Variational method for finding periodic orbits in a general flow

Yueheng Lan∗ and Predrag Cvitanović†

Center for Nonlinear Science, School of Physics,
Georgia Institute of Technology, Atlanta 30332-0430, U.S.A
(Dated: October 2, 2018)
A variational principle for determining unstable periodic orbits of flows as well as unstable spatio-
temporally periodic solutions of extended systems is proposed and implemented. An initial loop
approximating a periodic solution is evolved in the space of loops toward a true periodic solution by
a minimization of local errors along the loop. The “Newton descent” partial differential equation
that governs this evolution is an infinitesimal step version of the damped Newton-Raphson iteration.
The feasibility of the method is demonstrated by its application to the Hénon-Heiles system, the
arXiv:nlin/0308008v1 [nlin.CD] 5 Aug 2003

circular restricted three-body problem, and the Kuramoto-Sivashinsky system in a weakly turbulent
regime.

PACS numbers: 95.10.Fh, 02.70.Bf, 47.52.+j, 05.45.+a

Keywords: periodic orbits, variational methods, spatio-temporal chaos, turbulence, cost function minimiza-
tion, Kuramoto-Sivashinsky system, restricted three-body problem, Hénon-Heiles system

I. INTRODUCTION odic orbits is based on devising a new dynamical system

which possesses the desired orbit as an attracting fixed
point with a sizable basin of attraction. Beyond that,
The periodic orbit theory of classical and quantum
there is much freedom in constructing such system.
chaos [1, 2] is one of the major advances in the study
of long-time behavior of chaotic dynamical systems. The For example, the multipoint shooting method elimi-
theory expresses all long time averages over chaotic dy- nates the long-time exponential instability of unstable or-
namics in terms of cycle expansions [3], sums over pe- bits by splitting an orbit into a number of short segments,
riodic orbits (cycles) ordered hierarchically according to each with a controllable expansion rate. The multiple
the orbit length, stability, or action. If the symbolic dy- shooting combined with the Newton-Raphson method is
namics is known, and the flow is hyperbolic, longer cycles an efficient tool for locating periodic orbits of maps [16].
are shadowed by the shorter ones, and cycle expansions A search for periodic orbits of a continuous time flow can
converge exponentially or even super-exponentially with be reduced to a multiple shooting search for periodic or-
the cycle length [4]. bits of a set of maps by constructing a set of phase space
A variety of methods for determining all periodic orbits Poincaré sections such that an orbit leaving one section
up to a given length have been devised and successfully reaches the next one in a qualitatively predictable man-
implemented for low-dimensional systems [5, 6, 7, 8, 9]. ner, without traversing other sections along the way. In
For more complex dynamics, such as turbulent flows [10], turbulent, high-dimensional flows such sequences of sec-
nonlinear waves [11], or quantum fields [12, 13] with high tions are hard to come by. One solution might be a large
(or infinite) dimensional phase spaces and complicated set of Poincaré sections, with the intervening flight seg-
dynamical behavior, many of the existing methods be- ments short and controllable.
come unfeasible in practice. In the most computationally Here we follow a different strategy, and discard
demanding calculation carried out so far, Kawahara and Poincaré sections altogether; we replace maps between
Kida [14] have found two periodic solutions in a 15, 422- spatially fixed Poincaré sections, by maps induced by
dimensional discretization of a turbulent plane Couette discretizing the time evolution into small time steps. For
flow. The topology of high-dimensional flows is hard to sufficiently small time steps such maps are small defor-
visualize, and even with a decent starting guess for the mations of identity. We distribute many points along a
shape and location of a periodic orbit, methods like the smooth loop L, our initial guess of a cycle location and
Newton-Raphson method are likely to fail. In ref. [15] its topological layout. If both the time steps and the loop
we have argued that variational, cost-function minimiza- deformations are taken infinitesimal, a partial differential
tion methods offer a robust alternative. Here we derive, equation governs the “Newton descent”, a fictitious time
implement and discuss in detail one such new variational flow of a trial loop L into a genuine cycle p, with expo-
method for finding periodic orbits in general flows, and nential convergence in the fictitious time variable. We
specifically high-dimensional flows. then use methods developed for solving PDEs to get the
In essence, any numerical algorithm for finding peri- solution. Stated succinctly, the idea of our method is to
make an informed rough guess of what the desired cycle
looks like globally, and then use a variational method to
drive the initial guess towards the exact solution. For ro-
∗ Electronic address: [email protected] bustness, we replace the guess of a single orbit point by a
† Electronic address: [email protected] guess of an entire orbit. For numerical safety we replace
 2

the Newton-Raphson iteration by the “Newton descent”, The orientation of the s-velocity vector tangent to the
a differential flow that minimizes a cost function com- loop L
puted as deviation of the approximate flow from the true
flow along a smooth loop approximation to a cycle. dx̃
ṽ(x̃) =
In sect. II we derive the partial differential equation ds
which governs the evolution of an initial guess loop to-
is intrinsic to the loop, but its magnitude depends on the
ward a cycle and the corresponding cost function. An
(still to be specified) parametrization s of the loop.
extension of the method to Hamiltonian systems and sys-
At each loop point x̃n ∈ L we thus have two vectors,
tems with higher time derivatives is presented in sect. III.
the loop tangent ṽn = ṽ(x̃n ) and the flow velocity vn =
Simplifications due to symmetries and details of our nu-
v(x̃n ). Our goal is to deform L until the directions of ṽn
merical implementation of the method are discussed in
and vn coincide for all n = 1, . . . , N , N → ∞, that is
sect. IV. In sect. V we test the method on the Hénon-
L = p. To match their magnitude, we introduce a local
Heiles system, the restricted three body problem, and
time scaling factor
a weakly turbulent Kuramoto-Sivashinsky system. We
summarize our results and discuss possible improvements
λ(sn ) ≡ ∆tn /∆sn , (4)
of the method in sect. VI.
where ∆sn = sn+1 − sn , n = 1, . . . , N − 1 , ∆sN = 2π −
(sN − s1 ), and likewise for ∆tn . The scaling factor λ(sn )
II. THE NEWTON DESCENT METHOD IN ensures that the loop increment ∆sn is proportional to
LOOP SPACE its counterpart ∆tn + δtn on the cycle when the loop L
is close to the cycle p, with δtn → 0 as L → p.
A. A variational equation for the loop evolution Let x(t) = f t (x) be the state of the system at time
t obtained by integrating (2), and J(x, t) = dx(t)/dx(0)
A periodic orbit is a solution (x, T), x ∈ Rd , T ∈ R of be the corresponding Jacobian matrix obtained by inte-
the periodic orbit condition grating

dJ ∂vi
f T (x) = x , T>0 (1) = AJ , Aij = , with J(x, 0) = 1 . (5)
dt ∂xj
for a given flow or discrete time mapping x 7→ f t (x).
Our goal is to determine periodic orbits of flows defined Since the point xn = x̃n + δx̃n is on the cycle,
by first order ODEs
f ∆tn +δtn (x̃n + δx̃n ) = x̃n+1 + δx̃n+1 . (6)
dx
= v(x) , x ∈ M ⊂ Rd , (x, v) ∈ TM (2) Linearization
dt
in many (even infinitely many) dimensions d. Here M is f δt (x) ≈ x + v(x)δt , f t (x + δx) ≈ x(t) + J(x, t)δx ,
the phase space (or state space) in which evolution takes
place, TM is the tangent bundle [17], and the vector field of (6) about the loop point x̃n and the time interval ∆tn
v(x) is assumed to be smooth (sufficiently differentiable) to the next cycle point leads to the multiple shooting
almost everywhere. Newton-Raphson equation, for any step size ∆tn :
We make our initial guess at the shape and the location
δx̃n+1 − J(x̃n , ∆tn )δx̃n − vn+1 δtn = f ∆tn (x̃n ) − x̃n+1 .
of a cycle p by drawing a loop L, a smooth, differentiable
(7)
closed curve x̃(s) ∈ L ⊂ M, where s is a loop parameter.
Provided that the initial guess is sufficiently good, the
As the loop is periodic, we find it convenient to restrict
Newton-Raphson iteration of (7) generates a sequence of
s to [0, 2π], with the periodic condition x̃(s) = x̃(s + 2π).
loops L with a decreasing cost function [15]
Assume that L is close to the true cycle p, pick N pairs
of nearby points along the loop and along the cycle N
N X ∆tn
F 2 (x̃) ≡ (f (x̃n ) − x̃n+1 )2 , x̃N +1 = x̃1 .
x̃n = x̃(sn ) , 0 ≤ s1 < . . . < sN < 2π , (2π)2 i=1
xn = x(tn ) , 0 ≤ t1 < . . . < tN < T p , (3) (8)
The prefactor N/(2π)2 makes the definition of F 2 consis-
and denote by δx̃n the deviation of a point xn on the tent with (13) in the N → ∞ limit. If the flow is locally
periodic orbit p from the nearby point x̃n , strongly unstable, the neighborhood in which the lin-
earization is valid could be so small that the full Newton
xn = x̃n + δx̃n . step would overshoot, rendering F 2 bigger rather than
smaller. In this case the step-reduced, damped Newton
The deviations δx̃ are assumed small, vanishing as L ap- method is needed. As proved in ref. [18], under conditions
proaches p. satisfied here, F 2 decreases monotonically if appropriate
 3

L( )=p by x̃ = x̃(s, τ ) ∈ L(τ ), where s denotes the position

along the loop, and the fictitious time τ parametrizes
the deformation of the loop, see Fig. 1(a). We refer to
this infinitesimal step version of the damped Newton-
Raphson method as the “Newton descent”.
The important feature of this equation is that a de-
L( τ) creasing cost functional exists. Rewriting (10) as

x ∂
(ṽ − λv) = −(ṽ − λv) , (11)
∂τ
we have

(a) L(0) ṽ − λv = e−τ (ṽ − λv)|τ =0 , (12)

v
v so the fictitious time τ flow decreases the cost functional
L( τ )

1
I
2
F 2 [x̃] = ds (ṽ(x̃) − λv(x̃)) (13)
2π L(τ )

p monotonically as the loop evolves toward the cycle.

At each iteration step the differences of the loop tan-
gent velocities and the dynamical flow velocities are re-
duced by the Newton descent. As τ → ∞, the fictitious
(b) time flow alignes the loop tangent ṽ with the dynami-
cal flow vector ṽ = λv, and the loop x̃(s, τ ) ∈ L(τ ), see
FIG. 1: (a) An annulus L(τ ) swept by the Newton descent Fig. 1(b), converges to a genuine periodic orbit p = L(∞)
flow dx̃/dτ , connecting smoothly the initial loop L(0) to the of the dynamical flow ẋ = v(x). Once the cycle p is
periodic orbit p = L(∞). (b) In general the loop velocity field dt
reached, by (4), λ(s, ∞) = ds (x̃(s, ∞)), and the cycle
ṽ(x̃) does not coincide with λv(x̃); for a periodic orbit p, it
does so at every x ∈ p. period is given by
Z 2π
Tp = λ(x̃(s, ∞))ds .
step size is taken. If infinitesimal steps are taken, de- 0
crease of F 2 is ensured. We parametrize such continuous Of course, as at this stage we have already identified the
deformations of the loop by a fictitious time τ . cycle, we may pick instead an initial point on p and cal-
We fix ∆sn and proceed by δτ each step of the itera- culate the period by a direct integration of the dynamical
tion, that is, multiply the right hand side of (7) by δτ . equations (2).
According to (4), the change of ∆tn with respect to τ is
equal to δtn = ∂λ ∂
∂τ (sn , τ )δτ ∆sn . As δx̃n = ∂τ x̃(sn , τ )δτ ,
dividing both sides of (7) by δτ yields B. Marginal directions and accumulation of loop
points
dx̃n+1 dx̃n ∂λ
− J(x̃n , ∆tn ) − vn+1 (sn , τ )∆sn
dτ dτ ∂τ
Numerically, two perils lurk in a direct implementation
= f ∆tn (x̃n ) − x̃n+1 . (9) of the Newton descent (10).
In the N → ∞ limit, the stepsizes ∆sn , ∆tn = First, when a cycle is reached, it remains a cycle under
O(1/N ) → 0, and we have a cyclic permutation of the representative points, so on
the cycle the operator
vn+1 ≈ vn , x̃n+1 ≈ x̃n + ṽn ∆sn ,
∂
J(x̃n , ∆tn ) ≈ 1 + A(x̃n )∆tn , f ∆tn (x̃n ) ≈ x̃n + vn ∆tn . Ā = − λA
∂s
Substituting into (9) and using the scaling relation (4), has a marginal eigenvector v(x̃(s)) with eigenvalue 0. If λ
we obtain is fixed, as the loop approaches the cycle, (10) approaches
its limit
∂ 2 x̃ ∂ x̃ ∂λ
− λA −v = λv − ṽ. (10) ∂x
∂s∂τ ∂τ ∂τ Ā = 0.
∂τ
This PDE, which describes the evolution of a loop L(τ )
toward a periodic orbit p, is the central result of this Therefore, on the cycle, the operator Ā−1 becomes sin-
paper. The family of loops so generated is parametrized gular and the numerical woes arise.
 4

The second potential peril hides in the freedom of is, the action) of the rubber band is extremal (maximal
choosing the loop (re-)parametrization. Since s is re- or minimal under infinitesimal changes of the boundary
lated to the time t by the yet unspecified factor λ(s, τ ), points). Note that the extremization of action requires
uneven distributions of the sampling points over the loop only D configuration coordinate variations, not the full
L could arise, with the numerical discretization points x̃n 2D-dimensional phase space variations.
clumping densely along some segments of L and leaving Can we exploit this property of the Newtonian me-
big gaps elsewhere, thus degrading the numerical smooth- chanics to reduce the dimenionality of our variational
ness of the loop. calculations? The answer is yes, and easiest to under-
We remedy these difficulties by imposing constraints stand in terms of the Hamilton’s variational principle
on (10). In our calculation for Kuramoto-Sivashinsky which states that classical trajectories are extrema of the
system of sect. V, the first difficulty is dealt with by in- Hamilton’s principal function (or, for fixed energy E, the
troducing one Poincaré section, for example, by fixing action S = R + Et)
one coordinate of one of the sampling points, x̃1 (s1 , τ ) = Z t1
const. This breaks the translational invariance along the R(q1 , t1 ; q0 , t0 ) = dt L(q(t), q̇(t), t) ,
cycle. Other types of constraints might be better suited t0
to a specific problem at hand. For example, we can where L(q, q̇, t) is the Lagrangian. Given a loop L(τ ) we
demand that the average displacement of the sampling
can compute not only the tangent “velocity” vector ṽ,
points along the loop vanishes, thus avoiding a spiraling but also the local loop curvature or “acceleration” vector
descent towards the desired cycle.
We deal with the second potential difficulty by choos- ∂ 2 x̃
ã = ,
ing a particularly simple loop parametrization. So far, ∂s2
the parametrization s is arbitrary and there is much
and indeed, as many s derivatives as needed. Matching
freedom in choosing the best one for our purposes. We
the dynamical acceleration a(x̃) (assumed to be functions
pick s− and τ −independent constant scaling λ(s, τ ) = λ.
of x̃ and v(x̃)) with the loop “acceleration” ã(x̃) results
With uniform grid size ∆sn = ∆s and fixed λ, the loop
in a new cost function and the corresponding PDE (11)
parameter s = t/λ is proportional to time t, and the dis-
for the evolution of the loop
cretization (10) distributes the sampling points along the
loop evenly in time. As the loop approaches a cycle, ∂∂τx̃ ∂
(ã − λ2 a) = −(ã − λ2 a) .
is numerically obtainable from (10), and on the cycle the ∂τ
period is given by Tp = 2πλ.
We use λ2 instead of λ in order to keep the notation
Even though this paper focuses on searches for pe-
consistent with (4), that is t = λ s. Expressed in terms
riodic orbits, the Newton descent is a general method.
of the loop variables x̃(s), the above equation becomes
With appropriate modifications of boundary conditions
and scaling of time, (10) can be adapted to determination ∂ 3 x̃ ∂a ∂ 2 x̃
 
2 ∂a ∂ x̃ ∂a ∂ x̃ ∂λ
of homoclinic or heteroclinic orbits between equilibrium 2
−λ −λ + − 2λa
∂ s∂τ ∂v ∂s∂τ ∂ x̃ ∂τ ∂v ∂s ∂τ
points or periodic orbits of a flow, or more general bound-
ary value problems. Applied to 2-point boundary value = λ2 a − ã , (14)
problems, Newton descent is similar to the quasilineariza- ∂ x̃
where v = λ∂s . Although (14) looks more complicated
tion [19] but has the advantage that the free parameter than (10), in numerical fictitious time integrations, we
λ(s, τ ) is available for adjusting scales in the problem are rewarded by having to keep only half of the phase
and that searches can be restricted to phase space sub- space variables.
manifolds of interest. A simple example of a restriction More generally, if a differential equation has the form:
to a submanifold are searches for cycles of a given en-
ergy, constrained to the H(q, p) = E energy shell in the
phase space of a Hamiltonian system. Furthermore, as x(m) = f (x, x(1) , · · · , x(m−1) ) , (15)
we shall show now, the symplectic structure of Hamil- k

ton’s equations greatly reduces the dimensionality of the where x(k) = ddtkx , k = 1, · · · , m and x ∈ Rd , the same
submanifold that we need to consider. technique can be used to match the highest derivatives
λm x(m) and x̃(m) ,
∂ (m)
(x̃ − λm x(m) ) = −(x̃(m) − λm x(m) ) ,
III. EXTENSIONS OF NEWTON DESCENT ∂τ
m
∂
with x̃(m) = ∂s m x̃(s) calculated directly from x̃(s) on the
In classical mechanics particle trajectories are also so- loop by differentiation. In loop variables x̃(s) we have,
lutions of a variational principle, the Hamilton’s varia-
m
tional principle. For example, one can determine a pe- ∂ m+1 x̃ X ∂f ∂ ∂ k x̃ ∂λ
riodic orbit of a billiard by wrapping around a rubber m
− λm (k)
· − mλm−1 x̃(m)
∂s ∂τ ∂x ∂τ λk ∂sk ∂τ
band of a roughly correct topology, and then moving k=0

the points along the billiard walls until the length (that = λm x(m) − x̃(m) , (16)
 5
k
where x = x(0) and x̃(k) = λ∂k ∂x̃k s , k = 1, · · · , m − 1 located at the top-right and bottom-left corners take care
are assumed. Conventionally, (15) is converted to a sys- of the periodic boundary condition.
tem of md first order differential equations, whose dis- The discretized version of (10) with a fictitious time
cretized derivative (see (17) below) are banded matrices Euler step δτ is
with band width of 5md. Using (16), we only need d
λv̂ − ṽˆ
    
equations for the same accuracy and the corresponding  v̂ δx̂
= δτ , (18)
band width is (m + 4)d. The computing load has been â 0 δλ 0
greatly reduced, the more so the larger m is. Neverthe-
less, choice of a good initial loop guess and visualization where
of the dynamics are always aided by a plot of the orbit
in the full md-dimensional phase space, where loops can- Â = D̂ + diag[A1 , A2 , · · · , AN ] ,
not self-intersect and topological features of the flow is
with An = A(x̃(sn )) defined in (5), and
exhibited more clearly.
v̂ = (v1 , v2 , · · · , vN ) , with vn = v(x̃(sn )) ,
IV. IMPLEMENTATION OF NEWTON ṽˆ = (ṽ1 , ṽ2 , · · · , ṽN ) , with ṽn = ṽ(x̃(sn )) ,
DESCENT
are the two vector fields that we want to match every-
where along the loop. â is an N d dimensional row vector
As the loop points satisfy a periodic boundary con-
which imposes the constraint on the coordinate varia-
dition, it is natural to employ truncated discrete Fast
tions δx̂ = (δx̃1 , δx̃2 , · · · , δx̃N ). The discretized New-
Fourier Transforms (FFT) in numerical integrations of
ton descent (18) is an infinitesimal time step variant of
(10). Since we are interested only in the final, stationary
the multipoint (Poincaré section) shooting equation for
cycle p, the accuracy of the fictitious time integration is
flows [16]. In order to solve for the deformation of the
not crucial; all we have to ensure is the smoothness of the
loop coordinates and period, δx̂ and δλ, we need to in-
loop throughout the integration. The Euler integration
vert the [(N d + 1)×(N d + 1)] matrix on the left hand
with fairly large time steps δτ suffices. The computation-
side of (18).
ally most onerous step in implementation of the Newton
In our numerical work, this matrix is inverted using
descent is the inversion of large matrix Ā in (10). When
the banded LU decomposition on the embedded band-
the dimension of the dynamical phase space of (2) is high,
diagonal matrix, and the Woodbury formula [21] on the
the inversion of Ā needed to get ∂∂τx̃ takes most of the inte-
cyclic, border terms. The LU decomposition takes most
gration time, making the evolution extremely slow. This
of the computational time and considerably slows down
problem is partially solved if the finite difference meth-
the fictitious time integration. We speed up the inte-
ods are used. The large matrix Ā then becomes sparse
gration by a new inversion scheme which relies on the
and the inversion can be done far more quickly.
smoothness of the flow in the loop space. It goes as fol-
lows. Once we have the LU decomposition at one step,
A. Numerical implementation we use it to approximately invert the matrix in the next
step, with accurate inversion achieved by the iterative
approximate inversions [21]. In our applications we find
In a discretization of a loop, numerical stability re-
that a single LU decomposition can be used for many δτ
quires accurate discretization of loop derivatives such as
evolution steps. The further we go, the more iterations at
∂ x̃ each step are needed to implement the inversion. After
ṽn ≡ ≈ (D̂x̃)n .
∂s x̃=x̃(sn )
the number of such iterations exceeds some given fixed
maximum number, we perform another LU decomposi-
In our numerical work we use the four-point approxima- tion and proceed as before. The number of integration
tion [20], steps following one decomposition is an indication of the

0 8 −1 1−8

smoothness of the evolution, and we adjust accordingly
 −8 0 8 −1 1  the integration step size δτ : the greater the number, the
 1 −8 0 8 −1 bigger the step size. As the loop approaches a cycle, the
 
1 

D̂ =

··· evolution becomes so smooth that the step size can be
12h 
 
 1 −8 0 8 −1  brought all the way up to δτ = 1, the full undamped
 −1 1 −8 0 8 
Newton-Raphson iteration step. In practice, one can
8 −1 1 −8 0
start with a small but reasonable number of points, in
(17)
order to get a coarse solution of relatively low accuracy.
where h = 2π/N . Here, each entry represents a [d × d]
After achieving that, the refined guess loop can be con-
matrix, 8 → 81, etc., with blank spaces are filled with
structed by interpolating more points, and proceed with
zeros. The two [2d×2d] matrices
for a more accurate calculation in which δτ can be set as
1 −81 −1 0
   
large as the full Newton step δτ = 1, recovering the rapid
M1 = , M2 = ,
0 1 81 −1 quadratic convergence of the Newton-Raphson method.
 6

It is essential that the smoothness of the loop is main- (2). However, if the value of cost functional is not equal
tained throughout the calculation. We monitor the to zero at the minimum while the gradient is zero, (18)
smoothness by checking the Fourier spectrum of x̃(·, τ ). yields a singular matrix Â. In such cases the search has
An unstable difference scheme for loop derivatives might to be abandoned and restarted with a new initial loop
lead to unbounded sawtooth oscillations [22]. A heuristic guess. In the periodic orbit searches of sect. V starting
local linear stability analysis (described in [23]) indicates with blind initial guesses (guesses unaided by a symbolic
that our scheme is stable, and that the high frequency dynamics partition), such local minima were encountered
components do not generate instabilities. in about 30% of cases.

B. Initialization and convergence C. Symmetry considerations

As in any other method, a qualitative understanding of The system under consideration often possesses certain
the dynamics is a prerequisite to successful cycle searches. symmetries. If this is the case, the symmetry should be
We start by numerical integration with the dynamical both be feared for possible marginal eigen-directions, and
system (2). Numerical experiments reveal regions where be embraced as a guide to possible simplifications of the
a trajectory spends most of its life, giving us the first numerical calculation.
hunch as to how to initialize a loop. We take the FFT If the dynamical system equations (2) are invariant
of some nearly recurred orbit segment and keep only the under a discrete symmetry, the concept of fundamental
lowest frequency components. The inverse Fourier trans- domain [5, 26] can be utilized to reduce the length of the
form back to the phase space yields a smooth loop that we initial loop when searching for a cycle of a given symme-
use as our initial guess. Since any generic orbit segment is try. In this case, we need discretize only an irreducible
not closed and might exhibit large gaps, the Gibbs phe- segment of the loop, decreasing significantly the dimen-
nomenon can take the initial loop so constructed quite sionality of the loop representation. Other parts of the
far away from the region of interest. We deal with this loop are replicated by symmetry operations, with the full
problem by manually deforming the orbit segment into loop tiled by copies of the fundamental domain segment.
a closed loop before performing the FFT. Searching for The boundary conditions are not periodic any longer, but
longer cycles with multiple circuits requires more delicate all that we need to do is modify the cyclic terms. Instead
initial conditions. The hope is that a few short cycles can of using M1 and M2 in (17), we use M1 Q and M2 Q−1 ,
help us establish an approximate symbolic dynamics, and where Q is the relevant symmetry operation that maps
guess for longer cycles can be constructed by cutting and the fundamental segment to the neighbor that precedes
glueing the short, known ones. For low dimensional sys- it. In this way, a fraction of the points represent the cycle
tems, such methods yield quite good systematic initial with the same accuracy, speeding up the search consid-
guesses for longer cycles [24]. erably.
An alternative way to initialize the search is by utiliz- If a continuous symmetry is present, it may compli-
ing adiabatic deformations of dynamics, or the homotopy cate the situation at first glance but becomes something
evolution [25]. If the dynamical system (2) depends on a that we can take advantage of after careful checking. For
parameter µ, short cycles might survive as µ varies pass- example, for a Hamiltonian system unstable cycles may
ing through a family of dynamical systems, giving in the form continuous families [27, 28], with one or more mem-
process birth to new cycles through sequences of bifur- bers of a family belonging to a given constant energy sur-
cations. Most short unstable cycles vary little for small face. In order to cope with the marginal eigendirection
changes of µ. So, a cycle existing for parameter value µ1 associated with such continous family, we search for a
can be chosen as the initial trial loop for a nearby cycle cycle on a particular energy surface by replacing the last
surviving a small change µ1 → µ2 . In practice, one or row of equation (18) by an energy shell constraint [16].
two iterations often suffice to find the new cycle. We put one point of the loop, say x̃2 , on the constant
A good choice of the initial loop significantly expe- energy surface H(x̃) = E, and impose the constraint
dites the computation, but there are more reasons why ▽H(x̃2 ) · δx̃2 = 0, so as to keep x̃2 on the surface for
good initial loops are crucial. First of all, if we break the all τ . The integration of (10) then automatically brings
translational invariance by imposing a constraint such as all other loop points to the same energy surface. Alter-
x̃1 (s1 , τ ) = c, we have to make sure that both the ini- natively, we can look for a cycle of given fixed period T
tial loop and the desired cycle intersects this Poincaré by fixing λ and dropping the constraint in the bottom
plane. Hence, the initial loop cannot be wildly different line of (18). These two approaches are conjugate to each
from the desired cycle. Second, in view of (12), the loop other, both needed in applications. In most cases, they
always evolves towards a local minimum of the cost func- are equivalent. One exception is the harmonic oscillator
tional (13), with discretization points moving along the for which the oscillations have identical period but dif-
ṽ−λv fixed direction, determined by the initial condition. ferent energy. Note that in both cases the translational
If the local minimum corresponds to a zero of the cost invariance is restored, as we have discarded the Poincaré
functional, we obtain a true cycle of the dynamical flow section condition of sect. II B. As explained in [6], this
 7

causes no trouble in numerical calculations. 0.6

0.2
V. APPLICATIONS

y
We have checked that the iteration of (18) yields −0.2

quickly and robustly the short unstable cycles for stan-

dard models of low-dimensional dissipative flows such as −0.6
the Rössler system [29]. A more daunting challenge are −0.6 −0.2 0.2 0.6
(a) x
searches for cycles in Hamiltonian flows, and searches for
0.6
spatio-temporally periodic solutions of PDEs. In all nu-
merical examples that follow, the convergence condition
is F 2 < 10−5 . 0.2

y
A. Hénon-Heiles system and restricted three-body −0.2
problem

−0.6
First, we test the Hamiltonian version of the Newton −0.6 −0.2 0.2 0.6
(b) x
descent derived in Sect. III by applying the method to
two Hamiltonian systems, both with two degrees of free- 10

dom. In both cases, our initial loop guesses are rather

5
arbitrary combinations of trigonometric functions. Nev-
ertheless, the observed convergence is fast. 0

ln(F)
The Hénon-Heiles system [30] is a standard model in
celestial mechanics, described by the Hamiltonian −5

1 2 y3 −10
H= (px + p2y + x2 + y 2 ) + x2 y − . (19)
2 3 −15
0 2 4 6 8 10
It has a time reversal symmetry and a three-fold dis- (c) τ
crete spatial symmetry. Figure 2 shows a typical appli-
cation of (14), with the Newton descent search restricted FIG. 2: The Hénon-Heiles system in a chaotic region: (a)
to the configuration space. The initial loop, Fig. 2(a), An initial loop L(0), and (b) the unstable periodic orbit p
of period T = 13.1947 reached by the Newton descent (14).
is a rather coarse initial guess. We fix arbitrarily the
(c) The exponential decrease of the cost function, ln(F 2 ) ≈
scaling λ = 2.1, that is, we search for a cycle p of the −2.0502 τ + 6.0214.
fixed period Tp = 13.1947, with no constraint on the en-
ergy. Figure 2(b) shows the cycle found by the Newton
descent, with energy E = 0.1794, and the full discrete p p
symmetry of the Hamiltonian. This cycle persists adia- where r1 = (x + µ)2 + y 2 , r2 = (x − 1 + µ)2 + y 2 .
batically for a small range of values of λ; with λ changed These equations describe the motion of a test particle in
much, the Newton descent takes the same initial loop into a rotating frame under the influence of the gravitational
other cycles. Figure 2(c) verifies that the cost functional force of two heavy bodies with masses 1 and µ ≪ 1 fixed
F 2 decreases exponentially with slope -2 throughout the at (−µ, 0) and (1 − µ, 0) in the (x, y) coordinate frame.
τ = [0, 10] integration interval, as predicted by (12). The The stationary solutions of (20) are called the Lagrange
points get more and more sparse as τ increases, because points, corresponding to a circular motion of the test
our numerical implementation adaptively chooses bigger particle in phase with the rotation of the heavy bodies.
and bigger step sizes δτ . The periodic solutions in the rotating frame correspond
In the Hénon-Heiles case, the accelerations ax , ay de- to periodic or quasi-periodic motion of the test particle
pend only on the configuration variables x, y. More gen- in the inertial frame. Figure 3 shows an initial loop and
erally, the accelerations could also depend on ẋ, ẏ. Con- the cycle to which it converges, in the rotating frame.
sider as an example the equations of motion for the re- Although the cycle looks simple, the Newton descent re-
stricted three-body problem [31], quires advancing in small δτ steps in order for the initial
loop to converge to it.
x+µ x−1+µ In order to successfully apply the Hamiltonian version
ẍ = 2ẏ + x − (1 − µ) −µ ,
r13 r23 of the Newton descent (14), we have to ensure that the
y y test particle keeps a finite distance from the origin. If
ÿ = −2ẋ + y − (1 − µ) 3 − µ 3 , (20)
r1 r2 a cycle passes very close to one of the heavy bodies,
 8

0.3 where a−k = −ak ∈ R . In terms of the Fourier com-

ponents, PDE (21) becomes an infinite ladder of ODEs,

0.1 ∞
X
a˙k = (k 2 − νk 4 )ak − k am ak−m . (23)
y m=−∞
−0.1
In numerical simulations we work with the Galerkin trun-
cations of the Fourier series since in the neighborhood of
−0.3 the strange attractor the magnitude of ak decreases very
0.7 0.9 1.1 fast with k, high frequency modes playing a negligible
(a) x role in the asymptotic dynamics. In this way Galerkin
truncations reduce the dynamics to a finite but large
0.3
number of ODEs. We work with d = 32 dimensions in
our numerical calculations. In ref. [24], multipoint shoot-
0.1
ing has been successfully applied to obtain periodic orbits
close to the onset of spatiotemporal chaos (ν = 0.03). In
y

this regime, our method is so stable that big time steps

−0.1 δτ can be employed even at the initial guesses, leading to
extremely fast convergence. We attribute this robustness
to the simplicity of the structure of the attractor at high
−0.3 viscosity values.
0.7 0.9 1.1 The challenge comes with decreasing ν, with the dy-
(b) x namics turning more and more turbulent. Already at
ν = 0.015, the system is moderately turbulent and the
FIG. 3: (a) An initial loop L(0), and (b) the unstable periodic phase space portraits of the dynamics reveal a complex
orbit p of period Tp = 2.7365 reached by the Newton descent labyrinth of “eddies” of different scales and orientations.
(14), the restricted three body problem (20) in the chaotic While the highly unstable nature of orbits and intricate
regime, µ = 0.04. structure of the invariant set hinder applications of con-
ventional cycle-search routines, in this setting our vari-
ational method shines through. We design rather arbi-
the acceleration can become so large that our scheme trary initial loops from numerical trajectory segments,
of uniformly distributing the loop points in time might and the calculation proceeds as before, except that now
fail to represent the loop faithfully. Another distribu- a small δτ has to be used initially to ensure numerical
tion scheme is required in this case, for example, making stability. Topologically different loops are very likely to
the density of points proportional to the magnitude of result in different cycles, while some initial loop guesses
acceleration. my lead to local nonzero minima of the cost functional
F 2 . As explained in Sect. IV, in such cases the method
diverges, and the search is restarted with a new inital
B. Periodic orbits of Kuramoto-Sivashinsky system
loop guess.
Two initial loop guesses are displayed in Fig. 4, along
The Kuramoto-Sivashinsky equation arises as an am- with the two periodic orbits detected by the Newton de-
plitude equation for interfacial instability in a variety of scent. In discretization of the initial loops, each point has
contexts [32, 33]. In 1-dimensional space, it reads to be specified in all d dimensions; here the coordinates
{a1 , a2 } are picked so that topological similarity between
ut = (u2 )x − uxx − νuxxxx, (21) initial and final loops is visually easy to identify. Other
projections from d = 32 dimensions to subsets of 2 coor-
where ν is a “super-viscosity” parameter which controls
dinates appear to make the identification harder, if not
the rate of dissipation and (u2 )x is the nonlinear con-
impossible. In both calculations, we molded segments
vection term. As ν decreases, the system undergoes a
of typical trajectories into smooth closed loops by the
series of bifurcations, leading to increasingly turbulent,
Fourier filtering method of Sect. IV. As the desired orbit
spatio-temporally chaotic dynamics.
becomes longer and more complex, more sampling points
If we impose the periodic boundary condition u(t, x +
are needed to represent the loop. We use N = 512 points
2π) = u(t, x) and choose to study only the odd solutions
to represent the loop in the (a)-(b) case and N = 1024
u(−x, t) = −u(x, t), u(x, t) can be expanded in a discrete
points in the (c)-(d) case. The space-time evolution of
spatial Fourier series [24],
u(x, t) for these two unstable spatio-temporally periodic
∞
X solutions is displayed in Fig. 5. As u(x, t) is antisymmet-
u(x, t) = i ak (t)eikx , (22) ric on [−π, π], it suffices to display the solutions on the
k=−∞ x ∈ [0, π] interval.
 9

0.2
1
−0.2

2
a
−0.6 0.6

t
−1
0.2
−1.4
−1.6 −1.2 −0.8 −0.4
a 0 1 2 3
(a) 1
(a) x
0.2

1
−0.2
2
a

−0.6
0.6

t
−1

0.2
−1.4
−1.6 −1.2 −0.8 −0.4
a1
(b) 0 1 2 3
0 (b) x

FIG. 5: Level plot of the space-time evolution u(x, t) for

−0.5 the two spatiotemporally periodic solutions of Fig. 4: (a) the
evolution of p1 , with start of a repeat after the cycle period
a2

T1 = 0.744892, and (b) one full period T2 = 1.184668 in the

−1 evolution of p2 .

−1.5
−2 −1.5 −1 −0.5 0
a
(c) 1 of minimizing the cost functional (13). This equation de-
0 scribes the fictitious time τ flow in the space of loops
which decreases the cost functional at uniform exponen-
tial rate (see (12)). Variants of the method are presented
−0.5 for special classes of systems, such as Hamiltonian sys-
tems. An efficient integration scheme for the PDE is de-
2
a

vised and tested on the Kuramoto-Sivashinsky system,

−1
the Hénon-Heiles system and the restricted three-body
problem.
−1.5
−2 −1.5 −1 −0.5 0 Our method uses information from a large number of
a
(d) 1 points in phase space, with the global topology of the
desired cycle protected by insistence on smoothness and
FIG. 4: The Kuramoto-Sivashinsky system in a spatio- a uniform discretization of the loop. The method is quite
temporally turbulent regime (viscosity parameter ν = 0.015, robust in practice.
d = 32 Fourier modes truncation). (a) An initial guess L1 ,
and (b) the periodic orbit p1 of period T1 = 0.744892 reached The numerical results presented here are only a proof
by the Newton descent. (c) Another initial guess L2 , and (d) of principle. We do not know to what periodic orbit the
the resulting periodic orbit p2 of period T2 = 1.184668. flow (10) will evolve for a given dynamical system and a
given initial loop. Empirically, the flow goes toward the
“nearest” periodic orbit, with the largest topological re-
VI. DISCUSSION semblance. Each particular application will still require
much work in order to elucidate and enumerate relevant
In order to cope with the difficulty of finding periodic topological structures. The hope is that the short spatio-
orbits in high-dimensional chaotic flows, we have devised temporally periodic solutions revealed by the Newton
the Newton descent method, an infinitesimal step variant descent searches will serve as the basic building blocks
of the damped Newton-Raphson method. Our main re- for systematic investigations of chaotic and perhaps even
sult is the PDE (10) which solves the variational problem “turbulent” dynamics.
 10

Acknowledgments

We would like to thank Cristel Chandre for careful

reading of the manuscript and numerous suggestions.

[1] D. Ruelle, Statistical Mechanics, Thermodynamic For- [18] H. B. Keller, Numerical methods for two-point boundary-
malism (Addison-Wesley, Reading MA, 1978). value problems (Dover, New York, 1992).
[2] M. Gutzwiller, Chaos in Classical and Quantum Mechan- [19] J. Stoer and R. Bulirsch, Introduction to Numerical Anal-
ics (Springer-Verlag, New York, 1990). ysis (Springer-Verlag, New York, 1983).
[3] R. Artuso, E. Aurell, and P. Cvitanović, Nonlinearity 3, [20] A. Brandenburg, in Advances in nonlinear dynamos,
325 (1990). edited by A. Ferriz-Mas and M. Núñez (Taylor & Francis,
[4] H. Rugh, Nonlinearity 5, 1237 (1992). London, 2003), astro-ph/0109497.
[5] P. Cvitanović, R. Artuso, R. Mainieri, G. Tanner, and [21] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P.
G. Vattay, Chaos: Classical and Quantum (Niels Bohr Flannery, Numerical Recipes in C (Cambridge University
Institute, Copenhagen, 2003), www.nbi.dk/ChaosBook. Press, Cambridge, 1992).
[6] D. Viswanath, unpublished. [22] J. W. Thomas, Numerical Partial Differential Equations
[7] F. Diakonos, P. Schmelcher, and O. Biham, Phys. Rev. (Spring-Verlag, New York, 1995).
Lett. 81, 4319 (1998), chao-dyn/9810022. [23] B. Roždestvenskiǐ and N. Janenko, Systems of Quasilin-
[8] O. Biham and W. Wenzel, Phys. Rev. Lett. 63, 819 ear Equations and Their Applications to Gas Dynamics
(1989). (AMS, Providence, 1983).
[9] R. L. Davidchack and Y.-C. Lai, Phys. Rev. E 60, 6172 [24] F. Christiansen, P. Cvitanović, and V. Putkaradze, Non-
(1999). linearity 10, 55 (1997).
[10] U. Frisch, Turbulence (Cambrige University Press, Cam- [25] C. Conley, Isolated Invariant Sets and the Morse Index
brige, 1996). (AMS, Providence, 1978).
[11] I. S. Aranson and L. Kramer, 74, 99 (2002). [26] P. Cvitanović and B. Eckhardt, Nonlinearity 6, 277
[12] T. S. Biró, S. G. Matinyan, and B. Müller, Chaos and (1993), chao-dyn/9303016.
Gauge Field Theory (World Scientific, Singapore, 1994). [27] M. Hénon, Generating Families in the Restricted Three-
[13] P. Cvitanović, Physica A 288, 61 (2000), Body Problem II. Quantitative Study of Bifurcations
nlin.CD/0001034. (Springer, New York, 2001).
[14] G. Kawahara and S. Kida, J. Fluid Mech. 449, 291 [28] J.-A. Sepulchre and R. Mackay, Nonlinearity 10, 679
(2001). (1997).
[15] P. Cvitanović and Y. Lan, in Proceed. of 10. Intern. [29] O. E. Rössler, Phys. Lett. A 57, 397 (1976).
Workshop on Multiparticle Production: Correlations and [30] M. Hénon and C. Heiles, Astron. J. 69, 73 (1964).
Fluctuations in QCD, edited by N. Antoniou (World Sci- [31] V. Szebehely, Theory of Orbits (Academic Press, New
entific, Singapore, 2003), nlin.CD/0308006. York, 1967).
[16] F. Christiansen, chapter “Fixed points, and how to get [32] Y. Kuramoto and T. Tsuzuki, Progr. Theor. Physics 55,
them”, in ref. [5]. 365 (1976).
[17] V. I. Arnol’d, Ordinary Differential Equations (Springer, [33] G. I. Sivashinsky, Acta Astr. 4, 1177 (1977).
New York, 1992).