Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Matlab code for Lyapunov exponents of fractional order systems

Matlab code for Lyapunov exponents of fractional order systems April 5, 2018 0:22 LE-FO-arxiv MARIUS-F. DANCA Romanian Institute od Science and Technology, 400487 Cluj-Napoca, Romania, danca@rist.ro NIKOLAY KUZNETSOV Dept. of Applied Cybernetics, Saint-Petersburg State University, Russia and Dept. of Mathematical Information Technology, University of Jyv¨askyl¨a, Finland nkuznetsov239@gmail.com Received (to be inserted by publisher) In this paper the Benettin-Wolf algorithm to determine all Lyapunov exponents for a class of fractional-order systems modeled by Caputo's derivative and the corresponding Matlab code are presented. First it is proved that the considered class of fractional-order systems admits the necessary variational system necessary to find the Lyapunov exponents. The underlying nu- merical method to solve the extended system of fractional order, composed of the initial value problem and the variational system, is the predictor-corrector Adams-Bashforth-Moulton for fractional differential equations. The Matlab program prints and plots the Lyapunov exponents as function of time. Also, the programs to obtain Lyapunov exponents as function of the bifur- cation parameter and as function of the fractional order are described. The Matlab program for Lyapunov exponents is developed from an existing Matlab program for Lyapunov exponents of integer order. To decrease the computing time, a fast Matlab program which implements the Adams-Bashforth-Moulton method, is utilized. Four representative examples are considered. Keywords: Lyapunov exponents, Benettin-Wolf algorithm, Fractional-order dynamical system 1. Introduction Despite a long history, the doubts that fractional- ity of solutions of FDEs with general nonlinearities order (FO) derivatives have no clear geometrical in- see, for example, [Oldham & Spanier, 1974; Caputo, terpretations (see e.g. [Podlubny, 2002]), was one 1967], and [Diethelm & Ford, 2002; Kilbas & Tru- of the several reasons that fractional calculus was jillo, 2001; Podlubny, 1999]. not used in physics or engineering. However, dur- The Lyapunov exponents (LEs) measure the aver- ing the last more than 10 years, fractional calcu- age rate of divergence or convergence of orbits start- lus starts to attract increasing attention. There are ing from nearby initial points. Therefore, they can nowadays more and more works on FO systems be used to analyze the stability of limits sets and and their related applications in physics, engineer- to check sensitive dependence on initial conditions, ing, mathematics, finance, chemistry, and so on. For that is, the presence of potential chaotic attractors. the theory on the existence, uniqueness, continuous On the other side, in [Cvitanovi´ c et al., 2016] Cvi- dependence on parameters and asymptotic stabil- tanovi´ c et al. do not recommend the evaluation arXiv:1804.01143v1 [physics.comp-ph] 29 Mar 2018 April 5, 2018 0:22 LE-FO-arxiv 2 M.-F. Danca, N.V. Kuznetsov of the LEs and recommend to: “Compute stabil- tractors in the case of multistability (see, e.g. such ity exponents and the associated covariant vectors corresponding examples for the classical Lorenz sys- instead. Cost less and gets you more insight. [...] we tem and Henon map [Leonov et al., 2016; Kuznetsov are doubtful of their utility as means of predicting et al., 2018b]). any observables of physical significance”. Moreover, In order to study the chaoticity of an attractor in we additionally note here, Perron’s counterexample numerical experiments, one has to consider a grid [Leonov & Kuznetsov, 2007] which shows actually of point covering the attractor and compute corre- that the use of LEs, obtained via the linearization sponding finite-time local LEs for a certain time. procedure, for the study of the behaviour of non- Remark that while the time series obtained from linear system requires a rigorous justification. How- a physical experiment are assumed to be reliable ever, determining LEs remains the subject of many on the whole considered time interval, the time se- works and grown into a real software industry for ries produced by the integration of mathematical modern nonlinear physics (see, e.g.[Hegger et al., dynamical model can be reliable on a limited time 1999; Barreira & Pesin, 2001; Skokos, 2010; Czornik interval only due to computational errors. et al., 2013; Pikovsky & Politi, 2016; Vallejo & San- Taking into account the above discussion further juan, 2017] and others). we consider finite-time local LEs and their compu- Nowadays there are two widely used definitions , tation in Matlab by the analog of the Bennetin-Wolf of the LEs: via the exponential growth rates of algorithm. norms of the fundamental matrix columns [Lya- punov, 1892] and via the exponential growth rates 2. Benettin-Wolf Algorithm for LEs of the sigular values of fundamental matrix [Os- of FO eledets, 1968]. Corresponding approaches for the The determination of LEs of a system of integer LEs computation and their difference are discussed, or fractional order with the Benettin-Wolf algo- e.g., in [Kuznetsov et al., 2018a, 2016]. rithm, requires the numerical integration of differen- Remark that in numerical experiments we can con- tial equations of integer or fractional order. Because sider only finite time, and, thus, the numerically the purpose of this paper is to present a Matlab computed values of LEs can differ significantly from code for LEs of systems of FO, in the next subsec- the limit values (e.g. if the considered trajectory tions, only the most important steps (such as the belongs to a transient chaotic set ), and are often existence of variational equations of FO) used to referred as finite-time LEs. implement the algorithm in Matlab language are Applying the statistical physics approach and as- presented (theoretical details can be found in the suming the ergodicity (see, e.g. [Oseledets, 1968]), related references). the LEs for a given dynamical system are often es- timated by local LEs along a “typical” trajectory. However, in numerical experiments, the rigorous use 2.1. Numerical integration of FDEs of the ergodic theory is a challenging task (see, e.g. The autonomous FO systems considered in this pa- [Cvitanovi´ c et al., 2016, p.118]). If the LEs are the per are modeled by the following Initial Value Prob- same for any trajectory, then Frederickson et al. lem (IVP) with Caputo’s derivative [Frederickson et al., 1983, p.190] suggested to call D x = f (x), them as absolute ones and wrote that such abso- (1) x(0) = x , lute values rarely exist. For example, substantially n n different values of the local LEs can be obtained for t ∈ [0, T ], q ∈ (0, 1), f : R → R and D , Ca- along trajectories on coexisting nonsymmetric at- puto’s differential operator of order q with starting Relying on the Oseledec ergodic theorem [Oseledets, 1968] the above definitions often do not differ (see, e.g. Eckmann & Ruelle [Eckmann & Ruelle, 1985, p.620,p.650], Wolf et al. [Wolf et al., 1985a, p.286,p.290-291], and Abarbanel et al. [Abarbanel et al., 1993, p.1363,p.1364]), however in general case, they may lead to different values [Kuznetsov et al., 2018a],[Bylov et al., 1966, p.289],[Leonov & Kuznetsov, 2007, p.1083]. While trivial attractors (stable equilibrium points) can be easily found analytically or numerically, the search of all periodic and chaotic attractors for a given system is a challenging problem. See, e.g. famous 16th Hilbert problem [Hilbert, 1901-1902] on the number of coexisting periodic attractors in two dimensional polynomial systems, which was formulated in 1900 and is still unsolved, and its generalization for multidimensional systems with chaotic attractors [Leonov & Kuznetsov, 2015]. Based on philosophical arguments rather than a mathematical point of view, some researchers questioned the appropriateness April 5, 2018 0:22 LE-FO-arxiv Matlab code for Lyapunov exponents of fractional order systems 3 Let us next assume that we are working on a uni- point 0 form grid {t = jh : n = 0, 1, ..., N + 1} with some integer N with the step size h and the case q −q 0 D x(t) = (t− τ ) x (τ )dτ, of q ∈ (0, 1). Then, the predictor form, x , at the Γ(1− q) point t , is the fractional variant of the Adams- j+1 with Γ the known Euler function. Bashforth method Properties of the Caputo’s differential operator, D , are discussed in [Podlubny, 1999; Gorenflo & x (t ) = x + b f (x (t )) , n+1 0 j,n+1 j Γ (q) Mainardi, 1997]. j=0 Under Lipschitz continuity of the function f , the while the corrector formula (the fractional variant IVP (1) admits a unique solution [Diethelm et al., of the one-step implicit Adams Moulton method) 2002]. reads Remark 2.1. i ) In the case of an integer-order dynamical sys- x (t ) = x + f x (t ) n+1 0 n+1 Γ(q + 2) tem, denoting the solution of the underlying IVP as x(t, x ), one has x ◦ x = x (see e.g. [Zhou, h 0 s t t+s + a f (x (t )) , j,n+1 j 2016]). Due to the memory dependence of the Γ(q + 2) j=0 derivatives, this does not hold in the case of sys- where a and b are the corrector and predictor tems modeled by FDEs. However, motivated by the weights respectively given by the following formula numerical character of this paper, the definition of integer order dynamical systems which states that if the underlying IVP admits unique solutions exist- q+1 n − (n− q) (n + 1) if j = 0, ing on infinite time interval, the problem defines a  q+1 q+1 (n− j + 2) + (n− j) dynamical system (see [Stuart & Humphries, 1998, a = j,n+1 q+1 −2 (n− j + 1) if 1 ≤ j ≤ n, Definition 2.1.2]) is adopted. 1 if j = n + 1, ii ) Even fractional-order dynamics describe a real object more accurately than classical integer-order and dynamics, systems modeled by the IVP (1) cannot q q q b = ((n + 1− j) − (n− j) ) . have any non-constant periodic solution (see e.g. j,n+1 [Tavazoei & Haeri, 2009]). However, a solution may be asymptotically periodic [Danca et al., 2018a]. The predictor-corrector ABM method has an error These trajectories are called numerical ly periodic, which is roughly proportional to h . Thus, to ob- in the sense that the trajectory, from numerical tain an error of, e.g., 1.0E − 6, a step size close to point of view, can be an extremely-near periodic h = 1.0E − 3 should be considered. with respect to, e.g., Euclidean norm [Danca et al., Because due to the long memory processes, 2018a]. A numerically periodic trajectory refers to the utilized ABM method as described in [Diethelm as a closed trajectory in the phase space in the sense et al., 2002], is time consuming. Therefore, a fast op- that the closing error is within a given bound of timized ABM method, FDE12.m [Garrappa, 2012] is 1E − n, with n being a sufficiently large positive used. integer. FDE12.m is called by the following command line: The numerical integrations required by the LEs al- [t,x] = FDE12(q,fcn,t0,tf,x0,h); gorithm for FO systems are performed in this pa- per with the predictor-corrector Adams-Bashforth- Moulton (ABM) method for FDEs, proposed Di- where q represents the commensurate fractional- ethelm et al. [Diethelm et al., 2002] which is con- order, fcn.m is the file with the function to be inte- structed for the fully general set of equations with- grated, t0 and tf define the time span, x0 are the out any special assumptions, being easy to imple- initial conditions, and h represents the integration ment in any language. step-size. of using initial conditions of the classical form in the Caputo derivative [Hartley et al., 2013]. However, it should be emphasized that, in practical (physical) problems, physically interpretable initial conditions are necessary and Caputo's derivative is a fully justified tool [Diethelm, 2014]. April 5, 2018 0:22 LE-FO-arxiv 4 M.-F. Danca, N.V. Kuznetsov For example, consider the integration of the FO first work to propose a Gram-Schmidt orthogonal- Rabinovich-Fabrikant (RF) system [Danca, 2016], ization procedure to compute LEs for continuous which has the following Matlab function, RF.m systems of integer order, as described in [Eckmann & Ruelle, 1985]), and by Wolf et al. [Wolf et al., 1985b] (see also [Eckmann et al., 1986]). function dx = RF(t,x) dx=[x(2)*(x(3)-1+x(1)*x(1))+0.1*x(1); The algorithm to find all LEs, described as a For- x(1)*(3*x(3)+1-x(1)*x(1))+0.1*x(2); tran code by Wolf et al. [Wolf et al., 1985b], and also -2*x(3)*(p+x(1)*x(2))]; as a Basic code in [Baker & Gollub, 1990], solves the equations of motion under perturbations and peri- with the bifurcation parameter p = 0.98. odic orthonormalization. With the following parameters, one obtains the To note that the accuracy and reliability of nu- chaotic attractor in Fig. 1 (a): merically determined LEs depend on initial condi- tions, on the selection of the perturbations, perfor- [t,x]=FDE12(0.999,@RF,0,1500,[0.1;0.1;0.1],... mances of the utilized integration numerical method 0.01); and also on orthonormalization step size. A long- time numerical calculation of the leading Lyapunov exponent requires rescaling the distance between To note that FDE12.m requires the entry, x0, as nearby trajectories, in order to keep the separation a column vector [x10,x20,x30]’ or, similarly, within the linearized flow range. To avoid overflow, [x10;x20;x30]. Also, the returned exit, x, is a col- one calculates the divergence of nearby trajectories umn vector. for finite timesteps and renormalizes to unity af- ter a finite number of steps (Gram-Schmidt pro- 2.2. Algorithm for LEs of the FO cedure [Eckmann & Ruelle, 1985; Christiansen & system (1) Rugh, 1997]). Therefore, the main steps to determine numerically Notation 2.1. Hereafter the finite-time local LEs of the LEs are: numerical integration of the FO sys- FO are called LEs. tem (1) together with the variational system (2) The existence of the variational equations necessary (i.e. the extended system), Gram-Schmidt proce- to determine LEs is ensured by the following Theo- dure and picking up the exponents during the renor- rem [Li et al., 2010] malization procedure, the LEs being determined as the average of the logarithm of the stretching factor Theorem 1. System (1) has the fol lowing varia- of each perturbation, steps presented in Algorithm tional equations which define the LEs D Φ(t) = D f (x)Φ(t), ∗ x (2) Φ(0) = I, 3. The Matlab code for LEs where Φ is the matrix solution of the system (1), D Consider the following general assumptions: is the Jacobian of f and I is the identity matrix. Further we assume that for the matrix Φ(t) the co- • The considered systems, modeled by the IVP (1), cycle property takes place. are autonomous; Therefore, the algorithm to determine the LEs of • The system (1) is of commensurate order: q = the system (1) becomes similar to the case of inte- q = ... = q = q; 2 ne ger order. • In the case of chaotic behavior, the fractional- Lyapunov exponents measure the exponential order has been chosen close to 1, such that chaos growth, or decay, of infinitesimal phase-space per- is significant. turbations of a chaotic dynamical system. • The right hand side of system (1), f , is smooth The algorithm for numerical evaluation of LEs uti- enough. lized in this paper, has been proposed in the seminal • Because of the space restrictions, only the main works of Benettin et al. [Benettin et al., 1980] (see programs code are presented, and indications on also [Shimada & Nagashima, 1979]), one of the the how to write the other ones. The incommensurate case can be treated similarly, the only difference referring to the utilized numerical method for FDEs. April 5, 2018 0:22 LE-FO-arxiv Matlab code for Lyapunov exponents of fractional order systems 5 Algorithm 1: Algorithm for LEs of FO system (1) Input: -ne . number of equations -x start . n initial conditions of (1) -t start, t end . time span -h norm . Normalization step-size n it ← (t end− t start)/h norm . iterations number for i ← ne + 1 to ne(ne + 1) do x(i) = 1.0 . initial conditions of (2) end t ← t start for i ← 1 to n it do x ← integration of FO systems (1)-(2) t ← t + h norm zn(1), ..., zn(ne) ← Gram-Schmidt procedure s(1) ← 0 for k ← 1 to ne do s(k) ← s(k) + log(zn(k)) . vector magnitudes LE(k) ← s(k)/(t− t start) . LEs end end Output:LE A simple way to build in Matlab language the algo- it is necessary to solve the extended system (1)-(2) rithm for FO systems, the program FO Lyapunov.m of FO (yellow line) which is given in the function (Appendix A), was to modify either some exist- ext fcn.m. In this file, beside the right hand side ing program, e.g., the program lyapunov.m [Gov- function f of the system (1), the Jacobi matrix J orukhin, 2004], which is a Matlab variant of the should be included (see Appendix B where the func- original LEs algorithm proposed in [Benettin et al., tion for the RF system is presented). 1980] or [Wolf et al., 1985b] or, similarly, to trans- The program plots the time evolutions of the LEs. late in Matlab the BASIC program [Baker & Gol- lub, 1990], a close variant of the original algorithm) 4. Numerical tests or, also, the Fortran code proposed by Wolf et all in [Wolf et al., 1985b], and modify it for FDEs. Beyond numerical artifacts that might occur when The program, called FO Lyapunov.m, is launched numerically integrating a system of ODEs of integer with the following command line: order, notions such as “shadowing time” and “maxi- mally effective computational time” reveal that it is [t,LE]=FO_Lyapunov(ne,@ext_fcn,t_start,h_norm,... possible to have reliable numerical simulations only t_end,x_start,h,q,out); on a relative finite-time interval (see, e.g., [Sarra & Meador, 2011; Wang et al., 2012]). The case of FO systems is even more delicate. In this paper we have where ne represents the equations (and state vari- considered generally t ∈ [0, 300] and, for the Lorenz ables) number, ext fcn.m the function containing system, t ∈ [0, 500]. the extended system (1)-(2), t start and t end the time span, h norm the normalization step, x start 1. Let consider the function for the RF sys- the initial condition (as column vector), h the step tem, LE RF.m, which include the extended sys- size of FDE12.m, and out indicates the steps num- tem (1)-(2) (Appendix B). Because ne=3, be- ber when intermediate values of time and LEs are side the 3 variables x(1),x(2),x(3) required printed (for out=0, no intermediate results will be by the numerical solution of the original sys- printed out). tem (1), the matrix solution of the system (2) As shown in the algorithm for LEs (Algorithm 1), requires other more ne×ne=9 variables from April 5, 2018 0:22 LE-FO-arxiv 6 M.-F. Danca, N.V. Kuznetsov the total of ne(ne+1)=12 variables: x(1:12), ter p = 200, after some neglected transients, one f=zeros(size(x))=zeros(12), where f loads obtains an apparently stable cycle (see Fig. 2 (a) and Remark 2.1 (ii )). the first ne=3 righthand side expressions of sys- tem (1), and the ne×ne=9 righthand side expres- To obtain the LEs one write the following com- sions of the variational system (2). mand line: For example, for the RF system, with the follow- ing command line: [t,LE]=FO_Lyapunov(3,@LE_Lorenz,0,5,500,... [0.1;0.1;0.1],0.001,0.985,10); [t,LE]=FO_Lyapunov(3,@LE_RF,0,0.02,300,... [0.1;0.1;0.1],0.005,0.999,1000); which gives LE=(-0.0026, -0.0870, -1.6225). The function LE Lorenz.m can be one obtains the intermediary results printed ev- obtained similarly with LE RF.m. The time evo- ery out=1000 h norm steps, presented in Table lution of the LEs is presented in Table 2 (see 1 (see also Fig. 1 (b) where the dynamics of the also Fig. 2 (b)). LEs are drawn). 50.00 0.1759 -0.1591 -1.5683 100.00 0.0611 -0.1108 -1.6050 10.00 0.1611 0.0660 -2.1614 150.00 0.0346 -0.0927 -1.6300 20.00 0.1923 0.0069 -2.0503 200.00 0.0215 -0.0877 -1.6288 30.00 0.0984 -0.7397 -1.1817 250.00 0.0135 -0.0866 -1.6269 40.00 0.0248 0.0542 -1.9761 300.00 0.0082 -0.0865 -1.6255 50.00 0.0440 -0.0168 -1.9867 350.00 0.0043 -0.0866 -1.6244 60.00 0.0697 -0.0006 -2.0701 400.00 0.0014 -0.0867 -1.6236 70.00 0.0354 -0.0116 -2.0167 450.00 -0.0008 -0.0869 -1.6230 80.00 0.0331 -0.0618 -1.9360 500.00 -0.0026 -0.0870 -1.6225 90.00 0.0201 0.0125 -1.9754 100.00 0.0429 0.0112 -1.9794 Table 2. Time evolution of the LEs for the Lorenz 110.00 0.0337 0.0252 -1.9698 system. LE=(-0.0026, -0.0870, -1.6225). 120.00 0.0262 -0.0213 -1.9038 130.00 0.0624 -0.0043 -1.9537 3. Consider next the non-smooth 4-dimensional 140.00 0.0660 -0.0029 -1.9811 system [Danca et al., 2018a] 150.00 0.0645 -0.0010 -2.0008 160.00 0.0743 0.0018 -2.0304 D x = −x + x , ∗ 1 1 2 170.00 0.0710 -0.0009 -2.0394 D x = −x sgn(x ) + x , ∗ 2 3 1 4 180.00 0.0583 0.0139 -2.0548 (4) D x = |x |− a, ∗ 3 1 190.00 0.0678 0.0042 -2.0665 D x = −bx , 200.00 0.0662 -0.0217 -2.0498 ∗ 4 2 210.00 0.0628 -0.0155 -2.0622 with a = 1, b = 0.5 and q = 0.98. 220.00 0.0602 -0.0125 -2.0715 230.00 0.0570 -0.0222 -2.0666 With the command line: 240.00 0.0643 -0.0222 -2.0813 250.00 0.0639 -0.0079 -2.1020 [t,LE]=FO_Lyapunov(4,@LE_4d,0,0.02,300,... 260.00 0.0623 0.0045 -2.1121 [0.1;0.1;0.1;0.1],0.005,0.98,1000); 270.00 0.0630 0.0006 -2.0987 280.00 0.0551 -0.0036 -2.0771 290.00 0.0761 0.0009 -2.0937 one obtains LE=(0.1262, 0.0846, 0.0778, 300.00 0.0749 0.0018 -2.0850 -1.5244) (Fig. 3). For this example, ne=4, the number of the Table 1. For t ∈ [0, 300], the RF system has LE = variables is 20 and, therefore, compared with (0.0749, 0.0018, -2.0850) (last line, blue). systems with ne=3, the function file, LE 4d.m, 2. If one considers the Lorenz system must be modified accordingly (compare the red q line in LE RF.m). Also, in order to obtain the D x = σ(x(2)− x(1)), ∗ 1 q printed intermediated values of LEs, in the D x = −x(1)x(3) + px(1)− x(2) (3) ∗ 2 q FO Lyapunov.m, the fprintf command, must be D x = x(1)x(2)− βx(3); ∗ 3 modified by adding one supplementary specifier with q = 0.985 and the standard parameters %10.3f. σ = 10, β = 8/3 and the bifurcation parame- Because from the four LEs, the system admits April 5, 2018 0:22 LE-FO-arxiv Matlab code for Lyapunov exponents of fractional order systems 7 three positive LEs, it can be considered as hy- tional order q (program run Lyapunov q.m in perchaotic (see [Danca et al., 2018b] for a dis- Appendix E). The used program FO Lyapunov q, cussion about the number of positive LEs of hy- is obtained from FO Lyapunov with the following perchaotic systems). modifications: (a) The header line of the code is replaced with Remark 4.1. Note that because the system is the following line (to note that the input out is not smooth and also discontinuous, the correct no more necessary); numerical integration required for this system cannot be done without a previous smooth ap- function [t,LE]=FO_Lyapunov_q(ne,ext_fcn,... proximation [Danca et al., 2018a]. Thus, like all t_start,h_norm,t_end,x_start,h,q); numerical methods for FDEs which are designed for continuous dynamical systems, the integra- tor FDE12 cannot be utilized in this case with- (b) the printing and plotting lines (*) (red out the mentioned approximation. Also, without color) are deleted; a smooth approximation, the Jacobian matrix, (c) the rest of the lines remain the same; utilized in the Benettin-Wolf algorithm, cannot The function containing the extended system, be determined. ext fcn.m, does not require any modification. For example, for the RF system, with the com- 4. LEs can be also plotted as function of the mand bifurcation parameter p for p ∈ [p , p ] min max (program run Lyapunov p.m, Appendix C). run_FO_Lypaunov_q(3,@LE_RF,0,0.05,150,... The program uses a slightly modified vari- [0.1;0.1;0.1],0.002,0.9,1,800) ant of FO Lyapunov.m, which is called FO Lyapunov p.m. To obtain FO Lyapunov p.m from FO Lyapunov.m, the following modifications one obtains the LEs plotted in Fig. 4 (b). have to be done: Remark 4.2. The presented programs can (a) The header line of the code FO Lyapunov.m be optimized especially in the cases of is replaced with the following line (to note that FO Lyapunov p.m and FO Lyapunov q.m. A sim- the output t and the input out are no more nec- ple improvement was to use while loop (in- essary) stead for) which for these two programs re- duce substantially the computational time. How- function LE=FO_Lyapunov_p(ne,ext_fcn_p,... ever, every parameter p (or order q) step, several t_start,h_norm,t_end,x_start,h,q,p); constant parameters are shared between pro- grams, fact which slows significantly the pro- (b) the printing and plotting lines (*) (red grams. Therefore, supplementary optimization color) are deleted; should be done. (c) the rest of the lines remain the same; 6. Another potential application of the proposed The m-file containing the extended system, LEs algorithm, is to represent the LEs as func- ext fcn p.m, for the RF system, is presented in tion of two variables: the order q and the bifur- Appendix D. Passing the parameter p between cation parameter p. these codes, can be easily realized due of the fa- Let the Chen system cilities of the program FDE12 (see FDE12). D x = a(x − x ), ∗ 1 2 1 For example, for the RF system, for p ∈ [1.1, 1.3], D x = (p− a)x − x x + cx , ∗ 2 1 1 3 2 for 1000 values of p (n = 1000), with the follow- D x = x x − bx , ∗ 3 1 2 3 ing command with parameters a = 35, b = 3 and q and p run_FO_Lyapunov_p(3,@LE_RF_p,0,0.02,200,... variables. By considering S := LE(q, c), for [0.1;0.1;0.1],0.002,0.998,1.1,1.3,800) i = 1, 2, 3, for q ∈ [0.9, 1] and p ∈ [20, 30], the ob- tained surfaces are plotted in Fig. 5 (see [Danca one obtains the evolution of the LEs drawn in et al., 2018b] where the algorithm to obtain LEs Fig. 4 (a). as surfaces is described). One can see that there 5. LEs can be plotted also as function of the frac- exists a unique positive LE (red surface, S ) for 1 April 5, 2018 0:22 LE-FO-arxiv 8 REFERENCES all values of the considered parameter c but only ments of numerical integration are implemented, for some fractional order values q, for which S is with Benettin-Wolf algorithm, for integer but also situated over the horizontal plane LE=0, where for fractional order systems, two or three decimals LEs are zero. Also, one can see that S , which is are the maximal expected number of decimals. almost identic to S when the underlying LE1,2, One of the most important algorithm variable, are negative, becomes zero for a relative large is h norm, which determines the normalization values of q and p, when S separates from S moments influences the results. This dependence 2 1 and identify (with the underlying numerical er- can be also deduced from the example of the ror) with the plane LE=0. RF system. With h=0.005 and h norm=0.005 one obtains LEs=(0.0631, 0.0031, -2.0774), while 5. Conclusion and discussion with the same stapsize, h=0.005, but h norm=0.5, LEs=(0.0643, 0.0026, -1.8254). Starting from an existing variant for integer-order Since on our best knowledge, there is not criterion system, in this paper, based on the Benettin-Wolf to choose precisely h norm, the recommended way algorithm, we proposed a Matlab program to deter- would be to realize several tests to choose the value mine numerically finite-time local LEs for FO sys- for which slightly modifications does not change sig- tems. nificantly the results. Due to inherent numerical errors, the algorithm In the case of fixed step-size integration numerical should be utilized with precaution. As known, methods, like the considered FDE12, h norm can be among the numerical errors, the results depends chosen as depending on the step size h. Therefore, strongly on the initial conditions, time integration as Fig. 6 shows, h norm could be multiple of inte- interval and, especially, on the renormalization step gration step size h (ε represents the perturbation size h norm. between nearly two trajectories x and x ¯). The relative large numerical errors of the Benettin- Regarding the Gram-Schmidt procedure, in or- Wolf algorithm, which for the considered examples der to speed up the code, one can use the QR was of order of 1.E − 2, can be observed in the decomposition based on the Householder trans- cases when one knows that the system presents a formation: [Q, R] = qr(A). To have matrix R numerically periodic cycle (see Remark 2.1 (ii )), with positive diagonal elements, one can ad- when the maximum LE should be zero. For exam- ditionally use Q = Q*diag(sign(diag(R))); R = ple, the Lorenz system, which for p = 200 presents R*diag(sign(diag(R))). See also [Ramasubrama- such apparently stable cycle (see Fig. 2 (a)), has the nian & Sriram, 2000]. maximum LE zero only with two precise decimals. The integration step-size of FDE12, h, plays an im- Actually, in general, three precise decimals for zero portant role. As the documentation of the program LEs are extremely rare. Note that this zero value, specifies, there exists the possibility to increase the appears only for h norm=5 and a smaller integra- performances of the numerical integration, but in tion step size, h = 0.001 and only for a larger time time integration detriment. interval [0,500]. Therefore, if no special improve- Acknowledgements The work is done within Russian Science Foundation project 14-21-00041. References Abarbanel, H., Brown, R., Sidorowich, J. & Tsimring, L. [1993] “The analysis of observed chaotic data in physical systems,” Reviews of Modern Physics 65, 1331–1392. Baker, G. & Gollub, P. [1990] Chaotic Dynamics: An Introduction (Cambridge University Press, Cam- bridge). Barreira, L. & Pesin, Y. [2001] “Lectures on Lyapunov exponents and smooth ergodic theory,” Proc. of. Symposia in Pure Math. (AMS), pp. 3–89. Benettin, G., Galgani, L., Giorgilli, A. & Strelcyn, J.-M. [1980] “Lyapunov characteristic exponents for smooth dynamical systems and for hamiltonian systems. A method for computing all of them. Part II: Numerical application,” Meccanica 15, 21–30. April 5, 2018 0:22 LE-FO-arxiv REFERENCES 9 Bylov, B. E., Vinograd, R. E., Grobman, D. M. & Nemytskii, V. V. [1966] Theory of characteristic exponents and its applications to problems of stability (in Russian) (Nauka, Moscow). Caputo, M. [1967] “Linear models of dissipation whose Q is almost frequency independent-II,” Geophysical Journal of the Royal Astronomical Society 13, 529–539. Christiansen, F. & Rugh, H. [1997] “Computing Lyapunov spectra with continuous gram-schmidt orthonor- malization,” Nonlinearity 10, 1063–1072. Cvitanovi´ c, P., Artuso, R., Mainieri, R., Tanner, G. & Vattay, G. [2016] Chaos: Classical and Quantum (Niels Bohr Institute, Copenhagen), http://ChaosBook.org. Czornik, A., Nawrat, A. & Niezabitowski, M. [2013] “Lyapunov exponents for discrete time-varying sys- tems,” Studies in Computational Intel ligence 440, 29–44. Danca, M.-F. [2016] “Hidden transient chaotic attractors of Rabinovich–Fabrikant system,” Nonlinear Dynamics 86, 1263–1270. Danca, M.-F., Feˇ ckan, M., Kuznetsov, N. & Chen, G. [2018a] “Complex dynamics, hidden attractors and continuous approximation of a fractional-order hyperchaotic PWC system,” Nonlinear Dynamics 91, 2523–2540. Danca, M.-F., Feˇ ckan, M., Kuznetsov, N. V. & Chen, G. [2018b] “Fractional-order PWC systems without zero Lyapunov exponents,” Nonlinear Dynamics Doi:10.1007/s11071-018-4108-2. Diethelm, K. [2014] “An extension of the well-posedness concept for fractional differential equations of caputoŠs type,” Applicable Analysis 93, 2126–2135. Diethelm, K. & Ford, N. [2002] “Analysis of fractional differential equations,” Journal of Mathematical Analysis and Applications 265, 229–248. Diethelm, K., Ford, N. & Freed, A. [2002] “A predictor-corrector approach for the numerical solution of fractional differential equations,” Nonlinear Dynamics 29, 3–22. Eckmann, J. P., Kamphorst, S. O., Ruelle, D. & Ciliberto, S. [1986] “Liapunov exponents from time series,” Phys. Rev. A 34, 4971–4979. Eckmann, J.-P. & Ruelle, D. [1985] “Ergodic theory of chaos and strange attractors,” Reviews of Modern Physics 57, 617–656. Frederickson, P., Kaplan, J., Yorke, E. & Yorke, J. [1983] “The Liapunov dimension of strange attractors,” Journal of Differential Equations 49, 185–207. Garrappa, R. [2012] “Predictor-corrector PECE method for fractional differen- tial equations,” https://www.mathworks.com/matlabcentral/fileexchange/32918-predictor-corrector- pece-method-for-fractional-differential-equations?focused=5235405&tab=function. Gorenflo, R. & Mainardi, F. [1997] “Fractional Calculus,” Fractals and Fractional Calculus in Continuum Mechanics. International Centre for Mechanical Sciences (Courses and Lectures), Vol. 8 (Springer, Vienna). Govorukhin, V. [2004] “Calculation Lyapunov exponents for ODE,” http://www.math.rsu.ru/mexmat/kvm/matds/. Hartley, T., Lorenzo, C., Trigeassou, J. & Maamri, N. [2013] “Equivalence of history-function based and infinite-dimensional-state initializations for fractional-order operators,” ASME. J. Comput. Nonlinear Dynam. 8, art. num. 041014. Hegger, R., Kantz, H. & Schreiber, T. [1999] “Practical implementation of nonlinear time series methods: The TISEAN package,” Chaos 9, 413–435. Hilbert, D. [1901-1902] “Mathematical problems,” Bul l. Amer. Math. Soc. , 437–479. Kilbas, A. & Trujillo, J. [2001] “Differential equations of fractional order:methods results and problem Ůi,” Applicable Analysis 78, 153–192. Kuznetsov, N., Alexeeva, T. & Leonov, G. [2016] “Invariance of Lyapunov exponents and Lyapunov dimension for regular and irregular linearizations,” Nonlinear Dynamics 85, 195–201, doi:10.1007/ s11071-016-2678-4. Kuznetsov, N., Leonov, G., Mokaev, T., Prasad, A. & Shrimali, M. [2018a] “Finite-time Lya- punov dimension and hidden attractor of the Rabinovich system,” Nonlinear dynamics https://doi.org/10.1007/s11071-018-4054-z. Kuznetsov, N., Leonov, G., Mokaev, T., Prasad, A. & Shrimali, M. [2018b] “Finite-time Lya- April 5, 2018 0:22 LE-FO-arxiv 10 REFERENCES punov dimension and hidden attractor of the Rabinovich system,” Nonlinear dynamics https://doi.org/10.1007/s11071-018-4054-z. Leonov, G. & Kuznetsov, N. [2007] “Time-varying linearization and the Perron effects,” International Journal of Bifurcation and Chaos 17, 1079–1107, doi:10.1142/S0218127407017732. Leonov, G. & Kuznetsov, N. [2015] “On differences and similarities in the analysis of Lorenz, Chen, and Lu systems,” Applied Mathematics and Computation 256, 334–343, doi:10.1016/j.amc.2014.12.132. Leonov, G., Kuznetsov, N., Korzhemanova, N. & Kusakin, D. [2016] “Lyapunov dimension formula for the global attractor of the Lorenz system,” Communications in Nonlinear Science and Numerical Simulation 41, 84–103, doi:10.1016/j.cnsns.2016.04.032. Li, C., Gong, Z., Qian, D. & Chen, Y. [2010] “On the bound of the Lyapunov exponents for the fractional differential systems,” Chaos 20, art. num. 016001CHA. Lyapunov, A. [1892] The General Problem of the Stability of Motion (in Russian) (Kharkov), [English transl.: Academic Press, NY, 1966]. Oldham, K. & Spanier, J. [1974] The Fractional Calculus: Theory and Applications of Differentiation and Integration to Arbitrary Order (Academic Press, New York). Oseledets, V. [1968] “A multiplicative ergodic theorem. Characteristic Ljapunov, exponents of dynamical systems,” Trudy Moskovskogo Matematicheskogo Obshchestva (in Russian) 19, 179–210. Pikovsky, A. & Politi, A. [2016] Lyapunov Exponents: A Tool to Explore Complex Dynamics (Cambridge University Press). Podlubny, I. [1999] Fractional Differential Equations: An Introduction to Fractional Derivatives, Fractional Differential Equations, to Methods of their Solution and some of their Applications, Lecture Notes in Applied and Computational Mechanics (Academic Press, San Diego). Podlubny, I. [2002] “Geometric and physical interpretation of fractional integration and fractional differ- entiation,” Fractional Calculus and Applied Analysis 5, 367–386. Ramasubramanian, K. & Sriram, M. [2000] “A comparative study of computation of Lyapunov spectra with different algorithms,” Physica D: Nonlinear Phenomena 139, 72–86. Sarra, S. & Meador, C. [2011] “On the numerical solution of chaotic dynamical systems using extend precision floating point arithmetic and very high order numerical methods,” Nonlinear Analysis 16, 340Ű352. Shimada, I. & Nagashima, T. [1979] “A numerical approach to ergodic problem of dissipative dynamical systems,” Progress of Theoretical Physics 61, 1605–1616. Skokos, C. [2010] “The Lyapunov characteristic exponents and their computation,” Lecture Notes in Physics 790, 63–135. Stuart, A. & Humphries, A. [1998] Dynamical Systems and Numerical Analysis (Cambridge University Press). Tavazoei, M. & Haeri, M. [2009] “A proof for non existence of periodic solutions in time invariant fractional order systems,” Automatica 45, 1886 – 1890. Vallejo, J. & Sanjuan, M. [2017] Predictability of Chaotic Dynamics: A Finite-time Lyapunov Exponents Approach (Springer, Cham, Switzerland). Wang, P., Li, J. & Li, Q. [2012] “Computational uncertainty and the application of a high-performance multiple precision scheme to obtaining the correct reference solution of lorenz equations,” Numerical Algorithms 59, 147–159. Wolf, A., Swift, J. B., Swinney, H. L. & Vastano, J. A. [1985a] “Determining Lyapunov exponents from a time series,” Physica D: Nonlinear Phenomena 16, 285–317. Wolf, A., Swift, J. B., Swinney, H. L. & Vastano, J. A. [1985b] “Determining Lyapunov exponents from a time series,” Physica D: Nonlinear Phenomena 16, 285–317. Zhou, Y. [2016] Fractional Evolution Equations and Inclusions: Analysis and Control (Academic Press). April 5, 2018 0:22 LE-FO-arxiv REFERENCES 11 Fig. 1. (a) A chaotic attractor of the RF system of FO, for q = 0.999. (b) Dynamics of the LEs. Fig. 2. (a) An apparently stable cycle of the generalized Lorenz system of FO, for q = 0.985. (b) Dynamics of the LEs. April 5, 2018 0:22 LE-FO-arxiv 12 REFERENCES Fig. 3. Dynamics of the LEs of the 4-dimensional system of FO (4). Fig. 4. Perturbation and rescaling of a nearby trajectory, after every h steps, considered as multiple of h (here h = norm norm 2h, sketch). April 5, 2018 0:22 LE-FO-arxiv REFERENCES 13 Fig. 5. LEs of RF system. (a) LEs as function of q for q ∈ [0.9, 1]. (b) LEs as function of p for p ∈ [1.1, 1.3]. Fig. 6. LEs of Chen’s system of FO represented by function of two variables: q and parameter p. Surfaces S , for i = 1, 2, 3, represents LE(q, p). April 5, 2018 0:22 LE-FO-arxiv 14 REFERENCES Appendices A. Program for LEs of FO % Memory allocation function [t,LE]=FO_Lyapunov(ne,ext_fcn,t_start, x=zeros(ne*(ne+1),1); h_norm,t_end,x_start,h,q,out); x0=x; c=zeros(ne,1); gsc=c; zn=c; % Program to compute LEs of systems of FO. n_it = round((t_end-t_start)/h_norm); % The program uses a fast variant of the % Initial values % predictor-corrector Adams-Bashforth-Moulton, x(1:ne)=x_start; % "FDE12.m" for FDEs, by Roberto Garrappa: i=1; while i<=ne % https://goo.gl/XScYmi x((ne+1)*i)=1.0; i=i+1; % m-file required: FDE12 and end % the function containing the extended system t=t_start; % (see e.g. LE_RF.m). % Main loop it=1; % FO_Lyapunov.m was developed, by while it<=n_it % modifying the program Lyapunov.m, % Solution of extended ODE system % by V.N. Govorukhin: [T,Y] = FDE12(q,ext_fcn,t,t+h_norm,x,h); t=t+h_norm; % https://goo.gl/wZVCtg Y=transpose(Y); x=Y(size(Y,1),:); %solution at t+h_norm % FO_Lyapunov.m, FDE12.m and LE_RF.m i=1; % must be in the same folder. while i<=ne j=1; % How to use it: while j<=ne; % [t,LE]=FO_Lyapunov(ne,ext_fcn,t_start,... x0(ne*i+j)=x(ne*j+i); % h_norm,t_end,x_start,h,q,out); j=j+1; end; % Input: i=i+1; % ne - system dimension; end; % ext_fcn - function containing the extended % orthonormal Gram-Schmidt basis % system; zn(1)=0.0; % t_start, t_end - time span; j=1; % h_norm - step for Gram-Schmidt while j<=ne % renormalization; zn(1)=zn(1)+x0(ne*j+1)*x0(ne*j+1); % x_start - initial condition; j=j+1; % outp - priniting step of LEs; end; % ioutp==0 - no print. zn(1)=sqrt(zn(1)); j=1; % Output: while j<=ne % t - time values; x0(ne*j+1)=x0(ne*j+1)/zn(1); % LE Lyapunov exponents to each time value. j=j+1; end % Example of use for the RF system: j=2; % [t,LE]=FO_Lyapunov(3,@LE_RF,0,0.02,300,... while j<=ne % [0.1;0;1;0.1],0.005,0.999,1000); k=1; while k<=j-1 % The program is presented in: gsc(k)=0.0; l=1; % Marius-F. Danca and N. Kuznetsov, while l<=ne; % Matlab code for Lyapunov exponents of gsc(k)=gsc(k)+x0(ne*l+j)*x0(ne*l+ % fractional order systems k); l=l+1; end hold on; k=k+1; end April 5, 2018 0:22 LE-FO-arxiv REFERENCES 15 k=1; B. Function LE RF.m while k<=ne l=1; function f=LE_RF(t,x) while l<=j-1 x0(ne*k+j)=x0(ne*k+j)-gsc(l)*x0( %Output data must be a column vector ne*k+l); f=zeros(size(x)); l=l+1; end %variables allocated to the variational equations k=k+1; X= [x(4), x(7), x(10); end; x(5), x(8), x(11); zn(j)=0.0; x(6), x(9), x(12)]; k=1; while k<=ne %RF equations zn(j)=zn(j)+x0(ne*k+j)*x0(ne*k+j); f(1)=x(2)*(x(3)-1+x(1)*x(1))+0.1*x(1); k=k+1; f(2)=x(1)*(3*x(3)+1-x(1)*x(1))+0.1*x(2); end f(3)=-2*x(3)*(0.98+x(1)*x(2)); zn(j)=sqrt(zn(j)); k=1; %Jacobian matrix while k<=ne J=[2*x(1)*x(2)+0.1, x(1)*x(1)+x(3)-1, x(2); x0(ne*k+j)=x0(ne*k+j)/zn(j); -3*x(1)*x(1)+3*x(3)+1,0.1,3*x(1); k=k+1; -2*x(2)*x(3),-2*x(1)*x(3),-2*(x(1)*x(2)+0.98)]; end j=j+1; %Righthand side of variational equations end f(4:12)=J*X; % To be modified if ne>3 % update running vector magnitudes k=1; while k<=ne; C. Program for LEs as function on p c(k)=c(k)+log(zn(k)); k=k+1; function run_Lyapunov_p(ne,ext_fcn,t_start,h_norm, end; % normalize exponent t_end,x_start,h,q,p_min,p_max,n); k=1; hold on; while k<=ne p_step=(p_max-p_min)/n LE(k)=c(k)/(t-t_start); p=p_min; k=k+1; while p<=p_max end LE=FO_Lyapunov_p(ne,ext_fcn,t_start,h_norm,t_end i=1; ,x_start,h,q,p); while i<=ne p=p+p_step j=1; plot(p,LE); while j<=ne; end x(ne*j+i)=x0(ne*i+j); j=j+1; end D. Function LE RF p.m i=i+1; end; x=transpose(x); function f=LE_RF_p(t,x,p) it=it+1; %p is the parameter % print and plot the results f=zeros(size(x)); if (mod(it,out)==0) % (*) X= [x(4), x(7), x(10); fprintf(’%10.2f %10.4f %10.4f... x(5), x(8), x(11); %10.4f\n’,[t,LE]); % (*) x(6), x(9), x(12)]; end; % (*) %RF equations plot(t,LE) % (*) f(1)=x(2)*(x(3)-1+x(1)*x(1))+0.1*x(1); end f(2)=x(1)*(3*x(3)+1-x(1)*x(1))+0.1*x(2); % displays the box outline around axes f(3)=-2*x(3)*(p+x(1)*x(2)); xlabel(’t’,’fontsize’,16) % (*) %Jacobian matrix ylabel(’LEs’,’fontsize’,14) % (*) J=[2*x(1)*x(2)+0.1, x(1)*x(1)+x(3)-1, x(2); set(gca,’fontsize’,14)% (*) -3*x(1)*x(1)+3*x(3)+1,0.1,3*x(1); box on;% (*) -2*x(2)*x(3),-2*x(1)*x(3),-2*(x(1)*x(2)+p)]; line([0,t],[0,0],’color’,’k’)% (*) f(4:12)=J*X; % To be modified if ne>3 April 5, 2018 0:22 LE-FO-arxiv 16 REFERENCES E. Program for LEs as function of q while q<q_max [t,LE]=FO_Lyapunov_q(ne,ext_fcn,t_start,h_norm, t_end,x_start,h,q); q=q+q_step; function run_FO_Lyapunov_q(ne,ext_fcn,t_start, fprintf(’q=%10.4f\n %10.4f’, q); h_norm,t_end,x_start,h,q_min,q_max,n); plot(q,LE); hold on; end q_step=(q_max-q_min)/n; q=q_min; http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Nonlinear Sciences arXiv (Cornell University)

Matlab code for Lyapunov exponents of fractional order systems

Nonlinear Sciences , Volume 2018 (1804) – Mar 29, 2018

Loading next page...
 
/lp/arxiv-cornell-university/matlab-code-for-lyapunov-exponents-of-fractional-order-systems-5100K0gl1N

References (50)

ISSN
0218-1274
eISSN
ARCH-3338
DOI
10.1142/S0218127418500670
Publisher site
See Article on Publisher Site

Abstract

April 5, 2018 0:22 LE-FO-arxiv MARIUS-F. DANCA Romanian Institute od Science and Technology, 400487 Cluj-Napoca, Romania, danca@rist.ro NIKOLAY KUZNETSOV Dept. of Applied Cybernetics, Saint-Petersburg State University, Russia and Dept. of Mathematical Information Technology, University of Jyv¨askyl¨a, Finland nkuznetsov239@gmail.com Received (to be inserted by publisher) In this paper the Benettin-Wolf algorithm to determine all Lyapunov exponents for a class of fractional-order systems modeled by Caputo's derivative and the corresponding Matlab code are presented. First it is proved that the considered class of fractional-order systems admits the necessary variational system necessary to find the Lyapunov exponents. The underlying nu- merical method to solve the extended system of fractional order, composed of the initial value problem and the variational system, is the predictor-corrector Adams-Bashforth-Moulton for fractional differential equations. The Matlab program prints and plots the Lyapunov exponents as function of time. Also, the programs to obtain Lyapunov exponents as function of the bifur- cation parameter and as function of the fractional order are described. The Matlab program for Lyapunov exponents is developed from an existing Matlab program for Lyapunov exponents of integer order. To decrease the computing time, a fast Matlab program which implements the Adams-Bashforth-Moulton method, is utilized. Four representative examples are considered. Keywords: Lyapunov exponents, Benettin-Wolf algorithm, Fractional-order dynamical system 1. Introduction Despite a long history, the doubts that fractional- ity of solutions of FDEs with general nonlinearities order (FO) derivatives have no clear geometrical in- see, for example, [Oldham & Spanier, 1974; Caputo, terpretations (see e.g. [Podlubny, 2002]), was one 1967], and [Diethelm & Ford, 2002; Kilbas & Tru- of the several reasons that fractional calculus was jillo, 2001; Podlubny, 1999]. not used in physics or engineering. However, dur- The Lyapunov exponents (LEs) measure the aver- ing the last more than 10 years, fractional calcu- age rate of divergence or convergence of orbits start- lus starts to attract increasing attention. There are ing from nearby initial points. Therefore, they can nowadays more and more works on FO systems be used to analyze the stability of limits sets and and their related applications in physics, engineer- to check sensitive dependence on initial conditions, ing, mathematics, finance, chemistry, and so on. For that is, the presence of potential chaotic attractors. the theory on the existence, uniqueness, continuous On the other side, in [Cvitanovi´ c et al., 2016] Cvi- dependence on parameters and asymptotic stabil- tanovi´ c et al. do not recommend the evaluation arXiv:1804.01143v1 [physics.comp-ph] 29 Mar 2018 April 5, 2018 0:22 LE-FO-arxiv 2 M.-F. Danca, N.V. Kuznetsov of the LEs and recommend to: “Compute stabil- tractors in the case of multistability (see, e.g. such ity exponents and the associated covariant vectors corresponding examples for the classical Lorenz sys- instead. Cost less and gets you more insight. [...] we tem and Henon map [Leonov et al., 2016; Kuznetsov are doubtful of their utility as means of predicting et al., 2018b]). any observables of physical significance”. Moreover, In order to study the chaoticity of an attractor in we additionally note here, Perron’s counterexample numerical experiments, one has to consider a grid [Leonov & Kuznetsov, 2007] which shows actually of point covering the attractor and compute corre- that the use of LEs, obtained via the linearization sponding finite-time local LEs for a certain time. procedure, for the study of the behaviour of non- Remark that while the time series obtained from linear system requires a rigorous justification. How- a physical experiment are assumed to be reliable ever, determining LEs remains the subject of many on the whole considered time interval, the time se- works and grown into a real software industry for ries produced by the integration of mathematical modern nonlinear physics (see, e.g.[Hegger et al., dynamical model can be reliable on a limited time 1999; Barreira & Pesin, 2001; Skokos, 2010; Czornik interval only due to computational errors. et al., 2013; Pikovsky & Politi, 2016; Vallejo & San- Taking into account the above discussion further juan, 2017] and others). we consider finite-time local LEs and their compu- Nowadays there are two widely used definitions , tation in Matlab by the analog of the Bennetin-Wolf of the LEs: via the exponential growth rates of algorithm. norms of the fundamental matrix columns [Lya- punov, 1892] and via the exponential growth rates 2. Benettin-Wolf Algorithm for LEs of the sigular values of fundamental matrix [Os- of FO eledets, 1968]. Corresponding approaches for the The determination of LEs of a system of integer LEs computation and their difference are discussed, or fractional order with the Benettin-Wolf algo- e.g., in [Kuznetsov et al., 2018a, 2016]. rithm, requires the numerical integration of differen- Remark that in numerical experiments we can con- tial equations of integer or fractional order. Because sider only finite time, and, thus, the numerically the purpose of this paper is to present a Matlab computed values of LEs can differ significantly from code for LEs of systems of FO, in the next subsec- the limit values (e.g. if the considered trajectory tions, only the most important steps (such as the belongs to a transient chaotic set ), and are often existence of variational equations of FO) used to referred as finite-time LEs. implement the algorithm in Matlab language are Applying the statistical physics approach and as- presented (theoretical details can be found in the suming the ergodicity (see, e.g. [Oseledets, 1968]), related references). the LEs for a given dynamical system are often es- timated by local LEs along a “typical” trajectory. However, in numerical experiments, the rigorous use 2.1. Numerical integration of FDEs of the ergodic theory is a challenging task (see, e.g. The autonomous FO systems considered in this pa- [Cvitanovi´ c et al., 2016, p.118]). If the LEs are the per are modeled by the following Initial Value Prob- same for any trajectory, then Frederickson et al. lem (IVP) with Caputo’s derivative [Frederickson et al., 1983, p.190] suggested to call D x = f (x), them as absolute ones and wrote that such abso- (1) x(0) = x , lute values rarely exist. For example, substantially n n different values of the local LEs can be obtained for t ∈ [0, T ], q ∈ (0, 1), f : R → R and D , Ca- along trajectories on coexisting nonsymmetric at- puto’s differential operator of order q with starting Relying on the Oseledec ergodic theorem [Oseledets, 1968] the above definitions often do not differ (see, e.g. Eckmann & Ruelle [Eckmann & Ruelle, 1985, p.620,p.650], Wolf et al. [Wolf et al., 1985a, p.286,p.290-291], and Abarbanel et al. [Abarbanel et al., 1993, p.1363,p.1364]), however in general case, they may lead to different values [Kuznetsov et al., 2018a],[Bylov et al., 1966, p.289],[Leonov & Kuznetsov, 2007, p.1083]. While trivial attractors (stable equilibrium points) can be easily found analytically or numerically, the search of all periodic and chaotic attractors for a given system is a challenging problem. See, e.g. famous 16th Hilbert problem [Hilbert, 1901-1902] on the number of coexisting periodic attractors in two dimensional polynomial systems, which was formulated in 1900 and is still unsolved, and its generalization for multidimensional systems with chaotic attractors [Leonov & Kuznetsov, 2015]. Based on philosophical arguments rather than a mathematical point of view, some researchers questioned the appropriateness April 5, 2018 0:22 LE-FO-arxiv Matlab code for Lyapunov exponents of fractional order systems 3 Let us next assume that we are working on a uni- point 0 form grid {t = jh : n = 0, 1, ..., N + 1} with some integer N with the step size h and the case q −q 0 D x(t) = (t− τ ) x (τ )dτ, of q ∈ (0, 1). Then, the predictor form, x , at the Γ(1− q) point t , is the fractional variant of the Adams- j+1 with Γ the known Euler function. Bashforth method Properties of the Caputo’s differential operator, D , are discussed in [Podlubny, 1999; Gorenflo & x (t ) = x + b f (x (t )) , n+1 0 j,n+1 j Γ (q) Mainardi, 1997]. j=0 Under Lipschitz continuity of the function f , the while the corrector formula (the fractional variant IVP (1) admits a unique solution [Diethelm et al., of the one-step implicit Adams Moulton method) 2002]. reads Remark 2.1. i ) In the case of an integer-order dynamical sys- x (t ) = x + f x (t ) n+1 0 n+1 Γ(q + 2) tem, denoting the solution of the underlying IVP as x(t, x ), one has x ◦ x = x (see e.g. [Zhou, h 0 s t t+s + a f (x (t )) , j,n+1 j 2016]). Due to the memory dependence of the Γ(q + 2) j=0 derivatives, this does not hold in the case of sys- where a and b are the corrector and predictor tems modeled by FDEs. However, motivated by the weights respectively given by the following formula numerical character of this paper, the definition of integer order dynamical systems which states that if the underlying IVP admits unique solutions exist- q+1 n − (n− q) (n + 1) if j = 0, ing on infinite time interval, the problem defines a  q+1 q+1 (n− j + 2) + (n− j) dynamical system (see [Stuart & Humphries, 1998, a = j,n+1 q+1 −2 (n− j + 1) if 1 ≤ j ≤ n, Definition 2.1.2]) is adopted. 1 if j = n + 1, ii ) Even fractional-order dynamics describe a real object more accurately than classical integer-order and dynamics, systems modeled by the IVP (1) cannot q q q b = ((n + 1− j) − (n− j) ) . have any non-constant periodic solution (see e.g. j,n+1 [Tavazoei & Haeri, 2009]). However, a solution may be asymptotically periodic [Danca et al., 2018a]. The predictor-corrector ABM method has an error These trajectories are called numerical ly periodic, which is roughly proportional to h . Thus, to ob- in the sense that the trajectory, from numerical tain an error of, e.g., 1.0E − 6, a step size close to point of view, can be an extremely-near periodic h = 1.0E − 3 should be considered. with respect to, e.g., Euclidean norm [Danca et al., Because due to the long memory processes, 2018a]. A numerically periodic trajectory refers to the utilized ABM method as described in [Diethelm as a closed trajectory in the phase space in the sense et al., 2002], is time consuming. Therefore, a fast op- that the closing error is within a given bound of timized ABM method, FDE12.m [Garrappa, 2012] is 1E − n, with n being a sufficiently large positive used. integer. FDE12.m is called by the following command line: The numerical integrations required by the LEs al- [t,x] = FDE12(q,fcn,t0,tf,x0,h); gorithm for FO systems are performed in this pa- per with the predictor-corrector Adams-Bashforth- Moulton (ABM) method for FDEs, proposed Di- where q represents the commensurate fractional- ethelm et al. [Diethelm et al., 2002] which is con- order, fcn.m is the file with the function to be inte- structed for the fully general set of equations with- grated, t0 and tf define the time span, x0 are the out any special assumptions, being easy to imple- initial conditions, and h represents the integration ment in any language. step-size. of using initial conditions of the classical form in the Caputo derivative [Hartley et al., 2013]. However, it should be emphasized that, in practical (physical) problems, physically interpretable initial conditions are necessary and Caputo's derivative is a fully justified tool [Diethelm, 2014]. April 5, 2018 0:22 LE-FO-arxiv 4 M.-F. Danca, N.V. Kuznetsov For example, consider the integration of the FO first work to propose a Gram-Schmidt orthogonal- Rabinovich-Fabrikant (RF) system [Danca, 2016], ization procedure to compute LEs for continuous which has the following Matlab function, RF.m systems of integer order, as described in [Eckmann & Ruelle, 1985]), and by Wolf et al. [Wolf et al., 1985b] (see also [Eckmann et al., 1986]). function dx = RF(t,x) dx=[x(2)*(x(3)-1+x(1)*x(1))+0.1*x(1); The algorithm to find all LEs, described as a For- x(1)*(3*x(3)+1-x(1)*x(1))+0.1*x(2); tran code by Wolf et al. [Wolf et al., 1985b], and also -2*x(3)*(p+x(1)*x(2))]; as a Basic code in [Baker & Gollub, 1990], solves the equations of motion under perturbations and peri- with the bifurcation parameter p = 0.98. odic orthonormalization. With the following parameters, one obtains the To note that the accuracy and reliability of nu- chaotic attractor in Fig. 1 (a): merically determined LEs depend on initial condi- tions, on the selection of the perturbations, perfor- [t,x]=FDE12(0.999,@RF,0,1500,[0.1;0.1;0.1],... mances of the utilized integration numerical method 0.01); and also on orthonormalization step size. A long- time numerical calculation of the leading Lyapunov exponent requires rescaling the distance between To note that FDE12.m requires the entry, x0, as nearby trajectories, in order to keep the separation a column vector [x10,x20,x30]’ or, similarly, within the linearized flow range. To avoid overflow, [x10;x20;x30]. Also, the returned exit, x, is a col- one calculates the divergence of nearby trajectories umn vector. for finite timesteps and renormalizes to unity af- ter a finite number of steps (Gram-Schmidt pro- 2.2. Algorithm for LEs of the FO cedure [Eckmann & Ruelle, 1985; Christiansen & system (1) Rugh, 1997]). Therefore, the main steps to determine numerically Notation 2.1. Hereafter the finite-time local LEs of the LEs are: numerical integration of the FO sys- FO are called LEs. tem (1) together with the variational system (2) The existence of the variational equations necessary (i.e. the extended system), Gram-Schmidt proce- to determine LEs is ensured by the following Theo- dure and picking up the exponents during the renor- rem [Li et al., 2010] malization procedure, the LEs being determined as the average of the logarithm of the stretching factor Theorem 1. System (1) has the fol lowing varia- of each perturbation, steps presented in Algorithm tional equations which define the LEs D Φ(t) = D f (x)Φ(t), ∗ x (2) Φ(0) = I, 3. The Matlab code for LEs where Φ is the matrix solution of the system (1), D Consider the following general assumptions: is the Jacobian of f and I is the identity matrix. Further we assume that for the matrix Φ(t) the co- • The considered systems, modeled by the IVP (1), cycle property takes place. are autonomous; Therefore, the algorithm to determine the LEs of • The system (1) is of commensurate order: q = the system (1) becomes similar to the case of inte- q = ... = q = q; 2 ne ger order. • In the case of chaotic behavior, the fractional- Lyapunov exponents measure the exponential order has been chosen close to 1, such that chaos growth, or decay, of infinitesimal phase-space per- is significant. turbations of a chaotic dynamical system. • The right hand side of system (1), f , is smooth The algorithm for numerical evaluation of LEs uti- enough. lized in this paper, has been proposed in the seminal • Because of the space restrictions, only the main works of Benettin et al. [Benettin et al., 1980] (see programs code are presented, and indications on also [Shimada & Nagashima, 1979]), one of the the how to write the other ones. The incommensurate case can be treated similarly, the only difference referring to the utilized numerical method for FDEs. April 5, 2018 0:22 LE-FO-arxiv Matlab code for Lyapunov exponents of fractional order systems 5 Algorithm 1: Algorithm for LEs of FO system (1) Input: -ne . number of equations -x start . n initial conditions of (1) -t start, t end . time span -h norm . Normalization step-size n it ← (t end− t start)/h norm . iterations number for i ← ne + 1 to ne(ne + 1) do x(i) = 1.0 . initial conditions of (2) end t ← t start for i ← 1 to n it do x ← integration of FO systems (1)-(2) t ← t + h norm zn(1), ..., zn(ne) ← Gram-Schmidt procedure s(1) ← 0 for k ← 1 to ne do s(k) ← s(k) + log(zn(k)) . vector magnitudes LE(k) ← s(k)/(t− t start) . LEs end end Output:LE A simple way to build in Matlab language the algo- it is necessary to solve the extended system (1)-(2) rithm for FO systems, the program FO Lyapunov.m of FO (yellow line) which is given in the function (Appendix A), was to modify either some exist- ext fcn.m. In this file, beside the right hand side ing program, e.g., the program lyapunov.m [Gov- function f of the system (1), the Jacobi matrix J orukhin, 2004], which is a Matlab variant of the should be included (see Appendix B where the func- original LEs algorithm proposed in [Benettin et al., tion for the RF system is presented). 1980] or [Wolf et al., 1985b] or, similarly, to trans- The program plots the time evolutions of the LEs. late in Matlab the BASIC program [Baker & Gol- lub, 1990], a close variant of the original algorithm) 4. Numerical tests or, also, the Fortran code proposed by Wolf et all in [Wolf et al., 1985b], and modify it for FDEs. Beyond numerical artifacts that might occur when The program, called FO Lyapunov.m, is launched numerically integrating a system of ODEs of integer with the following command line: order, notions such as “shadowing time” and “maxi- mally effective computational time” reveal that it is [t,LE]=FO_Lyapunov(ne,@ext_fcn,t_start,h_norm,... possible to have reliable numerical simulations only t_end,x_start,h,q,out); on a relative finite-time interval (see, e.g., [Sarra & Meador, 2011; Wang et al., 2012]). The case of FO systems is even more delicate. In this paper we have where ne represents the equations (and state vari- considered generally t ∈ [0, 300] and, for the Lorenz ables) number, ext fcn.m the function containing system, t ∈ [0, 500]. the extended system (1)-(2), t start and t end the time span, h norm the normalization step, x start 1. Let consider the function for the RF sys- the initial condition (as column vector), h the step tem, LE RF.m, which include the extended sys- size of FDE12.m, and out indicates the steps num- tem (1)-(2) (Appendix B). Because ne=3, be- ber when intermediate values of time and LEs are side the 3 variables x(1),x(2),x(3) required printed (for out=0, no intermediate results will be by the numerical solution of the original sys- printed out). tem (1), the matrix solution of the system (2) As shown in the algorithm for LEs (Algorithm 1), requires other more ne×ne=9 variables from April 5, 2018 0:22 LE-FO-arxiv 6 M.-F. Danca, N.V. Kuznetsov the total of ne(ne+1)=12 variables: x(1:12), ter p = 200, after some neglected transients, one f=zeros(size(x))=zeros(12), where f loads obtains an apparently stable cycle (see Fig. 2 (a) and Remark 2.1 (ii )). the first ne=3 righthand side expressions of sys- tem (1), and the ne×ne=9 righthand side expres- To obtain the LEs one write the following com- sions of the variational system (2). mand line: For example, for the RF system, with the follow- ing command line: [t,LE]=FO_Lyapunov(3,@LE_Lorenz,0,5,500,... [0.1;0.1;0.1],0.001,0.985,10); [t,LE]=FO_Lyapunov(3,@LE_RF,0,0.02,300,... [0.1;0.1;0.1],0.005,0.999,1000); which gives LE=(-0.0026, -0.0870, -1.6225). The function LE Lorenz.m can be one obtains the intermediary results printed ev- obtained similarly with LE RF.m. The time evo- ery out=1000 h norm steps, presented in Table lution of the LEs is presented in Table 2 (see 1 (see also Fig. 1 (b) where the dynamics of the also Fig. 2 (b)). LEs are drawn). 50.00 0.1759 -0.1591 -1.5683 100.00 0.0611 -0.1108 -1.6050 10.00 0.1611 0.0660 -2.1614 150.00 0.0346 -0.0927 -1.6300 20.00 0.1923 0.0069 -2.0503 200.00 0.0215 -0.0877 -1.6288 30.00 0.0984 -0.7397 -1.1817 250.00 0.0135 -0.0866 -1.6269 40.00 0.0248 0.0542 -1.9761 300.00 0.0082 -0.0865 -1.6255 50.00 0.0440 -0.0168 -1.9867 350.00 0.0043 -0.0866 -1.6244 60.00 0.0697 -0.0006 -2.0701 400.00 0.0014 -0.0867 -1.6236 70.00 0.0354 -0.0116 -2.0167 450.00 -0.0008 -0.0869 -1.6230 80.00 0.0331 -0.0618 -1.9360 500.00 -0.0026 -0.0870 -1.6225 90.00 0.0201 0.0125 -1.9754 100.00 0.0429 0.0112 -1.9794 Table 2. Time evolution of the LEs for the Lorenz 110.00 0.0337 0.0252 -1.9698 system. LE=(-0.0026, -0.0870, -1.6225). 120.00 0.0262 -0.0213 -1.9038 130.00 0.0624 -0.0043 -1.9537 3. Consider next the non-smooth 4-dimensional 140.00 0.0660 -0.0029 -1.9811 system [Danca et al., 2018a] 150.00 0.0645 -0.0010 -2.0008 160.00 0.0743 0.0018 -2.0304 D x = −x + x , ∗ 1 1 2 170.00 0.0710 -0.0009 -2.0394 D x = −x sgn(x ) + x , ∗ 2 3 1 4 180.00 0.0583 0.0139 -2.0548 (4) D x = |x |− a, ∗ 3 1 190.00 0.0678 0.0042 -2.0665 D x = −bx , 200.00 0.0662 -0.0217 -2.0498 ∗ 4 2 210.00 0.0628 -0.0155 -2.0622 with a = 1, b = 0.5 and q = 0.98. 220.00 0.0602 -0.0125 -2.0715 230.00 0.0570 -0.0222 -2.0666 With the command line: 240.00 0.0643 -0.0222 -2.0813 250.00 0.0639 -0.0079 -2.1020 [t,LE]=FO_Lyapunov(4,@LE_4d,0,0.02,300,... 260.00 0.0623 0.0045 -2.1121 [0.1;0.1;0.1;0.1],0.005,0.98,1000); 270.00 0.0630 0.0006 -2.0987 280.00 0.0551 -0.0036 -2.0771 290.00 0.0761 0.0009 -2.0937 one obtains LE=(0.1262, 0.0846, 0.0778, 300.00 0.0749 0.0018 -2.0850 -1.5244) (Fig. 3). For this example, ne=4, the number of the Table 1. For t ∈ [0, 300], the RF system has LE = variables is 20 and, therefore, compared with (0.0749, 0.0018, -2.0850) (last line, blue). systems with ne=3, the function file, LE 4d.m, 2. If one considers the Lorenz system must be modified accordingly (compare the red q line in LE RF.m). Also, in order to obtain the D x = σ(x(2)− x(1)), ∗ 1 q printed intermediated values of LEs, in the D x = −x(1)x(3) + px(1)− x(2) (3) ∗ 2 q FO Lyapunov.m, the fprintf command, must be D x = x(1)x(2)− βx(3); ∗ 3 modified by adding one supplementary specifier with q = 0.985 and the standard parameters %10.3f. σ = 10, β = 8/3 and the bifurcation parame- Because from the four LEs, the system admits April 5, 2018 0:22 LE-FO-arxiv Matlab code for Lyapunov exponents of fractional order systems 7 three positive LEs, it can be considered as hy- tional order q (program run Lyapunov q.m in perchaotic (see [Danca et al., 2018b] for a dis- Appendix E). The used program FO Lyapunov q, cussion about the number of positive LEs of hy- is obtained from FO Lyapunov with the following perchaotic systems). modifications: (a) The header line of the code is replaced with Remark 4.1. Note that because the system is the following line (to note that the input out is not smooth and also discontinuous, the correct no more necessary); numerical integration required for this system cannot be done without a previous smooth ap- function [t,LE]=FO_Lyapunov_q(ne,ext_fcn,... proximation [Danca et al., 2018a]. Thus, like all t_start,h_norm,t_end,x_start,h,q); numerical methods for FDEs which are designed for continuous dynamical systems, the integra- tor FDE12 cannot be utilized in this case with- (b) the printing and plotting lines (*) (red out the mentioned approximation. Also, without color) are deleted; a smooth approximation, the Jacobian matrix, (c) the rest of the lines remain the same; utilized in the Benettin-Wolf algorithm, cannot The function containing the extended system, be determined. ext fcn.m, does not require any modification. For example, for the RF system, with the com- 4. LEs can be also plotted as function of the mand bifurcation parameter p for p ∈ [p , p ] min max (program run Lyapunov p.m, Appendix C). run_FO_Lypaunov_q(3,@LE_RF,0,0.05,150,... The program uses a slightly modified vari- [0.1;0.1;0.1],0.002,0.9,1,800) ant of FO Lyapunov.m, which is called FO Lyapunov p.m. To obtain FO Lyapunov p.m from FO Lyapunov.m, the following modifications one obtains the LEs plotted in Fig. 4 (b). have to be done: Remark 4.2. The presented programs can (a) The header line of the code FO Lyapunov.m be optimized especially in the cases of is replaced with the following line (to note that FO Lyapunov p.m and FO Lyapunov q.m. A sim- the output t and the input out are no more nec- ple improvement was to use while loop (in- essary) stead for) which for these two programs re- duce substantially the computational time. How- function LE=FO_Lyapunov_p(ne,ext_fcn_p,... ever, every parameter p (or order q) step, several t_start,h_norm,t_end,x_start,h,q,p); constant parameters are shared between pro- grams, fact which slows significantly the pro- (b) the printing and plotting lines (*) (red grams. Therefore, supplementary optimization color) are deleted; should be done. (c) the rest of the lines remain the same; 6. Another potential application of the proposed The m-file containing the extended system, LEs algorithm, is to represent the LEs as func- ext fcn p.m, for the RF system, is presented in tion of two variables: the order q and the bifur- Appendix D. Passing the parameter p between cation parameter p. these codes, can be easily realized due of the fa- Let the Chen system cilities of the program FDE12 (see FDE12). D x = a(x − x ), ∗ 1 2 1 For example, for the RF system, for p ∈ [1.1, 1.3], D x = (p− a)x − x x + cx , ∗ 2 1 1 3 2 for 1000 values of p (n = 1000), with the follow- D x = x x − bx , ∗ 3 1 2 3 ing command with parameters a = 35, b = 3 and q and p run_FO_Lyapunov_p(3,@LE_RF_p,0,0.02,200,... variables. By considering S := LE(q, c), for [0.1;0.1;0.1],0.002,0.998,1.1,1.3,800) i = 1, 2, 3, for q ∈ [0.9, 1] and p ∈ [20, 30], the ob- tained surfaces are plotted in Fig. 5 (see [Danca one obtains the evolution of the LEs drawn in et al., 2018b] where the algorithm to obtain LEs Fig. 4 (a). as surfaces is described). One can see that there 5. LEs can be plotted also as function of the frac- exists a unique positive LE (red surface, S ) for 1 April 5, 2018 0:22 LE-FO-arxiv 8 REFERENCES all values of the considered parameter c but only ments of numerical integration are implemented, for some fractional order values q, for which S is with Benettin-Wolf algorithm, for integer but also situated over the horizontal plane LE=0, where for fractional order systems, two or three decimals LEs are zero. Also, one can see that S , which is are the maximal expected number of decimals. almost identic to S when the underlying LE1,2, One of the most important algorithm variable, are negative, becomes zero for a relative large is h norm, which determines the normalization values of q and p, when S separates from S moments influences the results. This dependence 2 1 and identify (with the underlying numerical er- can be also deduced from the example of the ror) with the plane LE=0. RF system. With h=0.005 and h norm=0.005 one obtains LEs=(0.0631, 0.0031, -2.0774), while 5. Conclusion and discussion with the same stapsize, h=0.005, but h norm=0.5, LEs=(0.0643, 0.0026, -1.8254). Starting from an existing variant for integer-order Since on our best knowledge, there is not criterion system, in this paper, based on the Benettin-Wolf to choose precisely h norm, the recommended way algorithm, we proposed a Matlab program to deter- would be to realize several tests to choose the value mine numerically finite-time local LEs for FO sys- for which slightly modifications does not change sig- tems. nificantly the results. Due to inherent numerical errors, the algorithm In the case of fixed step-size integration numerical should be utilized with precaution. As known, methods, like the considered FDE12, h norm can be among the numerical errors, the results depends chosen as depending on the step size h. Therefore, strongly on the initial conditions, time integration as Fig. 6 shows, h norm could be multiple of inte- interval and, especially, on the renormalization step gration step size h (ε represents the perturbation size h norm. between nearly two trajectories x and x ¯). The relative large numerical errors of the Benettin- Regarding the Gram-Schmidt procedure, in or- Wolf algorithm, which for the considered examples der to speed up the code, one can use the QR was of order of 1.E − 2, can be observed in the decomposition based on the Householder trans- cases when one knows that the system presents a formation: [Q, R] = qr(A). To have matrix R numerically periodic cycle (see Remark 2.1 (ii )), with positive diagonal elements, one can ad- when the maximum LE should be zero. For exam- ditionally use Q = Q*diag(sign(diag(R))); R = ple, the Lorenz system, which for p = 200 presents R*diag(sign(diag(R))). See also [Ramasubrama- such apparently stable cycle (see Fig. 2 (a)), has the nian & Sriram, 2000]. maximum LE zero only with two precise decimals. The integration step-size of FDE12, h, plays an im- Actually, in general, three precise decimals for zero portant role. As the documentation of the program LEs are extremely rare. Note that this zero value, specifies, there exists the possibility to increase the appears only for h norm=5 and a smaller integra- performances of the numerical integration, but in tion step size, h = 0.001 and only for a larger time time integration detriment. interval [0,500]. Therefore, if no special improve- Acknowledgements The work is done within Russian Science Foundation project 14-21-00041. References Abarbanel, H., Brown, R., Sidorowich, J. & Tsimring, L. [1993] “The analysis of observed chaotic data in physical systems,” Reviews of Modern Physics 65, 1331–1392. Baker, G. & Gollub, P. [1990] Chaotic Dynamics: An Introduction (Cambridge University Press, Cam- bridge). Barreira, L. & Pesin, Y. [2001] “Lectures on Lyapunov exponents and smooth ergodic theory,” Proc. of. Symposia in Pure Math. (AMS), pp. 3–89. Benettin, G., Galgani, L., Giorgilli, A. & Strelcyn, J.-M. [1980] “Lyapunov characteristic exponents for smooth dynamical systems and for hamiltonian systems. A method for computing all of them. Part II: Numerical application,” Meccanica 15, 21–30. April 5, 2018 0:22 LE-FO-arxiv REFERENCES 9 Bylov, B. E., Vinograd, R. E., Grobman, D. M. & Nemytskii, V. V. [1966] Theory of characteristic exponents and its applications to problems of stability (in Russian) (Nauka, Moscow). Caputo, M. [1967] “Linear models of dissipation whose Q is almost frequency independent-II,” Geophysical Journal of the Royal Astronomical Society 13, 529–539. Christiansen, F. & Rugh, H. [1997] “Computing Lyapunov spectra with continuous gram-schmidt orthonor- malization,” Nonlinearity 10, 1063–1072. Cvitanovi´ c, P., Artuso, R., Mainieri, R., Tanner, G. & Vattay, G. [2016] Chaos: Classical and Quantum (Niels Bohr Institute, Copenhagen), http://ChaosBook.org. Czornik, A., Nawrat, A. & Niezabitowski, M. [2013] “Lyapunov exponents for discrete time-varying sys- tems,” Studies in Computational Intel ligence 440, 29–44. Danca, M.-F. [2016] “Hidden transient chaotic attractors of Rabinovich–Fabrikant system,” Nonlinear Dynamics 86, 1263–1270. Danca, M.-F., Feˇ ckan, M., Kuznetsov, N. & Chen, G. [2018a] “Complex dynamics, hidden attractors and continuous approximation of a fractional-order hyperchaotic PWC system,” Nonlinear Dynamics 91, 2523–2540. Danca, M.-F., Feˇ ckan, M., Kuznetsov, N. V. & Chen, G. [2018b] “Fractional-order PWC systems without zero Lyapunov exponents,” Nonlinear Dynamics Doi:10.1007/s11071-018-4108-2. Diethelm, K. [2014] “An extension of the well-posedness concept for fractional differential equations of caputoŠs type,” Applicable Analysis 93, 2126–2135. Diethelm, K. & Ford, N. [2002] “Analysis of fractional differential equations,” Journal of Mathematical Analysis and Applications 265, 229–248. Diethelm, K., Ford, N. & Freed, A. [2002] “A predictor-corrector approach for the numerical solution of fractional differential equations,” Nonlinear Dynamics 29, 3–22. Eckmann, J. P., Kamphorst, S. O., Ruelle, D. & Ciliberto, S. [1986] “Liapunov exponents from time series,” Phys. Rev. A 34, 4971–4979. Eckmann, J.-P. & Ruelle, D. [1985] “Ergodic theory of chaos and strange attractors,” Reviews of Modern Physics 57, 617–656. Frederickson, P., Kaplan, J., Yorke, E. & Yorke, J. [1983] “The Liapunov dimension of strange attractors,” Journal of Differential Equations 49, 185–207. Garrappa, R. [2012] “Predictor-corrector PECE method for fractional differen- tial equations,” https://www.mathworks.com/matlabcentral/fileexchange/32918-predictor-corrector- pece-method-for-fractional-differential-equations?focused=5235405&tab=function. Gorenflo, R. & Mainardi, F. [1997] “Fractional Calculus,” Fractals and Fractional Calculus in Continuum Mechanics. International Centre for Mechanical Sciences (Courses and Lectures), Vol. 8 (Springer, Vienna). Govorukhin, V. [2004] “Calculation Lyapunov exponents for ODE,” http://www.math.rsu.ru/mexmat/kvm/matds/. Hartley, T., Lorenzo, C., Trigeassou, J. & Maamri, N. [2013] “Equivalence of history-function based and infinite-dimensional-state initializations for fractional-order operators,” ASME. J. Comput. Nonlinear Dynam. 8, art. num. 041014. Hegger, R., Kantz, H. & Schreiber, T. [1999] “Practical implementation of nonlinear time series methods: The TISEAN package,” Chaos 9, 413–435. Hilbert, D. [1901-1902] “Mathematical problems,” Bul l. Amer. Math. Soc. , 437–479. Kilbas, A. & Trujillo, J. [2001] “Differential equations of fractional order:methods results and problem Ůi,” Applicable Analysis 78, 153–192. Kuznetsov, N., Alexeeva, T. & Leonov, G. [2016] “Invariance of Lyapunov exponents and Lyapunov dimension for regular and irregular linearizations,” Nonlinear Dynamics 85, 195–201, doi:10.1007/ s11071-016-2678-4. Kuznetsov, N., Leonov, G., Mokaev, T., Prasad, A. & Shrimali, M. [2018a] “Finite-time Lya- punov dimension and hidden attractor of the Rabinovich system,” Nonlinear dynamics https://doi.org/10.1007/s11071-018-4054-z. Kuznetsov, N., Leonov, G., Mokaev, T., Prasad, A. & Shrimali, M. [2018b] “Finite-time Lya- April 5, 2018 0:22 LE-FO-arxiv 10 REFERENCES punov dimension and hidden attractor of the Rabinovich system,” Nonlinear dynamics https://doi.org/10.1007/s11071-018-4054-z. Leonov, G. & Kuznetsov, N. [2007] “Time-varying linearization and the Perron effects,” International Journal of Bifurcation and Chaos 17, 1079–1107, doi:10.1142/S0218127407017732. Leonov, G. & Kuznetsov, N. [2015] “On differences and similarities in the analysis of Lorenz, Chen, and Lu systems,” Applied Mathematics and Computation 256, 334–343, doi:10.1016/j.amc.2014.12.132. Leonov, G., Kuznetsov, N., Korzhemanova, N. & Kusakin, D. [2016] “Lyapunov dimension formula for the global attractor of the Lorenz system,” Communications in Nonlinear Science and Numerical Simulation 41, 84–103, doi:10.1016/j.cnsns.2016.04.032. Li, C., Gong, Z., Qian, D. & Chen, Y. [2010] “On the bound of the Lyapunov exponents for the fractional differential systems,” Chaos 20, art. num. 016001CHA. Lyapunov, A. [1892] The General Problem of the Stability of Motion (in Russian) (Kharkov), [English transl.: Academic Press, NY, 1966]. Oldham, K. & Spanier, J. [1974] The Fractional Calculus: Theory and Applications of Differentiation and Integration to Arbitrary Order (Academic Press, New York). Oseledets, V. [1968] “A multiplicative ergodic theorem. Characteristic Ljapunov, exponents of dynamical systems,” Trudy Moskovskogo Matematicheskogo Obshchestva (in Russian) 19, 179–210. Pikovsky, A. & Politi, A. [2016] Lyapunov Exponents: A Tool to Explore Complex Dynamics (Cambridge University Press). Podlubny, I. [1999] Fractional Differential Equations: An Introduction to Fractional Derivatives, Fractional Differential Equations, to Methods of their Solution and some of their Applications, Lecture Notes in Applied and Computational Mechanics (Academic Press, San Diego). Podlubny, I. [2002] “Geometric and physical interpretation of fractional integration and fractional differ- entiation,” Fractional Calculus and Applied Analysis 5, 367–386. Ramasubramanian, K. & Sriram, M. [2000] “A comparative study of computation of Lyapunov spectra with different algorithms,” Physica D: Nonlinear Phenomena 139, 72–86. Sarra, S. & Meador, C. [2011] “On the numerical solution of chaotic dynamical systems using extend precision floating point arithmetic and very high order numerical methods,” Nonlinear Analysis 16, 340Ű352. Shimada, I. & Nagashima, T. [1979] “A numerical approach to ergodic problem of dissipative dynamical systems,” Progress of Theoretical Physics 61, 1605–1616. Skokos, C. [2010] “The Lyapunov characteristic exponents and their computation,” Lecture Notes in Physics 790, 63–135. Stuart, A. & Humphries, A. [1998] Dynamical Systems and Numerical Analysis (Cambridge University Press). Tavazoei, M. & Haeri, M. [2009] “A proof for non existence of periodic solutions in time invariant fractional order systems,” Automatica 45, 1886 – 1890. Vallejo, J. & Sanjuan, M. [2017] Predictability of Chaotic Dynamics: A Finite-time Lyapunov Exponents Approach (Springer, Cham, Switzerland). Wang, P., Li, J. & Li, Q. [2012] “Computational uncertainty and the application of a high-performance multiple precision scheme to obtaining the correct reference solution of lorenz equations,” Numerical Algorithms 59, 147–159. Wolf, A., Swift, J. B., Swinney, H. L. & Vastano, J. A. [1985a] “Determining Lyapunov exponents from a time series,” Physica D: Nonlinear Phenomena 16, 285–317. Wolf, A., Swift, J. B., Swinney, H. L. & Vastano, J. A. [1985b] “Determining Lyapunov exponents from a time series,” Physica D: Nonlinear Phenomena 16, 285–317. Zhou, Y. [2016] Fractional Evolution Equations and Inclusions: Analysis and Control (Academic Press). April 5, 2018 0:22 LE-FO-arxiv REFERENCES 11 Fig. 1. (a) A chaotic attractor of the RF system of FO, for q = 0.999. (b) Dynamics of the LEs. Fig. 2. (a) An apparently stable cycle of the generalized Lorenz system of FO, for q = 0.985. (b) Dynamics of the LEs. April 5, 2018 0:22 LE-FO-arxiv 12 REFERENCES Fig. 3. Dynamics of the LEs of the 4-dimensional system of FO (4). Fig. 4. Perturbation and rescaling of a nearby trajectory, after every h steps, considered as multiple of h (here h = norm norm 2h, sketch). April 5, 2018 0:22 LE-FO-arxiv REFERENCES 13 Fig. 5. LEs of RF system. (a) LEs as function of q for q ∈ [0.9, 1]. (b) LEs as function of p for p ∈ [1.1, 1.3]. Fig. 6. LEs of Chen’s system of FO represented by function of two variables: q and parameter p. Surfaces S , for i = 1, 2, 3, represents LE(q, p). April 5, 2018 0:22 LE-FO-arxiv 14 REFERENCES Appendices A. Program for LEs of FO % Memory allocation function [t,LE]=FO_Lyapunov(ne,ext_fcn,t_start, x=zeros(ne*(ne+1),1); h_norm,t_end,x_start,h,q,out); x0=x; c=zeros(ne,1); gsc=c; zn=c; % Program to compute LEs of systems of FO. n_it = round((t_end-t_start)/h_norm); % The program uses a fast variant of the % Initial values % predictor-corrector Adams-Bashforth-Moulton, x(1:ne)=x_start; % "FDE12.m" for FDEs, by Roberto Garrappa: i=1; while i<=ne % https://goo.gl/XScYmi x((ne+1)*i)=1.0; i=i+1; % m-file required: FDE12 and end % the function containing the extended system t=t_start; % (see e.g. LE_RF.m). % Main loop it=1; % FO_Lyapunov.m was developed, by while it<=n_it % modifying the program Lyapunov.m, % Solution of extended ODE system % by V.N. Govorukhin: [T,Y] = FDE12(q,ext_fcn,t,t+h_norm,x,h); t=t+h_norm; % https://goo.gl/wZVCtg Y=transpose(Y); x=Y(size(Y,1),:); %solution at t+h_norm % FO_Lyapunov.m, FDE12.m and LE_RF.m i=1; % must be in the same folder. while i<=ne j=1; % How to use it: while j<=ne; % [t,LE]=FO_Lyapunov(ne,ext_fcn,t_start,... x0(ne*i+j)=x(ne*j+i); % h_norm,t_end,x_start,h,q,out); j=j+1; end; % Input: i=i+1; % ne - system dimension; end; % ext_fcn - function containing the extended % orthonormal Gram-Schmidt basis % system; zn(1)=0.0; % t_start, t_end - time span; j=1; % h_norm - step for Gram-Schmidt while j<=ne % renormalization; zn(1)=zn(1)+x0(ne*j+1)*x0(ne*j+1); % x_start - initial condition; j=j+1; % outp - priniting step of LEs; end; % ioutp==0 - no print. zn(1)=sqrt(zn(1)); j=1; % Output: while j<=ne % t - time values; x0(ne*j+1)=x0(ne*j+1)/zn(1); % LE Lyapunov exponents to each time value. j=j+1; end % Example of use for the RF system: j=2; % [t,LE]=FO_Lyapunov(3,@LE_RF,0,0.02,300,... while j<=ne % [0.1;0;1;0.1],0.005,0.999,1000); k=1; while k<=j-1 % The program is presented in: gsc(k)=0.0; l=1; % Marius-F. Danca and N. Kuznetsov, while l<=ne; % Matlab code for Lyapunov exponents of gsc(k)=gsc(k)+x0(ne*l+j)*x0(ne*l+ % fractional order systems k); l=l+1; end hold on; k=k+1; end April 5, 2018 0:22 LE-FO-arxiv REFERENCES 15 k=1; B. Function LE RF.m while k<=ne l=1; function f=LE_RF(t,x) while l<=j-1 x0(ne*k+j)=x0(ne*k+j)-gsc(l)*x0( %Output data must be a column vector ne*k+l); f=zeros(size(x)); l=l+1; end %variables allocated to the variational equations k=k+1; X= [x(4), x(7), x(10); end; x(5), x(8), x(11); zn(j)=0.0; x(6), x(9), x(12)]; k=1; while k<=ne %RF equations zn(j)=zn(j)+x0(ne*k+j)*x0(ne*k+j); f(1)=x(2)*(x(3)-1+x(1)*x(1))+0.1*x(1); k=k+1; f(2)=x(1)*(3*x(3)+1-x(1)*x(1))+0.1*x(2); end f(3)=-2*x(3)*(0.98+x(1)*x(2)); zn(j)=sqrt(zn(j)); k=1; %Jacobian matrix while k<=ne J=[2*x(1)*x(2)+0.1, x(1)*x(1)+x(3)-1, x(2); x0(ne*k+j)=x0(ne*k+j)/zn(j); -3*x(1)*x(1)+3*x(3)+1,0.1,3*x(1); k=k+1; -2*x(2)*x(3),-2*x(1)*x(3),-2*(x(1)*x(2)+0.98)]; end j=j+1; %Righthand side of variational equations end f(4:12)=J*X; % To be modified if ne>3 % update running vector magnitudes k=1; while k<=ne; C. Program for LEs as function on p c(k)=c(k)+log(zn(k)); k=k+1; function run_Lyapunov_p(ne,ext_fcn,t_start,h_norm, end; % normalize exponent t_end,x_start,h,q,p_min,p_max,n); k=1; hold on; while k<=ne p_step=(p_max-p_min)/n LE(k)=c(k)/(t-t_start); p=p_min; k=k+1; while p<=p_max end LE=FO_Lyapunov_p(ne,ext_fcn,t_start,h_norm,t_end i=1; ,x_start,h,q,p); while i<=ne p=p+p_step j=1; plot(p,LE); while j<=ne; end x(ne*j+i)=x0(ne*i+j); j=j+1; end D. Function LE RF p.m i=i+1; end; x=transpose(x); function f=LE_RF_p(t,x,p) it=it+1; %p is the parameter % print and plot the results f=zeros(size(x)); if (mod(it,out)==0) % (*) X= [x(4), x(7), x(10); fprintf(’%10.2f %10.4f %10.4f... x(5), x(8), x(11); %10.4f\n’,[t,LE]); % (*) x(6), x(9), x(12)]; end; % (*) %RF equations plot(t,LE) % (*) f(1)=x(2)*(x(3)-1+x(1)*x(1))+0.1*x(1); end f(2)=x(1)*(3*x(3)+1-x(1)*x(1))+0.1*x(2); % displays the box outline around axes f(3)=-2*x(3)*(p+x(1)*x(2)); xlabel(’t’,’fontsize’,16) % (*) %Jacobian matrix ylabel(’LEs’,’fontsize’,14) % (*) J=[2*x(1)*x(2)+0.1, x(1)*x(1)+x(3)-1, x(2); set(gca,’fontsize’,14)% (*) -3*x(1)*x(1)+3*x(3)+1,0.1,3*x(1); box on;% (*) -2*x(2)*x(3),-2*x(1)*x(3),-2*(x(1)*x(2)+p)]; line([0,t],[0,0],’color’,’k’)% (*) f(4:12)=J*X; % To be modified if ne>3 April 5, 2018 0:22 LE-FO-arxiv 16 REFERENCES E. Program for LEs as function of q while q<q_max [t,LE]=FO_Lyapunov_q(ne,ext_fcn,t_start,h_norm, t_end,x_start,h,q); q=q+q_step; function run_FO_Lyapunov_q(ne,ext_fcn,t_start, fprintf(’q=%10.4f\n %10.4f’, q); h_norm,t_end,x_start,h,q_min,q_max,n); plot(q,LE); hold on; end q_step=(q_max-q_min)/n; q=q_min;

Journal

Nonlinear SciencesarXiv (Cornell University)

Published: Mar 29, 2018

There are no references for this article.