Downloaded from rsta.royalsocietypublishing.

org on April 19, 2010

Phil. Trans. R. Soc. A (2010) 368, 21892204 doi:10.1098/rsta.2010.0032

Synchronization engineering: tuning the phase relationship between dissimilar oscillators using nonlinear feedback
BY CRAIG G. RUSIN1, *, HIROSHI KORI2 , ISTVN Z. KISS3 AND JOHN L. HUDSON1
of Chemical Engineering, University of Virginia, Charlottesville, VA 22902, USA 2 Division of Advanced Sciences, Ochadai Academic Production, Ochanomizu University, Tokyo 112-8610, Japan 3 Department of Chemistry, Saint Louis University, St Louis, MO 63103, USA
A mild, nonlinear, time-delayed feedback signal was applied to two heterogeneous oscillators in order to synchronize their frequencies with an arbitrary and controllable phase difference. The feedback was designed using phase models constructed from experimental measurements of the intrinsic dynamical properties of the oscillators. The feedback signal produced an interaction function that corresponds to the desired collective behaviour. The synchronized phase difference between the elements can be tuned to any value on the interval 0 and 2p by shifting the phase of the interaction function using the feedback delay. Numerical simulations were conducted and experiments carried out with electrochemical oscillators.
Keywords: synchronization; phase models; nonlinear feedback
1 Department

1. Introduction
The collective dynamical behaviour of a rhythmic population is dependent on the interactions between individual elements. Coupling among rhythmic elements can lead to synchronization, where a number of rhythmic elements organize into a single group with a uniform phase and frequency. Such synchronization can lead to coherent light emission from laser systems (Petrov et al. 1997; Oliva & Strogatz 2001), neuronal clustering in the suprachiasmatic nucleus (Iglesia et al. 2000; Yamaguchi et al. 2003), oscillations in chemical systems (Epstein & Pojman 1998), epileptic events (Traub & Wong 1982) and Parkinsonian tremors (Tass 1999). By manipulating the interactions between individual elements, it is possible to steer the collective behaviour of the system to a desired state. A number of methods have been developed to control the collective behaviour of rhythmic systems (Battogtokh & Mikhailov 1996; Popovych et al. 2006). Proportional integralderivative controller-based feedback has been used to create phase-locked states (Di Donato et al. 2007), feedback based on mutual information has been
*Author for correspondence (cgrusin@virginia.edu). One contribution of 10 to a Theme Issue Experiments in complex and excitable dynamical systems.

2189

This journal is 2010 The Royal Society

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2190

C. G. Rusin et al.

shown effective in controlling the characteristic time scales of oscillators (Belykh et al. 2005) and tunable heterogeneities have been used to steer the behaviour of excitable media (Mikhailov & Showalter 2006). We have developed a general engineering methodology based on phase models that allows precise control over the collective dynamical behaviour of a rhythmic system using global nonlinear time-delayed feedback. A key advantage of this method is that it can be tailored to the unique dynamical behaviour of a physical system. This is accomplished by using experimental measurements to calibrate the phase model to the target system, eliminating the need for physico-chemical models. Mild feedback is used to gently steer the collective behaviour of the target system towards the desired state, ensuring that the rhythmic behaviour of the individual elements of the system is preserved. This method has been used to generate dynamical behaviours such as phase synchronization, phase desynchronization, clustering and itinerant clustering within a population of rhythmic electrochemical elements (Kiss et al. 2007; Kori et al. 2008). In this paper, we use our methodology to address the general question of how to tune the stationary phase difference between two dissimilar oscillators without prior knowledge of their underlying dynamic behaviour. This twooscillator system represents a canonical model for a large class of rhythmic biological systems (Winfree 1980; Iglesia et al. 2000) and incorporates many of the non-idealities that are present in typical experimental systems, yet has a simple analytical solution. Heterogeneities are common in experimental and natural rhythmic systems where the intrinsic oscillator frequency is not under experimental control. We demonstrate how to choose and obtain an interaction function that produces an arbitrary phase difference between two rhythmic elements, and we demonstrate the application of the method using an experimental chemical system.

2. General methodology
The phase behaviour of a population of oscillating elements can be approximated in the limit of weak global interactions as (Kuramoto 1984) K dfi = ui + dt N
N

H (fj fi ),
j=1

(2.1)

where fi and ui are the phase and inherent frequency of element i, K is the interaction strength, N is the number of elements in the population and H (Df) is an interaction function that describes the effect one element has on all other elements in the system. The interaction function can be derived from measurable quantities of the experimental system as H (Df) = 1 2p
2p 0

Z (f)h x(f + Df) df,

(2.2)

where Z (f) is the response function, x(f) is the waveform and h(x) is the coupling function between the elements. The response function, Z (f), represents the sensitivity of a single element to perturbations along its limit cycle. This function is an intrinsic property of the oscillator and can be experimentally measured
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

Synchronization engineering

2191

(Galan et al. 2005; Kiss et al. 2005; Tsubo et al. 2007). The coupling function, h(x), represents the form of the physical connection between the elements in the system. For the simulations and experiments described in this work, the interactions among elements are produced by global nonlinear time-delayed feedback of the form
S

h(x(t)) =
n=0

kn x(t tn Dt)n ,

(2.3)

where t is the time, S is the polynomial feedback order, kn is the nth order polynomial feedback coefcient, tn is the nth order polynomial feedback delay and Dt is a feedback delay that is common to all polynomial terms. By manipulating the shape of the interaction function, it is possible to control the collective behaviour of the system. In general, this can be accomplished using a three-step process: determine a target interaction function that will produce the desired global behaviour, measure the response function of the rhythmic elements to be used and numerically optimize the shape of the coupling function to achieve the target interaction function.

3. Tuning the phase difference between two rhythmic elements (a) Determining the required interaction function
Applying the Kuramoto phase approximation (equation (2.1)) to a system of two non-identical elements yields equations for the two phase variables {f1 , f2 } which when subtracted give a relationship for the phase difference, Df = f1 f2 , between the elements dDf K = Du + [H (Df) H (Df)], dt 2 (3.1)

where Du is the difference between the natural frequencies of the two oscillators. Equation (3.1) can be simplied to dDf = Du KH (Df), dt where H is the odd part of the interaction function, dened as 2H (Df) = H (Df) H (Df). (3.3) (3.2)

From equation (3.2), it is seen that stationary solutions will occur at phase differences that satisfy Du = H (Df). (3.4) K For a homogeneous system (identical elements, Du = 0), the stationary solutions correspond to the roots of H (Df). The behaviour of a slightly heterogeneous system (Du = 0) can be brought close to that of a homogeneous system by increasing the interaction strength between the elements. In the limit as K ,
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2192

C. G. Rusin et al.

the xed point solutions approach the roots of H (Df). However, caution must be used with this approach since an excessively large interaction strength will disrupt the waveform and invalidate the phase approximation. For a given interaction function H (Df), the stationary phase difference, Df , is obtained from the roots of its odd part, H (Df ) 2H (Df ) = H (Df ) H (Df ) = 0. (3.5) By shifting the phase of the interaction function by an amount Dt, Df can be tuned to take on values between 0 and 1 rad/2p. (As shall be seen below, the phase shift, Dt, of H (Df) is equal to the common feedback delay in the experiments to be discussed.) The stationary phase differences between the two oscillating elements are then given by the solutions of H Df + Dt H Df + Dt = 0. (3.6)

By denition, H (Df) is a 2p periodic function; therefore equation (3.6) must always have roots at 0 and 0.5 rad/2p. These roots represent the trivial inphase and anti-phase synchronization states, respectively. The non-trivial roots of equation (3.6) are governed by the higher order terms of H (Df). An interaction function with second-order harmonics is sufcient to obtain a non-trivial, stationary phase difference between the two oscillators. A convenient form for this purpose is the function H (Df) = sin(Df + Dt) + R sin(2(Df + Dt)). (3.7)

Figure 1a illustrates the interaction function dened in equation (3.7) setting R = 0.1 and Dt = 0, while gure 1b illustrates its associated synchronized states as a function of Dt. The non-trivial bifurcation branches illustrated in gure 1b are relatively narrow, indicating that small changes in the parameter Dt will cause large changes in the synchronized state. To reduce this sensitivity, the parameter R was selected to maximize the width of the non-trivial bifurcation branches. To simplify the analysis, the parameters of equation (3.7) were constrained such that only supercritical bifurcations occur. To determine the optimum value of R, a linear stability analysis of equation (3.2) was performed, indicating that stable stationary states occur when dH (Df)/d Df > 0. Therefore, the phase-locked solution, Df*, is stable if H (Df + Dt) H (Df + Dt) < 0. (3.8) From this, it can be seen that the in-phase synchronization state (Df = 0) is stable if (3.9) H (Dt) > 0, while the anti-phase state is stable if H (Dt + p) > 0. Equations correspond bifurcation correspond bifurcation (3.10)

(3.9) and (3.10) indicate that the trivial bifurcation branches to the region where H (Df) > 0. Since the trivial and non-trivial branches are mutually exclusive, the non-trivial branch must to the region where H (Df) < 0. Therefore, a wide non-trivial branch will occur when H (Df) has a large domain of negative slope.

Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

Synchronization engineering
(a) 1.5 1.0 H (rad s1) 0.5 0 0.5 1.0 1.5 Df* (rad/2p) (b) 1.00 0.75 0.50 0.25 0

2193

(c) 1.5 1.0 H (rad s1) 0.5 0 0.5 1.0 1.5 0 0.2 0.4 0.6 Df (rad/2p) 0.8 1.0

(d) 1.00 Df* (rad/2p) 0.75 0.50 0.25 0 0.1 0.2 0.3 Dt (rad/2p) 0.4 0.5

Figure 1. Relationship between H (Df) and the time asymptotic phase behaviour of a twooscillator system. (a) Interaction function H (Df) = sin(Df + Dt) + R sin(2(Df + Dt)), R = 0.1, Dt = 0 rad/2p, and (b) the calculated oscillator phase difference. Solid lines denote stable states, and dotted lines denote unstable states. (c) Optimized interaction function (R = 0.5, Dt = 0 rad/2p) and (d) the calculated oscillator phase difference.

The value for the parameter R of equation (3.7) was selected to maximize the width of the non-trivial branch of the bifurcation diagram, that is where H (Df) < 0. The value of R was constrained to the interval 0 R 0.5 owing to the presence of subcritical bifurcations outside this interval. Numerical analysis determined that the width of the region where H (Df) < 0 monotonically increased with the value of R in this interval. As a result, the widest non-trivial bifurcation branch occurred when R = 0.5. The target interaction function can be seen in gure 1c for Dt = 0. With these parameters, equation (3.7) can be analytically solved for H (Df ) = 0, yielding Df = 0, p, arccos(cos(Dt)/ cos(2Dt)). The stable and unstable stationary states are shown as a function of Dt in gure 1d.

(b) Numerical simulation of phase difference tuning

Before conducting experiments on the experimental system, the methodology was veried with numerical simulations using the Brusselator oscillator, a simple two-variable ODE system that exhibits self-sustained oscillations (Glansdorff & Prigogine 1971). The phase difference between two interacting Brusselator
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2194
(a) x (f) 1.0 0.5 0 0.5 0

C. G. Rusin et al.
(b) 4 Z (f) 0.5 f (rad/2p) 1.0 2 0 2 0 0.5 f (rad/2p) 1.0

Figure 2. (a) Waveform and (b) phase response function of the Brusselator oscillator. The parameter values are A = 1.0 and B = 2.3, resulting in T = 6.43 s and u = 0.977 rad s1 .

oscillators is tuned to values between 0 and 2p. The dynamical equations for an interacting homogeneous Brusselator population under global feedback are K N dxi = (B 1)xi + A2 xi + f (xi , yi ) + h(xj ) dt N j=1 dyi 2 (3.11) for i = 1, . . . , N , dt = Bxi A yi f (xi , yi ) f (x, y) = B x 2 + 2Axy + x 2 y A where h(x) is the feedback function dened in equation (2.3) that is constructed from and applied to the variables xi . The variables xi and yi are transformed such that the xed point is shifted to (x, y) = (0, 0). For a single uncoupled oscillator, a Hopf bifurcation occurs at B = Bc 1 + A2 . The parameters of equation (3.11) were chosen to be A = 1.0 (so that Bc = 2.0) and B = 2.3. The waveform x(f) and the response functions Z (f) along the x-direction are displayed in gure 2. The response function was calculated using the XPPAUT program (Ermentrout 2002). A feedback parameter set {kn , tn } is chosen to create a tunable non-trivial solution such that an arbitrary stationary phase difference Df can be achieved. The target interaction function was selected to be equation (3.7), where R = 0.5. The phase shift of the interaction function can be directly controlled using the common feedback delay parameter Dt. For the purpose of determining the feedback parameters, the common feedback delay can be set to zero. Second-order feedback is used in order to produce the required rst and second harmonics in the target interaction function. The constant term in equation (2.3) is arbitrary and can be safely neglected (k0 = 0). Although both the linear and quadratic terms of the feedback function contain weak third (and higher) order harmonics, their effect on the interaction function is negligible since the third (and higher) order harmonics of the response function are small. The feedback parameters are found by numerically solving for h(x) (Kori et al. 2008), with one solution being (k1 , k2 , t1 , t2 ) (2.56, 4.68, 5.34, 1.98). As expected, the magnitudes of the higher harmonics are very small, |h3 Z3 | 0.01 and |h4 Z4 | 0.004, (3.13) (3.12)

for the parameter set (equation 3.12).

Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

Synchronization engineering
p

2195

Df*

p/2

0 0.8

1.2

1.6 a (rad)

2.0

2.4

Figure 3. Numerical simulation using a system of two Brusselator oscillators. Stationary phase difference Df as a function of a, where a uDt. The lines show the theoretical prediction (stable, solid; unstable, dotted) from the phase model. Data points are numerically obtained, where Df for p < Df < 2p is plotted as 2p Df for convenience.

The parameter set in equation (3.12) is veried by direct numerical simulations of two coupled Brusselators (equation (3.11) with N = 2), starting from random initial conditions. The feedback parameters of equation (2.3) were set to be (k1 , k2 , t1 , t2 , Dt) = (2.56, 4.68, 5.34, 1.98, u1 a), where u is the natural frequency of the oscillator and a is a variable control parameter. After a short transient period, the phase difference between the elements achieves a nearly constant value. Figure 3 displays Df for 0 < Df p and 2p Df for p < Df 2p, which shows excellent agreement with theoretical predictions. Other parameter sets found numerically from equation (2.2) give quantitatively similar results (data not shown).

4. Experimental demonstration of phase difference tuning (a) Experimental setup

Experiments were conducted using an apparatus consisting of an electrochemical cell, a set of zero resistance ammeters (ZRAs) and a real-time data acquisition system (gure 4). The cell consisted of two Ni electrodes (99.99% pure, 1 mm diameter) in a 3 M H2 SO4 solution, a Pt mesh counter electrode and an Hg/Hg2 SO4 /K2 SO4 (sat) reference electrode. The cell was enclosed in a jacketed glass vessel held at a constant temperature of 11 C. An EG&G potentiostat was used to adjust the circuit potential, V0 , of the cell, causing the nickel electrodes to undergo transpassive dissolution. The dissolution current of each electrode, Im (t), was measured by a ZRA. A 650 U resistor (Rp ) was attached to each channel to induce oscillations in the electrode potential (Zhai et al. 2004). A Labview-based real-time data acquisition computer was used to acquire data from the ZRA, send the data to a host machine, calculate a global feedback signal from these data and apply the feedback signal to the potentiostat at 250 Hz. The measurements of Offset dissolution current were scaled such that the mean value of each channel (Im ) was removed and then normalized by the amplitude of its oscillation (Am ) relative
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2196
counter electrode

C. G. Rusin et al.

reference 3 M H2SO4 working electrode potentiostat

V(t) = V0 + dV Rp ZRA I(t) computer feedback (dV)

Figure 4. Schematic diagram of the experimental apparatus. Rp represents a set of resistors, one for each element in the population. The computer is a real-time data acquisition computer, which calculates the feedback signal, dV .

to the mean amplitude of the population (Amean ): Amean Offset (Im (t) Im ). I m (t) = Am (4.1)

The host machine was used to continuously determine the offset and amplitude of each rhythmic element in the population over time. The feedback signal, dV , fed back to the potentiostat was calculated using the equation dV = K N
N

h(xm (t)),
m=1

(4.2)

where K is the overall feedback gain (i.e. interaction strength), h(x) is equation (2.3) and xm (t) is the potential drop across the double layer for the mth electrode, calculated as (4.3) xm (t) = V (t) I m (t)Rp , where V (t) is the applied voltage and Rp is the channel resistance. Before the nickel electrode array was placed into the system, it was polished with a wet rotary polisher to remove any initial oxide layer that may have been present. Six polishing discs were used, decreasing in roughness from 180 to 4000 grit. This ensured that the electrodes started from roughly identical conditions and that these initial conditions were approximately identical for each experiment. After polishing, the electrodes were placed in the acid solution and the potential was ramped from 0.68 to 1.25 V and back to 0 V without any resistors present. This caused a thin passive oxide layer to form on each electrode. After this cycle was complete, the resistors were reconnected and the system was brought up to the desired operating voltage. The system was allowed to line out at this voltage
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

Synchronization engineering
(a) 1.15 1.10 E (V) 1.05 1.00 0.95 0.90 0 2 4 6 time (s) 8 10 Df (rad/2p) (b) 1.4 1.2 1.0 0.8 0.6 0 6 12 18 24 time (s) 30

2197

36

Figure 5. (a) Time series of electrode potential, E, during electrochemical dissolution of two nickel wires in 3 M sulphuric acid without feedback (V = 1.165 V, Rp = 650 U, u1 = 0.4796 Hz, u2 = 0.5105 Hz). (b) Unbounded phase difference between the two waveforms seen in (a).

for approximately 1.52.0 h, as this was found to reduce the amount of drift in the inherent frequency of the oscillators. The electrode potential waveform of both elements can be seen in gure 5a. In the absence of feedback, the phase difference between the two elements increases over time (gure 5b).

(b) Determining the response function

The response function can be determined from equation (2.2), provided that the interaction function and feedback function are both known. Work by Miyazaki & Kinoshita (2006) was extended to provide a method for measuring the interaction function of a rhythmic system composed of two non-identical oscillators under weak global feedback. Physically, the interaction function represents the change in the frequency of a rhythmic element as a function of its phase relative to all other interacting elements in the system. Therefore, the interaction function can be determined by recording the frequency (or period) of two interacting oscillators as a function of their phase difference. Two electrochemical oscillators were created, with different inherent frequencies, and a weak feedback signal was applied to the system. The magnitude of the feedback was selected such that the system was unable to synchronize, allowing the phase difference between the elements to grow over time (gure 6a). The weak interaction caused the period of the two oscillators to uctuate as the phase difference between the elements changed (gure 6b). By integrating the instantaneous frequency of one of the oscillators over the observed period of a single cycle and equating this to 2p rad (Miyazaki & Kinoshita 2006), the interaction function can be determined using the equation H (Df) = 2p [P(Df) Pbase ], 2 KPbase (4.4)

where Pbase is the intrinsic period of the oscillator and K is the overall feedback gain. Measuring interaction functions under different feedback conditions allowed the response function (gure 6d) to be calculated from equation (2.2). These sets of data were used in a multi-objective optimization to nd the single best
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2198
(a) 4
3 Df (rad/2p) 2 1

C. G. Rusin et al.
(b) 2.10

period (s)

2.05

2.00

50

100 time (s)

150

200

1.95

50

100 time (s)

150

200

(c) 1.0
H(Df) (rad s1) 0.5

(d ) 40
20 Z (rad/2p) 0 20 40 E (V)

1.15

0.85

0 0.5 1.0

2p

0.25

0.50 Df (rad/2p)

0.75

1.00

0.25

0.50 f (rad/2p)

0.75

1.00

Figure 6. Measuring a response function. (a) Unbounded phase difference of the heterogeneous smooth oscillator seen for the experimental system under rst-order feedback. K = 0.07, k0 = 0 V, k1 = 1, t1 = 0.012 rad/2p. (b) Fluctuations in the period of a single oscillator owing to feedback. (c) Measured interaction function obtained using equation (4.4). (d) Optimized response function Z (f) and waveform (inset) of a single oscillator.

response function that could reproduce all the measured interaction functions simultaneously. This tting was necessary to reduce the effects of experimental noise in the interaction function measurements. The Fourier coefcients of the response function were used as the optimization parameters. The number of Fourier terms to be optimized was determined by the number of higher harmonics present in the waveform and the measured interaction functions. Close to the Hopf bifurcation point (approx. 1.105 V), only one to two terms were necessary since the system was largely sinusoidal. At higher circuit potentials (approx. 1.20 V), the system became relaxational and the response function required approximately seven terms. At the experimental circuit potential of 1.165 V, four Fourier terms were used since the system was at a slightly higher voltage than the Hopf bifurcation point, but still exhibited relatively smooth oscillations. Additionally, it was found that increasing the number of coefcients did not signicantly alter the shape of the response function. A simplex optimization algorithm was used to determine the value of the Fourier coefcients by minimizing an objective function of the form
N Pts 1/2

error =
n=1 i=1

(|Tin

Din | ) 2

(4.5)

Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

Synchronization engineering
(a) 2
1 H (rad s1) 0.5 0 1 1.0 2 0 0.25 0.75 0.50 Df (rad/2p) 1.00 1.5 0 0.25 0.75 0.50 Df (rad/2p) 1.00 0 0.5

2199

(b) 1.5
1.0

Figure 7. (a) The target interaction function (dotted line) and the optimized interaction function (solid line) determined by feedback parameter optimization. (b) Interaction function measured using the experimental system (dots) and phase model prediction (line) using the response function in gure 6 and feedback parameters (K = 0.03, k0 = 0.03 V, k1 = 1.72, k2 = 4.6816 V1 , t1 = 0.012 rad/2p, t2 = 0.143 rad/2p).

where Tin is the ith data point of the nth measured interaction function, D is the interaction function calculated using the optimized response function, N is the number of measured interaction function datasets and Pts is the number of measured points in the function. The initial conditions were taken to be the Fourier coefcients of a sine wave, since the response function approaches a sine wave as the system approaches the Hopf bifurcation point. On average, the optimization required approximately 1200 iterations to converge, corresponding to approximately 20 min.

(c) Calculating the feedback signal

The target interaction function (gure 7a) was engineered into the experimental system using the nonlinear time-delayed feedback dened in equation (2.3). The feedback parameters kn and tn can be calculated using equation (2.2), provided that the response function of the system and the target interaction function are known. The common feedback delay, Dt, was set to zero for this calculation. A simplex optimization algorithm was used to minimize the error between the calculated and target interaction functions by manipulating the feedback parameters. The objective function for this optimization was
Pts 1/2

error =
i=1

(|Ti Hi |)

(4.6)

where Ti is the ith data point of the target interaction function, H is the optimized interaction function and Pts is the number of data points in the functions T and H . Second-order feedback was used, since this is the highest order harmonic present in the target interaction function. The initial conditions for the optimization were taken to be kn = 10n1 and tn = 0.01 rad/2p. This allows each polynomial feedback term to be of the same order of magnitude, since |x| < 1. The optimization successfully found a feedback parameter set that produced the
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2200

C. G. Rusin et al.

target interaction function with sufcient accuracy: (K , k0 , k1 , k2 , t1 , t2 ) = (0.03, 0.03 V, 1.72, 4.6816 V1 , 0.012 rad/2p, 0.143 rad/2p). (4.7) Figure 7a shows the optimized interaction function (solid line) compared with the target interaction function (dashed line). There is good agreement between the target interaction function and the optimized interaction function. Since the optimized function has a large domain of negative slope, it will produce a large non-trivial synchronization region. The feedback parameters were experimentally validated by measuring the interaction function produced by the optimized feedback conditions. Figure 7b illustrates the agreement between the predicted interaction function and the measured interaction function, indicating that the optimized parameters were successful in producing the desired interaction function.

(d) Experimental results

The feedback parameters (equation (4.7)) were applied to the experimental system and the common feedback delay, Dt, was adjusted to achieve the desired phase difference between the two rhythmic elements. The phase of an element was linearly interpolated between adjacent peaks in the observed waveform; the phases of the peaks were dened as 0 and 2p rad, respectively. When Dt = 0 rad/2p, the elements phase-synchronized with Df = 0 rad/2p (gure 8a,d). Under these conditions, the system exhibited only one stable stationary state, corresponding to the root of H (Df) located at Df = 0 (gure 8g). Increasing Dt to 0.23 rad/2p caused the interaction function to shift, changing the synchronized phase difference to Df = 0.73 rad/2p (gure 8b,e). In this case, the system exhibited bi-stability, in which two stable stationary states (Df = 0.73 and Df = 0.28 rad/2p) coexist simultaneously (gure 8h). Further increasing Dt to 0.5 rad/2p caused the elements to synchronize in an anti-phase conguration, where Df = 0.5 rad/2p (gure 8c,f ). This anti-phase conguration corresponded to the root of H (Df) located at Df = 0.5 rad/2p (gure 8i), which was previously unstable. A more quantitative comparison between the phase model predictions and experimental results can be seen in gure 9. The bifurcation diagram in gure 9a was generated by tracking the roots of H (Df) (and their associated derivatives) as a function of Dt. Therefore, the diagram represents all the possible stationary states (both stable and unstable) for the experimental system at different values of Dt. To determine the accuracy of these predictions, the stationary phase difference between two electrochemical elements was experimentally measured with different values of Dt. These measurements were made by rst synchronizing the system on the upper non-trivial branch of the bifurcation diagram using the appropriate delay and initial conditions. When the initial phase difference between the elements was 0.5 < Df < 1 rad/2p, the system approached the upper bifurcation branch, otherwise it approached the lower branch. Once on the upper branch, Dt was adjusted to probe the extent of the branch. The global feedback was not removed during this process, ensuring that the system remained on the upper branch during these experiments. Once the upper branch was mapped,
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

Synchronization engineering
(a) 1.15
1.10 E (V) 1.05 1.00 0.95 0.90

2201
(c)

(b)

(d) 1.00
Df (rad/2p) 0.75 0.50 0.25 0 2 time (s) 4 6

(e)

(f)

2 time (s)

2 time (s)

(g) 1.5
1.0 0.5 0 0.5 1.0 1.5 H(rad s1)

(h) 0.6
0.4 0.2 0 0.2 0.4 0 0.25 0.50 0.75 Df (rad/2p) 1.00 0 0.25 0.50 0.75 Df (rad/2p)

(i) 1.5
1.0 0.5 0 0.5 1.0 1.5

1.00

0.25 0.50 0.75 Df (rad/2p)

1.00

Figure 8. (ac) Time series of the electrode potential, (df ) phase difference and (gi) H (Df) of a system of two elements with second-order global feedback (K = 0.03, k0 = 0.03 V, k1 = 1.72, k2 = 4.6816 V1 , t1 = 0.012 rad/2p, t2 = 0.143 rad/2p). (a,d,g) In-phase synchronization (Dt = 0 rad/2p). (b,e,h) Representative out-of-phase synchronization (Dt = 0.23 rad/2p). (c,f ,i) Antiphase synchronization (Dt = 0.5 rad/2p). The phase loop diagrams indicate the relative position of the elements in the system and the direction of rotation. The roots of H (Df) are indicated with the associated stationary state stability (black circles denote stable states, grey squares denote unstable states).

(a) 1.00 0.75 0.50 0.25

(b) 1.00 phase model (rad/2p) 0.1 0.2 0.3 Dt (rad/2p) 0.4 0.5 0.75 0.50 0.25

Df*(rad/2p)

0.25 0.50 0.75 experimental data (rad/2p)

1.00

Figure 9. (a) Stationary phase difference values of a system of two rhythmic elements under second-order feedback (K = 0.03, k0 = 0.03 V, k1 = 1.72, k2 = 4.6816 V1 , t1 = 0.012 rad/2p, t2 = 0.143 rad/2p). Lines represent phase model predictions of the stable stationary states (solid) and unstable stationary states (dotted). Circles represent experimental measurements with positive feedback, and triangles represent measurements with negative feedback. (b) Parity plot of experimental measurements versus phase model predictions.
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2202

C. G. Rusin et al.

the feedback was removed and the system was allowed to drift into the range of initial conditions required to reach the lower branch (0 < Df < 0.5 rad/2p). Upon reaching these conditions, the feedback was reapplied and the lower branch was mapped using the same process. The unstable stationary states were experimentally probed using the negative feedback; the overall feedback gain, K , was multiplied by 1, while all the other feedback parameters remained constant. According to the linear stability analysis of equation (3.2), changing the sign of K inverts the stability of the xed points. Therefore, the unstable states under positive feedback become stable under negative feedback, allowing for experimental observation. Excellent agreement was found between experimental measurements and theoretical predictions for both the stable and unstable stationary states (gure 9b).

5. Conclusions
We have demonstrated the application of a synchronization methodology by controlling the stationary phase difference between two oscillators with nonlinear feedback; we demonstrate the method in both numerical simulations and chemical experiments. Our methodology is based on the fact that the existence and stability conditions of dynamical states in weakly coupled oscillators are characterized by the interaction function in the phase model. Thus, by knowing the interaction function that results in a desired collective behaviour, the only remaining issue is how to construct the physical interaction that yields the phase interaction function. An interaction function can be constructed using the proposed feedback function (equation (2.3)). The choice of the specic form of the feedback function was motivated by the exible application of the imposed interaction function for synchronization engineering. It has been shown that the nth order term of the feedback signal effectively enhances the nth Fourier components of the interaction function. The time delay of the nth term is used to adjust the balance of even and odd parts of the Fourier components. As a result, all the harmonics in the interaction function can be tuned by adjusting the appropriate polynomial gain and time delay (Kori et al. 2008). The proposed methodology is applied to tune the phase difference of uncoupled oscillators through feedback perturbations. The methodology is still applicable when the oscillators are relatively weakly coupled (even with time varying strengths) because the feedback signal can dominate the relatively weak inherent coupling effects within the limitations of the phase model description; a similar approach was previously used (Kiss et al. 2007) for the problem of desynchronization of a population of coupled oscillators. Extension of the proposed method is required for the application to strongly interacting oscillators and to a population of oscillators, especially when there is a range of inherent coupling strengths. Nonetheless, the general framework proposed here should serve as a guideline for tackling these important problems that are expected to arise in biology. A major advantage of this methodology is that the feedback, which results in a target interaction function, can be designed through the knowledge of the macroscopic observable of an isolated oscillator, that is, the waveform and the phase response function. This point is benecial when applications to biological
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

Synchronization engineering

2203

systems are considered; it is usually a formidable task to construct an appropriate detailed mathematical model of a biological system; however, the investigation of the phase response function is often possible.
This work was supported in part by the National Science Foundation through grant CBET-0730597. I.Z.K. thanks the organizers for nancial aid to attend ECC10. H.K. acknowledges nancial support from the Grants-in-Aid for Young Scientists (no. 19800001) and the Sumitomo Foundation (no. 071019).

References
Battogtokh, D. & Mikhailov, A. 1996 Controlling turbulence in the complex GinzburgLandau equation. Physica D 90, 8495. (doi:10.1016/0167-2789(95)00232-4) Belykh, V. N., Osipov, G. V., Kucklander, N., Blasius, B. & Kurts, J. 2005 Automatic control of phase synchronization in coupled complex oscillators. Physica D 200, 81104. (doi:10.1016/ j.physd.2004.10.008) Di Donato, P. F. A., Macau, E. E. N. & Grebogi, C. 2007 Phase locking control in the circle map. Nonlinear Dyn. 47, 7582. (doi:10.1007/s11071-006-9055-7) Epstein, I. R. & Pojman, J. A. 1998 Chemical dynamics: oscillations, waves, patterns, and chaos. Oxford, UK: Oxford University Press. Ermentrout, G. B. 2002 Simulating, analyzing, and animating dynamical systems: a guide to XPPAUT for researchers and students. Philadelphia, PA: SIAM. Galan, R. F., Ermentrout, G. B. & Urban, N. N. 2005 Efcient estimation of phase-resetting curves in real neurons and its signicance for neural-network modeling. Phys. Rev. Lett. 94, 158 101. (doi:10.1103/PhysRevLett.94.158101) Glansdorff, P. & Prigogine, I. 1971 Thermodynamic theory of structure, stability, and uctuations. London, UK: Wiley. Iglesia, H. O., Meyer, J., Carpino, A. & Schwartz, W. J. 2000 Antiphase oscillation of the left and right suprachiasmatic nuclei. Science 290, 799801. (doi:10.1126/science.290.5492.799) Kiss, I. Z., Zhai, Y. M. & Hudson, J. L. 2005 Predicting mutual entrainment of oscillators with experiment-based phase models. Phys. Rev. Lett. 94, 248 301. (doi:10.1103/PhysRevLett. 94.248301) Kiss, I. Z., Rusin, C. G., Kori, H. & Hudson, J. L. 2007 Engineering complex dynamical structures: sequential patterns and desynchronization. Science 316, 18861889. (doi:10.1126/ science.1140858) Kori, H., Rusin, C. G., Kiss, I. Z. & Hudson, J. L. 2008 Synchronization engineering: theoretical framework and application to dynamical clustering. Chaos 18, 026 111. (doi:10.1063/1.2927531) Kuramoto, Y. 1984 Chemical oscillations, waves and turbulence. New York, NY: Springer. Mikhailov, A. S. & Showalter, K. 2006 Control of waves, patterns and turbulence in chemical systems. Phys. Rep. 425, 79194. (doi:10.1016/j.physrep.2005.11.003) Miyazaki, J. & Kinoshita, S. 2006 Method for determining a coupling function in coupled oscillators with application to BelousovZhabotinsky oscillators. Phys. Rev. E 74, 056 209. (doi:10.1103/PhysRevE.74.056209) Oliva, R. A. & Strogatz, S. H. 2001 Dynamics of a large array of globally coupled lasers with distributed frequencies. Int. J. Bifurcation Chaos 11, 23592374. (doi:10.1142/ S0218127401003450) Petrov, V., Ouyang, Q. & Swinney, H. L. 1997 Resonant pattern formation in a chemical system. Nature 388, 655657. (doi:10.1038/41732) Popovych, O. V., Hauptmann, C. & Tass, P. A. 2006 Control of neuronal synchrony by nonlinear delayed feedback. Biol. Cybern. 95, 6985. (doi:10.1007/s00422-006-0066-8) Tass, P. A. 1999 Phase resetting in medicine and biology. Stochastic modeling and data analysis. Berlin, Germany: Springer. Traub, R. D. & Wong, R. K. S. 1982 Cellular mechanism of neuronal synchronization in epilepsy. Science 216, 745747. (doi:10.1126/science.7079735)
Phil. Trans. R. Soc. A (2010)

Downloaded from rsta.royalsocietypublishing.org on April 19, 2010

2204

C. G. Rusin et al.

Tsubo, Y., Takada, M., Reyes, A. & Fukai, T. 2007 Layer and frequency dependencies of phase response properties of pyramidal neurons in rat motor cortex. Eur. J. Neurosci. 25, 34293441. (doi:10.1111/j.1460-9568.2007.05579.x) Winfree, A. T. 1980 The geometry of biological time. New York, NY: Springer. Yamaguchi, S., Isejima, H., Matsuo, T., Okura, R., Yagita, K., Kobayashi, M. & Okamura, H. 2003 Synchronization of cellular clocks in the suprachiasmatic nucleus. Science 302, 14081412. (doi:10.1126/science.1089287) Zhai, Y., Kiss, I. Z. & Hudson, J. L. 2004 Emerging coherence of oscillating chemical reactions on arrays: experiments and simulations. Ind. Eng. Chem. Res. 43, 315326. (doi:10.1021/ ie030164z)

Phil. Trans. R. Soc. A (2010)