Global Stochastic Synchronization of Chaotic Oscillators
Global Stochastic Synchronization of Chaotic Oscillators
Global Stochastic Synchronization of Chaotic Oscillators
2
Westin Seattle Hotel, Seattle, Washington, USA
June 11-13, 2008
Abstract—We study synchronization of two chaotic oscillators chaotic systems with time-varying coupling. We focus on
in a master-slave configuration. The two dynamic systems are the general case where the intermittent coupling changes
coupled via a directed feedback that randomly switches among randomly over time. Intermittent coupling is made possi-
a finite set of given constant functions at a prescribed time
rate. We use stochastic Lyapunov stability theory and partial ble through a switching function, that changes randomly
averaging techniques to show that global synchronization is over time, assuming values among a finite set of constant
possible if the switching period is sufficiently small and if functions. The synchronization problem is transformed into
the two systems globally exponentially synchronize under an a nonlinear stochastic stability problem, and it is studied
average feedback coupling. The approach is applied to the using partial averaging techniques [26], [27], [28], nonlinear
synchronization of two Chua’s circuits.
Index Terms—Global synchronization, master-slave synchro- system theory [29], and stochastic stability theory [30], [31].
nization, stochastic synchronization, chaos, Chua’s circuit, ex- We associate to the stochastic system a partially averaged
ponential synchronization system characterized by a constant coupling. This auxiliary
system can be studied using well-established and manageable
I. I NTRODUCTION Lyapunov based techniques as those presented in [20]. Under
Chaos synchronization is a topic of great interest, due to suitable regularity conditions, we show that if the partially
its observation in a huge variety of phenomena of different averaged system is globally exponentially stable and the
nature. In many biological systems, synchronization plays an switching period is sufficiently small, the stochastic system
important role in self-organization of organisms’ groups [1]. globally synchronizes.
Examples of synchronization include communication among The type of intermittent coupling considered in this paper
fireflies [2], locomotion of animals [3], molecular and cellular has been also analyzed in the framework of consensus theory
activity [4], and cardiac stimulation [5]. The study of neural [32], [33]. We note that, in consensus theory, the individual
activity [6] is a correlated issue as well. Other examples and systems’ dynamics is linear while in the present case the
applications can also be found in meteorology [7], chemistry coupled systems are strongly nonlinear. Partial averaging
[4], and optics [8]. Many reviews on chaos synchronization techniques have been used in the synchronization literature by
are currently available [9], [10], [11], [12], [13]. [15], [16]. Both these efforts deal with peer-to-peer coupling
In the literature, different paradigms have been proposed and only local asymptotic synchronization results based on
to employ synchronization of two or more chaotic oscillators. linearized dynamics are presented. In this paper, the inherent
We mention, among the others, peer-to-peer coupling [14], non linear nature of the coupled systems is retained and
[15], [16], back-stepping [17], generalized synchronization global results are presented.
[18], phase synchronization [10], and master-slave synchro- The system studied in this paper finds many practical
nization [19], [20], [21]. In this work, we focus on master- applications. For example, in communication and signal pro-
slave synchronization. In this case, one system acts as a “mas- cessing, chaotic behavior can be used for message encryption
ter” by driving the other system that behave consequently as and secure communication [34], [35]. Higher communica-
a “slave”. tion efficiency can potentially be achieved through sporadic
Most of the research efforts on master-slave synchroniza- transmission of the driving signal. This is particularly useful
tion focus on time invariant coupling, see [19], [20], [22] when the access to a communication network is limited to
among the others. Nevertheless, experimental and numerical assigned time-slots. Furthermore, in many biological systems,
evidences show that synchronization can also be achieved interactions happen only sporadically and randomly [36] of
using intermittent time-varying master-slave coupling [23], neurons and fireflies.
[24], [25]. In [23], experimental results on synchronization We organize the paper as follows. In Section II, we define
of two periodically coupled chaotic circuits are presented. In the master-slave synchronization problem in the case of a
[24], the slave system is driven by a sequence of samples stochastic linear feedback. In Section III, we present a few
of the master’s state (impulsive synchronization). In [25], the general results on stability of nonlinear stochastic systems.
signal transmission from the master to the slave system is In Section IV, we apply these results to the synchronization
adaptively controlled (selective synchronization). problem. As a sample case, in Section V we consider the
The main goal of this work is to establish sufficient case of two stochastically coupled Chua’s circuits. Section
conditions for global synchronization of master-slave coupled VI is left for conclusions.
512
By taking the conditional expected value of (10) we obtain Inequality (19) can also be used to find an upper bound
for kx(σk̂ )k in terms of the initial conditions. In fact, since
E[E[V (x(σk̂+2 ), σk̂+2 )|x(σk̂+1 )]|x(σk̂ )] t0 ∈ [σk̂−1 , σk̂ ) according to the definition of k̂, (19) holds
≤ (1 − ν)E[V (x(σk̂+1 ), σk̂+1 )|x(σk̂ )] (11) for t = t0 and t ≥ t0 . Since σbk ≥ t0 , we obtain
Using the smoothing lemma (see for example [40] Lemma kx(σk̂ )k ≤ γ 2 kx(t0 )k (21)
1.1 p. 474) and substituting (9) in (11), we find
Finally, using (21) to bound the right side of (20), we obtain
E[V (x(σk̂+2 ), σk̂+2 )|x(σk̂ )] ≤ (1 − ν)2 V (x(σk̂ ), σk̂ ) (12)
E[kx(t)k2 ] ≤ αkx(t0 )k2 e−β(t−t0 ) (22)
Recalling that equation (5) describes a Markov process and
4 −2
iterating the argument above for any n > k̂, we obtain where α = γ (1 − ν) λmax /λmin and β = −ln(1 − ν)/ε.
Therefore, according to Definition 1, the system (5) is glob-
E[V (x(σn ), σn )|x(σk̂ )] ≤ (1 − ν)n−k̂ V (x(σk̂ ), σk̂ ) (13) ally mean square exponentially stable.
By using the bounds in (7), equation (13) gives We associate to (5) the partially averaged system
E[kx(σn )k2 |x(σk̂ )] ≤ λmax /λmin (1 − ν)n−k̂ kx(σk̂ )k2 ẋ(t) = f (x(t), t) = E[f (x(t), t, Ω)] (23)
(14)
Equation (23) represents a deterministic, time-varying, non
Equation (14) can be used to derive an upper bound for
linear system. We notice that f (0, t) = 0, ∀t ∈ R+ . If (23)
the unconditioned expected value, that is needed to assess
is globally exponentially stable, by the Converse Theorem
the global mean square exponential stability according to
of Lyapunov (see [29], Theorem 3.12) we know that there
Definition 1. Since k̂ is a given instant of time and x0 is
exists a Lyapunov function V (x, t), whose time derivative
a prescribed initial condition, x(σk̂ ) is a finite-state random
is negative definite along its trajectories. In the following
variable taking values in {x1 (σk̂ ), ..., xN (σk̂ )}. From the
theorem, we show that if V (x, t) satisfies further regularity
definition of conditional expectation, see for example [41],
conditions and the switching period is sufficiently small, the
we have
original system (5) is globally mean square exponentially
XN
2 stable.
E[kx(σn )k ] = E[kx(σn )k2 |xi (σk̂ )]P {xi (σk̂ )} (15)
i=1
Theorem 2: Consider the system (5) and the associated
where P {xi (σk̂ )} is the probability that xi (σk̂ ) is the realiza- partially averaged system (23) and suppose that there exists
tion of the random variable x(σk̂ ). Hence, using inequality a Lyapunov function V (x, t) which satisfies the following
(14), equation (15) yields conditions:
N
X λmax 1. V (0, t) = 0, ∀t ∈ R+ and there exist positive numbers
E[kx(σn )k2 ] ≤ (1 − ν)n−k̂ kxi (σk̂ )k2 P {xi (σk̂ )} λmin , λmax such that for every (x, t) ∈ Rn × R+ ,
i=1
λmin
(16) λmin kxk2 ≤ V (x, t) ≤ λmax kxk2 (24)
In order to assess global mean square exponential stability
we need to analyze the system dynamics inside every switch- 2. there exists a positive number w such that for every (x, t) ∈
ing interval. Given a generic switching interval [σk , σk+1 ) Rn × R+
and an instant t ∈ [σk , σk+1 ), using the triangle inequality, ∂V ∂V
(x, t) + (x, t)f (x, t) ≤ −wkxk2 (25)
∀t ≥ t in [σk , σk+1 ), equation (5) yields ∂t ∂x
Z t ∂V ∂V
3. ∀t ∈ R+ , ∂x (0, t) = 0 and ∂x (x, t) is globally
kx(t)k ≤ kx(t)k + kf (x(ξ), ξ, Ω)kdξ (17)
t Lipschitz, with Lipschitz constant Cv . Moreover, ∀t ∈ R+ ,
Since the functions fω are globally Lipschitz in R+ and ∂2V ∂2V
∂x∂t (0, t) = 0, and ∂x∂t (x, t) is globally Lipschitz, with
all the corresponding Lipschitz constants are bounded by a Lipschitz constant Cvt .
constant L, equation (17) yields
Z t Then, there exists an ε∗ > 0 such that ∀ε < ε∗ system
kx(t)k ≤ kx(t)k + Lkx(ξ)kdξ (18) (5) is globally mean square exponentially stable.
t
Using the Gronwall-Bellman inequality, see for example [29], Proof: Consider the Lyapunov function V (x, t). Its
we have derivative along the solution of (5) is
kx(t)k ≤ γkx(t)k (19)
∂V ∂V
with γ = eLε . Therefore, using (19) in (16), we find that V̇ (x(t), t) =(x(t), t) + (x(t), t)f (x(t), t, Ω) (26)
∂t ∂x
∀t ∈ [σn , σn+1 )
For every nonnegative integer k, we define
N
X λmax
E[kx(t)k2 ] ≤ γ (1 − ν)n−k̂ kxi (σk̂ )k2 P {xi (σk̂ )} ∆V (σk+1 , σk ) = E[V (x(σk+1 ), σk+1 )|x(σk )]
i=1
λmin
(20) − V (x(σk ), σk ) (27)
513
From (5), (26) and (27) we have For clarity we restate here the main theorem of [20].
·Z σk+1 ¸
∆V (σk+1 , σk ) = E V̇ (x(t), t)dt = Theorem 3: The system (32) is globally exponentially
·Z σk+1 σk stable, if the feedback gain matrix K ∗ is chosen such that
∂V
E (x(t), t)f (x(t), t, Ω) li (ξ, t) ≤ −w < 0, i = 1, 2, ...n (33)
σk ∂x
¸
∂V
− (x(σk ), t)f (x(σk ), t, Ω)dt for every ξ ∈ Rn and t ∈ R+ , where li (ξ, t)’ s are the
∂x eigenvalues of the matrix
·Z σk+1 ¸
∂V ∂V
+E (x(t), t) − (x(σk ), t)dt
∂t ∂t Q(ξ, t) = (A − K ∗ + Mx(t),x(t)−ξ )T P
·Z σk+1 σk ¸
∂V ∂V + P (A − K ∗ + Mx(t),x(t)−ξ ) (34)
+E (x(σk ), t) + (x(σk ), t)f (x(σk ), t, Ω)dt
σk ∂t ∂x
and P is a positive definite symmetric constant matrix. A
Lyapunov function for (32) can be constructed as
(28)
V (e(t)) = e(t)T P e(t) (35)
We seek an upper bound for the absolute values of the
three terms in the summation above. We start our analysis with
by considering the first and the second terms. Using the
Lipschitz conditions on fω and on the first and second V̇ (e(t)) = e(t)T Q(ξ, t)e(t) ≤ −wke(t)k2 < 0 (36)
derivatives
PN of V , considering that for each realization ω of
Ω i=1 P {Ω = ωi } = 1 and Lω,ε < L for each ω), and
following the arguments of [28] (see parts II and III of the We note that, for i = 1, ..., N , the function f (e(t), t, Ki ) =
proof of Theorem 2 ), we find that the absolute value of (A + Mx,x−e − Ki )e(t) is globally Lipschitz in R+ with
the first term of the summation (28) is less than or equal to Lipschitz constant Li = kAk + m + kKi k, where kM k ≤
2L2 Cv e2εL ε2 kx(σk )k2 . Similarly, we find that absolute value m. The Lipschitz constants Li are bounded by L = kAk +
of second term is less than or equal to LCvt e2εL ε2 kx(σk )k2 . m + max1≤i≤N {kKi k}. We further notice that f (0, t, Ki ) =
As a result, we have that (25) provides a bound for the 0, ∀t ∈ R+ . We associate to the system (3) the partially
third term in the right side of (28). averaged system
Similarly to all the above boundaries we obtain
ė(t) = (A + Mx(t),x(t)−ξ )e(t) − Ke(t) (37)
∆V (σk+1 , σk ) ≤ [g(ε) − wǫ]kx(σk )k2 (29) PN
where K = E[K(t)] = i=1 pi Ki . Here, pi indicates
where the function g(ε) is defined by the probability of K(t) assuming value Ki , that is pi =
g(ε) = (2L2 Kv e2εL + LKvt e2εL )ε2 (30) P {K(t) = Ki }.
Since K is constant, we can apply Theorem 3 to (37).
Noticing that g(0) = 0 and g ′ (0) = 0, we have that there The Lyapunov function (35) constructed for the partially
exists ε∗ > 0 such that averaged system, can be used to assess the stability of the
∆V (σk+1 , σk ) ≤ −wkx(σk )k2 (31) stochastic system. In fact V (0, t) = 0 and (24) holds for
λmin = min{λ(P )} and λmax = max{λ(P )}, since P is
where w = [wǫ − g(ε)] > 0, for every ε < ε∗ . a constant matrix (here, λ(•) indicates the spectrum of the
Therefore, if the switching period ε is sufficiently small, (31) matrix). Furthermore, (36) is equivalent to (25) and Condition
and (27) imply that hypotheses of Theorems 1 are satisfied. 3 of Theorem 2 is satisfied with Cv = 2kP k and Cvt = 0.
Thus the claim follows. Thus equation (31), specified for the case at hand, reads
IV. S TOCHASTIC CHAOS SYNCHRONIZATION 2L2 Cv e2Lε ε − w = 0 (38)
In this section, we combine the general findings of Section
3 on stochastic stability of non linear systems with available and it yields the sought value of ε∗ . By applying Theorem
results on synchronizability of deterministic systems, to pro- 2, we claim that the system (3) is globally mean square
vide sufficient conditions for the global synchronization of exponentially stable ∀ε < ε∗ . We summarize the above
system (5). In particular, we make use of the results of [20], arguments in the following Corollary.
where a criterion for global exponential synchronization of
(1) and (2) is given in the case of constant feedback gain. Corollary 1: Consider the system (3) and the correspond-
The error system equation (3), when K(t) equals the constant ing partially averaged system (37). If the feedback gain
K ∗ , becomes matrix K(t) is chosen such that
514
for every ξ ∈ Rn and t ∈ R+ , where li (ξ, t)’ s are the Here, we present a few numerical results that illustrate
eigenvalues of the matrix how two stochastically coupled Chua’s circuits synchronize
for a sufficiently fast switching rate. We select a = 9.78,
Q(ξ, t) = (A − K + Mx(t),x(t)−ξ )T P b = 14.97, m0 = −1.31, and m1 = −0.75 in order to gen-
+ P (A − K + Mx(t),x(t)−ξ ) (40) erate chaotic behavior [20]. We let K(t) switching randomly
between the two constant matrices K1 and K2 , where K1
and P is a positive definite symmetric constant matrix, then is the zero matrix and K2 = diag[20 27.5 20]. For these
there exists an ε∗ > 0 such that such that ∀ε < ε∗ system parameters we have kAk = 18.8859, max1≤i≤2 {kKi k} =
(3) is mean square globally exponentially stable. kK2 k = 27.5, and L = 59.1977. Selecting p1 = 0.6 and
p2 = 0.4, w can be chosen from (45) to be equal to 0.5.
V. C ASE STUDY: SYNCHRONIZATION OF TWO CHAOTIC From equation (38) we have that for ε ≤ ε∗ = 3.5629 · 10−5
C HUA’ S CIRCUIT the system synchronizes globally mean square exponentially.
Fig. 1 and Fig. 2 depicts the trajectories of the master and
As an example, we apply our results to synchronization
slave systems on the x1 x2 and x1 x3 planes for a switching
of Chua’s circuits, see for example [42]. A Chua’s circuit is
period ε = 10−5 .
described in terms of vector state x(t) = [x1 (t) x2 (t) x3 (t)]T
by
ẋ(t) = Ax(t) + g(x(t)) (41) 1
Master
Slave
with 0.8
−a a 0 −ah(x1 ) 0.6
where K = diag[k 1 k 2 k 3 ]. 0
515
stochastic unidirectional feedback, that switches randomly [22] L. M. Pecora and T. L. Carroll, “Master stability functions for syn-
among a finite set of constant values. Using tools based on chronized coupled systems,” Physical Review Letters, vol. 80, no. 10,
pp. 2109–2112, March 1998.
Lyapunov stability theory and partially averaging techniques [23] L. Fortuna, M. Frasca, and A. Rizzo, “Experimental pulse synchroniza-
we showed that, under suitable regularity conditions, the tion of two chaotic circuits,” Chaos, Solitons and Fractals, vol. 17, no.
synchronization characteristics of the partially averaged 2-3, pp. 335–361, 2003.
[24] M. Itoh, T. Yang, and L. O. Chua, “Conditions for impulsive synchro-
system are inherited by the stochastic system. Our findings nization of chaotic and hyperchaotic systems,” International Journal
are illustrated through numerical simulations on Chua’s of Bifurcation and Chaos, vol. 11, pp. 551–560, 2001.
circuits. [25] M. Zochowski, “Intermittent dynamical control,” Physica D, vol. 145,
pp. 181–190, 2000.
[26] G. Grammel and I. Maizurna, “Exponential stability and partial aver-
aging,” Journal of Mathematical Analysis and Applications, vol. 283,
R EFERENCES pp. 276–286, 2003.
[27] J. K. Hale, Ordinary Differential Equations. New York: John Wiley
[1] S. Camazine, W. Ristine, and M. E. Didion, Self-Organization in & Sons, 1969.
Biological Systems. Princeton University Press, 2003. [28] J. Peuteman and D. Aeyels, “Exponential stability of nonlinear time-
[2] J. Buck and E. Buck, “Synchronous fireflies,” Scientific American, vol. varying differential equations and partial averaging,” Mathematics of
234, no. 5, pp. 75–85, 1976. Control, Signals and System, vol. 15, no. 1, pp. 42–70, 2002.
[3] J. J. Collins and I. Stewart, “Coupled nonlinear oscillators and the [29] H. K. Khalil, Nonlinear Systems, 3rd ed. Upper Saddle River, NJ:
symmetries of animal gaits,” Journal of Nonlinear Science, vol. 3, Prentice Hall, 2002.
no. 1, pp. 349–392, 1993. [30] H. Kushner, Introduction to Stochastic Control. New York, NY: Holt,
[4] A. Goldbeter, Biochemical Oscillations and Cellular Rhythms: The Rinehart and Winston, Inc., 1971.
Molecular Bases of Periodic and Chaotic Behaviour. Cambridge [31] H. J. Kushner, Stochastic Stability and Control. New York: Academic
University Press, 1996. Press, 1967.
[5] A. Garfinkel, M. Spano, , W. Ditto, and J. Weiss, “Controlling cardiac [32] Y. Hatano and M. Mesbahi, “Agreement over random networks,” IEEE
chaos,” Science, vol. 257, pp. 1230–1235, 1992. Transactions on Automatic Control, vol. 50, no. 11, pp. 1867 – 1872,
[6] P. Rapp, T. R. Bashore, J. M. Martinerie, A. M. Albano, I. D. 2005.
Zimmerman, and M. A. I., “Dynamics of brain electrical activity,” [33] M. Porfiri and D. J. Stilwell, “Consensus seeking over random
Brain Topography, vol. 2, no. 1-2, pp. 99–118, 1989. weighted directed graphs,” IEEE Transactions on Automatic Control,
[7] G. S. Duane, P. J. Webster, and J. B. Weiss, “Go-occurrence of accepted for pubblication.
northern and southern hemisphere blocks as partially synchronized [34] K. M. Cuomo, V. A. Oppenheim, and S. H. Strogatz, “Synchronization
chaos,” Journal of Atmospheric sciences, vol. 56, no. 24, pp. 4183– of lorentz-based chaotic circuits with application to communications,”
4205, 1999. IEEE Transactions on Circuits and Systems II, vol. 40, no. 10, pp.
[8] G. D. Van Wiggeren and R. Roy, “Communication with chaotic lasers,” 626–633, 1993.
Science, vol. 279, pp. 1198–1200, 1998. [35] B. Jovic, C. P. Unsworth, G. S. Sandhu, and S. M. Berber, “A robust
[9] B. R. Andrievskii and A. L. Fradkov, “Control of chaos: Methods and sequence synchronization unit for multi-user ds-cdma chaos-based
applications,” Automation and Remote Control, vol. 64(5), pp. 673– communication systems,” Signal Processing, vol. 87, no. 7, pp. 1692–
713, 2003. 1708, 2007.
[10] S. Boccaletti, J. Kurths, G. Osipov, D. L. Valladares, and C. S. Zhou, [36] I. V. Belykh, V. N. Belykh, and M. Hasler, “Blinking model and syn-
“The synchronization of chaotic systems,” Physics Reports, vol. 366, chronization in small-world networks with a time-varying coupling,”
no. 1, pp. 1–101, 2002. Physica D: Nonlinear Phenomena, vol. 195, no. 1-2, pp. 188–206,
[11] G. Chen and X. Yu, Eds., Chaos Control Theory and Applications, ser. 2004.
Lecture Notes in Control and Information Sciences. Berlin: Springer, [37] O. L. V. Costa, M. D. Fragoso, and R. P. Marques, Discrete-Time
2003, vol. 292. Markov Jump Linear systems. London, UK: Springer, 2005.
[12] E. Mosekilde, Y. Maistrenko, and D. Postnov, Chaotic Synchronization: [38] X. Feng, K. A. Loparo, Y. Ji, and H. J. Chizeck, “Stochastic stability
Applications to Living systems. New Jersey: World Scientific, 2002. properties of jump linear systems,” IEEE Transactions on Automatic
Control, vol. 37(1), pp. 38–53, 1992.
[13] L. M. Pecora, T. L. Carroll, G. A. Johnson, and D. J. Mar, “Fundamen-
[39] D. Aeyels and J. Peuteman, “On exponential stability of nonlinear
tals of synchronization in chaotic systems, concepts, and applications,”
time-varying differential equations,” Automatica, vol. 35, no. 6, pp.
Chaos, vol. 7, no. 4, pp. 520–543, 1997.
1091–1100, 1999.
[14] Z. Ge and Y. Chen, “Synchronization of mutual coupled chaotic
[40] A. Gut, Probability: A Graduate Course. New York, NY: Springer,
systems via partial stability theory,” Chaos, Solitons and Fractals,
2005.
vol. 34, no. 3, pp. 787–794, 2007.
[41] P. Bremaud, Markov Chains, Gibbs Fields, Monte Carlo Simulation,
[15] M. Porfiri, D. J. Stilwell, E. M. Bollt, and J. D. Skufca, “Random
and Queues, Springer-Verlag, Ed., 1999.
talk: Random walk and synchronizability in a moving neighborhood
[42] L. P. Shil’nikov, “Chua’s circuit: rigorous results and future problems,”
network,” Physica D, vol. 224, pp. 102–113, 2006.
International Journal of Bifurcation and Chaos, vol. 4, no. 10, pp.
[16] D. J. Stilwell, E. M. Bollt, and D. G. Roberson, “Sufficient conditions
489–519, 1993.
for fast switching synchronization in time varying network topologies,”
SIAM Journal on Applied Dynamical Systems, Accepted for publica-
tion.
[17] S. Bowong and F. M. M. Kakmeni, “Synchronization of uncertain
chaotic system via backstepping approach,” Chaos Solitons and Frac-
tals, vol. 21, pp. 999–1011, 2004.
[18] R. Zhang, Z. Xu, S. X. Yang, and X. He, “Generalized synchronization
via impulsive control,” Chaos, solitons and Fractals, vol. Article in
Press, 2006.
[19] T. L. Carroll and L. M. Pecora, “Synchronizing chaotic circuits,” IEEE
Transaction on Circuits and Systems, vol. 38, no. 4, pp. 453–456, 1991.
[20] G. P. Jiang, W. K. S. Tang, and G. Chen, “A simple global synchro-
nization criterion for coupled chaotic system,” Chaos, Solitons and
Fractals, vol. 15, no. 5, pp. 925–935, 2003.
[21] A. I. Lerescu, N. Contandache, S. Oancea, and I. Grosu, “Collection
of master-slave synchronized chaotic systems,” Chaos, Solitons and
Fractals, vol. 22, no. 3, pp. 599–604, 2004.
516