2008 5th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE 2008) 1 Modi ed Carleman Linearization and its use in oscillators Joaquín Collado, Irving Sánchez * Automatic Control Dept. CINVESTAV Av. IPN 2508 Col. Zacatenco 07360 México, D. F. MEXICO Abstract— The standard Carleman Linearization states that every analytic n-dimensional nonlinear systems is equivalent to an in nite dimensional linear system. In this paper we truncate this linearization and introduce some modi cation which reduces the error in the truncation process. We applied this Modi ed Carleman Linearization to two van der Pol oscillators with slight different frequencies. Each one is approximate by a 14th order linear system; then we coupled this two linear oscillators and look for a synchronization. We conclude that even the approximation is very good, is not possible synchronize high order linear oscillators as the nonlinear oscillators do. redundancies in its application to the Linearization. Section III presents the original Carleman Linearization, and Section IV the modi ed version. The modi ed Carleman Linearization is applied to the van der Pol oscillator in Section V. In Section VI two linearized van der Pol oscillators with slightly different frequency are coupled in a symmetric form and it is shown that it is impossible to synchronize linear oscillators with diffusively symmetric coupling. Keywords - Carleman Linearization, synchronization, van der Pol oscillators. II. P RELIMINARIES In this section we introduced some de nitions and preliminaries required in the sequel. I. I NTRODUCTION De nition 1: Given A 2 Rn Kronecker Product [4] or [21] is 2 a11 B a12 B 6 a21 B a22 B 6 AeB = 6 .. .. .. 4 . . . The Carleman Linearization was proposed for autonomous nonlinear analytic systems almost 80 years ago [6], the linearized system is equivalent to the original nonlinear system; however the procedure gives us an in nite dimensional system, to be used is required to truncate at some nite dimensional point. Bellman proposed, but he did not developed [1], that this truncation be performed in conjunction with some modi cation. This approximates the deleted terms in the Taylor series by a linear combination of the terms that we keep in the truncation. This paper work out this modi cation based on sampling the limit cycle trajectory of the van der Pol oscillator and using least squares method to calculate the parameters introduced in this modi cation. This modi ed Carleman linearization does not guarantee any stability property, then we reassign the eigenvalues to the j!-axis and these eigenvalues satisfy an integer relation, see Fig. 1. This last condition is required in order to have a periodic solution of the approximate linear system. an1 B and B 2 Rs t , their a1m B a2m B .. . anm B 3 7 7 7 2 Rns 5 mt Remark 2: The usual notation for the Kronecker Product is , but we will reserve for a reduced Kronecker Product, as explained in the sequel. 2 Let the vector x = xex = x1 x2 xn T , then 4x21 ; x1 x2 ; x1 x3 ; : : : ; x1 xn ; x2 x1 ; x22 ; x2 x3 ; : : : ; x2 xn ; {z } | {z } | n terms n terms 3 2 x3 x1 ; x3 x2 ; x23 ; : : : ; x3 xn ; : : : ; xn x1 ; xn x2 ; : : : ; x2n 5 2 Rn {z } | {z } | The organization of the paper is as follows: in section II are described the Kronecker Product of matrices and a Reduced Kronecker Product is introduced in order to remove n terms n terms Remark 3: Note that all mixed products xi xj for all i 6= j, appear duplicated in x e x. Let the operation x x be de ned as x e x removing all the duplicated terms, i.e. * E-mail: fjcollado,isanchezg@ctrl.cinvestav.mx IEEE Catalog Number: CFP08827-CDR ISBN: 978-1-4244-2499-3 Library of Congress: 2008903800 978-1-4244-2499-3/08/$25.00 ©2008 IEEE an2 B m 13 2008 5th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE 2008) x 2 6 x = 4x21 ; x1 x2 ; x1 x3 ; : : : ; x1 xn ; x22 ; x2 x3 ; : : : ; x2 xn ; {z } | {z } | n terms d h (2) i x dt (n 1) terms 3 7 x2 ; x3 x4 ; : : : ; x3 xn ; : : : ; x2n 5 2 Rn(n+1)=2 {z } |{z} |3 (n 2) terms d [x x] dt N X1 = [Ak In + In 2 = Ak ] x(k+1) k=1 1 term Hence, x(2) satis es a differential relation that depends on x ; : : : ; x(N ) but not on x(1) , however this equation has the same structure than (3). In order to avoid redundancy in our linearized model,we will use through the paper x x, for this reduced Kronecker Product. This product has been known for a long time, its use in Control Theory was introduced by Brockett [5]. Qiu and Davison named bialternate product in [15] and they used it for Robust Stability. (2) Continuing in this form yields a differential relation for x(j) to degree N of the form NX j+1 d h (j) i = Aj;k x(k+j x dt 1) (4) k=1 III. BASIC C ARLEMAN L INEARIZATION where A1;k = Ak and for j > 1; Let be a nonlinear autonomous dynamic system described by the Ordinary Differential Equation (ODE): Aj;k + x x] + mk (2) , mk , Then x_ = 1 X x k Ak x(k) {z times x 2 Rmk . } 6 6 x ,6 4 where dim(x ) Ak x x(2) .. . x(N ) 3 7 7 7 5 Thus we can write a nite dimensional linear differential equation, which is an approximation of the original ODE (1) d x dt Ak x(k) k=1 IEEE Catalog Number: CFP08827-CDR ISBN: 978-1-4244-2499-3 Library of Congress: 2008903800 978-1-4244-2499-3/08/$25.00 ©2008 IEEE In = M = n + m2 + ::: + mN (3) If we rewrite (1) after truncate the power series to N from (3) Then we obtain In 2 is a power series representation equivalent to the original nonlinear system (1) and (2), see [6] or [16]. N X In Now we de ne k=1 x_ = Ak 1 Kronecker products in each term, and Remark 4: Note that the index in Ak matrices corresponds to the degree in x of the Taylor series and mk it is the number of different elements in x(k) . We introduce the notation x(k) = x | In + In } Remark 5: Note that the differential relations (4) do not dx(j) depends on x(k) for constitute an ODE, because each dt k j. n+k 1 , k k = 1; 2; :::, and means the Modi ed Kronecker product as de ned in the section of Preliminaries. where the matrices Ak 2 Rn {z + In where there are j j terms. We assume that f (x) is analytic in Rn , thus we can write (1) in Taylor series form [16] as: x] + A3 [x In | (j 1) times (1) x_ = f (x) f : Rn ! R n f (x) = A1 x + A2 [x = Ak 2 6 6 6 = 6 6 4 A11 0 0 .. . A12 A21 0 .. . ::: ::: ::: .. . A1N A2;N A3;N .. . 0 0 ::: AN 1 ~ = Ax 14 1 2 3 7 7 7 7x 7 5 (5) 2008 5th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE 2008) II) If we want to approximate a particular solution, as the limit cycle in a van der Pol oscillator, sample this trajectory equidistantly and again apply least squares to minimize the same criteria J. This state equation is called a Truncated Carleman Linearization of the nonlinear-analytic state equation (1)1 . Remark 6: Truncation of the in nite dimensional vector x(1) to the M -dimensional vector x , is equivalent to keep all the terms of degree less or equal to N . In this way, the equation (7) represents a, better in general, nite order linearization of degree M of (1). IV. M ODIFIED C ARLEMAN L INEARIZATION In (5) the high order terms have been deleted, Bellman [1] proposed to make an approximation of the deleted terms by a weighted linear terms that depends on the variables in x , this improves in general the linear approximation of (1), thus we can write the state equation of x as d x dt 2 6 6 where G(x ) = 6 4 V. L INEARIZED VAN DER P OL O SCILLATOR The Modi ed Carleman linearization can be used for any non-linear systems whose right hand side of de ning equations be analytic; we are going to apply the above procedure to the well known van der Pol Oscillator [13] and [20], the equation that describes this system is ~ + G(x ) = Ax (6) 3 g1 (x) g2 (x) 7 7 7 contains the terms of order .. 5 . gN (x) higher than N , and each gi (x) is polynomial of degree N + 1. Partitions on G(x ) are compatible with partitions in ~ A. If we make approximations gi (x) by N X (k) , i;k x z + "(z 2 where We de ne x1 , z, x2 , z and x , 2 Rmi mk , see [2]; then the nite order linearization of degree M of (1) becomes: d x dt A~ + R x = (7) where 6 6 R =6 4 1;1 1;2 2;1 2;2 .. . N;1 .. . N;2 ::: ::: .. . ::: 1;N 2;N .. . N;N 3 7 7 7; 5 i;k 2R mi mk x1 x2 x = A1 x + A2 x(2) + A3 x(3) + A4 x(4) 0 1 A1 = ; A2 = A4 = 0; !2 " 0 0 0 0 A3 = 0 " 0 0 (8) Not necessarily all the row blocks i; for i = 1; : : : ; N need to be different from zero. They may contain up to M 2 parameters. There are two forms to determine this set of parameters: T , then (10) Then using (5) the linearization of van der Pol Oscillator for N = 4 is I) Choosing a region D Rn of the original state space where our system is know to be operating, sample enough number of points in D and minimize J = N X (k) gi (x) using least squares [3], [17], [18] and i;k x d x dt k=1 [19] to calculate the free parameters introduced. 1 Note that the initial conditions cannot be arbitrary, because the conditions for x(i) ; i 2; depend directly of initial conditions of x, since x(i) , (i 2) is the Reduced Kronecker Product of x with itself. IEEE Catalog Number: CFP08827-CDR ISBN: 978-1-4244-2499-3 Library of Congress: 2008903800 978-1-4244-2499-3/08/$25.00 ©2008 IEEE (9) rewriting (9) in Taylor series form and using the Modi ed Kronecker Product as de ned in the Preliminaries Section for x, we have = Ax 2 1)z + ! 2 z = 0 where " is the non-linear damping parameter (when " = 0 the van der Pol oscillator reduces to an harmonic linear oscillator, see [20]) and ! it is the frequency in radians. k=1 i;k 3 2 A1;1 6 0 = 6 4 0 0 ~ = Ax 0 A2;1 0 0 the vector state x 2 R14 , and 15 A1;3 0 A3;1 0 3 0 A2;3 7 7x 0 5 A4;1 (11) 2008 5th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE 2008) A1;1 A2;1 A2;3 A3;1 A4;1 = 2 = 4 2 0 = 4 0 0 2 6 = 6 4 2 6 6 = 6 6 4 0 1 ; A1;3 = !2 " 3 0 2 0 !2 " 1 5 2 0 2! 2" 0 " 0 0 !2 0 0 0 !2 0 0 0 0 0 2" 3 " 2! 2 0 4 " 2! 2 0 0 0 0 0 " 3 0 0 0 0 5 0 0 3 0 0 2 0 7 7 2" 1 5 3! 2 3" 0 0 3 0 2" 2 3! 2 3" 0 4! 2 0 0 0 0 1 7:9848 1:3921 0:0137 0:0011 0:004 1:9669 4:4836 2:3702 0:2828 0:004 0:0021 0:0056 0:0019 0:0009 = In this form we obtain (8) for the van der Pol oscillator, then we have a nite order linearization of the van der Pol Oscillator, which is an unstable 14th order system, because some eigenvalues of the matrix A = A~ + R are on the right hand side of the complex plane. 0 0 0 1 4" For the linear oscillator to be marginally stable and the have a periodic solution, the eigenvalues of the state matrix must lie on the imaginary axis and they should be a multiple of some fundamental frequency., i. e. let the spectra of A be (A) = f 1 ; 2 ; : : : ; 14 g then 3 7 7 7 7 5 Re ( i ) = 0 and b = 0 0 0 0 0 v 0 g1 (x) g2 (x) g3 (x) 0 g4 (x) g5 (x) g6 (x) g7 (x) = = = = = = = = jm! 0 8i = 1; : : : ; 14 (12) 2 = R14 ; b= Kx ; 0 0 ::: 1 K 2 R1 T 14 The b used above is such that the pair (A; b) must be controllable [7]. Then K must be such that the matrix A = A bK has its eigenvalues on the vertical axis of the complex plane and these satisfy multiplicity described in (12), see [7] and [9]. The effect of this state feedback is exempli ed in Fig. T where g1 (x) g2 (x) g3 (x) g4 (x) g5 (x) g6 (x) g7 (x) i To make this we use a control signal such that d x = Ax + bv dt where However, as we saw in the previous section, (11) is a truncated linearization, where the higher order terms have been deleted, in fact exists a G(x) of the form G(x) 4 "x41 x2 2"x31 x22 3"x21 x32 "x51 x2 2"x41 x22 3"x31 x32 4"x21 x42 1 Then the equation that describes the real dynamics of x is d x dt ~ + G(x) = Ax Now, we can calculate i;k , for which we use least square method for each gi (x) above, the following example shows a particular case of this linearization for the van der Pol Oscillator. From the van der Pol Oscillator equation in (9) when ! = " = 1, we calculate the value of i;k using the least squares method as described above. As example g1 (x) is approximate by g1 (x) Fig. 1. Eigenvalue assignment in order to satisfy the constraint in (12). 1x IEEE Catalog Number: CFP08827-CDR ISBN: 978-1-4244-2499-3 Library of Congress: 2008903800 978-1-4244-2499-3/08/$25.00 ©2008 IEEE 16 2008 5th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE 2008) 5 Example 7: The simple projection of the eigenvalues to the imaginary axis does not guarantee that the eigenvalues be multiple of the lowest one, then we may interpret as a nonlinear projection as show in gure 1. Hence K must be such that one of the eigenvalues for the matrix A must be the corresponding to the natural frequency of (9) on the vertical axis and the rest, multiples of this one. This fact makes that the linearized van der Pol systems oscillates. Then (A) 0 = f ij = !0 j i = n 0; n 2 N; i = 1; :::; 7g A way to obtain ! 0 when " is large is in [8], otherwise we can the less precise formula in (9). Fig. 3. Symmetric coupled linearized van der Pol oscillators. In this way the system d x = Ax (13) dt is an stable, but not asymptotically stable, nite order linearization of (9). The Fig. 2 shows the waveforms of the original and the linearized van der Pol oscillators. Notice how well approximates the linear system to the van der Pol oscillator. Then we add a control signal in (13) for both oscillators such that. d x dt i yi yi0 (14) = Ai xi + Bui = Cxi = xi;1 B 2 R14 1 ; C 2 R1 14 ; ui = k(yi yj ); i 6= j; B = . . . 01..010..0100..01000 = . . . 01..010..0100..01000 C Fig. 2. First state of the original van der Pol oscillator compared with the rst component of the 14th order linear approximation. T ui 2 R k 2 R; k>0 Thus VI. D IFFUSED COUPLED LINEARIZED VAN DER P OL d x dt i O SCILLATORS 1)z 1 + z1 = 0 z1 (0) = 2; z 1 (0) = 0 z 2 + (z22 1)z 2 + (1:01) z2 = 0 2 z2 (0) = 2; z 2 (0) = 0 kBxj In this case the both pairs (Ai ; B) i = 1; 2; are controllable and both pairs (C; Ai ) are observable respectively. i. e., ! 1 = "1 = "2 = 1 and ! 2 = 1:01; we can use A1 from example above and obtain A2 using the method described in the previous section. In this way the behavior of x1 and x2 depends directly from the value of k, in fact both systems become unstable for any value of k as show in the gure 4 For coupling the two oscillators, we use following diagram, see [12], [14] and [11]: IEEE Catalog Number: CFP08827-CDR ISBN: 978-1-4244-2499-3 Library of Congress: 2008903800 978-1-4244-2499-3/08/$25.00 ©2008 IEEE kBC)xi these values for C and B correspond to the velocity z in the original van der Pol oscillator . It is routine verify that the pairs (Ai ; B) and (C; Ai ), for i = 1; 2; are controllable and observable respectively. In (13) we have a stable nite 14th order linearization of the van der Pol Oscillator for ! = " = 1. Now apply our procedure for two original van der Pol oscillators with slightly different frequencies as z 1 + (z12 = (Ai 17 2008 5th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE 2008) Fig. 4 The coupled oscillators cannot synchronize Fig. 1. The fact that the amplitude increases is due to during the state feedback through ui = k(yi yj ) modi es the eigenvalues of the matrix (Ai kBC) to the left and to the right side on the complex plane. The Fig. 5 shows the roots locus of the eigenvalues of the matrix (A1 kBC). 6 Two nonlinear van der Pol oscillators synchronized VII. C ONCLUSIONS In the rst part of the paper we propose a Modi ed version of the Truncated Carleman Linearization method, then we applied this novel procedure to the van der Pol oscillator and used to try to synchronize to linearized van der Pol oscillators with coupling proportional to velocity. We conclude that it is not possible to synchronize them as they do in the original systems using the same coupling. R EFERENCES [1] Bellman, R. Perturbation Techniques in Mathematics, Physics, and Engineering. Holt, Rinehart and Winston, Inc., pp. 69-74, 1966. [2] Bellman, R. & Richardson, J.M., “On some questions arising in the approximative solution of nonlinear differential equations”, Quart. Appl. Math, Vol. 20, pp. 333-339, 1963. [3] A. Björck, "Numerical Methods for Least Squares Problems", SIAM, Philadelphia, 1996. [4] Brewer, J. "Kronecker products and matrix calculus in system theory". IEEE Trans. on Circuits and Systems, Volume CAS-25, Issue 9, pp. 772 - 781, Sep 1978. [5] Brockett, R. "Lie Algebras and Lie Groups in Control Theory". In Geometric Methods In Systems Theory. D. Mayne and R. Brockett (Eds.), Reidel Dordrecht, 1973. [6] Carleman T. "Application de la Théorie des Équations Intégrales Linéaires aux Systémes D 'Équations Différentialles Non Linéaires", Acta. Math., Vol. 59, pp. 63-87, 1932. [7] Chen, C. T. Linear System Theory and Design, Third Edition, Oxford University Press, pp. 235-239, 1999. [8] Dorodnicyn, A. A. "Asymptotic Solution of van der Pol´s Equation", Engl. transl. in Amer. Math. Soc. Transl., Series 1, 88, pp. 1-24, 1953. [9] Kailath, T., Linear Systems, Prentice-Hall, Inc., Englewood Cliffs, N. J. 07632, 1980. [10] H. K. Khalil. "Nonlinear Systems", 3rd. Ed. Upper Saddle River:Prentice-Hall, USA, 2002. [11] Kuramoto Y., "Chemical Oscillations,Waves, and Turbulence", New York: Springer, 1984. [12] Mimila P. O., "Sincronización de Osciladores", Tesis de maestría, Automatic Control Dept. Cinvestav-IPN, Sep. 2006. [13] Nayfeh, A. H. Nonlinear Oscillations, John Wiley & Sons Inc., pp. 3-6, 1979. [14] Pikovsky, A., M. Rosenblum, and J. Kurths. "Synchronization: A universal concept in nonlinear sciences". Cambridge University Press, Cambridge, 2001. [15] Qiu, L. and E. J. Davison. "The Stability Robustness Determination of State Space Models with Real Unstructured Uncertainty". Math. of Contr., Signals and Systems, Vol. 4, pp. 247-267, 1991. [16] Rugh, W. J. Non Linear System Theory The Volterra/Wiener Approach. The Johns HopkinsUniversity Press, pp. 103-116, 1981. Fig. 5. Root Locus of coupled Linearized 14th order van der Pol oscillators. This show that the eigenvalues turn on the right side at the same moment when k is different of zero, thus we can say it is no possible to synchronize the linear oscillators with diffusively symmetric coupling. Notice this behavior is not for the coupling proposed, but for any linear coupling, because to synchronize linear oscillator is required that eigenvalues coincide on the imaginary axis, that keeps the multiplicity property and that these conditions holds for a range of values of k greater that a threshold [14]. The Fig. 6 shows how two original nonlinear van der Pol oscillators can be synchronized with diffusively symmetric coupling, the natural frecuencies are the same that the example with linear oscillators above IEEE Catalog Number: CFP08827-CDR ISBN: 978-1-4244-2499-3 Library of Congress: 2008903800 978-1-4244-2499-3/08/$25.00 ©2008 IEEE 18 2008 5th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE 2008) [17] Stewart, G. W. Introduction to Matrix Computations, Academic Press, pp. 217-221, 1973. [18] Stewart, G. W. Matrix Algorithms: Basic Decompositions, SIAM, Philadelphia, 1998. [19] L. N. Trefethen and D. Bau, III, "Numerical Linear Algebra", SIAM, Philadelphia, 1997. [20] van der Pol, B. L. . "The nonlinear theory of electric oscillators", Proc of the IRE, Vol. 22, Num. 9, pp. 1051-1056, September 1934. [21] Charles F. Van Loan, "The ubiquitous Kronecker product". Journal of Computational and Applied Mathematics, 123 (2000) 85–100. IEEE Catalog Number: CFP08827-CDR ISBN: 978-1-4244-2499-3 Library of Congress: 2008903800 978-1-4244-2499-3/08/$25.00 ©2008 IEEE 19 7