JAM Journal of Applied Mathematics 1687-0042 1110-757X Hindawi Publishing Corporation 236281 10.1155/2012/236281 236281 Research Article A New Family of Phase-Fitted and Amplification-Fitted Runge-Kutta Type Methods for Oscillators Chen Zhaoxia 1 You Xiong 1, 2 Shu Xin 1, 3 Zhang Mei 1 Vigo-Aguiar Jesus 1 Department of Applied Mathematics Nanjing Agricultural University Nanjing 210095 China njau.edu.cn 2 State Key Laboratory for Novel Software Technology at Nanjing University Nanjing 210093 China nju.edu.cn 3 Department of Computer Science and Engineering Shanghai Jiaotong University Shanghai 200240 China sjtu.edu.cn 2012 23 10 2012 2012 19 04 2012 17 07 2012 23 07 2012 2012 Copyright © 2012 Zhaoxia Chen et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

In order to solve initial value problems of differential equations with oscillatory solutions, this paper improves traditional Runge-Kutta (RK) methods by introducing frequency-depending weights in the update. New practical RK integrators are obtained with the phase-fitting and amplification-fitting conditions and algebraic order conditions. Two of the new methods have updates that are also phase-fitted and amplification-fitted. The linear stability and phase properties of the new methods are examined. The results of numerical experiments on physical and biological problems show the robustness and competence of the new methods compared to some highly efficient integrators in the literature.

1. Introduction

A vast of problems in applied fields, such as elastics, mechanics, astrophysics, electronics, molecular dynamics, ecology, and biochemistry, can be cast into the form of an initial value problem of the system of first-order ordinary differential equations as follows: (1.1)y=f(x,y),y(x0)=y0, where yd,  f:×dd is a sufficiently smooth function. Traditionally, the problem (1.1) has been solved numerically by Runge-Kutta methods (RK) or linear multistep methods (LMMs) (see ). In many applications, the solution to the problem (1.1) is oscillatory or periodic. But the general-purpose methods do not take into account the oscillatory feature of the problem and the numerical results they produce are generally not satisfactory.

Recently, some authors have proposed to adapt traditional integrators to the oscillatory character of the solution to the problem (1.1) (see ). Bettis  constructs a three-stage method and a four-stage method which solve the equation y'=iωy    (i2=-1) without truncation error. Paternoster  develops a class of implicit methods of Runge-Kutta (RK) and Runge-Kutta-Nyström (RKN) types by the trigonometric fitting technique. For oscillatory problems which can be put in the form of a second-order equation y′′=f(x,y), Franco  improves the update of the classical RKN methods and proposes a family of explicit RKN methods adapted to perturbed oscillators (ARKN) and a class of explicit adapted RK methods in . Anastassi and Simos  construct a phase-fitted and amplification-fitted RK method of “almost” order five. Van de Vyver  investigates phase-fitted and amplification-fitted two step hybrid methods (FTSH). For other important work on frequency dependent integrators for general second-order oscillatory equations, the reader is referred to . Many authors focus on effective numerical integration of specific categories of oscillatory problems. For example, Vigo-Aguiar and Simos  construct an exponentially fitted and trigonometrically fitted method for orbital problems. The papers  have designed highly efficient integrators for the Schrödinger equation.

Based on the previous work, we consider, in this paper, phase-fitted and amplification-fitted RK type integrators whose coefficients in the update depend on the product of the fitting frequency and the step size. In Section 2, we present a result on order conditions for frequency-depending modified Runge-Kutta type methods. Section 3 introduces the notion of phase-fitted and amplification-fitted RK type methods (FRK) and derives the phase-fitting and amplification-fitting conditions. In Section 4 two FRK methods of order four and two FRK methods of order five are constructed. Their error coefficients and error constants are also calculated. The linear stability and phase properties of the new FRK methods are analyzed in Section 5. In Section 6, numerical experiments are carried out to illustrate the effectiveness and superiority of our new methods compared to two well-known highly efficient integrators we have chosen from the recent literature. Section 7 is devoted to conclusive comments.

2. Order Conditions for RK Type Methods with Frequency-Dependent Weights

Assume that the principal frequency of the problem (1.1) is known or can be accurately estimated in advance. This estimated frequency is denoted by ω, which is also called the fitting frequency. We consider the following s-stage modified Runge-Kutta method: (2.1)ki=f(x0+cih,y0+hj=1saijkj),i=1,,s,y1=y0+hi=1sbi(ν)ki, where h is the step size, ci,aij,i,j=1,,s are real numbers, and bi(ν),i=1,,s are even functions of ν=hω. The scheme (2.1) can be represented by the Butcher tableau as follows: (2.2)cAbT=c1a11a1scsas1assb1(ν)bs(ν), or simply by (A,b,c). Conventionally, we assume that (2.3)j=1saij=ci,i=1,,s.

As explained in , the order conditions for the modified RK method (2.1) can be derived by just considering the autonomous equation y'=f(y). Using Butcher's rooted tree theory in , the exact solution to the problem (1.1) and the numerical solution produced by the modified RK type method (2.1) have the B-series expressions as follows: (2.4)y(x0+h)=j=0hjj!ρ(t)=jα(t)(t)(y0),y1=j=0hjj!ρ(t)=jα(t)γ(t)bTΦ(t)(t)(y0), where the trees t, the functions ρ(t) (order), α(t) (number of equivalent trees), γ(t) (density), Φ(t)=(Φ1(t),,Φs(t))T (the s×1 vector of elementary weights), and the elementary differentials (t)(y0) are defined in . Then the local truncation error has the following series expansion: (2.5)LTE=j=0hjj!ρ(t)=jα(t)(1-γ(t)bTΦ(t))(t)(y0). If for any (p+1)th differentiable function f(y), when the scheme (2.1) is applied to the problem (1.1), the local truncation error LTE=y(x0+h)-y1=𝒪(hp+1), then the method (2.1) is said to have (algebraic) order p.

Theorem 2.1 (see Franco [<xref ref-type="bibr" rid="B11">11</xref>]).

The modified RK type method (2.1) has order p if and only if the following conditions are satisfied: (2.6)bTΦ(t)=1γ(t)+𝒪(νp-ρ(t)+1),ρ(t)=1,2,,p.

If the method (2.1) is of order p, then we have (2.7)LTE=hp+1(p+1)!ρ(t)=p+1α(t)(1-γ(t)bTΦ(t))(t)(y0)+𝒪(νp+2):=hp+1(p+1)!ρ(t)=p+1α(t)ϵ(t)(t)(y0)+𝒪(νp+2), where ϵ(t)=1-γ(t)bTΦ(t) is called the error coefficient associated with the tree t of order p+1 which is involved in the leading term of the local truncation error. Denote (2.8)ECp+1(ν)=(ρ(t)=p+1ϵ(t)2)1/2. The positive number (2.9)ECp+1=limν0ECp+1(ν) is called the error constant of the method.

3. Phase-Fitted and Amplification-Fitted Conditions

For oscillatory problems, as suggested by Paternoster  and Van der Houwen and Sommeijer , apart from the algebraic order, the analysis of phase lag and dissipation is important. To this end, we consider the following linear scalar equation: (3.1)y=iωy. The exact solution of this equation with the initial value y(x0)=y0 satisfies (3.2)y(x0+h)=R0(z)y0, where R0(z)=exp(z),  z=iν. This means that after a period of time h, the exact solution experiences a phase advance ν=hω and the amplification remains constant.

An application of the modified RK method (2.1) to (3.1) yields (3.3)y1=R(z)y0, where (3.4)R(z)=1+zbT(I-zA)-1e,z=iν,e=(1,,1)T. Thus, after a time step h, the numerical solution obtains a phase advance argR(z) and the amplification factor |R(iν)|. R(z) is called the stability function of the method (2.1). Denote the real and imaginary part of R(z) by U(ν) and V(ν), respectively. Then, for small h, we have (3.5)U(ν)=1-ν2(bTAe)+ν4(bTA3e)-ν6(bTA5e)+ν8(bTA7e)-,V(ν)=ν(bTe)-ν3(bTA2e)+ν5(bTA4e)-ν7(bTA6e)+. For small h, argR(z)=tan-1(V(ν)/U(ν)) and |R(z)|=U2(ν)+V2(ν). The above analysis leads to the following definition.

Definition 3.1 (see [<xref ref-type="bibr" rid="B22">22</xref>]).

The quantities (3.6)P(ν)=ν-argR(iν),D(ν)=1-|R(iν)| are called the phase lag (or dispersion) and the error of amplification factor (or dissipation) of the method, respectively. If (3.7)P(ν)=𝒪(νq+1),D(ν)=𝒪(νp+1), then the method is called dispersive of order q and dissipative of order p, respectively. If (3.8)P(ν)=0,D(ν)=0, the method is called phase-fitted (or zero-dispersive) and amplification-fitted (or zero-dissipative), respectively. A modified RK method which is phase-fitted and amplification-fitted is called in short an FRK method.

It is interesting to consider the phase properties of the update of the scheme (2.1). Suppose that the internal stages have been exact for the linear equation (3.1), that is, Yi=exp(iciν)y0, then the update gives (3.9)y1=Ru(z)y0, where (3.10)Ru(z)=1+zi=1sbi(ν)exp(ciz),z=iν. Denote the real and imaginary part of Ru(z) by Uu(ν) and Vu(ν), respectively. Then, for small h, (3.11)Uu(ν)=1-νi=1sbi(ν)sin(ciν),Vu(ν)=νi=1sbi(ν)cos(ciν).

Definition 3.2.

The quantities (3.12)Pu(ν)=ν-argRu(iν),Du(ν)=1-|Ru(iν)| are called the phase lag (or dispersion) and the error of amplification factor (or dissipation) of the update of the method, respectively. If (3.13)Pu(ν)=𝒪(νq+1),Du(ν)=𝒪(νp+1), then the update of the method is called dispersive of order q and dissipative of order p, respectively. If (3.14)Pu(ν)=0,Du(ν)=0, the update of the method is called phase-fitted (or zero-dispersive) and amplification-fitted (or zero-dissipative) in the update, respectively.

Generally speaking, a traditional RK method with constant coefficients inevitably carries a nonzero phase lag and a nonzero error of amplification factor when applied to the linear oscillatory equation (3.1). Therefore, they are neither phase-fitted nor amplification-fitted. So does the update. For example, the classical RK method of order four with constant coefficients (denoted as RK4, see ) given by (3.15)01/21/21/21/2111/62/62/61/6 has a phase lag and an error of amplification factor as follows: (3.16)P(ν)=1120ν5+𝒪(ν7),D(ν)=1144ν6+𝒪(ν8). For the update, (3.17)Pu(ν)=-12880ν5+𝒪(ν7),Du(ν)=-15760ν6+𝒪(ν8). Therefore, both the method and its update are dispersive of order four and dissipative of order five.

The following theorem gives the necessary and sufficient conditions for a modified RK method and its update to be phase-fitted and amplification-fitted, respectively.

Theorem 3.3.

(i) The method (2.1) is phase-fitted and amplification-fitted if and only if (3.18)U(ν)=cos(ν),V(ν)=sin(ν).

(ii) The update of the method (2.1) is phase-fitted and amplification-fitted if and only if (3.19)Uu(ν)=cos(ν),Vu(ν)=sin(ν).

The proof of this theorem is immediate.

4. Construction of New Methods

Now we proceed to construct modified RK type methods that are both phase-fitted and amplification-fitted based on the internal coefficients of two classical RK methods. For convenience we restrict ourselves to explicit methods.

4.1. Fourth-Order Methods

Consider a four-stage modified RK method with the following Butcher tableau: (4.1)01/21/21/21/211b1b2b3b4. For s=4, the phase-fitting and amplification-fitting conditions (3.18) have the following form: (4.2)1-(12b2(ν)+12b3(ν)+b4(ν))ν2+14b4(ν)ν4=cos(ν),(b1(ν)+b2(ν)+b3(ν)+b4(ν))ν-(14b3(ν)+12b4(ν))ν3=sin(ν). On the other hand, the first-order and second-order conditions become (4.3)bTe=b1(ν)+b2(ν)+b3(ν)+b4(ν)=1,bTc=12b2(ν)+12b3(ν)+b4(ν)=12. Here, for the purpose of deriving the weights of the method, the higher order terms in the order conditions are omitted. We will follow this convention in the sequel. Solving (4.2) and (4.3), we obtain (4.4)b1(ν)=2(-2+ν2+2cos(ν))ν4,b2(ν)=ν3-4ν+4sin(ν)ν3,b3(ν)=-4(-2+2cos(ν)+νsin(ν))ν4,b4(ν)=b1(ν). As ν0, bi(ν) have the following Taylor expansions: (4.5)b1(ν)=16-1180ν2+110080ν4-1907200ν6+1119750400ν8+,b2(ν)=13+130ν2-11260ν4+1907200ν6-19979200ν8+,b3(ν)=13-145ν2+11680ν4-1134400ν6+111975040ν8+,b4(ν)=16-1180ν2+110080ν4-1907200ν6+1119750400ν8+. By direct calculation, we obtain the following expansions: (4.6)bTc2=-8+4ν2+ν4+8cos(ν)4ν4=13+𝒪(ν2),bTAc=ν-sin(ν)ν3=16+𝒪(ν2),bTc3=-24+12ν2+ν4+24cos(ν)8ν4=14+𝒪(ν),bT(c·Ac)=-2+2ν2+2cos(ν)-νsin(ν)2ν4=18+𝒪(ν),bT(Ac2)=ν-sin(ν)2ν3=112+𝒪(ν),bTA2c=-2+ν2+2cos(ν)2ν4=124+𝒪(ν), where the dot “·” between two vectors indicates componentwise product, and c2 and c3 indicate the componentwise square of the vector. We will follow the convention in the sequel. Thus, the coefficients given by (4.1) and (4.4) satisfy all the conditions of order four in Theorem 2.1. But they cannot satisfy all the conditions of order five. For instance, (4.7)bTc4=524-71440ν2+111520ν4-11036800ν615+𝒪(ν). Therefore, this method is of order four. The method was originally obtained by Simos  and we denote it by Simos4.

Corresponding to the nine fifth-order rooted trees t5j,j=1,,9, the error coefficients of Simos4 are given by (4.8)e(t51)=280-140ν2+11ν4-280cos(ν)16ν4,e(t52)=30-20ν2+2ν4-30cos(ν)+5νsin(ν)2ν4,e(t53)=30-30ν2+4ν4-30cos(ν)+15νsin(ν)4ν4,e(t54)=1-15(-2+ν2+2cos(ν))ν4,e(t55)=10-10ν2+ν4-10cos(ν)+5νsin(ν)ν4,e(t56)=-5ν+ν3+5sin(ν)ν3,e(t57)=1-10(-2+ν2+2cos(ν))ν4,e(t58)=1-15(-2+ν2+2cos(ν))ν4,e(t59)=1. Then for Simos4, (4.9)EC5(ν)=K16ν4, where (4.10)K=739200-798400ν2+296160ν4-40200ν6+2169ν8+739200cos2(ν)+320ν(185-205ν2+26ν4)sin(ν)+18000ν2sin2(ν)-80cos(ν)(18480-9980ν2+797ν4+740νsin(ν)).

Now we require that the update of the method (4.1) is phase-fitted and amplification-fitted. By Theorem 3.3 (ii), we have (4.11)b1(ν)+b2(ν)cos(12ν)+b3(ν)cos(12ν)+b4(ν)cos(ν)=sin(ν)ν,b2(ν)sin(12ν)+b3(ν)sin(12ν)+b4(ν)sin(ν)=1-cos(ν)ν. Solving (4.2) and (4.11), we obtain (4.12)b1(ν)=4(ν-2sin(ν/2))sin(ν/2)ν2(-4+ν2+4cos(ν/2)),b2(ν)=Lν4(-4+ν2+4cos(ν/2)),b3(ν)=-8(νcos(ν/2)-2sin(ν/2))sin(ν/2)ν4,b4(ν)=b1(ν), where (4.13)L=2sin(ν2)(8ν-4ν3+ν5+4ν(-4+ν2)cos(ν2)+8νcos(ν)+32sin(ν2)-8ν2sin(ν2)-16sin(ν)+4ν2sin(ν)). It can be verified that the method given by (4.1) and (4.12) is of order four. We denote the method by FRK4.

The error coefficients of FRK4 are given by (4.14)e(t51)=W8ν2Q,e(t52)=1+5sin(ν/2)(ν(-4+ν2)cos(ν/2)  +2(ν-2ν3+νcos(ν)+(4+3ν2)sin(ν/2)-2sin(ν)))/(ν4Q),e(t53)=1-15sin(ν/2)(-νcos(ν/2)Q+2(ν3-(4+ν2)sin(ν/2)+2sin(ν)))2ν4Q,e(t54)=1-30(ν-2sin(ν/2))sin(ν/2)ν2Q,e(t55)=1+10sin(ν/2)(ν(-4+ν2)cos(ν/2)  +2(ν-ν3+νcos(ν)+(4+ν2)sin(ν/2)-2sin(ν)))/(ν4Q),e(t56)=1-40cos(ν/4)sin3(ν/4)(ν3+4νcos(ν/2)-8sin(ν/2))ν4Q,e(t57)=1-20(ν-2sin(ν/2))sin(ν/2)ν2Q,e(t58)=1-30(ν-2sin(ν/2))sin(ν/2)ν2Q,e(t59)=1, where (4.15)W=160-32ν2+8ν4+2(-5+16ν2)cos(ν2)-160cos(ν)+10cos(3ν2)-140νsin(ν2)-5ν3sin(ν2),Q=-4+ν2+4cos(ν2). Then for FRK4, we have (4.16)EC5(ν)=288148-6972882881ν2+129013768711151544322881ν4-132896821756033855311902310402881ν6+𝒪(ν8).

It can be seen that as ν0, both the methods Simos4 and FRK4 reduce to a classical RK method of order four (on page 138 of ), which we denote by RK4. The method RK4 is called the prototype method or the limit method of the methods Simos4 and FRK4. Moreover, as ν0, both Simos4 and FRK4 have the same error constant as that of RK4: EC5=2881/48.

4.2. Fifth-Order Methods

Consider the following modified RK type method with FSAL property, the prototype of which can be found on page 167 of  or on page 178 of , (4.17)01/51/53/103/409/404/544/45-56/1532/98/919372/6561-25360/218764448/6561-212/72919017/3168-355/3346732/524749/176-5103/186561b1b2b3b4b5b60b1b2b3b4b5b60.

For the method (4.17), the phase-fitting and amplification-fitting conditions (3.18) become (4.18)1-(15b2(ν)+310b3(ν)+45b4(ν)+89b5(ν)+b6(ν))ν2+(425b4(ν)+4241215b5(ν)+2155b6(ν))ν4-7550b6(ν)ν6=cos(ν),(b1(ν)+b2(ν)+b3(ν)+b4(ν)+b5(ν)+b6(ν))ν-(9200b3(ν)+825b4(ν)+3281b5(ν)+12b6(ν))ν3-(84818225b5(ν)+14275b6(ν))ν5=sin(ν). Combining these two equations and the following sufficient conditions in Theorem 2.1 corresponding the trees of orders one, two, and three, we have (4.19)bTe=b1(ν)+b2(ν)+b3(ν)+b4(ν)+b5(ν)+b6(ν)=1,bTc=15b2(ν)+310b3(ν)+45b4(ν)+89b5(ν)+b6(ν)=12,bTc2=125b2(ν)+9100b3(ν)+1625b4(ν)+6481b5(ν)+b6(ν)=13,bTAc=9200b3(ν)+825b4(ν)+3281b5(ν)+12b6(ν)=16. Solving (4.18) and (4.19), we obtain (4.20)b1(ν)=28ν7-235ν5+28800sin(ν)-36600ν+7350ν3+7800νcos(ν)+1350ν2sin(ν)288(4+ν2)ν5,b2(ν)=0,b3(ν)=4(3550ν5+371ν7-186750sin(ν)+236400ν-46500ν3-49650νcos(ν)-9450ν2sin(ν))/(3339(4+ν2)ν5),b4(ν)=225ν5+22ν7+9000sin(ν)-10200ν+750ν3+1200νcos(ν)+1350ν2sin(ν)48(4+ν2)ν5,b5(ν)=-243(1800ν-1200sin(ν)-650ν3-600νcos(ν)+69ν5+150ν2sin(ν))1696(4+ν2)ν5,b6(ν)=11(600ν-450sin(ν)-150ν3-150νcos(ν)+11ν5)21(4+ν2)ν5. As ν0, we obtain the following expansions: (4.21)b1(ν)=35384-14948384ν2+8991161216ν4-211110948608ν6+523791086898176ν8+,b2(ν)=0,b3(ν)=5001113+977140238ν2-1433841428ν4+94067222136992ν6-3334133150306432ν8+,b4(ν)=125192-618064ν2+151193536ν4-228112773376ν6+8059181149696ν8+,b5(ν)=-21876784-24394976ν2+1791759808ν4-1021516715776ν6+1208779020032ν8+,b6(ν)=1184+111764ν2-18784672ν4+71127008ν6-13859906624ν8+.

By simple computation, we have (4.22)bTc3=ν(4800-1000ν2+200ν4+37ν6)-1050νcos(ν)-150(25+ν2)sin(ν)150ν5(4+ν2)=14+𝒪(ν2),bT(c·Ac)=12bTc3=18+𝒪(ν2),bTAc2=13bTc3=112+𝒪(ν2),bTA2c=-600+750ν2-75ν4+11ν6+600cos(ν)-450νsin(ν)150ν4(4+ν2)=124+𝒪(ν2),bTc4=ν(925200-195500ν2+20850ν4+2583ν6)-204450νcos(ν)-150(4805+179ν2)sin(ν)13500ν5(4+ν2)=15+𝒪(ν),bT(c2·Ac)=12bTc4=110+𝒪(ν),bT(c·(Ac2))=13bTc4=115+𝒪(ν),bT(c·A2c)=-2(-ν(7200+500ν2-150ν4+33ν6)+75νcos(ν)+75(95+23ν2)sin(ν))1125ν5(4+ν2)=130+𝒪(ν),bT(Ac·Ac)=14bTc4=120+𝒪(ν),bTAc3=ν(3000-1150ν2+155ν4+7ν6)-1050νcos(ν)+150(-13+2ν2)sin(ν)225ν5(4+ν2)=120+𝒪(ν),bTA(c·Ac)=ν(7200-2000ν2+300ν4+33ν6)-1950νcos(ν)+150(-35+ν2)sin(ν)1500ν5(4+ν2)=140+𝒪(ν),bTA2c2=23bTA(c·Ac)=160+𝒪(ν),bTA3c=-6ν+ν3+6sin(ν)6ν5=1120+𝒪(ν).

By Theorem 2.1, the coefficients given by (4.17) and (4.20) satisfy all the conditions of order five. But it cannot satisfy all the conditions of order six. For example, (4.23)bTc5=8995400+228839720000ν2+𝒪(ν4)16+𝒪(ν). Therefore, this method is of order five and we denote it as FRK5a.

Corresponding to the twenty sixth-order rooted trees t6j,j=1,,20, the error coefficients of FRK5a are given by (4.24)e(t61)=(ν(-124203600+26615500ν2-1403850ν4+18711ν6)+27724350νcos(ν)+150(643195+22177ν2)sin(ν))/(202500ν5(4+ν2)),e(t62)=e(t63)=e(t65)=e(t6,10)=e(t61),e(t64)=(-5493600ν+551000ν3+31800ν5-3879ν7+756600νcos(ν)  +600(7895+1028ν2)sin(ν))/(5625ν5(4+ν2)),e(t66)=(ν(-6667200+2224000ν2-194700ν4+6957ν6)+2084700νcos(ν)  -300(-15275+1393ν2)sin(ν))/(16875ν5(4+ν2)),e(t67)=(ν(-2296800+563000ν2-36600ν4+873ν6)+565800νcos(ν)  +600(2885+14ν2)sin(ν))/(5625ν5(4+ν2)),e(t68)=e(t67),e(t69)=(ν(57600-1600ν2-948ν4+75ν6)-4800νcos(ν)  -4800(11+2ν2)sin(ν))/(75ν5(4+ν2))e(t6,11)=e(t64),e(t6,12)=(ν(-652800+333000ν2-36400ν4+1533ν6)+290550νcos(ν)  -450(-805+283ν2)sin(ν))/(2250ν5(4+ν2)),e(t6,13)=(ν(-602400+300000ν2-32500ν4+1359ν6)+262650νcos(ν)  -450(-755+249ν2)sin(ν))/(2250ν5(4+ν2)),e(t6,14)=e(t6,16)=e(t6,13),e(t6,15)=ν(5280-360ν2-56ν4+5ν6)-600νcos(ν)-360(13+2ν2)sin(ν)5ν5(4+ν2),e(t6,17)=(ν(-352800+101000ν2-7800ν4+243ν6)+97800νcos(ν)  -600(-425+16ν2)sin(ν))/(375ν5(4+ν2)),e(t6,18)=ν(-3360+1320ν2-128ν4+5ν6)+1200νcos(ν)-360(-6+ν2)sin(ν)5ν5(4+ν2),e(t6,19)=e(t6,18),e(t6,20)=1+24(-600ν+150ν3-11ν5+150νcos(ν)+450sin(ν))5ν5(4+ν2) Then for FRK5a, we have (4.25)EC6(ν)=N202500ν5(4+ν2), where (4.26)N=200598105893906250+704003855089083750ν2-297809752687323750ν4+56298277495370000ν6-5785233488749200ν8+376641824701500ν10-15770219564700ν12+335861285301ν14+300ν2(-1028483389544400+258042555395500ν2-17441989606650ν4+394861660839ν6)cos(ν)-11250(17830942746125-509988761071ν2+154551250229ν4)cos(2ν)-1033593801921000000νsin(ν)+195145999800990000ν3sin(ν)-14464008183675000ν5sin(ν)+1060750694812500ν7sin(ν)-58620919851900ν9sin(ν)+115600689172687500νsin(2ν)-103692151057500ν3sin(2ν).

Now we require that the update of the method (4.17) is phase-fitted and amplification-fitted. By Theorem 3.3 (ii), we have (4.27)b1(ν)+b2(ν)cos(15ν)+b3(ν)cos(310ν)+b4(ν)cos(45ν)+cos(89ν)+cos(ν)=sin(ν)ν,b2(ν)sin(15ν)+b3(ν)sin(310ν)+b4(ν)sin(45ν)+sin(89ν)+sin(ν)=1-cos(ν)ν. We can solve the algebraic systems (4.18), (4.27), and the last two equations in (4.19) for bi(ν). The closed expressions for bi(ν) is too complicated. As ν0, they have the series expressions as follows: (4.28)b1(ν)=35384-2043758311904122880ν2-119367195208354081968707998823219200000ν4+950118849645188449321459792252269346255956397260800000000ν6-2870157856674050512150034054708268901403216681907224475027036857958400000000000ν8+,b2(v)=0,b3(ν)=5001113+339645238625839040ν2-7661109532360639140386979516958720000ν4-22790810005346940380832667326403096664437431009280000000ν6+17288060849199373243285599701968665711686983514654709393361771429888000000000ν8+,b4(ν)=125192-108309311984020480ν2+46882369940954332290266627440640000ν4+173699043980152150475225975075644875198546575360000000ν6-111781706153729849951569016265805372688111212714829833513579053056000000000ν8+,b5(ν)=-21876784+12242792596372480ν2+1875321113915112123475776834764800000ν4-195812790051300151076399318193914853157398118400000000ν6+1267931566087942517306241455372326619771597686823455499054848409600000000000ν8+,b6(ν)=1184+2401663868008960ν2-4489254073870875170634958247526400000ν4+12565766033461779855286229164227973164496820633600000000ν6-238480338530165061304454610608339212672837853551677391277706444800000000000ν8+.

It can be verified that the method given by (4.17) and (4.28) is of order five. We denote the method by FRK5b.

The error coefficients of FRK5b are given by (4.29)e(t61)=1900-156847283272092521600000ν2+76682205913855761116811228116725760000000ν4-383259943742292563771548045126393781401436989030400000000000ν6+370402254931781919511694075547402964972025083781900122921452752076800000000000000ν8+,e(t62)=e(t63)=e(t65)=e(t6,10)=e(t61),e(t64)=-131294307758125600000ν2+101770174398983783135143415014400000000ν4-50361593135536115530775449733160594484360806400000000000ν6+72224360822842722385742546232982312178753258185004780278718136320000000000000ν8+,e(t66)=-55576679174376800000ν2+33491511839833846512838011715302400000000ν4-1048979474294635763483385762842336670088069120000000000ν6+248300884277299965814264276737898589991181298872775071704180772044800000000000000ν8+,e(t67)=-22001910158125600000ν2+854131951646627753946003905100800000000ν4-92480070367754551230905137733160594484360806400000000000ν6+128724265121148448273612702030327511378753258185004780278718136320000000000000ν8+,e(t68)=e(t67),e(t69)=125+223705996876000ν2+7761616739041737883365875840000000ν4-4892287009909721519044696109671620703006720000000000ν6+389540085404399482872391415269600333281385757708532511613255680000000000000ν8+,e(t6,11)=e(t64),e(t6,12)=-1180+21701908346500480000ν2+1020783650901433441756803124080640000000ν4-3493913295664516557052587471759585426762465935360000000000ν6+2234301283265584634362936546350800879000372364000546317567787008000000000000ν8+,e(t6,13)=-1180+18584653946500480000ν2+930238779819453217756803124080640000000ν4-3221268919738174612962286671759585426762465935360000000000ν6+7281345860082294443565287484305877349315013032740019121114872545280000000000000ν8+,e(t6,14)=e(t6,16)=e(t6,13),e(t6,15)=7147843258336000ν2-583379511259816321022309002240000000ν4-211837357556206212072454348877372965624053760000000000ν6+691659579152868326008265024592122578750362020556086697635348480000000000000ν8+,e(t6,17)=-1495267298080000ν2+1575285628973489693043153920000000ν4-221743464671062174816067751959584086523904000000000ν6+718629211591722473564376192288838272019314312436020007146618880000000000000ν8+,e(t6,18)=844759258336000ν2+68710504690203833003187000320000000ν4-1533846637126032635046025948877372965624053760000000000ν6+3326834021203497429309815144900301418750362020556086697635348480000000000000ν8+,e(t6,19)=e(t6,18),e(t6,20)=-15-2183338611200ν2+4081140067155341700743633408000000ν4-11423423666783436232078391629245765520801792000000000ν6+23848033853016506130445461060833921291678734018536223254511616000000000000ν8+. Then for FRK5b, we have (4.30)EC6(ν)=33801900+220734462985341850432000033801ν2+103509564749048121419103379115113160786723706880000000033801ν4-69660490721569916026762009831470726461219046528761082257686935686021120000000000033801ν6+2435292716542829071193295619114291884728658932829824455954743137266530514091318708537909674780917760000000000000033801ν8+.

It can be seen that as ν0, both the methods FRK5a and FRK5b reduce to a classical RK method of order five (on page 178 of ), which we denote by RK5. The method RK5 is called the prototype method or the limit method of the methods FRK5a and FRK5b. Moreover, as ν0, both FRK5a and FRK5b have the same error constant as that of RK5: EC6=33801/900.

5. Analysis of Stability and Phase Properties

In order to investigate the linear stability of the methods (2.1) or any other frequency-depending method, we applied it to the linear test equation as (5.1)y=iλy,λ>0. One step of computation yields (5.2)y1=M(iμ,iν)y0,μ=hλ, where (5.3)M(iμ,iν)=1+iμbT(I-iμA)-1e=(1-μ2(bT(ν)Ae)+μ4(bT(ν)A3e)-μ6(bT(ν)A5e)+μ8(bT(ν)A7e)-)+i(μ(bT(ν)e)-μ3(bT(ν)A2e)+μ5(bT(ν)A4e)-μ7(bT(ν)A6e)+) is called the stability matrix of the method (2.1).

Definition 5.1.

The region in the μ-ν plane Ω={(μ,ν)μ>0,ν>0,|R(iμ,iν):1} is called the imaginary stability region of the method (2.1) and the closed curve defined by |R(iμ,iν)|=1 is called the stability boundary of the method.

The imaginary stability regions of the methods Simos4, FRK4, FRK5a and FRK5b are depicted in Figures 1 and 2.

The imaginary stability region of Simos4 (a) and FRK4a (b).

The imaginary stability region of FRK5a (a) and FRK5b (b).

Definition 5.2.

The quantities (5.4)ϕ(μ,ν)=μ-arg(M(iμ,iν)),d(μ,ν)=1-|M(iμ,iν)| are called the phase lag and the error of amplification factor of the method (2.1), respectively. If ϕ(μ,ν)=𝒪(μq+1), and d(μ,ν)=𝒪(μp+1), then the method (2.1) is said to be dispersive of order q and dissipative of order p, respectively (see Ozawa  and Van de Vyver ).

By definition, an FRK method has zero phase lag and zero dissipation when applied to the standard linear oscillator (5.1) whose frequency λ coincides with the fitting frequency ω. That is to say, ϕ(μ,μ)=0,  d(μ,μ)=0.

Suppose that the internal stages have been exact for the linear equation (5.1), that is, Yi=exp(iciμ)y0, then the update gives (5.5)y1=Mu(iμ,iν)y0, where (5.6)Mu(iμ,iν)=1+iμi=1sbi(ν)exp(ciμ)=(1+μi=1sbi(ν)sin(ciμ))+iμi=1sbi(ν)cos(ciμ). Accordingly, the quantities (5.7)ϕu(μ,ν)=μ-arg(Mu(iμ,iν)),du(μ,ν)=1-|Mu(iμ,iν)| are called the phase lag and the dissipation of the update of the method (2.1), respectively. If ϕu(μ,ν)=𝒪(μq+1), and du(μ,ν)=𝒪(μp+1), then the update of the method (2.1) is said to be dispersive of order q and dissipative of order p, respectively.

Letting ν=rμ, we obtain the phase lags and dissipations of the four FRK methods derived in Section 4.

Simos4: (5.8)ϕ(μ,ν)=1-r2120μ5+𝒪(μ7),d(μ,ν)=1-r2144μ6+𝒪(μ8),ϕu(μ,ν)=-1+4r22880μ5+𝒪(μ7),du(μ,ν)=-1+4r25760μ6+𝒪(μ8),ϕu(μ,μ)=-1576μ5+𝒪(μ7),du(μ,μ)=-11152μ6+𝒪(μ8).

FRK4: (5.9)ϕ(μ,ν)=12-17r2+5r41440μ5+𝒪(μ7),d(μ,ν)=4-5r2+r4576μ6+𝒪(μ8),ϕu(μ,ν)=-1-9r2+10r42880μ5+𝒪(μ7),du(μ,ν)=-1-9r2+10r45760μ6+𝒪(μ8),ϕu(μ,μ)=0,du(μ,μ)=0.

FRK5a: (5.10)ϕ(μ,ν)=r2-12100μ7+𝒪(μ9),d(μ,ν)=1-r23600μ6+𝒪(μ8),ϕu(μ,ν)=4021+144855r22449440000μ7+𝒪(μ9),du(μ,ν)=-7+570r24536000μ6+𝒪(μ8),ϕu(μ,μ)=531787480000μ7+𝒪(μ9),du(μ,μ)=-5774536000μ6+𝒪(μ8).

FRK5b: (5.11)ϕ(μ,ν)=-20666880+22148840r2-3027619r4+1545659r643400448000μ7+𝒪(μ9),d(μ,ν)=1722240-2154757r2+432517r46200064000μ6+𝒪(μ8),ϕu(μ,ν)=9618232+190446368r2-408728565r4+208663965r65859060480000μ7+𝒪(μ9),du(μ,ν)=-9568-422949r2+432517r46200064000μ6+𝒪(μ8),ϕu(μ,μ)=0,du(μ,μ)=0.

Therefore, when integrating the test equation (5.1), the methods Simos4 and FRK4 are dispersive of order four and dissipative of order five, and FRK5a and FRK5b are dispersive of order six and dissipative of order five. Similar conclusions for their updates can be drawn from the above analysis.

Note that if the update of (2.1) is phase-fitted and amplification-fitted, it must be true that ϕu(μ,μ)=0 and du(μ,μ)=0, as is verified above for the two methods FRK4 and FRK5b.

6. Numerical Experiments

In order to examine the effectiveness of the new FRK methods proposed in this paper, we apply them to four test problems. We also employ several highly efficient integrators from the literature for comparison. The numerical methods we choose for experiments are as follow:

FRK5a: the seven-stage RK method of order five given by (4.17) and (4.20) in Section 4 of this paper;

FRK5b: the seven-stage RK method of order five given by (4.17) and (4.28) in Section 4 of this paper;

Simos4: the RK method of order four presented in , that is, the four-stage FRK method of order four given by (4.1) and (4.12) in Section 4 of this paper;

FRK4: the four-stage RK method of order four given by (4.1) and (4.12) in Section 4 of this paper;

ARK4: the second four-stage adapted RK method of order four given in the Subsection 3.2 of Franco ;

EFRK4: the four-stage exponentially fitted RK method of order four given in ;

RK4: the classical RK method of order four presented in  (the prototype method of Simos4 and FRK4);

RK5: the classical RK method of order five presented in  (the prototype method of FRK5a and FRK5b).

In the figures showing the efficiency of these integrators, the horizontal axis stands for the number of function evaluations required and the vertical axis stands for the digital logarithm of the maximal global error (MGE).

Problem 1.

Consider the following orbit problem studied in : (6.1)p=-q+0.001cos(t),q=p,p(0)=0,q(0)=1. This equation is the real part of the “almost” periodic orbit problem as (6.2)y′′+y=0.001eit,y(0)=1,y(0)=0.9995i,y. The exact solution to this problem is q(t)=cos(t)+0.0005tsin(t). We select the fitting frequency ω=1.0007 and integrate the equation on the interval [0,1000] with step sizes h=1/2i,  i=0,1,2,3. The numerical results are presented in Figure 3(a).

Efficiency curves of Problem 1 (a) and Problem 2 (b).

Problem 2.

Consider the following linear problem studied in : (6.3)y′′+ω2y=(ω2-1)sin(t),y(0)=1,y(0)=ω+1. The exact solution to this problem is y(t)=cos(ωt)+sin(ωt)+sin(t). In this experiment, we take the parameter ω=20 and integrate the equation on the interval [0,100] with the step sizes h=1/(8i),  i=2,3,4,5. The numerical results are presented in Figure 3(b).

Problem 3.

Consider the prey-predator system in ecology (see ) as (6.4)u˙=u(α-βv),v˙=v(γu-δ), where u(t) is the number of prey at time t, v(t) is the number of predators at time t, u˙ and v˙ are first derivatives with respect to time, and α,β,γ,δ are positive constants. For given initial values u(0)=u0,v(0)=v0, the system has unique solution, but analytical solution is not available. However, the system has an invariant I(u,v)=δlnu+αlnu-γu-βv in the sense that (d/dt)I(u(t),v(t))0. The efficient of the methods will be tested by measuring the error growth of the invariant I(u(t),v(t)) against the number of function evaluations required.

In this experiment, we take the values of parameters α=2,β=γ=δ=1 and take initial data u0=1.6,  v0=2.2. We select the fitting frequency ω=1.0075 and integrate the equation on the interval [0,30] with step sizes h=1/2i,  i=2,3,4,5. The numerical results are presented in Figure 4(a).

Efficiency curves of Problem 3 (a) and Problem 4 (b).

Problem 4.

We consider the following two-gene regulatory system without self-regulation (see Widder et al.  and Polynikis et al. ): (6.5)r˙1=m1H+(p2;θ2,n2)-γ1r1,r˙2=m2H-(p1;θ1,n1)-γ2r2,p˙1=k1r1-δ1p1,p˙2=k2r2-δ2p2, where i=1,2, ri is the concentration of mRNA Ri produced by gene Gi, pi is the concentration of protein Pi, mi is the maximal transcription rate of Gi, ki is the translation rate of Ri, γi is the degradation rate of Ri, and δi is the degradation rate of Pi. The two functions (6.6)H+(p2;θ2,n2)=p2n2p2n2+θ2n2,H-(p1;θ1,n1)=θ1n1p1n1+θ1n1 are the Hill functions of activation and repression, respectively. The parameters n1,n2 are the Hill coefficients, θ1,θ2 are the expression thresholds.

The solution (r1*,r2*,p1*,p2*) of the system r˙1=r˙2=p˙1=p˙2=0 is an equilibrium of the system (6.5). Now we take the values of parameters as follows: (6.7)n1=n2=3,  m1=1.15,  m2=2.35,  k1=k2=1,γ1=γ2=1,δ1=δ2=1,θ1=θ2=0.21. For any initial point (r1(0),r2(0),p1(0),p2(0)) near the equilibrium, the system (6.5) has a stable limit cycle.

In this experiment, we take initial data (0.6,0.8,0.4,0.6) and select the fitting frequency ω=1.48. The problem is integrated on the interval [0,20] with step sizes h=1/2i,  i=1,2,3,4. The numerical results are presented in Figure 4(b).

It can be seen from Figures 3 and 4 that the FRK methods are more efficient than their prototype RK methods and are more efficient than the other frequency depending methods of the same algebraic order.

7. Conclusions

In this paper, classical Runge-Kutta methods are adapted to the time integration of initial value problems of first order differential equations whose solutions have oscillatory properties. The newly developed phase-fitted and amplification-fitted Runge-Kutta methods (FRK) adopt functions of the product ν=ωh of the fitting frequency ω and the step size h as weight coefficients in the update. FRK methods have zero dispersion error and zero dissipation error when applied to the standard linear oscillator y'=iωy. That is to say, they preserve initial phase and amplification with time. Therefore, FRK integrators are a kind of integrators which preserve the oscillation structure of the problem.

As the fitting frequency tends to zero, FRK methods reduce to their classical prototypes methods. Furthermore, an FRK method has the same algebraic order and the same error constant with its prototypes method. Numerical experiments illustrate the high efficiency of FRK methods compared with their prototype methods and some other frequency depending methods like exponentially fitted RK type methods.

Theorem 3.3 gives a pair of sufficient conditions for modified RK type methods to be phase-fitted and amplification-fitted. The coefficients of the methods Simos4 and FRK5a are obtained with these conditions combined with an appropriate number of order conditions associated to low order trees. Since the updates of the methods FRK4 and FRK5b are also phase-fitted and amplification-fitted, they may be more efficient than those methods whose updates do not have this property.

In practical computations of oscillatory problems, the true frequency is, in general, not available. The fitting frequency contained in an FRK method is just an estimate of the true frequency. Sometimes the choice of the value of the fitting frequency ω for the FRK method affects their effectiveness to some extent. For instance, in the experiment of Problem 1, we find that the value of ω=1.0007 is superior to the usual choice ω=1. For approaches to estimating principal frequencies we refer to the papers .

This paper provides a convenient approach to constructing FRK methods. Other effective approaches are possible. For example, the fitting frequency can also be incorporated into internal coefficients aij or into nodes ci just as in . Or frequency-dependent coefficients may be introduced to the terms yn in the modified (2.1) as in . In these cases, the internal stages can also be made phase-fitted and amplification-fitted.

Acknowledgments

This work was supported by the National Natural Science Foundation of China (Grant no. 11171155), the Fundamental Research Fund for the Central Universities (Grants no. Y0201100265 and KYZ201125), and the Research Fund for the Doctoral Program of Higher Education (Grant no. 20100091110033). The authors would like to thank the referees for their valuable comments and suggestions.

Butcher J. C. Numerical Methods for Ordinary Differential Equations 2008 2nd John Wiley & Sons 10.1002/9780470753767 2401398 ZBL1213.74276 Butcher J. C. Wanner G. Runge-Kutta methods: some historical notes Applied Numerical Mathematics 1996 22 1–3 113 151 10.1016/S0168-9274(96)00048-7 1424293 ZBL0867.65038 Hairer E. Nørsett S. P. Wanner G. Solving Ordinary Differential Equations I, Nonstiff Problems 1993 8 2nd Berlin, Germany Springer 1227985 Bettis D. G. Numerical integration of products of Fourier and ordinary polynomials Numerische Mathematik 1970 14 421 434 0266429 Gautschi W. Numerical integration of ordinary differential equations based on trigonometric polynomials Numerische Mathematik 1961 3 381 397 0138200 10.1007/BF01386037 ZBL0163.39002 Martín P. Ferrándiz J. M. Multistep numerical methods based on the Scheifele G-functions with application to satellite dynamics SIAM Journal on Numerical Analysis 1997 34 1 359 375 10.1137/S003614299426505X 1445740 Raptis A. D. Simos T. E. A four-step phase-fitted method for the numerical integration of second order initial value problems BIT Numerical Mathematics 1991 31 1 160 168 10.1007/BF01952791 1097489 ZBL0726.65089 Bettis D. G. Runge-Kutta algorithms for oscillatory problems Zeitschrift für Angewandte Mathematik und Physik 1979 30 4 699 704 10.1007/BF01590846 547658 ZBL0412.65038 Paternoster B. Runge-Kutta(-Nyström) methods for ODEs with periodic solutions based on trigonometric polynomials Applied Numerical Mathematics 1998 28 2–4 401 412 10.1016/S0168-9274(98)00056-7 1655173 ZBL0927.65097 Franco J. M. Runge-Kutta-Nyström methods adapted to the numerical integration of perturbed oscillators Computer Physics Communications 2002 147 3 770 787 10.1016/S0010-4655(02)00460-5 1925533 ZBL1019.65050 Franco J. M. Runge-Kutta methods adapted to the numerical integration of oscillatory problems Applied Numerical Mathematics 2004 50 3-4 427 443 10.1016/j.apnum.2004.01.005 2074013 ZBL1057.65043 Anastassi Z. A. Simos T. E. A dispersive-fitted and dissipative-fitted explicit Runge-Kutta method for the numerical solution of orbital problems New Astronomy 2004 10 1 31 37 2-s2.0-7444231046 10.1016/j.newast.2004.04.005 Van de Vyver H. Phase-fitted and amplification-fitted two-step hybrid methods for y''=f(x,y) Journal of Computational and Applied Mathematics 2007 209 1 33 53 10.1016/j.cam.2006.10.025 2384370 ZBL1141.65061 Vigo-Aguiar J. Ramos H. Dissipative Chebyshev exponential-fitted methods for numerical solution of second-order differential equations Journal of Computational and Applied Mathematics 2003 158 1 187 211 10.1016/S0377-0427(03)00473-4 2017377 ZBL1042.65053 Vigo-Aguiar J. Ferrandiz J. M. Higher-order variable step algorithms adapted to the accurate numerical integration of perturbed oscillators Computer Physics Communications 1998 12 5 467 470 Simos T. E. Vigo-Aguiar J. An exponentially-fitted high order method for long-term integration of periodic initial-value problems Computer Physics Communications 2001 140 3 358 365 2-s2.0-0035500501 10.1016/S0010-4655(01)00285-5 Tocino A. Vigo-Aguiar J. Symplectic conditions for exponential fitting Runge-Kutta-Nyström methods Mathematical and Computer Modelling 2005 42 7-8 873 876 10.1016/j.mcm.2005.09.015 2178516 ZBL1086.65120 Vigo-Aguiar J. Simos T. E. An exponentially fitted and trigonometrically fitted method for the numerical solution of orbital problems Astronomical Journal 2001 122 3 1656 1660 2-s2.0-4043144506 10.1086/322107 Avdelas G. Simos T. E. Vigo-Aguiar J. Embedded exponentially-fitted Runge-Kutta method for the numerical solution of the Schrodinger equation and related periodic initial-value problems Computer Physics Communications 2000 131 1 52 67 2-s2.0-0034274825 10.1016/S0010-4655(00)00080-1 Simos T. E. Vigo-Aguiar J. A symmetric high order method with minimal phase-lag for the numerical solution of the Schrodinger equation International Journal of Modern Physiscs C 2001 12 7 1035 1042 10.1142/S0129183101002292 Simos T. E. Vigo-Aguiar J. A dissipative exponentially-fitted method for the numerical solution of the Schrödinger equation and related problems Computer Physics Communications 2003 152 3 274 294 10.1016/S0010-4655(02)00871-8 1974705 ZBL1196.65123 van der Houwen P. J. Sommeijer B. P. Explicit Runge-Kutta (-Nyström) methods with reduced phase errors for computing oscillating solutions SIAM Journal on Numerical Analysis 1987 24 3 595 617 10.1137/0724041 888752 ZBL0624.65058 Simos T. E. An exponentially-fitted Runge-Kutta method for the numerical integration of initial-value problems with periodic or oscillating solutions Computer Physics Communications 1998 115 1 1 8 10.1016/S0010-4655(98)00088-5 1665072 ZBL1001.65080 Ozawa K. A four-stage implicit Runge-Kutta-Nyström method with variable coefficients for solving periodic initial value problems Japan Journal of Industrial and Applied Mathematics 1999 16 1 25 46 10.1007/BF03167523 1676362 Van de Vyver H. Stability and phase-lag analysis of explicit Runge-Kutta methods with variable coefficients for oscillatory problems Computer Physics Communications 2005 173 3 115 130 10.1016/j.cpc.2005.07.007 2187062 ZBL1196.65117 Vanden Berghe G. De Meyer H. Van Daele M. Van Hecke T. Exponentially fitted Runge-Kutta methods Journal of Computational and Applied Mathematics 2000 125 1-2 107 115 10.1016/S0377-0427(00)00462-3 1803185 ZBL0999.65065 Kalogiratou Z. Symplectic trigonometrically fitted partitioned Runge-Kutta methods Physics Letters A 2007 370 1 1 7 10.1016/j.physleta.2006.12.080 2412998 ZBL1209.65138 Murray J. D. Mathematical Biology 1993 19 3rd Berlin, Germany Springer 10.1007/b98869 1239892 Widder S. Schicho J. Schuster P. Dynamic patterns of gene regulation. I. Simple two-gene systems Journal of Theoretical Biology 2007 246 3 395 419 10.1016/j.jtbi.2007.01.004 2306875 Polynikis A. Hogan S. J. di Bernardo M. Comparing different ODE modelling approaches for gene regulatory networks Journal of Theoretical Biology 2009 261 4 511 530 2-s2.0-70350569876 10.1016/j.jtbi.2009.07.040 Vanden Berghe G. Ixaru L. G. Van Daele M. Optimal implicit exponentially-fitted Runge-Kutta methods Computer Physics Communications 2001 140 3 346 357 2-s2.0-0035500485 10.1016/S0010-4655(01)00279-X Vanden Berghe G. Ixaru L. Gr. De Meyer H. Frequency determination and step-length control for exponentially-fitted Runge-Kutta methods Journal of Computational and Applied Mathematics 2001 132 1 95 105 10.1016/S0377-0427(00)00602-6 1834805 ZBL0991.65062 Ixaru L. G. Vanden Berghe G. De Meyer H. Frequency evaluation in exponential fitting multistep algorithms for ODEs Journal of Computational and Applied Mathematics 2002 140 1-2 423 434 2-s2.0-0036502977 10.1016/S0377-0427(01)00474-5 Vigo-Aguiar J. Simos T. E. Ferrándiz J. M. Controlling the error growth in long-term numerical integration of perturbed oscillations in one or several frequencies Proceedings of The Royal Society of London Series A 2004 460 2042 561 567 10.1098/rspa.2003.1210 2034655 ZBL1041.65058 Van de Vyver H. Frequency evaluation for exponentially fitted Runge-Kutta methods Journal of Computational and Applied Mathematics 2005 184 2 442 463 10.1016/j.cam.2005.01.020 2157338 ZBL1077.65082 Ramos H. Vigo-Aguiar J. On the frequency choice in trigonometrically fitted methods Applied Mathematics Letters 2010 23 11 1378 1381 10.1016/j.aml.2010.07.003 2718515 ZBL1197.65082