Phase locking in chains of multiple-coupled oscillators

Physica D 143 (2000) 56–73
Phase locking in chains of multiple-coupled oscillators
Liwei Ren, Bard Ermentrout∗
Department of Mathematics, University of Pittsburgh, Pittsburgh, PA 15260, USA
Abstract
Phase locking in chains of weakly coupled oscillators with coupling beyond nearest neighbors is studied. Starting with a
piecewise linear coupling function, a homotopy method is applied to prove the existence of phase locked solutions. Numerical
examples are provided to illustrate the existence and the properties of the solutions. Differences between multiple coupling
and nearest neighbor coupling are also discussed. © 2000 Elsevier Science B.V. All rights reserved.
Keywords: Phase locking; Multiple-coupled oscillators; Coupling beyond nearest neighbor
1. Introduction
Weakly coupled oscillator arrays arise in many physical and biological systems. In particular, one-dimensional
chains of oscillators have been used to model a variety of biological systems such as the swim generator in the
lamprey [1] and olfactory waves in the procerebral lobe of the garden slug [5]. These models arise from general
systems of coupled oscillators under the assumption that the interactions between oscillators are sufﬁciently weak.
Under this “weak coupling” assumption, each oscillator is reducible to a single variable that describes the phase.
The most general form that these phase equations can take is
θi = ωi + Hi(θ1 − θi, . . . , θn − θi), i = 1, . . . , n,
where the functions Hi are 2π-periodic in each of their arguments and the parameters ωi are the local variations
in uncoupled frequency. Typically, we are interested in solutions that are periodic, i.e., θi(t + T ) = θi(t) + 2π.
The stability of solutions for general coupling was studied in [4], however, the structure of the solutions is never
discussed.
The most comprehensive results concern either globally coupled all-to-all systems of oscillators, e.g., [2] or [3],
or chains of oscillators with nearest-neighbor coupling [6,7,14]. In the latter papers, phase locked solutions were
analyzed which correspond to traveling waves. Such waves have been observed in several central nervous system
preparations using imaging of the electrical potentials [9,13]. Recent experimental work, however, indicates that
the coupling in the lamprey spinal cord cannot be regarded as nearest neighbor [11]. Similarly, local application
of nitric oxide in the slug procerebral lobe indicates that coupling between oscillators extends beyond the nearest
∗ Corresponding author. Tel.: +1-412-624-8324; fax: +1-412-624-8397.
E-mail address: bard@math.pitt.edu (B. Ermentrout)
0167-2789/00/$ – see front matter © 2000 Elsevier Science B.V. All rights reserved.
PII: S0167-2789(00)00096-8

L. Ren, B. Ermentrout / Physica D 143 (2000) 56–73 57
neighbors [8]. Thus, it is important to determine under what conditions oscillator chains that have coupling beyond
nearest neighbors can lead to phase locked solutions such as waves, which is the subject of the present paper.
Weak coupling in a chain of neurons or neural circuits simplifies the general structure of the phase models
considerably. Suppose the coupling strength depends only on the distance between two circuits. Since inputs to
neurons are treated independently and sum in a linear fashion, the resulting phase models have the general form
θi = ωi +
m
j=1
H+
j (θi+j − θi) +
m
j=1
H−
j (θi−j − θi), (1)
where i = 1, . . . , n + 1, θi is the phase and ωi is the frequency of the ith oscillator, and H±
j are 2π-periodic
functions of their arguments. We delete terms in the sum whenever i + j > n + 1 or i − j < 0 so that the
“boundary conditions” are those of a finite chain. The boundary effects are crucial and they make the analysis of
these equations difficult. We are interested in phase locked solutions, i.e., solutions for which θi is independent of
i and t. The equivalent equations, with the variables {θi} replaced by {φi = θi+1 − θi}, are considered. If θi =
which is the unknown frequency of the phase locked ensemble of the oscillators, then (1) becomes
= ωi +
m
j=1
H+
j


j
k=1
φi+k−1

 +
m
j=1
H−
j

−
j
k=1
φi−k

 . (2)
It was shown [6,7] that phase locked solutions of chains with nearest neighbor coupling could be approximated,
when there is a large number of oscillators, by passing to a continuum limit and analyzing the solutions of the
resulting singularly perturbed second-order two-point boundary value problem (BVP). Thus, over much of the
chain, the solution behaves like a solution to a first-order “outer equation”. The particular “outer equation” is
determined by the boundary conditions for the BVP. In [14], we considered chains with finitely many oscillators. It
was shown that under weak assumptions on the coupling functions, the phase lags between successive oscillators
have the property of monotonicity provided that the frequency difference between any two successive oscillators is
a sufficiently small constant along the chain. This implies that most chains of locally coupled oscillators that phase
lock will form traveling wave solutions similar to those found in the limit of large n in [6,7].
Kopell et al. [10] considered the problem of chains with m neighbors in the limit as the number of oscillators
tends to infinity. In this limit, phase locked solutions of (2) may be viewed as a one-parameter family of (2m −
1)th-order discrete dynamical systems, where the independent variable is the position along the chain and whose
dependent variable is the phase difference between successive oscillators. In [10] it was shown that for each value
of the parameter in some range, the (2m − 1)th-order system has a one-dimensional hyperbolic global center
manifold. This was done by using the theory of exponential dichotomies to show the system “shadows” a simple
one-dimensionalsystem.Forafinitechain,thedynamicalsystemisconstrainedbymanifoldsofboundaryconditions.
It was shown that for open sets of such conditions, the solution to the equation for phase locking in long chains
stays close to the center manifold except near the boundaries. These facts were used to show that a multiply coupled
system behaves, except near the boundaries, as a modified nearest-neighbor system. The existence of asymptotically
stable phase locked solutions was proven provided that the chain is long enough and the frequencies of oscillators
are sufficiently close.
In this paper, a special form of Eqs. (1) is considered for chains with finitely many oscillators, i.e., we do not
require that the length of the chain to tend to infinity. For simplicity, we assume n ≥ 2m + 1 (as a matter of fact, all
the results will also be true as long as n ≥ m + 2). The equations have the following form:
θi = ωi +
m
j=1
α+
j H+
(θi+j − θi) +
m
j=1
α−
j H−
(θi−j − θi), (3)

58 L. Ren, B. Ermentrout / Physica D 143 (2000) 56–73
where α±
1 ≥ α±
2 ≥ · · · ≥ α±
m > 0 and H± are 2π-periodic functions of their arguments. This particular form
is not unreasonable for neural models. If we assume that each local region oscillates in a similar manner and that
the coupling depends on the distance between units, then this form is quite natural. With these assumptions (2)
becomes
= ωi +
m
j=1
α+
j H+


j
k=1
φi+k−1

 +
m
j=1
α−
j H−

−
j
k=1
φi−k

 . (4)
Note that the terms are omitted from (3) and (4) if i + j or i − j goes beyond 1, 2, . . . , n + 1. This form will allow
us to prove the existence of stable solutions to (4) via a simple constructive method. Our strategy will be to first
consider a piecewise linear model for the functions H±. In this case, the existence of solutions is reduced to finding
a solution to a linear matrix equation. We then smoothly move from the piecewise linear version of the functions
H± to the desired version by using the implicit function theorem.
Crucial to our continuation of argument are certain hypotheses on the functions H±(φ). We define two functions
f and g as f (φ) + g(φ) = H+(φ) and f (φ) − g(φ) = H−(−φ). We assume the following hypotheses on f and
g in a sufficiently large interval around φ = 0:
(H1) g (φ) > |f (φ)| for φ ∈ J.
(H2) There exists a unique solution φL (respectively φR) to f (φ) = g(φ) (respectively f (φ) = −g(φ)) for some
φ ∈ J.
Note that if H+(φ) = H−(φ) = H(φ), i.e., the coupling is isotropic, then g(φ) is just the odd part of the function
H(φ) and f (φ) is the even part. This set of conditions is exactly the same as in [7] and is a subset of those in [6]. In
addition, φL = φR should be imposed. It can be shown that φR < 0 < φL when f (0) > |g(0)| and φL < 0 < φR
when f (0) < −|g(0)|. We can restate these hypotheses in terms of the functions H±:
(H1 ) H± (φ) > 0 for φ ∈ J.
(H2 ) There exists a unique solution φL (respectively φR) to H−(−φ) = 0 (respectively H+(φ) = 0) for some
φ ∈ J.
Hypothesis (H1 ) is analogous to the hypothesis made in [4]. The second hypothesis is required in order to get
some bounds on the behavior of the ends of the chain.
The numbers, φL, φR and the hypotheses on the interaction functions can be understood intuitively by looking at
the case of just a pair of mutually coupled oscillators. Consider a pair of coupled oscillators:
θ1 = ω + H+
(θ2 − θ1), θ2 = ω + H−
(θ1 − θ2).
The phase difference between them, φ = θ2 − θ1 satisfies
φ = H−
(−φ) − H+
(φ) = −2g(φ).
Thus, phase locked solutions are just roots of g(φ) = 0. If the coupling is only forward, i.e., H+ ≡ 0 then the phase
locked solution is φ = φL. Furthermore, it is a stable phase locked solution since we have assumed that φL ∈ J
and that g (φ) > 0 in the interval J. Thus, φL is the unique stable phase locked solution for a forwardly coupled
pair of oscillators. Similarly, φR is the the unique phase locked solution for a pair of backwardly coupled (H− ≡ 0)
oscillators. For H+ and H− nonzero, the unique phase locked solution is between φL and φR. It is stable since both
H± > 0 in an interval containing φL, φR.
A simple example is H± = α±H, H(φ) = A cos φ + B sin φ where B > 0, A = 0, α± > 0. Furthermore, A
should not be too large in magnitude.
We now introduce equations for the local phase differences. If we let φi = θi+1−θi, βi = ωi+1−ωi, i = 1, . . . , n,
then (3) leads to

φi = βi +
m
j=1
α+
j [f + g]


j
k=1
φi+k

 +
m
j=1
α−
j [f − g]


j
k=1
φi−k+1


−
m
j=1
α+
j [f + g]


j
k=1
φi+k−1

 −
m
j=1
α−
j [f − g]


j
k=1
φi−k

 . (5)
Again the terms out of index range will be ignored. Through most part of this paper, we study the case of βi ≡ 0
(which means that all the oscillators have the same frequency). Then (5) can be rewritten as
φ i =
m
j=1
α+
j [f + g]


j
k=1
φi+k

 +
m
j=1
α−
j [f − g]


j
k=1
φi−k+1


−
m
j=1
α+
j [f + g]


j
k=1
φi+k−1

 −
m
j=1
α−
j [f − g]


j
k=1
φi−k

 . (6)
For phase locked solutions, we have φi = 0 so that
m
j=1
α+
j [f + g]


j
k=1
φi+k−1

 +
m
j=1
α−
j [g − f ]


j
k=1
φi−k+1


=
m
j=1
α+
j [f + g]


j
k=1
φi+k

 +
m
j=1
α−
j [g − f ]


j
k=1
φi−k

 , (7)
where i = 1, . . . , n. Note that the terms containing φi are placed on the left-hand side and the terms without φi are
put on the right-hand side. This arrangement simplifies the analysis below.
In Section 2, H± are chosen to be piecewise linear functions. The reason for this is that we can explicitly find
solutions with these simple functions. Then a “bridge” can be built from the simple to the general case based on the
information collected from the simple case.
Section 3 provides a way to construct the “bridge”. That is, we set up a homotopy path starting with the solution
which we obtain in Section 2. Under very general assumptions, this homotopy path will lead to the solution of (7).
The solution is a unique asymptotically stable solution of (6) for a wide range of functions.
Numerical experiments are shown in Section 4. They confirm the results obtained from Section 3.
2. Piecewise linear coupling functions
We consider piecewise linear systems in this section in order to collect the information we need. Two piecewise
linear 2π-periodic coupling functions are constructed as H±(φ) = H±
E (φ) + H±
O (φ) with H±
E and H±
O (as even
parts and odd parts of H±, respectively) are defined as
H±
E (φ) ≡ b±
, H±
O (φ) =



φ, 0 ≤ φ ≤ c,
c(π − φ)
π − c
, c < φ ≤ π,
−H±
O (−φ), −π ≤ φ < 0,
where −π < −c < min(b−, −b+) < 0 < max(b−, −b+) < c < π.

Then if we choose c such that b± ∈ J = (−c, c) , the hypotheses (H1) and (H2) hold. We can also deduce that
φL = b− and φR = −b+.
Note that if |φi| ≤ c/m for the solution of the Eqs. (7), we have f (φ) = 1
2 (b+ +b−) and g(φ) = φ + 1
2 (b+ −b−)
in Eqs. (7). Then (7) yields
m
j=1
α+
j

b+
+
j
k=1
φi+k−1

 +
m
j=1
α−
j


j
k=1
φi−k+1 − b−


=
m
j=1
α+
j

b+
+
j
k=1
φi+k

 +
m
j=1
α−
j


j
k=1
φi−k − b−

 , (8)
where i = 1, . . . , n and the out-of-range terms are ignored as before.
More speciﬁcally, (8) can be reduced to


m
j=1
α+
j +
i
j=1
α−
j

 φi = α−
i φ0 +
i−1
j=1
(α−
j − α−
i )φi−j +
m
j=1
α+
j φi+j (9)
for 1 ≤ i ≤ m,


m
j=1
α+
j +
m
j=1
α−
j

 φi =
m
j=1
α−
j φi−j +
m
j=1
α+
j φi+j (10)
for m + 1 ≤ i ≤ n − m,


n+1−i
j=1
α+
j +
m
j=1
α−
j

 φi =
m
j=1
α−
j φi−j +
n−i
j=1
(α+
j − α+
n+1−i)φi+j + α+
n+1−iφn+1 (11)
for n − m + 1 ≤ i ≤ n, where φ0 = b− and φn+1 = −b+, i.e., φ0 = φL and φn+1 = φR.
From this, (9)–(11) can be written as a matrix equation
B = S, (12)
where
S = [α−
1 φ0, . . . , α−
mφ0, 0, . . . , 0, α+
mφn+1, . . . , α+
1 φn+1]T
,
= (φ1, . . . , φn)T and B = D − L − U. Here D is a diagonal matrix and L (respectively, U ) is lower tri-
angular (respectively, upper triangular ) with zero entries on the diagonal. D, L and U are matrices with nonnegative
entries.
Lemma 2.1. Assume that min(φL, φR) < 0 < max(φL, φR), then Eq. (12) has a unique solution ¯ . ¯ satisﬁes
min(φL, φR) < ¯φi < max(φL, φR), i = 1, . . . , n.
Proof. Without loss of generality, we only consider the case when φR < 0 < φL , i.e., φn+1 < 0 < φ0. To show
(12) has a unique solution, we only need to verify that B is nonsingular. By the special form of Eqs. (9)–(11), we
have bii ≥ j=i |bij |, i = 1, . . . , n and there is at least one “>”.

Also it is quite clear that B is irreducible. Thus B is irreducibly diagonally dominant. Any irreducibly diagonally
dominant matrix is nonsingular (see [12]). Hence (12) has a unique solution ¯ .
In order to show φR < ¯φi < φL, i = 1, . . . , n, we need to construct an iterative process. That is
(0)
= (0, . . . , 0)T
, (l+1)
= D−1
S + D−1
(L + U) (l)
, (13)
where l = 0, 1, . . . .
Let A = D−1(L + U) and Q = D−1S, then (l+1) = A (l) + Q. Thus,
l = Al
0 +
l−1
k=0
Ak
Q. (14)
It can be shown [12] that the spectral radius, ρ(A), is less than 1. Thus, the sums in (14) converge and the iteration
(13) converges. That is (l) → ¯ = (I − A)−1Q as l → ∞.
We claim that for i = 1, . . . , n, we have
φn+1 < φ
(l)
i < φ0. (15)
By referring to (9)–(11), the iteration (13) can be written as


m
j=1
α+
j +
i
j=1
α−
j

 φ
(l+1)
i = α−
i φ0 +
i−1
j=1
(α−
j − α−
i )φ
(l)
i−j +
m
j=1
α+
j φ
(l)
i+j ,
for 1 ≤ i ≤ m,


m
j=1
α+
j +
m
j=1
α−
j

 φ
(l+1)
i =
m
j=1
α−
j φ
(l)
i−j +
m
j=1
α+
j φ
(l)
i+j ,
for m + 1 ≤ i ≤ n − m,


n+1−i
j=1
α+
j +
m
j=1
α−
j

 φ
(l+1)
i =
m
j=1
α−
j φ
(l)
i−j +
n−i
j=1
(α+
j − α+
n+1−i)φ
(l)
i+j + α+
n+1−iφn+1,
for n − m + 1 ≤ i ≤ n.
We prove (15) by induction on l. For l = 0, (15) holds. Suppose (15) holds for l, then
φ
(l+1)
i <
α−
i φ0 + i−1
j=1(α−
j − α−
i )φ0 + m
j=1α+
j φ0
m
j=1α+
j + i
j=1α−
j
< φ0,
where 1 ≤ i ≤ m. Similarly, we can get φ
(l+1)
i < φ0 for m + 1 ≤ j ≤ n. So φ
(l+1)
i < φ0, i = 1, . . . , n. By
similar arguments, we have φn+1 < φ
(l+1)
i , i = 1, . . . , n. Hence (15) holds for any l ∈ N . Then we must have
φn+1 ≤ ¯φi ≤ φ0, since (l) → ¯ as l → ∞.
We know that (φ0, . . . , φ0) is not the solution, so there is at least an index i0 such that ¯φi0 < φ0. Then by (9)–(11)
and φn+1 ≤ ¯φi ≤ φ0, we can get ¯φi < φ0 for all i. Similarly, we have φn+1 < ¯φi for all i. Hence φn+1 < ¯φi < φ0,
i.e., φR < ¯φi < φL for i = 1, . . . , n.
Theorem 2.1. Assume that
−
c
m
≤ min(φL, φR) < 0 < max(φL, φR) ≤
c
m
(16)

for the piecewise linear functions f and g. Then the system (6) has an asymptotically stable equilibrium
¯ = ( ¯φ1, . . . , ¯φn) such that min(φL, φR) < ¯φi < max(φL, φR).
Proof. The existence and boundedness of ¯ have been proven. The linearized system of (6) around ¯ is = B .
It was shown in Lemma 2.1 that B is nonsingular so that B has no zero eigenvalue. For each i, bii ≥ j=i|bij |.
If we apply the Gerschgorin disk theorem, all the eigenvalues of B stay in Re(z) < 0, i.e., all the eigenvalues have
negative real parts such that ¯ is asymptotically stable.
The condition (16) in Theorem 2.1 will be violated for large m. We would like to modify it since most of ¯φi are not
necessarily close to φL or φR (only those which are near the two ends might be close to φL and φR). The key point
that guarantees that we can stably continue the solution is that the phase differences between any two oscillators
that are connected should lie in a region such that H± is increasing (i.e., within the interval (−c, c)). For then, we
can apply the results in [4]. The theorem gives sufﬁcient conditions which guarantee all these phase differences
lie in the interval (−c, c) but they are rather stringent. Thus, we can more directly give conditions looking at the
total phase lag between any two connected oscillators. Note that since φi = θi+1 − θi, the total phase lag between
oscillators i and i + l is just the sum of the local phase differences. Hence we have the following theorem.
Theorem 2.2. Assume that the solution ¯ in Lemma 2.1 satisﬁes the following conditions:
−c ≤
l
j=0
¯φi+j ≤ c, l = 0, 1, . . . , m − 1, (17)
for i = 1, . . . , n (note that if i + j is out of range of {1, . . . , n}, the term ¯φi+j is ignored in the sum), then ¯ is an
asymptotically stable equilibrium of (6). Also min(φL, φR) < ¯φi < max(φL, φR).
Remarks.
1. As noted above, the sums in (17) are nothing more than the total phase lags θi − θi±l so that this condition
is an assertion that the maximal phase lag between any pair of oscillators that are coupled lies in the interval
J = (−c, c).
2. From (9)–(11), each ¯φi seems to be the average of its 2m neighbors in some sense. For m + 1 ≤ i ≤ n − m,
i.e. in the middle of the chain, the average is the weighted average. But on the two ends, the averages have some
portions lost (or gained). This is the boundary effect and the reason why there exists nonzero values of φi in the
chain.
3. General coupling functions
In this section, we assume that H± satisfy (H1) and (H2). In addition, we assume that either φR < 0 < φL or
φL < 0 < φR.
Let b− = φL and b+ = −φR. We choose c ∈ (0, π) such that J ⊂ [−c, c]. Then the piecewise linear functions
in Section 2 can be constructed. We denote them as H+
0 , H−
0 , f0 and g0, respectively.
Withthesepreliminaries,wecanconstructtwohomotopycouplingfunctionsH±
λ (φ)asH±
λ (φ) = (1−λ)H±
0 (φ)+
λH±(φ), 0 ≤ λ ≤ 1. Then H±
λ (φ) = H±
0 when λ = 0 and H±
λ (φ) = H±(φ) when λ = 1. Accordingly, we have
the corresponding fλ and gλ. They are fλ(φ) = (1 − λ)f0(φ) + λf (φ) and gλ(φ) = (1 − λ)g0(φ) + λg(φ). As we
can see, the corresponding two numbers are φL(λ) and φR(λ). Luckily, we have φL(λ) = φL and φR(λ) = φR for
0 ≤ λ ≤ 1.

For the newly constructed coupling functions H±
λ , we have new versions of (3), (6) and (7), respectively, i.e.,
θi = ω +
m
j=1
α+
j H+
λ (θi+j − θi) +
m
j=1
α−
j H−
λ (θi−j − θi), (18)
φ i =
m
j=1
α+
j [fλ + gλ]


j
k=1
φi+k

 +
m
j=1
α−
j [fλ − gλ]


j
k=1
φi−k+1


−
m
j=1
α+
j [fλ + gλ]


j
k=1
φi+k−1

 −
m
j=1
α−
j [fλ − gλ]


j
k=1
φi−k

 , (19)
m
j=1
α+
j [fλ + gλ]


j
k=1
φi+k−1

 +
m
j=1
α−
j [gλ − fλ]


j
k=1
φi−k+1


=
m
j=1
α+
j [fλ + gλ]


j
k=1
φi+k

 +
m
j=1
α−
j [gλ − fλ]


j
k=1
φi−k

 . (20)
We ﬁrst prove a useful lemma.
Lemma 3.1. In Eq.(19), if all the sums of φi in the form of
j
k=1 are in J, then the Jacobian matrix of the right-
hand side has only eigenvalues with negative real parts.
The proof is to apply Lemma 3.1 and Lemma 3.2 in [4] to the system (18).
Theorem 3.1. Assume min(φL, φR) < 0 < max(φL, φR). If mφL and mφR ∈ J, then the system (6) has
asymptotically stable equilibrium ¯ = ( ¯φ1, . . . , ¯φn) and min(φL, φR) < ¯φi < max(φL, φR), i = 1, . . . , n.
Also ¯ is the unique equilibrium of (6) in the n-dimensional box I × I × · · · × I where the interval I =
[min(φL, φR), max(φL, φR)].
Proof. Without loss of generality, we assume φR < 0 < φL. For convenience, we denote the right-hand sides of
(19) and (6) by Fλ( ) and F( ), respectively, where Fλ, F : Rn → Rn. Then Fλ( ) = (1−λ)(B −S)+λF( ).
Hence B and S are as in (12). The idea of the proof is to trace the homotopy path ¯ (λ), where ¯ (λ) is the solution
of Fλ( ) = 0, as λ varies from 0 to 1.
At λ = 0, Fλ( ) = B − S. By Lemma 2.1, Fλ( ) = 0 has a unique solution ¯ (λ) = ¯ (0) such that
φR < ¯φi(0) < φL. Then the eigenvalues of the Jacobian matrix DFλ(0) = B have negative real parts by Lemma
3.1. So DFλ is nonsingular. By the implicit function theorem, there exists λ0 ∈ (0, 1] such that Fλ( ) = 0 has a
solution ¯ (λ) with φR < ¯φi(λ) < φL for each λ ∈ [0, λ0]. And DFλ( ¯ (λ)) has only eigenvalues with negative real
parts by using Lemma 3.1 again.
Starting with λ0, there exists λ1 ∈ (λ0, 1] such that for each λ ∈ (λ0, λ1], Fλ( ) = 0 has a solution ¯ (λ) with
φR < ¯φi(λ) < φL. DFλ( ¯φ(λ)) has only eigenvalues with negative real parts. Keep iterating this process until the
extension cannot be continued. Then we get 0 < λ0 < λ1 < λ2 < · · · . The properties above hold for all λk. Since
{λk} is monotonically increasing and bounded above by 1, there is λ∗ ∈ [0, 1] such that λk → λ∗ as k → ∞.
We claim λ∗ = 1. Suppose λ∗ < 1 by contradiction. Then continuity tells us that Fλ( ) = 0 has a solution
¯ (λ∗) such that φR ≤ ¯φi ≤ φL. Then DFλ( ¯ (λ∗)) has only eigenvalues with negative real parts from Lemma 3.1
once more.

It can be veriﬁed that both (φR, . . . , φR) and (φL, . . . , φL) are not solutions of Fλ∗ ( ) = 0. Otherwise we would
have a contradiction.
We claim φR < ¯φi(λ∗) < φL, i = 1, . . . , n. Suppose that there is i0 ∈ {1, . . . , n} such that either ¯φi0 (λ∗) =
φL > ¯φi0−1(λ∗) or ¯φi0 (λ∗) = φL > ¯φi0+1(λ∗). If m + 1 ≤ i0 ≤ n − m, noting that
j
k=1
¯φi0+k(λ∗
) ≤
j
k=1
¯φi0+k−1(λ∗
),
j
k=1
¯φi0−k(λ∗
) ≤
j
k=1
¯φi0−k+1(λ∗
) forj = 1, . . . , m.
At least one inequality is strict and gλ ± fλ > 0 in J. Then by (20), we have
m
j=1
α+
j [fλ + gλ]


j
k=1
φi0+k−1(λ∗
)

 +
m
j=1
α−
j [gλ − fλ]


j
k=1
φi0−k+1(λ∗
)


=
m
j=1
α+
j [fλ + gλ]


j
k=1
φi0+k(λ∗
)

 +
m
j=1
α−
j [gλ − fλ]


j
k=1
φi0−k(λ∗
)


<
m
j=1
α+
j [fλ + gλ]


j
k=1
φi0+k−1(λ∗
)

 +
m
j=1
α−
j [gλ − fλ]


j
k=1
φi0−k+1(λ∗
)

 ,
which is a contradiction since the ﬁrst and third lines are the same.
If i0 = 1, then
m
j=1
α+
j [fλ + gλ]


j
k=1
¯φk(λ∗
)

 + α−
1 [gλ − fλ]( ¯φ1(λ∗
))
=
m
j=1
α+
j [fλ + gλ]


j
k=1
¯φk+1(λ∗
)

 <
m
j=1
α+
j [fλ + gλ]


j
k=1
¯φk(λ∗
)


such that [gλ − fλ]( ¯φ1(λ∗)) < 0. Then gλ(φL) < fλ(φL) since ¯φ1(λ∗) = φL. This is a contradiction.
If 2 ≤ i0 ≤ m, then
m
j=1
α+
j [fλ + gλ]


j
k=1
¯φi0+k−1(λ∗
)

 +
i0
j=1
α−
j [gλ − fλ]


j
k=1
¯φi0−k+1(λ∗
)


=
m
j=1
α+
j [fλ + gλ]


j
k=1
¯φi0+k(λ∗
)

 +
i0−1
j=1
α−
j [gλ − fλ]


j
k=1
¯φi0−k(λ∗
)

 ,
such that

m
j=1
α+
j [fλ + gλ]


j
k=1
¯φi0+k−1(λ∗
)

 +
i0
j=1
α−
j [gλ − fλ]


j
k=1
¯φi0−k+1(λ∗
)


−
i0−1
j=1
α−
i0
[gλ − fλ]


j
k=1
¯φi0−k(λ∗
)


=
m
j=1
α+
j [fλ + gλ]


j
k=1
¯φi0+k(λ∗
)

 +
i0−1
j=1
(α−
j − α−
i0
)[gλ − fλ]


j
k=1
¯φi0−k(λ∗
)


<
m
j=1
α+
j [fλ + gλ]


j
k=1
¯φi0+k−1(λ∗
)

 +
i0−1
j=1
(α−
j − α−
i0
)[gλ − fλ]


j
k=1
¯φi0−k+1(λ∗
)

 .
Then
α−
i0
[gλ − fλ]


i0
k+1
¯φi0−k+1(λ∗
)

 −
i0−1
j=1
α−
i0
[gλ − fλ]


j
k=1
¯φi0−k(λ∗
)


≤ −α−
i0
i0−1
j=1
[gλ − fλ]


j
k=1
¯φi0−k+1(λ∗
)

 ,
i.e.,
i0
j=1
[gλ − fλ]


j
k=1
¯φi0−k+1(λ∗
)

 ≤
i0−1
j=1
[gλ − fλ]


j
k=1
¯φi0−k(λ∗
)

 ,
i.e.,
[gλ − fλ]( ¯φi0 (λ∗
)) +
i0
j=2
[gλ − fλ]


j
k=1
¯φi0−k+1(λ∗
)

 ≤
i0−1
j=1
[gλ − fλ]


j
k=1
¯φi0−k(λ∗
)

 ,
i.e.
[gλ − fλ]( ¯φi0 (λ∗
)) +
i0−1
j=1
[gλ − fλ]

 ¯φi0 (λ∗
) +
j
k=1
¯φi0−k(λ∗
)

 ≤
i0−1
j=1
[gλ − fλ]


j
k=1
¯φi0−k(λ∗
)

 . (21)
Since ¯φi0 (λ∗) = φL > 0, then [gλ − fλ]( ¯φi0 (λ∗)) = 0 and [gλ − fλ]( ¯φi0 (λ∗) +
j
k=1
¯φi0−k(λ∗)) >
[gλ − fλ]
j
k=1
¯φi0−k(λ∗) , j = 1, . . . , i0 − 1. So (21) is not possible. We get a contradiction.
Similarly, if m + 1 ≤ i0 ≤ n−m, we get a contradiction. Therefore, there is no i0 such that ¯φi0 (λ∗) = φL. This
leads to ¯φi(λ∗) < φL, i = 1, . . . , n.
By similar arguments, ¯φi(λ∗) > φR, i = 1, . . . , n. So φR < ¯φi(λ∗) < φL, i = 1, . . . , n. Thus we can
extend λ beyond λ∗. This is a contradiction. So λ∗ = 1. And ¯ (1) is the solution to F( ) = 0, which satisﬁes
φR < φi(1) < φL.
Nowweneedtoproveuniqueness.Suppose ¯ and ˆ aretwosolutionstoF( ) = 0inIn = [φR, φL]n.Thenbythe
mean value theorem, 0 = F( ¯ )−F( ˆ ) =
1
0 DF( ˆ +τ( ¯ − ˆ ))( ¯ − ˆ ) dτ = [
1
0 DF( ˆ +τ( ¯ − ˆ )) dτ]( ¯ − ˆ ).

But
1
0 DF( ˆ +τ( ¯ − ˆ )) dτ has only eigenvalues with negative real parts . It is nonsingular implying that ¯ − ˆ = 0,
i.e., ¯ = ˆ .
Remarks.
1. For nearest neighbor coupling (i.e., m = 1), we proved in [14] the existence of an asymptotically stable
equilibrium. We did not prove the uniqueness there. The uniqueness is automatically obtained from Theorem
3.1.
2. We note that along the homotopy path, we can prove
min
1≤j≤n
( ¯φj (λ)) < ¯φi(λ) < max
1≤j≤n
( ¯φj (λ)), (22)
where i = m + 1, . . . , n − m. Furthermore numerical results show that min1≤j≤n( ¯φj (λ)) only occurs at i = 1
or n. So does max1≤j≤n( ¯φj (λ)).
The condition
(C) mφL ∈ J and mφR ∈ J
could be weakened for the same reason as in Section 2. Thus we introduce the following condition set:
(C1) φR + l
j=1
¯φi+j (λ) ∈ J, l = 1, . . . , min(m − 1, n − i) for i = 1, . . . , n.
(C2) φL + l
j=1
¯φi+j (λ) ∈ J, l = 1, . . . , min(m − 1, n − i) for i = 1, . . . , n.
(C3) l-1
j=0
¯φi+j (λ) + φR ∈ J, l = 0, . . . , min(m − 1, n − i) for i = 1, . . . , n.
(C4) l-1
j=0
¯φi+j (λ) + φL ∈ J, l = 0, . . . , min(m − 1, n − i) for i = 1, . . . , n.
As we see from the proof of Theorem 3.1, if the condition (C) holds, we have min(φR, φL) < ¯φi(λ) <
max(φR, φL) for each λ ∈ [0, 1]. Hence the condition set (C1)–(C4) is satisfied for each λ ∈ [0, 1] along the
homotopy path. We thus have the extension of Theorem 3.1.
Theorem 3.2. Assume min(φR, φL) < 0 < max(φR, φL). If the solution in Lemma 2.1, which is ¯ (λ) at λ = 0,
satisfies (C1)–(C4), then there is a maximal λ∗ ∈ (0, 1] such that for each λ ∈ [0, λ∗) the solution ¯ (λ) satisfies
(C1)–(C4) and min(φR, φL) < ¯φi(λ) < max(φR, φL). If λ∗ = 1, then ¯ (1) is an asymptotically stable equilibrium
of (6). It is unique in the region G = { | l
j=0φi+j ∈ J, l = 0, 1, . . . , min(m−1, n−i) for each i = 1, . . . , n} ⊂
Rn. Furthermore, G is a convex set.
Remarks.
1. The proof of Theorem 3.2 is to mimic each step in Theorem 3.1. The conditions (C1)–(C4) guarantee that all the
summation terms of ¯φi(λ) in (20) stay in J such that gλ ± fλ > 0 is insured.
2. For λ = λ∗, ¯ (λ∗) might not satisfy (C1)–(C4). But by continuity, if we substitute J by ¯J (i.e., the closed interval
of the open interval J) in (C1)–(C4), ¯ (λ∗) satisfies the modified (C1)–(C4) and min(φR, φL) < ¯φi(λ∗) <
max(φR, φL). So all the summation terms of ¯φi(λ∗) stay in J. Then the asymptotic stability of ¯ (λ∗) is also
obtained by Lemma 3.1. Also DF( ¯ (λ∗)) has only eigenvalues with negative real parts. Thus if λ∗ < 1, λ still
could be extended to an open neighbor (λ∗, λ∗ + δ) such that ¯ (λ) is an asymptotically stable equilibrium of
(19) for each λ ∈ (λ∗, λ∗ + δ). This is done by applying the implicit function theorem.
3. Theorem 3.2 provides us a way to verify (even though the conditions are only sufficient ones) whether there is an
asymptotically stable equilibrium of (6) in the convex domain G. This can be done using a numerical approach.
We can partition the interval [0, 1] into L subintervals 0 = λ0 < λ1 < · · · < λL = 1 such that λl = lh where
h = 1/L. It can be shown that if λ∗ = 1 and h is small, then ¯ (λl) will be in the asymptotically convergent
range of system (19) for λ = λl+1. Then we could take ¯ (λl) as an initial vector to solve the IVP (19). In such a
way, we can get ¯ (λl+1). If ¯ (λl) does not satisfy (C1)–(C4) somewhere, we stop. Otherwise we continue until

λl = 1. One important thing is how to get ¯ (0). This is done by the iteration (13) which is convergent as we saw
in Section 2.
All the results we have obtained are for the identical oscillators, i.e., ωi ≡ ω (i.e., βi ≡ 0). If ωi are sufficiently
close to each other, i.e., βi is close to zero for each i = 1, . . . , n, we can apply the implicit function theorem to get
asymptotically stable equilibria for the system (5).
Theorem 3.3. If the conditions in Theorem 3.2 hold, λ∗ = 1 and βi is sufficiently close to zero, then the system (5)
has an asymptotically stable equilibrium.
Remark. Theorem 3.3 is obtained from perturbing βi from zero. It is reasonable to assume that the coupling
strength between two oscillators far away is sufficiently small. Let m = m1 + m2. We assume α±
j are very small
for m1 + 1 ≤ j ≤ m1 + m2. Then if the system (5) has an asymptotically stable equilibrium for the case when
α±
j = 0, j ≥ m1 +1, then (5) with sufficiently small α±
j (j ≥ m1 +1) still has an asymptotically stable equilibrium.
We will see this in our numerical results.
4. Numerical results
In this section, we choose H±(φ) = H(φ) = 0.5 cos φ + sin φ. Then f (φ) = 0.5 cos φ and g(φ) = sin φ.
Hence φL, φR and the interval J can be determined. And φL = arctan(0.5), φR = − arctan(0.5) and J =
(− arctan 2, arctan 2).
Fig. 1. n + 1 = 40, βi = 0, the coupling strength sets are Cm and m = 2, 4, 8 and 16, respectively.

We choose two sets of coupling strengths for our numerical experiments. They are Cm = {α±
j = 1/j,
j = 1, . . . , m} and Em = {α±
j = exp(−j + 1), j = 1, . . . , m} where m = 1, 2, . . . .
For both Cm and Em, we always have α±
1 = 1 > α±
2 > · · · > α±
m > 0. α±
j are very small for α±
j ∈ Em if j is
large, e.g. j ≥ 4.
Fig. 2. n + 1 = 100, βi = 0, the coupling strength sets are (a) Cm and (b) Em with m = 2, 4, 8 and 16.

Remarks.
1. arctan(0.5) ≈ 0.464 and arctan 2 ≈ 1.107. Then if m ≥ 3, the condition (C) in Section 3 is violated. But as we
see in the numerical results, (C1)–(C4) are still fine.
2. If we choose H = a cos φ + sin φ, then the smaller |a| is, the larger m we can get to satisfy (C). For example,
let a = 0.1, then m can be as large as m = 14.
Fig. 1 shows the numerical results for the cases when n + 1 = 40, βi = 0 and Cm is the coupling strength
set. Here we take m = 2, 4, 8 and 16 respectively. We plot ¯φi versus i/(n + 1). As we can see, ¯φi lie in the
interval (φR, φL) and (C1)–(C4) (λ = 1) hold. These guarantee asymptotic stability. The figure shows that the
inequalities (22) hold for i = m + 1, . . . , n − m and ¯φ1 = maxi=1,... ,n ¯φi and ¯φn = mini=1,... ,n ¯φi. Also we
find that ¯φi is monotonically decreasing when i = m + 1, . . . , n − m. But on the two boundaries, i.e., i ≤ m
or i ≥ n − m + 1, the monotonicity can be destroyed. This observation matches our comment that at the ends
of the chain, the nonlinear averages have some portions lost (or gained ). So ¯φi can fall below the nonlinear
averages at the left end (except at i = 1) and ¯φi could jump above the nonlinear averages at the right end in this
example.
In Fig. 2(a), n+1 = 100, βi = 0 and Cm is the coupling strength set for m = 2, 4, 8, 16. In Fig. 2(b) n+1 = 100,
βi = 0 and Em is the coupling strength set for m = 2, 4, 8, 16.
We can see from Figs. 1 and 2(a) and (b) that coupling with more oscillators will reduce the phase lags ¯φi. That
is the observation in [10] and it was explained in the case of piecewise linear coupling functions in Section 2.
In Fig. 2(b), the conditions (C1)–(C4) for m = 8 and m = 16 at λ = 1 are violated as we can see. But since α±
j
is small for j ≥ 5, the comments in the end of Section 3 tell us that we still expect the existence of a stable solution.
This is confirmed by the numerical results in Fig. 2(b).
Fig. 3. n + 1 = 100, βi = 0, the coupling strength sets are Cm and Em with m = 4.

We now show how different coupling strengths will affect the phase lags. This is done by comparing the results
of Em and Cm. Fig. 3 shows this for m = 4. It shows that strong coupling will reduce the phase lags ¯φi. Note that
exp(−j + 1) < 1/j for j = 2, . . . , m. This means the Cm type coupling is “stronger” than the Em type.
In the following numerical experiments, we consider the non-isotropic cases when α+
j ∈ Cm and α−
j ∈ Em, i.e.,
α+
j = 1/j and α−
j = exp(−j + 1), j = 1, . . . , m.
Fig. 4. α+
j = 1/j and α−
j = exp(−j + 1), n + 1 = 100, (a) βi = 0, and (b) βi = −0.005, and m = 4, 8.

In Fig. 4(a) n + 1 = 100 and βi = 0 with m = 4 and 8. It conﬁrms the results of Theorem 3.2. In Fig. 4(b)
n+1 = 100 and ωi = 0.5(n+1−i)/(n+1)+ω (i.e., βi = −0.005) where ω can be any positive constant. m = 4
and 8 are chosen.
In Fig. 5 n + 1 = 100 and ωi = ω + δi where δi are randomly chosen from the interval (0, 0.5). Thus βi are
chosen from (−0.5, 0.5) randomly.
Fig. 5. α+
j = 1/j, α−
j = exp(−j + 1) and ωi = ω + δi, where δi are randomly chosen from the interval (0, 0.5), (a) n + 1 = 100 and m = 4,
(b) n + 1 = 100 and m = 8.

Figs. 4(b), 5(a) and (b) verify the results of Theorem 3.3. In the case of Fig. 4(b), there is a frequency gradient.
This gradient is small so that ωi stay close to each other. In the cases of Fig. 5(a) and (b), ωi are chosen randomly
and close to each other. Once more we mention that coupling with more oscillators will reduce the phase lags
[10].
Fig. 6. (a) H±(φ) = 0.8 cos φ + sin φ, coupling is α±
j = 1/(2m + 1), n + 1 = 40 and m = 1, 2, 4, 5. For m < 5 there is stable phase
locking. However, note the nonmonotonicity near the boundaries. For m > 4 oscillators at the edges “break away” and phase locking is lost.
(b) Bifurcation diagram showing the range of existence of phase locking for H±(φ) = a cos φ + sin φ as a function of the parameter a. Right
endpoint is the maximum value of a for stable phase locking.

In the previous numerical simulations, it was shown that one of the effects of multiple coupling is to reduce the
phase lags between successive oscillators. Thus, it would seem that increasing the coupling length encourages tighter
phase locking. However, if the interval J becomes too short relative to the roots φL, φR, then it may be possible to
achieve phase locking with short range coupling, but lose it with longer range coupling. Fig. 6 illustrates this. We
ﬁrst choose, H±(φ) = 0.8 cos φ + sin φ. The interval J is now (−0.896, 0.896) and φL = 0.674, φR = −0.674.
For m = 1 this is still in the range for which we expect phase locking with monotonically varying phase shifts, φi,
However, if m gets larger, there is no guarantee that there will be a locked solutions. In the ﬁgure, m = 1, 2, 4 all lead
to phase locked solutions. Note the phase differences away from the edges are compressed, but the behavior near
the edges oscillates. When m = 5, the oscillators at the ends “drift” away; they are no longer able to phase lock with
the interior oscillators. However, as m continues to increase, stable locking can occur again. This is shown Fig. 6(b).
Each curve represents the total phase lag, θ10 −θ1 as a function of the parameter a where H±(φ) = a cos φ +sin φ.
Note that for m = 1, the existence of phase locking extends to a = 1, while for m = 4, 8 it is considerably shortened.
However, for m = 16 the range is again quite large. If m = 40 then coupling is “all-to-all” and synchrony is stable for
any value of a thus we expect that the range of phase locking should be a nonmonotonic function of the connectivity,
m. What is somewhat surprising is that the “worst” case for locking occurs at about m = 10 or connectivity over a
quarter of the chain. The investigation of these phenomena remains an open problem.
References
[1] A.H. Cohen, G.B. Ermentrout, T. Kiemel, N. Kopell, K.A. Sigvardt, T.L. Williams, Modeling of intersegmental coordination in the lamprey
central pattern generator for motion, Trends Neurosci. 15 (1992) 434–438.
[2] J.D. Crawford, K.T.R. Davies, Synchronization of globally coupled phase oscillators: singularities and scaling for general couplings, Physica
D 125 (1999) 1–46.
[3] G.B. Ermentrout, Synchronization in a pool of mutually coupled oscillators with random frequencies, J. Math. Biol. 22 (1985) 1–9.
[4] G.B. Ermentrout, Stable periodic solutions to discrete and continuum arrays of weakly coupled nonlinear oscillators, SIAM J. Appl. Math.
52 (1992) 1665–1687.
[5] G.B. Ermentrout, J. Flores, A. Gelperin, Minimal model of oscillations and waves in the Limax olfactory lobe with tests of the model’s
predictive power, J. Neurophysiol. 79 (1998) 2677–2689.
[6] G.B. Ermentrout, N. Kopell, Symmetry and phase locking in chains of weakly coupled oscillators, Comm. Pure Appl. Math. 49 (1986)
623–660.
[7] G.B. Ermentrout, N. Kopell, Phase transitions and other phenomena in chains of coupled oscillators, SIAM J. Appl. Math. 50 (1990)
1014–1052.
[8] A. Gelperin, J. Flores, B. Ermentrout, Coupled oscillator network for olfactory processing in Limax: responses to local perturbations with
nitric oxide, Soc. Neurosci. Abst. 25 (1999) 125.
[9] D. Kleinfeld, K.R. Delaney, M. Fee, J.A. Flores, D.W. Tank, A. Gelperin, Dynamics of propagating waves in the olfactory network of a
terrestrial mollusc: an electrical and optical study, J. Neurophysiol. 72 (1994) 1402–1419.
[10] N. Kopell, W. Zhang, G.B. Ermentrout, Multiple coupling in chains of oscillators, SIAM J. Math. Anal. 21 (1990) 935–953.
[11] W.L. Miller, K.A. Sigvardt, Extent and role of multisegmental coupling in the Lamprey spinal locomotor pattern generator, J. Neurophysiol.
83 (2000) 465–476.
[12] J.M. Ortega, W.C. Rheinboldt, Iterative Solution of Nonlinear Equations in Several Variables, Academic Press, New York, 1970.
[13] J.C. Prectl, L.B. Cohen, B. Pesaren, P.P. Mitra, D. Kleinfeld, Visual stimuli induce waves of electrical activity in turtle cortex, PNAS USA
94 (1997) 7621–7626.
[14] L. Ren, G.B. Ermentrout, Monotonicity of phase locked solutions in chains and arrays of nearest-neighbor coupled oscillators, SIAM J.
Math. Anal. 29 (1998) 208.

Phase locking in chains of multiple-coupled oscillators

More Related Content

What's hot (20)

Similar to Phase locking in chains of multiple-coupled oscillators (20)

More from Liwei Ren任力偉 (20)

Recently uploaded (20)

Phase locking in chains of multiple-coupled oscillators