 1 .K en K e n 1 Ku e D K e K/u e n 1 .K en K e K/u; or un e e n 1 .K uDK e n /u: K (C.1) Therefore we conclude the following inequality: ! kS.g; k1 ; k2 / ! An .g; k1 ; k2 /k D ku fn un k kK 1 e kk.K e n /uksup : (C.2) K We recall that e n WD kK 1 therefore we write e n 1k kK e kk.K e n /K e n k < 1; K e 1 kkK e nk 1 C kK : 1 n (C.3) On the other hand, we have e n k 1 kk1 ksup C 2 kk3 ksup kK .1 kk1 kH C 2 kk3 kH / d Y .1 C 2.˛/j /: j D1 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 (C.4) 320 D. Rostamy, M. Jabbari and M. Gadirian Hence we write the following inequality: ku e n 1 kk.K e un ksup kK e n /uksup K e 1 k kK e nk 1 C kK e k.K 1 n e 1 C kK e n /uksup K 1=2 Q C 2 kk3 kH / jdD1 1 C 2.˛/j e 1 k k.K e K e n /K e nk kK 1 k. kk k 1 1 H 1 e K e n /uksup : k.K Moreover, we know that Z Z ! e K e n /u.x/ D 1 .K k1 .x; y/u.y/d y C 2 k2 .x; y/ ru.y/d y D D D n n 2 X ! 1 X k1 .x ti /u.ti / C k2 .x ti / ru.ti / n n i D1 i D1 X X UO 1x .h/ 2 UO 2x .h/ 1 h2Zd ¹0º hz0 .mod n/ where U1x D k1 .x UO 1x h2Zd ¹0º hz0 .mod n/ ! y/u.y/, U2x D k2 .x y/:ru.y/, Z D k1 .x y/u.y/e 2 ih:y d y D Z X 2 il:.x y/ O k1 .l/e u.y/e D D kO1 .l/e 2 i l:x X Z u.y/e 2 i.hCl/:y dy D l2Zd D dy l2Zd X D 2 i h:y kO1 .l/u.h O C l/e 2 il:x ; l2Zd and UO 2x D D Z ! k2 .x D y/ ru.y/e Z X ! kO2 .l/e 2 i l:.x D 2 i h:y y/ dy ru.y/e 2 ih:y dy l2Zd Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 321 Worst case error for integro-differential equations Z X! 2 il:x O D k2 .l/e ru.y/e l2Zd D 2 i.hCl/y dy G X! kO2 .l/e 2 il:x u.h O C l/2 i.h C l/: l2Zd Thus, we have e k.K ˇ ˇ e K n /uksup D sup ˇˇ1 x2D ˇ ˇ D sup ˇˇ1 x2D X X UO 1x .h/ C 2 h2Zd ¹0º hz0 .mod n/ X h2Zd ¹0º hz0 .mod n/ X ˇ ˇ O U2x .h/ˇˇ kO1 .l/u.h O C l/e 2 il:x h2Zd ¹0º l2Zd hz0 .mod n/ X C 2 ˇ X! ˇ 2 il:x O k2 .l/u.h O C l/e 2 i.h C l/ˇˇ h2Zd ¹0º l2Zd hz0 .mod n/ DI0 ‚ 1 X …„ X ƒ jkO1 .l/jju.h O C l/j h2Zd ¹0º l2Zd hz0 .mod n/ X C 2 X ! jkO2 .l/u.h O C l/2 i.h C l/j : h2Zd ¹0º l2Zd hz0 .mod n/ ƒ‚ „ Therefore we compute " X O I0 1 jk1 .l/j l2Zd X ju.h O C l/j r˛ .; h C l/ ¹0º hz0 .mod n/ X h2Zd ¹0º hz0 .mod n/ 1 l2Zd 1=2 2 h2Zd X … DII 0 jkO1 .l/j2 1 r˛ .; h C l/ 1=2 # 1=2 ju.h O C l/j r˛ .; h C l/ X 2 h2Zd ¹0º hz0 .mod n/ Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 322 D. Rostamy, M. Jabbari and M. Gadirian X X l2Zd D 1 X h2Zd ¹0º hz0 .mod n/ jkO1 .l/jkukH X l2Zd D 1 kukH 1 r˛ .; h C l/ l2Zd X 1=2 X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; h C l/ 1=2 jkO1 .l/j2 r˛ .; l/ l2Zd X l2Zd X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; l/ r˛ .; h C l/ 1=2 D 1 kukH kk1 kH Sn;d .z/ e 1 kk! 1 kK a1 ru gkH kk1 kH Sn;d .z/ 1 ˇSn;d .z/: Also, we have X ! kkO2 .l/k1 ku.h O C l/2 i.h C l/k1 X II0 2 h2Zd ¹0º l2Zd hz0 .mod n/ " 2 X ! kkO2 .l/k1 1=2 ku.h O C l/2 i.h C l/k21 r˛ .; h C l/ X h2Zd ¹0º hz0 .mod n/ l2Zd X h2Zd ¹0º hz0 .mod n/ 2 krukH;1 1 r˛ .; h C l/ 1=2 # X ! 1=2 kkO2 .l/k21 r˛ .; h/ l2Zd X l2Zd X h2Zd ¹0º hz0 .mod n/ 1 r˛ .; h/r˛ .; h C l/ 1=2 ! 2 krukH;1 kk2 kH;1 Sn;d .z/ 2 ˇkrukH;1 Sn;d .z/: (C.5) Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 323 Worst case error for integro-differential equations In (C.5), we assume krukH;1 1. Therefore we have e k.K e n /uksup .1 C 2 /ˇSn;d .z/: K (C.6) Using a similar argument, we obtain e .K e n /K e nu D K eK e nu K e nK e nu K Z n 1 X D 1 k1 .x y/ k1 .y n G tj /u.tj / j D1 ! n 2 X ! C k2 .y tj /ru.tj / d y n j D1 Z n ! 1 X C 2 rk1 .y tj /u.tj / k2 .x y/ n D j D1 ! n 2 X ! r k2 .y tj /ru.tj / d y C n j D1 1 n n X k1 .x ti / i D1 n 1 X k1 .ti n tj /u.tj / j D1 n 2 X ! k2 .ti C n ! tj /ru.tj / j D1 2 n n X ! k2 .x ti / i D1 n 1 X rk1 .ti n tj /u.tj / j D1 n 2 X ! r k2 .ti C n ! tj /ru.tj / j D1 Z n 1 X D 1 k1 .x n D y/k1 .y tj /u.tj /d y j D1 n 1 X k1 .x n ! ti /k1 .ti tj /u.tj / i D1 C Z n 2 X 1 k1 .x n D ! y/k2 .y tj /ru.tj /d y j D1 n 1 X k1 .x n ! ti /k2 .ti ! tj /ru.tj / i D1 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 324 D. Rostamy, M. Jabbari and M. Gadirian Z n ! 1 X 2 k2 .x C n G y/rk1 .y tj /u.tj /d y j D1 n 2 X ! k2 .x n ! ti /rk1 .ti tj /u.tj / i D1 C Z n ! 2 X 2 k2 .x n G ! y/r k2 .y tj /ru.tj /d y j D1 n 2 X ! k2 .x n ! ti /r k2 .ti ! tj /ru.tj / ; i D1 DW I C II C III C IV: We write I C II D Z n 1 X 1 k1 .x n D y/k1 .y tj /d y j D1 n 1 X k1 .x n ! ti /k1 .ti tj / u.tj / i D1 C Z n 2 X 1 k1 .x n G ! y/k2 .y tj /d y j D1 n 1 X k1 .x n ! ti /k2 .ti ! tj / :ru.tj /; i D1 ˇ Z n 1 X ˇˇ k1 .x kI C IIk D sup ˇ1 ˇ D x2D n y/k1 .y tj /d y j D1 n 1 X k1 .x n ti /k1 .ti ˇ ˇ ˇ tj /ˇ ˇ ! y/k2 .y tj /d y i D1 ˇ Z n 2 X ˇˇ k1 .x C sup ˇ 1 ˇ D x2D n j D1 n 1 X k1 .x n i D1 DW sup x2D 1 n n X j D1 ji j C sup x2D 2 n n X ! ti /k2 .ti ˇ ˇ ˇ tj / :ru.tj /ˇ ˇ ! jiij j D1 Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 325 Worst case error for integro-differential equations and ˇ Z n ! 1 X ˇˇ kIII C IVk D sup k2 .x ˇ2 ˇ D x2D n y/:rk1 .y tj /d y j D1 n 2 X ! k2 .x n ti /:rk1 .ti ˇ ˇ ˇ tj /ˇ ˇ ! y/:r k2 .y tj /d y i D1 ˇ Z n ! 2 X ˇˇ C sup k2 .x ˇ 2 ˇ D x2D n j D1 n 2 X ! k2 .x n ˇ ˇ ˇ tj / :ru.tj /ˇ ˇ ! ! ti /:r k2 .ti i D1 DW sup x2D 1 n n X jiiij C sup x2D j D1 n X 2 n jivj j D1 where n Z i D 1 D k1 .x y/k1 .y X 1 ¹0º hz0 .mod n/ X 1 h2Zd ¹0º hz0 .mod n/ D X 1 h2Zd D Z k1 .x ti /k1 .ti tj / y/k1 .y tj /e 2 ih:y dy D Z X D kO1 .l/e 2 i l:.x y/ k1 .y 2 i h:y tj /e dy l2Zd X ¹0º hz0 .mod n/ l2Zd X X 1 ! i D1 h2Zd D tj /d y D 1X k1 .x n kO1 .l/e 2 il:x Z k1 .y tj /e 2 i.hCl/:y dy D kO1 .l/e 2 il:x kO1 .h C l/e 2 i.hCl/tj ; h2Zd ¹0º l2Zd hz0 .mod n/ Z ii D 1 D 1 ! k1 .x y/k2 .y tj /ru.tj /d y D n ! 1X k1 .x ti /k2 .ti n ! tj /:ru.tj / i D1 X h2Zd ¹0º hz0 .mod n/ Z k1 .x ! y/k2 .y tj /:ru.tj /e 2 i h:y dy D Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 326 D D. Rostamy, M. Jabbari and M. Gadirian Z X X 1 h2Zd ¹0º hz0 .mod n/ D X 1 ¹0º hz0 .mod n/ l2Zd X X 1 y/ ! k2 .y tj /:ru.tj /e 2 i h:y dy l2Zd X h2Zd D D kO1 .l/e 2 i l:.x kO1 .l/e 2 il:x ! k2 .y Z tj /:ru.tj /e 2 i.hCl/:y dy D kO1 .l/e 2 il:x kO3 .h C l/e 2 i.hCl/:tj ; h2Zd ¹0º l2Zd hz0 .mod n/ iii D 2 D n ! k2 .x Z y/:rk1 .y Z X 2 h2Zd ¹0º hz0 .mod n/ D h2Zd ¹0º hz0 .mod n/ D h2Zd ¹0º hz0 .mod n/ ! k2 .x Z iv D 2 D 2 2 2 i h:y tj /e dy y/ rk1 .y tj /e 2 i h:y dy tj /e 2 i.hCl/:y dy D 2 i.hCl/:tj ; l2Zd ! y/:r k2 .y tj /:ru.tj /d y D 1 X! k2 .x n i D1 Z X h2Zd ¹0º hz0 .mod n/ D tj / l2Zd l2Z d n D y/:rk1 .y X! O 1 .h C l/e kO2 .l/e 2 il:x :rk X 2 ti /:rk1 .ti D Z X! 2 i l:x O k2 .l/e :rk1 .y X 2 ! k2 .x Z X ! kO2 .l/e 2 i l:.x X 2 ! i D1 h2Zd ¹0º hz0 .mod n/ D tj /d y D 1 X! k2 .x n X h2Zd ¹0º hz0 .mod n/ ! ti /:r k2 .ti ! k2 .x ! tj /ru.tj / ! y/:r k2 .y tj /:ru.tj /e 2 i h:y dy D Z X ! kO2 .l/e 2 i l:.x D y/ ! :r k2 .y tj /:ru.tj /e 2 i h:y l2Zd Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 dy 327 Worst case error for integro-differential equations D Z X! ! 2 il:x O k2 .l/e : r k2 .y X 2 h2Zd l2Zd ¹0º hz0 .mod n/ D ! X! kO2 .l/e 2 il:x : kO 4 .h C l/e X 2 tj /ru.tj /e 2 i.hCl/:y dy D 2 i.hCl/tj : h2Zd ¹0º l2Zd hz0 .mod n/ Therefore we conclude e k.K e n /K e nk K n ˇ X ˇ 21 ˇ sup ˇ n x2D X j D1 j D1 n ˇ X ˇ 1 2 ˇ C sup ˇ n x2D j D1 n ˇ X ˇ 22 ˇ sup ˇ n x2D j D1 X kO1 .l/kO1 .h C l/e 2 i l:x e ˇ ˇ ˇ 2 i.hCl/:tj ˇ h2Zd ¹0º l2Zd hz0 .mod n/ n ˇ X ˇ 1 2 ˇ C sup ˇ n x2D C X X X X kO1 .l/kO3 .h C l/e 2 i l:x e ˇ ˇ ˇ 2 i.hCl/:tj ˇ h2Zd ¹0º l2Zd hz0 .mod n/ X X! O 1 .h C l/e 2 il:x e kO2 .l/:rk ˇ ˇ ˇ 2 i.hCl/:tj ˇ h2Zd ¹0º l2Zd hz0 .mod n/ X X! ! kO2 .l/: kO 4 .h C l/e 2 il:x e ˇ ˇ ˇ 2 i.hCl/:tj ˇ h2Zd ¹0º l2Zd hz0 .mod n/ 21 jkO1 .l/jjkO1 .h C l/j C 1 2 jkO1 .l/jjkO3 .h C l/j h2Zd ¹0º l2Zd hz0 .mod n/ ! ! ! O 1 .h C l/j C 2 jkO2 .l/ kO 4 .h C l/j C 1 2 jkO2 .l/ rk 2 ! ! 21 kk C1 k2H C 1 2 kk1 kH kk2 kH;1 C 1 2 kk2 kH;1 krk1 kH;1 ! ! C 22 jk2 kH kr k2 kH Sn;d .z/; thus we have e k.K e n /K e n k .21 ˇ 2 C 21 2 ˇ 2 C 22 ˇ 2 /Sn;d .z/; K and we write e n WD kK 1 e k k.K e n /K e n k .1 C 2 /2 ˇ 2 Sn;d .z/: K Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 328 D. Rostamy, M. Jabbari and M. Gadirian 1 Also, if we put n < 1, then it is sufficient to demand that Sn;d .z/ < . C 2 2. 1 2/ ˇ Hence we write the following inequality: Q e 1 k.1 kk1 kH C 2 kk3 kH / d .1 C 2.˛/j / 1 C kK j D1 1 en k kK 1 n Q 1 C .1 C 2 /ˇ jdD1 .1 C 2.˛/j / 1 ˇ 2 .1 C 2 /2 Sn;d .z/ 1 d Y 1 C ˇ.1 C 2 / .1 C 2.˛/j /: ˇ 2 .1 C 2 /2 Sn;d .z/ j D1 On the other hand, from (C.1) we have ku e n 1 k k.K e un ksup kK e n /uksup K 1 d Y 1 C ˇ.1 C 2 / e .1 C 2.˛/j /kK ˇ 2 .1 C 2 /2 Sn;d .z/ 1 1 C ˇ.1 C 2 / ˇ 2 .1 C 2 /2 Sn;d .z/ j D1 d Y k k! a1 ru gkH kk1 kH ! C 2 krukH;1 kk2 kH;1 Sn;d .z/ e .1 C 2.˛/j / 1 kK j D1 e n /uksup K .1 C ˇ.1 C 2 // 1 ˇ 2 .1 1 Qd C j D1 .1 C 2.˛/j / .1 2 /2 Sn;d .z/ C 2 /ˇSn;d .z/: Acknowledgments. The authors express their gratitude to Professor Joe Stephen form University of Waikato for many helpful remarks and suggestions. During the preparation of the paper, we received helpful suggestions from Dr. Ali Abkar from Imam Khomeini International University. Therefore, the authors are very grateful to them for valuable comments. Bibliography [1] K. Atkinson and W. Han, Theoretical Numerical Analysis. A Functional Analysis Framework, 3nd ed., Springer-Verlag, New York, 2009. [2] J. Conway, A Course in Functional Analysis, 2nd ed., Springer-Verlag, New York, 1990. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 Worst case error for integro-differential equations 329 [3] A. Debbouche, Fractional evolution integro-differential systems with nonlocal conditions, Adv. Dyn. Syst. Appl. 5 (2010), 49–60. [4] J. Dick, P. Kritzer, F. Kuo and I. H. Sloan, Lattice-Nyström method for Fredholm integral equations of the second kind, J. Complexity 23 (2007), 752–772. [5] J. Dick and F. Pillichshammer, Digital Nets and Sequences. Discrepancy Theory and Quasi-Monte Carlo Integration, Cambridge University Press, Cambridge, 2010. [6] F. J. Hickernell, H. S. Hong, P. Ecuyer and C. Lemieux, Extensible lattice sequences for quasi-Monte Carlo quadrature, SIAM J. Sci. Comput. 22 (2001), 1117–1138. [7] F. J. Hickernell and H. Niederreiter, The existence of good extensible rank-1 lattices, J. Complexity 19 (2003), no. 3, 286–300. [8] M. Hill and I. Robinson, Quadrature using 64-bit IEEE arithmetic for integrands over Œ0; 1 with a singularity at 1, Theoret. Comput. Sci. 351 (2006), no. 1, 82–100. [9] Y. Lai, Intermediate rank lattice rules and applications to finance, Appl. Numer. Math. 59 (2009), 1–20. [10] N. M. Korobov, Properties and calculation of optimal coefficients, Dokl. Akad. Nauk SSSR 132 (1959), 1009–1012; translation in Sov. Math. Dokl. 1 (1960), 696–700. [11] W. McLean, I. H. Sloan and V. Thomée, Time discretization via Laplace transform of an integro-differential equation of parabolic type, Numer. Math. 102 (2006), 497–522. [12] W. McLean and V. Thomée, Time discretization of an evolution equation via Laplace transforms, IMA J. Numer. Anal. 24 (2004), 439–463. [13] E. Novak, Quantum complexity of integration, J. Complexity 17 (2001), 2–16. [14] E. Novak and H. Wózniakowski, Complexity of linear problems with a fixed output basis, J. Complexity 16 (2000), 333–362. [15] E. Novak and H. Wózniakowski, Intractability results for integration and discrepancy, J. Complexity 17 (2001), 388–441. [16] D. Nuyens and R. Cools, Fast algorithms for component-by-component construction of rank-1 lattice rules in shift-invariant reproducing kernel Hilbert spaces, Math. Comp. 75 (2006), 903–920. [17] A. Pedas and E. Tamme, A discrete collocation method for Fredholm integro-differential equations with weakly singular kernels, Appl. Numer. Math. 61 (2011), 738–751. [18] A. Pedas and E. Tamme, Product integration for weakly singular integro-differential equations, Math. Model. Anal. 16 (2011), 153–172. [19] V. Sinescu, Construction of lattice rules for mulitple integration based on a weighted discrepancy, Ph.D. thesis, Univerity of Waikato, 2008. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28 330 D. Rostamy, M. Jabbari and M. Gadirian [20] I. H. Sloan and H. Wózniakowski, Tractability of multivariate integration for weighted Korobov classes, J. Complexity 17 (2001), 697–721. [21] I. H. Sloan and H. Wózniakowski, When are quasi-Monte Carlo algorithms efficient for high dimensional integrals?, J. Complexity 14 (1998), 1–33. [22] J. F. Traub and H. Wózniakowski, Path integration on a quantum computer, Quantum Inf. Process 1 (2002), no. 5, 365–388. [23] G. W. Wasilkowski and H. Wózniakowski, On the power of standard information for weighted approximation, Found. Comput. Math. 1 (2001), 417–434. [24] G. F. Webb, Abstract Volterra integrodifferential equations and a class of reactiondiffusion equations, in: Volterra Equations (Otaniemi 1978), Lecture Notes in Math. 737, Springer-Verlag, Berlin (1979), 295–303. [25] H. Wózniakowski, Tractability and strong tractability of linear multivariate problems, J. Complexity 10 (1994), 96–128. [26] H. Wózniakowski, Efficiency of quasi-Monte Carlo algorithms for high dimensional integrals, in: Monte Carlo and Quasi-Monte Carlo Methods, Springer Verlag, Berlin (1998), 114–136. Received December 19, 2012; accepted September 23, 2013. Author information Davoud Rostamy, Department of Mathematics, Imam Khomeini International University, Qazvin, Iran. E-mail: [email protected] Mohammad Jabbari, Department of Mathematics, Imam Khomeini International University, Qazvin, Iran. Mahshid Gadirian, Department of Mathematics, Imam Khomeini International University, Qazvin, Iran. Bereitgestellt von | De Gruyter / TCS Angemeldet | 212.87.45.97 Heruntergeladen am | 11.12.13 05:28
© Copyright 2026 Paperzz