5 JOURNAL OF COMPUTATIONAL AND APPLIED MATHEMATICS ELSEVIER Journal of Computational and Applied Mathematics 138 (2002) 219-241 ‘wow elsevier.com/locate/cam Runge-Kutta methods for numerical solution of stochastic differential equations A. Tocino®**', R, Ardanuy? "Departamento de Matemiticas, Universidad de Salamanca, Plaza de la Merced, 1, 37008-Salamanea, Spain “Departamento de Estadistica, Universidad de Salamanca, Plaza de los Caidos, 37008-Salamanca, Spain Received 2 September 2000; received in revised form 22 February 2001 Abstract The way to obtain deterministic Runge-Kutta methods from Taylor approximations is generalized for stochastic differ- cential equations, now by means of stochastic truncated expansions about a point for sufficiently smooth functions of an 116 process. A class of explicit Runge-Kutta schemes of second order in the weak sense for systems of stochastic differential equations with multiplicative noise is developed, Also two Runge-Kutta schemes of third order have been obtained for scalar equations with constant diffusion coefficients. Numerical examples that compare the proposed schemes to standard cones are presented. © 2002 Elsevier Science B.V. All rights reserved MSC: 65U05; 6OH10; 34F05 Keywords: Stochastic differential equations; Weak approximation; Runge-Kutta methods; Weak numerical schemes; Explicit schemes 1. Introduction Stochastic differential equations are becoming increasingly important due to its application for modelling stochastic phenomena in different fields, e.g. physies or economies (see [7,3]). Unfortu- nately, in many cases analytic solutions of these equations are not available and we are forced to use numerical methods to approximate them. Roughly speaking, there are two basic ways to achieve these approximations. When sample paths of the solutions need to be approximated, mean-square convergence is used and the methods so obtained are called strong. When we are only interested in the moments or other functionals of the solution, which implies a much weaker form of convergence * Corresponding author E-mail addresses: bacon@gugu.usal.es (A. Tocino), raa@gugu.usal.es (R. Ardanuy). ‘This author was partially supported by CICYT under the project BFM2000-1115, 0377-0427/02/S -see front matter © 2002 Elsevier Science B.V. All rights reserved. PI: $0377-0427(01 }00380-6220A. Tocino, R Ardanuy | Journal of Computational and Applied Mathematies 138 (2002) 219-241 than that needed for pathwise approximations, the methods are called weak. In the present paper we will only refer to weak methods; for this reason words like “scheme”, “order” and so on mean “weak scheme”, “weak order”. In previous works various mean-square and weak numerical methods have been derived; extended presentations on this subject are given in Milstein [12] and Kloeden and Platen [7]. Analogously with the deterministic case, stochastic Taylor schemes are obtained by truncating stochastic Taylor expansions. The practical difficulty of employing Taylor approximations is that they require to de- termine many derivatives. In this investigation we are interested in methods of Runge-Kutta (RK) type, ie. one step methods which avoid the use of derivatives. RK schemes in the strong sense have been proposed, for instance, by Chang [3], Hernandez and Spigler [4,5], Klauder and Petersen {6], Mauthner [8], McShane [9] and Riimelin [14]; see also Kloeden and Platen [7], Milstein [12] and the references in Saito and Mitsui [15]. On the other hand, Milstein [10-12], Talay [16] and Kloeden and Platen [7] present RK schemes in the weak sense. In this paper a class of explicit second order and two explicit third order RK schemes are de- veloped. The second order class contains a scheme proposed by Milstein [10]. In analogy with the ordinary ease, these RK schemes have been obtained by matching their truncated stochastic expan- sion about a point with the corresponding Taylor approximation, The required truncated expansions about a point for functions of the solution of a stochastic differential equation have been derived from It6-Taylor expansions in Tocino and Ardanuy [17]. Weak approximations In this paper we consider a filtered probability space (Q,¥.F;,P), an m-dimensional Wiener process {17;=(WW},...,/P)}i20 and a d-dimensional stochastic differential equation (sde) AX, =a X de + OOK IAM, to << ay where a=(a',...,a) denotes the d-dimensional drift vector and b=(b’/) the d x m-diffusion matrix. Let’s denote b/ = (bY,...,b4), 7 =1,...,m. The functions a = a(t,x) and b/ = b/(t,x) are assumed to be defined and measurable in [%, 7] x R% and to satisfy both Lipschitz and linear growth bound conditions in x. These assumptions ensure the existence of a unique solution of the sde (1) with the initial condition X;, = Xo if Xo is Fq-measurable (see [1]). We shall suppose that all of the initial moments E[|Xo|"] < 00, = 1,2,... exist; so, the moments of every X; will exist (see [1]). Let X,. denote the solution of (1) starting at time # € [fT] at x € RY Let Gp denote the space of functions /(1,x) defined in [%, 7] x R4 which have polynomial growth (with respect to x) and let f the subspace of functions f € %p for which all partial derivatives up to order f= 1,2,... belong to @p. Next, to Eq. (1) we consider the one-step approximation a(t Hh) =x + A(t,x,h,2), @) where A is some R-valued function and £ a random vector. We shall say that the one-step approx- imation =2,, converges weakly to X =X, with order +1 if there exists a function K(x) € 6A. Tocino, R- Ardanuy! Journal of Computational and Applied Mathematies 138 (2002) 219-241-221 such that ; ! e | [ce — x) — [Jr — 9] < xeon, jt ml ped, LS 1... 2p + @) where 2 denotes the ith component of the vector z. From (3) it’s obvious that the differences between the moments (from the first up to (2/ + 2)th inclusively) of the vector X and the corresponding moments of its approximation X have f +1 order of smallness in #. The number f +1 will be called the local order of the approximation, Let be given an equidistant discretization {t,/1,....tv} of the time interval [fo, 7] with step size A=(T—1)/N. From the one-step approximation (2) we construct the discrete approximation (also called scheme): X=, oo - 4) Xqst =Xn t+ Ata, Xn, 4, En), 2 =O, We shall say that the discrete approximation J Xx} (based on a step size A) converges weakly to X with order B if for cach g € 62!” there exists a constant K, > 0 (not depending on A) such that {aX y) — (Xr I] < Ky". The number f in the above definition is the order of the scheme on an interval, The following theorem (see [11]) establishes the relation between the order of a one-step approximation and the order of the scheme generated by such approximation: Theorem 1. Suppose that the coefficients of Eq. (1) are continuous, and a linear growth bound and belong to 62!'*. Suppose that the o order +1 and for each 0 0 and r € NV such that LE{G(K s(t + 4) — (Xialt + A))I| < KC + |x? yHe* OO) and hence it has order ff +1. On the other hand, if the functions L*f, where (t,x) =x and a € Ip, grow at most linearly with respect to x then one can see that the one-step approximation (8) verifies conditions (5). We can conelude, by Theorem 1, that under the above conditions the scheme generated by (8) Kn =F t VON Van qo) ae converges weakly with order B. It’s called the order B weak Taylor scheme. The order 2 Taylor scheme was first proposed by Milstein [10]. Talay [16] proved it. Milstein [11] also proposed the third order Taylor scheme for systems with additive noises. Weak Taylor schemes of any order were constructed by Platen; see Kloeden and Platen [7] We shall say that two one-step approximations X,, and X,, are Brequivalent, X,.(t-+h)& 1), if there exists a function K(x) € Gp such that alt 1 : at Te’ -)- TI - it i IE 0 verifying 1 1 [: (Ih. - Ip. for all choices of multi-indices a € Fy — {v} with k= 1,. : =m=1 (scalar case) and Aiv, is any variable satisfying the moment conditions < Kahl (12) If B=2, [E[AW,]| + ELAM.) ] ~ Al + BAW.) + (EU AW, )*] —34"| + |E(AW,)°I] < Ka* (3) for some constant K > 0, it’s easy to prove that the variables /o,n=No).. =A, L(o,0,n=Ko.0,n=A°/2, Fajyn = AW Fo,1\n = Faoyn = $4 AW us Vaytyn = $(AW,Y — A) satisfy (12). Then we have the simplified order 2 Taylor scheme Fy = 2, + bAW, +04 + dbbo (AW) ~ A) + Abia + abo + $b*boz + ban AAW y + Haro + aan + $b7a02)4?, (14) where for a function g = g(t,x) with tx € R we have denoted og hij = a 15. 90 = Faqs" ) (5) and g = goo = 4(nXn). Notice that a Gaussian random variable AW, ~ N(0.4) of a three-point distributed random variable AW, with PAW, = V3.4) = P(AW, = —V34) =}, P(AW, =0) =? satisfy the moment conditions (13). In the multi-dimensional case, the kth component of the simplified order 2 Taylor scheme is given by Hey wiawiadasl> jet) 7 8h Shani nas 5S (Soe wis + Vin) ja im ict a ani eH), +300 away! | 440 “ee ry (16) i where AWW’, Vins f= Pin = A) .. AW" are independent random variables satisfying the moment conditions (13) and ysssamt, are independent two-point distributed random variables with =PVijn =—4) iff i224 A, Tocino, R Ardanuy | Journal of Computational and Applied Mathemuties 138 (2002) 219-241 If B=3, d=m=1 and the diffusion coefficient b is a constant, from the third order weak Taylor scheme one can construct the simplified order 3 weak Taylor scheme, given by Xu =F, + BAW, + aA + bag A2, + Hai + aao, + 6 a92)A? + k(baj, + 2bay, + 2abagy + Bays )LAW, + Lb aA(AWy) + haroao A> + Ladd, + 30?ag\ay2 + ay9 + 2aayy + Bayy + ab? aos + a?aq2 + Lb*ag4) 4°, (8) where Al’, and AZ, are correlated Gaussian random variables with AW, ~N(0,4), AZ, ~N(0,4°/3), E[AW,AZ,] = 47/2. (19) In the sequel, for simplicity of notation we shall often abbreviate /,,,, AW,, AZ,, etc. to ,, AW and AZ, respectively. Obviously, we can obtain f-equivalent schemes to the order f scheme given in (11) by replacing the /,’s by new variables [,’s satisfying ian for some constant K > 0. For example, the new family can contain all the variables of the old one except one of them, say /, replaced by /; if (20) holds (in this case it reduces to compare the < KAM ag eT y— {0}, I 2p +1 (20) products which contain / with the corresponding with Z) we shall write 27, In general, we shall say that / and 7 are p-equivalent, in symbols /7, if by replacing in an approximation the variable 7 by 7 the new approximation is f-equivalent to the old one. For example, in the scalar case, if AW’ is as in (13) and the function A(t,x,4, AH’) of the one-step approximation (2) is a linear combination of products of the form AAW), i,j =0,1,... (note that the simplified Taylor approximation (14) belongs to this class), by (20) it’s clear that (any 234aW, A(AWY (anys, Each variable 4\(AJV)! with mean-square order 5/2 (ie. i+ j/2= 5) is 2-equivalent to zero because it has zero mean and its product with every variable of the family has at least mean-square order 3. Obviously, the variables 4(AIW)/ with mean-square order 3,3,4,... are 2-equivalent to zero: 22) Qt) AM(AWYZ0 if i442 > 4S In the multi-dimensional case, when Al',..., Al?” are independent random variables satisfy- ing the moment conditions (13) and the function A(/,x, 4, AW’) is a linear combination of prod- ucts of the form 4(AN')/---(A")/= and variables VY, satisfying (17), we have the following 2-equivalences: ai, gia fa figs, saw’ aw! (23) 0 itisA. Tocino, R- Ardanuy! Journal of Computational and Applied Mathematics 138 (2002) 219-241 228 34" iti k tpi a gt ., AW AW AW? } gow! if j=kK AK (24) 0 fix RiFK IAL BAW YAW "20 ivi + AA 5 ;. (23) Similarly, in the scalar case, if AW and AZ verify (19) and the function A(z,x, 4, AW) which defines the approximation is a linear combination of products AAW)’, i,j =0,1,... and AZ, we have the following 3-equivalences: 23) 93 AAWPR3L AW: AAWY AAW 20 if 49/2 > (26) 3. Runge-Kutta schemes An important disadvantage of simplified Taylor schemes is that they require to determine many derivatives. Using the idea of the deterministic case we shall obtain RK schemes by replacing the derivatives in simplified Taylor schemes by new evaluations of the coefficients of the equation. An explicit s-stage stochastic Runge-Kutta scheme will be given by Kyat = Ky + AY aly + Amy) + PAWS PO + yA) +R, @7) ml i= ja where ju) =0, m1 =Xns = i; m i ioe ry =X AY hyalina + HeAsm) + SOAW AS HM la + HAs). = Toons a ima at and R is a fit term, The numerical constants 2), i, j4,Ay. 74) and the term R must be chosen so that the approximation (27) is f-equivalent to the simplified order # Taylor scheme. Since the truncated expansion of order fi of a process is, under appropriate conditions, /f-equivalent to the process, it suffices to choose the parameters and R so that a f-cquivalent approximation to the truncated expansion of order fi of (27) is equal to the simplified order f Taylor scheme. Note that the fit term is free and then it’s obvious that for every family of parameters it can be chosen so that the required equality is fulfilled. But our goal is to avoid the use of derivatives in the scheme. Then, for efficiency, the number of derivatives in R must be notoriously smaller than in the Taylor scheme. If for a family of parameters we have the equality with R =0, the scheme does not involve any derivative; it’s then a Runge-Kutta scheme in strict sense. If R contains one or more derivatives we shall say that it’s a Runge-Kutta type scheme. For example, the second order scheme in Milstein [12, pp. 116-117], is a RK type scheme with s=4 which contains one derivative.226A, Toeino, R. Ardanuy | Journal of Computational and Applied Marhematies 138 (2002) 219-241 Generalizing Butcher arrays, the coefficients occurring in (27) can be displayed 1 Yat Peso thet ho Bla Be where the first matrix contains the coefficients corresponding to the deterministic part and each of the remaining ones contains the coefficients corresponding to the stochastic part with respect to a Wiener component. As in the deterministic case, in order to match the truncated expansion of (27) with the simplified Taylor scheme we need an expression of the order fi truncated expansion of a process (t-+4,X; + AX) in terms of 4 and AX =X;,.,—X;. This expression has been obtained by Tocino and Ardanuy [17] for f/=2 in the multi-dimensional case and for ff = 3 in the scalar case. So, the formula f+ Se "ates @ of, Saf SOF AM + AN)S fb SA OAR 4 4 4 2 1 ac!) ast of & A sigh td (5 i) awaves ta DC Sragiamant ( 2 igaet ME ijkT=t “fn ff +k (s+ +3 > a (28) where all of the functions are evaluated at (1,X;), expresses the 2-equivalence between the process and its second order truncated expansion, In the scalar case, d =m = 1, using notation (15), (28) reduces to P(t+ A,X, +X) foo + fivd + fu AX bf + (r» +B fiz + B dor fos + fue) a a + (fu Fyn) sare fo (29) Similarly, in the scalar case, if (t,x) = is a constant, the formula “aa - a He AK + 00) for find + fa 8X + (fro ~ ru) 4 Ae pyaar Axy + foo! » es + (toot 5 fa + FO Fit hee) SA. Tocino, R. Ardanuy! Journal of Computational and Applied Mathematics 138 (2002) 219-241-227 2 oe PAX + (@ +P fist he) : ; : : + (fe + FJ) SO + fo A. (30) 2 6 shows the 3-equivalence between the process and its third order truncated expansion, 4, Two-stage Runge-Kutta schemes In the scalar case, for s=2 the RK scheme (27) takes the form Ray =Xyt {ara + araty + A,m)}A + {Bib + Bablty + uA ME AW + Ry BL where 1 =X, + Aad + ybAW; for simplicity, in the above expression and from now on, when a function in a scheme is evaluated at (i,,,) we shall omit such point; notice that we have also abbreviated AW’, 10 AW, The coefficients a and b are supposed to belong to @f. By (29) and the equivalences (21) and (22) we shall obtain a 2-equivalent approximation to (31) and we shall match this approximation with the simplified order 2 Taylor scheme (14). By (29) we get alt + WA, Xy + Jad + AW) 2a + ayn + an(dad + ybAW) 1 bf 2 +5 (om + Bair + Bboiaos + es) we B EE ., + fan + yas UA(Aad + ybAW) tH a + za0(2ad + AW) and then, using (21) and (22), we have Aly + WA Xy + 24 + AW )SZ ag dy AAW + ad $ ayopt? + ag AA? + dagyb?y? A?. (32) Similarly, we get = bAW + bho (AWS + boeddAW + aby AAW + 3b*boy? AAT (ty + WAX, + Za + AW)AW $B(biy + 4B? bos syd? + abby dy A? (33)228 A, Tocino, R Ardanuy | Journal of Computational and Applied Mathematies 138 (2002) 219-241 From (32) and (33) we see that approximation (31) is 2-equivalent to Kye =H, + (Bi + BAW + (m +.o2)a + Prybbo (AWS 7B* bp: AAW + ory? + (a27a01b + Boubio + Brdabor + 51 + (xahaaq, + $427 ao2b? + Bopeyb(bi + 36*bos) + Brzyabbo2)A? +R. (4) Now, we compare the above approximation with the simplified order 2 Taylor approximation (14). Firstly, let’s suppose that 0b/Cx=k is a constant. Since in this case by; = 103 =0, (34) and (14) coincide if the constants verify H+o=1, Bit f=, and bb A. Notice that the first column contains the equations which appear in the deterministic case (b = 0) and the second one contains the analogues for the stochastic part of the scheme. The equations in the third column contain both deterministic and stochactic parameters. The above system has the unique solution 4.= = y= 1, a = 22 =f = fy = 1/2, which leads to the scheme nt MOAW, + 1b, + A,X, +04 + bAW, AW, Xn + 444 + fat, + A,X, +a4 + bAW,)A -5b (35) By construction, this scheme is 2-equivalent to the simplified order 2 Taylor scheme. On the other hand, since b is continuous and we have supposed that b/@x is a constant, we have that b is a Lipschitz function which grows at most linearly and therefore it’s immediate to show that if a also grows at most linearly then approximation (35) verifies conditions (5). By Theorem 1, we deduce the following, ox Theorem 2, Suppose that the coefficients of Eq. (1) belong to 6%, that the drift a verifies both Lipschitz and linear growth conditions in x and that ébjéx is a constant. Then the scheme given in (35) is of second order in the weak sense. Since Cb/dx shall be a known constant, at each step, in scheme (35) we have to evaluate the drift @ at two points, the diffusion coefficient b at two points, as well as generating the random. variable AWV,; remember that in the Taylor scheme we have to evaluate a, b, Cb/ét, Ca/dx, Ga/ct and @a/éx? at one point and to generate AW,. Notice also that this scheme coincides (in the analysed case) with the second order RK scheme proposed by Milstein [10]. The Butcher array of the scheme would be 1 rm 1 —Sbbo Alt 4]h 4 To match (34) with (14) in a more general case we can take the same parameters, but R must change because new partial derivatives could appear. For example, if ¢7/éx°=0 then (14) containsA, Tocino, R. Ardanuy! Journal of Computational and Applied Mathematics 138 (2002) 219-241 229 the same terms that in the above case but in (34) it appears the term fj17bby A? = Lbb.14*; so we must take R= —1b(bq, +5114). Note that now the scheme contains two derivatives. And it’s clear that in more general cases we can continue in this way (i.e. with the same parameters and by increasing the number of derivatives in R) to derive second order schemes of the form (31). Notice that we have also derived that only if @b/ax is a constant there exists a second order Runge-Kutta scheme in strict sense as (27) with s = A class of second order schemes In view of the above results we must take s > 2 in (27) in order to obtain for the general case a second order RK scheme that does not include most of the derivatives participating in (14). So, let us consider schemes of the form nt {aya t analty + HAN} + {Bib + Brb(tn + WA) + Bsbtn + HA DAW +R, (36) where Using (29), (21) and (22) analogously as in the above section, we shall obtain that the approximation (36) is 2-equivalent to X,, + (Bi + Bo + Bs AW + (cn + 22)ad + (oF + B33 Raat CN + nya, bAAW + (Ba + Bs)ubioAAW + (Boi + BxA)abp SAW Hr +1 BaF + BsF)b(bi1 + 4B7Bos)A? + (BaF + Bs Fabby? +R. (7) This approximation will be equal to the simplified second order Taylor approximation (14) if the parameters satisfy the nonlinear system Bit Bit Bs=1, (Bo + Bse= 3 rit p=} + BaF Pb AAW + cpa? + aghaag A? + LopPanb A? (38)230A. Tocino, Re Ardanuy | Journal of Computational and Applied Marhematies 138 (2002) 219-241 and if R= $bbp,((AI)? — A). The left column system has the unique solution m= 03 G9) 1 On substituting (39) on the right column of (38), it’s easy to see that the system obtained has the one-parameter solution (40) sb AX, +04 —LbAW,) AW, + F tnt bn bad ~ 5 a A ieee o 7 1 ch 7 +504 + salt, + A,X, + aA + bAW,)4 + =b— (AW, ~ A), ap Pease Vex which, by construction, is 2-equivalent to the order 2 Taylor scheme. On the other hand, it’s easy to show that it verifies (5) if a, b and ab/éx have at most linear growth, and so, by Theorem 1, we have proved the following, Theorem 3. Suppose that the coefficients a, b of Eq. (1) satisfy a Lipschitz condition and belong to ©. If a, b and dbjéx have at most linear growth then the RK schemes of (41) are of second order in the weak sense. At each step in a scheme of the family (41) we have to evaluate the drift a at two points, the diffusion coefficient 6 at three points and the function 2b/dx at one point, as well as generating a random variable AW’, satisfying (13). Notice that the second order RK scheme proposed by Milstein [10] belongs to the above class, 6. Multi-dimensional generalization The construction of a second order schemes family in the multi-dimensional case can be ac~ complished by the procedure of the previous section, Now we shall consider schemes whose &th component can be written as pt aa +aa'(ty + wd} + DOBBY + PLB, + HA, A) + LON (y+ WA, DAW! +R, (42) jaA. Tocino, R- Ardanuy! Journal of Computational and Applied Mathematies 138 (2002) 219-241-231 where nak, + ad + BAW +--+ yb", HK, + Fad + 7b AW 4 45,0", nda + 5,BAW! $0 + Fb". Here the constants and the fit term R must be chosen so that the above scheme is 2-equivalent to the simplified order 2 Taylor scheme (16) By (28) and the equivalences (23)-(25) we obtain that a (. + HAR, + ad + | 4 rt Qatar saw! eo a nat fat jt od Bibra Pavements “ i ijt and that wy (. $d yt Fad SP wa’) AW Lea! yor, aw aw! Gpnaa’ tt oF, 7 AW AW AW (44) .m; besides, a similar equivalence to (44) can be obtained for each b(t, + 4, AW? with 7 and j, replaced by / and 7, respectively. Using these last three equivalences we get the 2-equivalent to (42) scheme ok =F + OBL + Bh + BOY AW + (+ on ah ‘Xx, = My ™ eat = , ed ea yay + y Us ADS 1 aa BAW AW +D ic arb saw Ft232A, Toeino, Re Ardanuy | Journal of Computational and Applied Marhematies 138 (2002) 219-241 coh nd DEB + BGAN! + 2 + Bs 7 flit is iss Y irpie +7 i+ Sagi AW AW A int + mats (45) By (24) we have ™ 4 2, = (BLE, + BGA, Ds oe bb AW AAW oe Say +e es a 1 ita a He + BE Y Bel AAW hee +BY areal baal" 4 gayle aH Dea ft bib AAW’; + x Ds Ti) + BS with this comnee scheme (45) shall be equal to the simplified order 2 Taylor scheme (16) if the parameters verify the systems a +a =1, Bi + Bb + Bs(46) (47) m OE j R=3> y ar ) (AW AW + V) mm 4 aise 1 digg rae saw! *LdG Sls +0) Dae Aa Ai ~LNes 7, + 8s vib AAW’. a For each j = 1,..., m system (46) is analogous to (38); so, we can deduce that (46) has the m-parametric family of solutions .6;/3y, where a 1, 0 € {41}, i and io!) AAW’. (48)234 A, Tocino, R Ardanuy | Journal of Computational and Applied Mathematies 138 (2002) 219-241 of schemes (. +A.Xy +04 +7 oy vas’) This family of solutions leads to the clas eee : =x 5 4 oN rl Saf +138 z x, 1s AW" ra he (1+ ate oa — 3 e alh )han : : “ th foted (4248 ors Sovan’) ban where y #0, a1 =1, o2,...,0m € {£1}, R is given by (48) and the variables AW” and Vj; satisfy (13) and (17), respectively. They are 2-equivalent to the simplified order 2 Taylor scheme; with the same reasoning employed in the scalar case we can conclude that these schemes are of second order in the weak sense if a and belong to {, verify a Lipschitz condition and, together with the partial derivatives Gb" /éx!, G2 /ax!@x!, i, 1.k od, +m, have at most linear growth. In particular, if a2 Gm =1 we have the one-parameter class of schemes 7 {w + : BN, + A,X, + aA + ybAW) 2 i+ 1 paw) \ aw’ 3y + Hak +ah(ty + AX, +ad +bAW A+R, na (« + 4,Xy + ad ~ where » # 0 and & is given by (48) with ¢)=1, j= 7. Third order RK schemes By the same technique employed in previous sections we will obtain now two third order RK. schemes for scalar stochastic differential equations (d =m = 1) with constant diffusion coefficient (1.x) = b. In this case a three-stage RK scheme of the form (27) can be written as Kner Ky + {ara t+ a2a(ty + p24,ne) + 25aCty + 154,93)}4 + BLAW + R, (49) where np =Xy + Aaiad + pbAW, 1s =X y+ Asad + Asnalty + W2A.n2)A + abAW. Here we have to find out the parameters and R in such a way that (49) is 3-equivalent to the simplified order 3 Taylor scheme (18). And, as before, it suffices to choose them after replacing (49) by its third order truncated expansion.A. Tocino, R- Ardanuy! Journal of Computational and Applied Mathematics 138 (2002) 219-241 235 We begin by evaluating the third order truncated expansion of a(t, + 4t.4.12)4. By using (30) and equivalences (26) we get ACty + HoA,Xy + Jad + bAW) ALAA + aybyrd AW 2 2,1) a2 7 tayo fad? + aay Aa A? + yuh AAW ye tanbo Paw 1 + aanbin LAW + Saba air $+ aaypirdy A + ps 1 oY 3 (= 7 Fa) 13a. (50) + (air + dangd?) e narha? aia + zavab* ) a Now [et’s calculate the third order truncated expansion of a(t, +J434,13)4. For simplicity, we denote M =n — 24 = dyad + Agralty + y,m)A + SAT. By (26) we have that 4420248 a, and then, by using (30) we get Aly + 134, 5 7 BY ba nF MAL ad + ayoHy? + aqy4 M+ («» - a) a 2 eye 3 aetna e & ) ae (1), The third order truncated expansion of IM can be obtained from (50); and from its value, using (26), it’s easy to show the equivalences AM2y3bAAW + (sy + 232)08? + an bis PAW + aiofaAs2A* + aay Asada) + Faqnb?dap)34°, (52) M2nbPAW + (Vai +in)ad, (53) S 4-MP2 AWA AW! + aby (Asi + APA + (ai + ds Pa? AP + 2aoib*As272734°, (4) 2 MPLA (55) 4 MRR AAWY +3ab*y3(As1 + da2)4*. (56) Substituting (52)~(56) into (51) and using the obtained expression together with (50) we have that the scheme (49) shall be 3-equivalent to x, n+ BDAW + (04 +02 +93)a4 + ao b(az72 + #73) AAW + ayo(Hapl2 + a5H3) 4? + aang (adry + a3(A31 + 450)? + danrb*(a2v3 + a573) AAW) + badyasdsrr2>d AW236A. Tocino, R Ardanuy | Journal of Computational and Applied Mathematies 138 (2002) 219-241 + ban (apy + aps ys LAW + aamb(ardriyr + aa 23 + indy) PAW + Saosb*(a2y} + 973) 2 AW + Saya + Saosb? Yb" (anunry3 + 93 4073)4* + paaagab?(a4a17} + 9a(4ai + Asa QA? + fadordord aadsz72(V2 + 273)4* + 4 (axo — Pb 4ao4) (cap + 233) 49 + ary (crapr221 + 9303(Aa1 + 232))A? + 4a? aon( 42 + aay tha A324 + Re + as(Asi + Asa) )4? + aoraro%stt22s24? Schemes (18) and (57) coincide if the constants satisfy the system a + +%=1, afta + asps = 4, aha + 3(Aa1 + 432) = 3 any + 043 = 4, a2plrhay + aap Aai + dao +ay(ds tanh = 4h, asptaks ashavdaa = b, B applay2 + sas Em adaitr + 93431 + Aris = 92 +0)? = 273 + 9373 = 2 2 aapays + 931875 + aa(dai + Aaa) asdarn(r +213) = 4, aay} + oy} = 4, 1 b wav 3439 (57) (58) and if R= bay, {AZ — (2x2 + 2373 )AAW} + 4bPaq2 42. The equations on the left column are the same which appear in the deterministic case. As it’s known, see REA [13], they reduce to the system. a +a2+%=1, afta + Hshls = 4}, ad tad =, ashindsa = b, Ja = ba, Bai + haa = psy Using (59), the equations on the right column of (58) reduce to (59)A. Tocino, R. Ardanuy! Journal of Computational and Applied Mathematics 138 (2002) 219-241 237 opt} + a4} Ho + 2p =3 (60) and R = bay (AZ — }AAW) + j567a247. The ordinary case system (59) has two one-parameter and one two-parameter families of solutions; see [2]. It’s easy to see that none of the solutions of the one-parameter families (one of the families corresponds to the case jis =0, ji2=2/3; the other ‘one to the case 2 = ty = 2) can be a solution of (60). On the other hand, since we have in the two-parameter family that 1 Hs = =e oy = 2! 573M i 5 = TE, Ais, ta #0, Ws AO, He polis ~ te) #3(H ~ Ha) % by imposing on a solution of this family to verify (60) we get 6n3 — 1714 + 1Sus — 4 =0, the roots of this equation are yy = 1, py = 1/2 and py = 4/3. If pty =1 then po = ps and, as we have said, these solutions do not verify (60). Each of the other roots leads to a solution of system. (59)-(60). If 43 = we have the solution and if jy =} we have m=5=7 w=1, w=h in=t i=, Each solution defines a scheme 3-cquivalent to the simplified order 3 Taylor scheme. The first one is given by Xn =H, + DAW, + had + halt, + 24,¥,, + 2AW, + 2ad)A B 8 aS 3 2 7 +a (« + zeke + ban 4 3p + yal + 24,8, + 2b40, 2) 4 9 + bay, (az, — Laaw, 1 and? 61 a0 in 3 n) + yy and’, (61) and the second one by a 3 n+ DAW, — xd +a (« +54 Fn + soa, + 54) A 1 4,248 _8 SAB See Ieee +a (i+ 3a ka + 3bAWy — 30d + 4a (+ Geka t bata + ies) 4) 4 + baw (az ~ 54a.) + pian (62)238 A. Tocino, R Ardanuy | Journal of Computational and Applied Mathematies 138 (2002) 219-241 Their Butcher arrays are, respectively, and @a/a: Since it’s immediate to show that these two schemes verify conditions (5) if a, Ga grow at most linearly, we have the following. Theorem 4. Suppose that the diffusion coefficient b in the sde (1) is a constant. Suppose also that the drift coefficient a belongs to 6}, verifies a Lipschitz condition and, together with the partial derivatives ajax and ajax, have at most linear growth. Then the RK schemes (61) and (62), where AW,, AZ, are variables satisfying (19), have order 3 in the weak sense. Both schemes require at each step to evaluate @ at three points, da/dr at one point and 02a/dx? at one point and to generate the variables AW’,, AZ, verifying (19) (note that the simplified Taylor scheme also requires to evaluate aip, ity a0. dos. 12. dos) 8. Numerical results In this section, numerical results from the implementation of the second and third order schemes proposed in the paper are compared to those from the implementation of well-known schemes of the same order; we have also used in each example a lower order scheme to contrast with them. The two-stage order 2 Runge-Kutta method proposed in (35) will be denoted by RK2-2st, we will denote by RK2-3st the particular three-stage order 2 Runge-Kutta method of the family (41) with = 1/3; the order 3 Runge-Kutta method proposed in (61) will be denoted by RK3. In order to compare with them, we shall use the Euler method, the simplified order 2 Taylor method (14), denoted by Taylor2, a second order Runge-Kutta method proposed by Platen (see [7, p. 485]), denoted by RK2-PL, and the simplified order 3 Taylor scheme (18), As test problems we have taken linear and nonlinear one-dimensional stochastic differential equa- tions (d =m = 1) for which the exact solution X; in terms of the Wiener process is known; our aim is to simulate the known value E[g(Xr)], where we have chosen g(x) =x or g(x) =.7. For each example we have used NV = $000 simulations for step sizes A=2-',...,2-§ to compute the approximated value of the known expectation, The mean and the standard deviation of the errors for each considered scheme are summarized in Tables 1-4. Example 1. Consider the nonautonomous linear equation dX = +X) d+? dM, X=.A. Tocino, R. Ardanuy! Journal of Computational and Applied Mathematics 138 (2002) 219-241 Table | Euler Taylor2 RK2-2st 4 Error St dev. Error St dev. Enor St. dev. 2! 6.82499 1.08507 -2,39532 2.63769 2.41669 3.4005 rt 4.69419 1.98406 0.897562 3.43742 0.897463 3.68215 22 2.82498 2.72174 0.202145 3.73732 0.198247 3.81124 2+ 1.61232 3.21608 00783977 «3.82159 0,0783575 3.83989 2% 0.812689 «3.51058 ——0.0364222-—«3.84255 0.0364302 3.84737 2% — 0457826 ——«3.71088_——0.0177697 «3.88478 0.017559 «3.88594 Table 2 Euler RK2-PL, RK2-3st 4 Error St.dev. Error St.dev. Error St. dev. 2! 15,8098 2.46831 5.7853, 622192 5.78145 6.23213 24 103977 4.45399 2.02754 01811 2.02479 8.01739 2? -SOTIT8 6.41039 0.436488 8.96982 0.437138 8.96707 2+ 333311 7.68503 0.174519 —9.21052—0.174723 9.20959 24 177872852312 0914871 9.34292 -——0.0920636 9.34149 2 0.833435 «9.08408 049148 «9452737 ——«0.04872236—_—9, 52715 Table 3 Euler RK2-PL. RK2-3st 4 Error St. dev. Error St. dev. Enor St. dev. 714,328 62.3239 336.823 300.969 336.525 305.969 530.347 168.103 130.174 465.036 130,043 467.544 342.693 309.168 28.8318 536.84 28.9169 557.816 201517 435.735 9.94652 600.808 9.97494 601.18 108.836 538,064 2.8628 633,786 2.9216 633.761 58.333 585.088 2.28967 637.75 231216 637.75 Table 4 Taylor Taylor3 RK3 4 Error St. dev. Eror St dev. Enror St. dev. 2! 132851 0.129494 0.566487 0.142544 00925534 0.145892 2? 0459119 0.137325 0.0976038 0.141423 0.0142057 0.141976 2-* 0.124983 0.145034 0,00685451 0.14627 0.00540199 0.146353 2-* 00366927 0.143237 0,00331031 0.143519 0,00165378 0.14353 2° 0,00940354 0.143731 0,000523082 0.143833 0.000307867 0.143834 239240A. Tocino, R Ardanuy | Journal of Computational and Applied Mathematies 138 (2002) 219-241 We estimate the exact value E[Xr]=2e? —(1+T) at the point T=2. Since 2b/éx=0 we can apply the second order RK2-2st scheme. In this case this scheme coincides with RK2-3st. We compare them with the Taylor2 scheme; the Euler scheme is also included. The values are summarized in Table 1. The results obtained using the RK2-2st scheme are similar to those from the second order Taylor scheme; the advantage of RK2-2st is that it is derivative free. Example 2. Consider the nonlinear stochastic differential equation ax= (i * 6x? ‘) ae 4X23 am, X= with solution X, =(2¢+ 1+ W,/3)°. Since the equation is autonomous, we can use the second order RK-PL scheme to estimate the exact value E[X;] =28 and compare the obtained values with those from the second order RK2-3st scheme. The results are summarized in Table 2 Example 3. Consider the nonlinear equation given in Example 2 above. Here we approximate the value E[X?] = 869 + & using RK-PL and RK2-3st schemes; the results obtained are shown in Table 3. In Examples 2 and 3 we observed practically no differences between the proposed second order Runge-Kutta scheme RK2-3st and the second order one proposed by Platen. RK-PL is completely derivative free; in opposition, RK2-3st has only one derivative, is simpler and valid for the nonau- tonomous case Example 4. Consider the linear nonautonomous stochastic differential equation dX, = (1X, + 10r)de + bd, Xo = 10, with constant diffusion coefficient b = 0.1. The exact value E[X;]=20e'? — 10 was estimated using Taylor3 and the RK3 schemes, both of third order. To simulate the Gaussian variables AW, and AZ, satisfying (19) we have taken AW” = YAU; and AZ = $4°7(U, + J-U2), where Ui and Us are two independent Gaussian variables with distribution N(0, 1). Although the results, summarized in Table 4, are not extensive enough to justify general conclusions, they suggest that RK3 is more efficient than Taylor3. References [1] L. Amold, Stochastic Differential Equations, Wiley, New York, 1974, [2] UC. Butcher, The Numerical Analysis of Ordinary Differential Equations. Runge-Kutta and General Linear Methods, Wiley, Chichester, 1987. [3] CC. Chang, Numerical solution of stochastic differential equations with constant diffusion coelficients, Math, Comput, 49 (1987) 523-542, [4] D.B. Hemandez, R. Spigler, A-stability of Runge-Kutta methods for systems with additive noise, BIT 32 (1992) 620-633,A. Tocino, R- Ardanuy! Journal of Computational and Applied Mathematies 138 (2002) 219-241 241 [5] DB. Hemandez, R. Spigler, Convergence and stability of implicit Runge-Kutta methods for systems with multiplicative noise, BIT 33 (1993) 654-669, [6] 1.R. Klauder, W.P. Petersen, Numerical integration of multiplicative-noise stochastic differential equations, SIAM J, ‘Numer, Anal, 22 (1985) 1153-1166, [7] P.E, Kloeden, E, Platen, Numerical Solution of Stochastic Differential Equations, Springer, Berlin, 1992. [8] S. Mauthner, Step size control in the numerical solution of stochastic differential equations, J. Comput. Appl. Math, 100 (1998) 93-109. [9] EJ. Meshane, Stochastic Calculus and Stochastic Models, Academic Press, New York, 1974, [10] GN. Milstein, A method of second order accuracy integration of stochastic differential equations, Theory Probab, Appl. 23 (1978) 396-401 [11] GN. Milstein, Weak approximation of solutions of systems of stochastic differential equations, Theory Probab. Appl 30 (1985) 750-766. [12] GN. Milstein, Numerical Integration of Stochastic Differential Equations, translated version of 1988 Russian work, Kluwer, Dordrecht, 1995. [13] REA (Research, Education Association), Advanced Methods for Solving Differential Equations, New York, 1982. [14] W. Rimelin, Numerical treatment of stochastic differential equations, SIAM J. Numer. Anal. 19 (1982) 604-613, [15] Y. Saito, T. Mitsui, Stability analysis of numerical schemes for stochastic differential equations, SIAM J. Numer. Anal. 33 (1996) 2254-2267. [16] D. Talay, Efficient numerical schemes for the approximation of expectations of functionals of the solution of a SDE. and applications, in: H. Korezlioglu, G. Mazziotto, J. Szpirglas (Eds.), Lecture Notes in Control and Inform. Sci., Vol. 61, Springer, Berlin, 1984, pp. 294-313. [17] A. Tocino, R. Ardanuy, Truncated 1t8—Taylor expansions, Stochastic Anal. Appl., to appear.