Physica A 288 (2000) 61– 80

www.elsevier.com/locate/physa

Chaotic eld theory: a sketch

Predrag Cvitanovic ∗
Department of Physics and Astronomy, Northwestern University, 2145 Sheridan Road,
Evanston, IL 60208, USA

Abstract
Spatio-temporally chaotic dynamics of a classical eld can be described by means of an in-
nite hierarchy of its unstable spatio-temporally periodic solutions. The periodic orbit theory
yields the global averages characterizing the chaotic dynamics, as well as the starting semi-
classical approximation to the quantum theory. New methods for computing corrections to the
semiclassical approximation are developed; in particular, a nonlinear eld transformation yields
the perturbative corrections in a form more compact than the Feynman diagram expansions.
c 2000 Elsevier Science B.V. All rights reserved.

PACS: 02.50.Ey; 03.20.+i; 03.65.Sq; 05.40.+j; 05.45.+b

Keywords: Periodic orbits; Field theory; Semiclassical quantization; Trace formulas

0. Introduction

Formulated in 1946 –1949 and tested through 1970s, quantum electrodynamics takes
free electrons and photons as its point of departure, with nonlinear eects taken in
account perturbatively in terms of Feynman diagrams, as corrections of order (=)n =
(0:002322819 : : :)n . QED is a wildly successful theory, with Kinoshita’s [1] calculation
of the electron magnetic moment
X   n
1
2 (g − 2) = {· · · + + · · ·}
n


agreeing with Dehmelt’s experiments [2] to 12 signicant digits.

Quantum chromodynamics perturbative calculations seemed the natural next step,
the only new feature being the gluon–gluon interactions. However, in this case the
∗ Address as of January 1, 2001: School of Physics, Georgia Institute of Technology, Atlanta, GA 30332,
USA.

0378-4371/00/$ - see front matter c 2000 Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 8 - 4 3 7 1 ( 0 0 ) 0 0 4 1 5 - 5
62 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

Feynman–diagrammatic expansions for observables such as the meson and hadron

masses
X
(observable) = (QCD )n {· · · + + · · ·}
n

failed us utterly, perhaps because the expansion parameter is of order 1. I say perhaps,
because more likely the error in this case is thinking in terms of quarks and gluons in
the rst place. Strongly nonlinear eld theories require radically dierent approaches,
and in 1970s, with a deeper appreciation of the connections between eld theory and
statistical mechanics, their re-examination led to path integral formulations such as the
lattice QCD [3]. In lattice theories quantum uctuations explore the full gauge group
manifold, and classical dynamics of Yang–Mills elds plays no role.
We propose to re-examine here the path integral formulation and the role that the
classical solutions play in quantization of strongly nonlinear elds. In the path integral
formulation of a eld theory the dominant contributions come from saddlepoints, the
classical solutions of equations of motion. Usually, one imagines one dominant saddle
point, the “vacuum”:

The Feynman diagrams of QED and QCD are nothing more than a scheme to com-
pute the correction terms to this starting semiclassical, Gaussian saddlepoint approxima-
tion. But there might be other saddles. That eld theories might have a rich repertoire
of classical solutions became apparent with the discovery of instantons [4], analytic
solutions of the classical SU (2) Yang–Mills equations of motion, and the realization
that the associated instanton vacua receive contributions from countable ∞’s of sad-
dles. What is not clear is whether these are the important classical saddles. Could it
be that the strongly nonlinear theories are dominated by altogether dierent classical
solutions?
The search for the classical solutions of nonlinear eld theories such as the Yang–
Mills and gravity has so far been neither very successful nor very systematic. In modern
eld theories the main emphasis has been on symmetries as guiding principles in writing
down the actions. But writing down a dierential equation is only the start of the story;
even for systems as simple as 3 coupled ordinary dierential equations one in general
has no clue what the nature of the long-time solutions might be.
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 63

These are hard problems, and in explorations of modern eld theories the dynamics
tends to be is neglected, and understandably so, because the wealth of the classical
solutions of nonlinear systems can be truly bewildering. If the classical behavior of
these theories is anything like that of the eld theories that describe the classical world
– the hydrodynamics, the magneto-hydrodynamics, the Ginzburg–Landau system – there
should be too many solutions, with very few of the important ones analytical in form;
the strongly nonlinear classical eld theories are turbulent, after all. Furthermore, there
is not a dimmest hope that such solutions are either beautiful or analytic, and there is
not much enthusiasm for grinding out numerical solutions as long as one lacks ideas
as what to do with them.
By late 1970s it was generally understood that even the simplest nonlinear systems
exhibit chaos. Chaos is the norm also for generic Hamiltonian ows, and for path
integrals that implies that instead of a few, or countably few saddles, classical solutions
populate fractal sets of saddles.

For the path-integral formulation of quantum mechanics such solutions were discov-
ered and accounted for by Gutzwiller [5] in late 1960s. In this framework the spectrum
of the theory is computed from a set of its unstable classical periodic solutions. The
new aspect is that the individual saddles for classically chaotic systems are nothing
like the harmonic oscillator degrees of freedom, the quarks and gluons of QCD – they
are all unstable and highly nontrivial, accessible only by numerical techniques.
So, if one is to develop a semiclassical eld theory of systems that are classically
chaotic or “turbulent”, the problem one faces is twofold
(1) Determine, classify, and order by relative importance the classical solutions of
nonlinear eld theories.
(2) Develop methods for calculating perturbative corrections to the corresponding
classical saddles.
Our purpose here is to give an overview over the status of this program – for details
the reader is referred to the literature cited.
The rst task, a systematic exploration of solutions of eld theory has so far been
implemented only for one of the very simplest eld theories, the one-dimensional
64 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

Kuramoto–Sivashinsky system. We sketch below how its spatio-temporally chaotic

dynamics can be described in terms of spatio-temporally recurrent unstable patterns.
For the second task, the theory of perturbative corrections, we shall turn to an even
simpler system; a weakly stochastic mapping in one-dimension. The new aspect of the
theory is that now the corrections have to be computed saddle by saddle. In Sections
3– 6, we discuss three distinct methods for their evaluation.

1. Unstable recurrent patterns in classical ÿeld theories

Field theories such as four-dimensional QCD or gravity have many dimensions,

symmetries, tensorial indices. They are far too complicated for exploratory forays into
this forbidding terrain. We start instead by taking a simple spatio-temporally chaotic
nonlinear system of physical interest, and investigate the nature of its solutions.
One of the simplest and extensively studied spatially extended dynamical systems is
the Kuramoto–Sivashinsky system [6,7]
ut = (u2 )x − uxx − uxxxx (1)
which arises as an amplitude equation for interfacial instabilities in a variety of contexts.
The “ ame front” u(x; t) has compact support, with x ∈ [0; 2] a periodic space coordi-
nate. The u2 term makes this a nonlinear system, t is the time, and  is a fourth-order
“viscosity” damping parameter that irons out any sharp features. Numerical simulations
demonstrate that as the viscosity decreases (or the size of the system increases), the
“ ame front” becomes increasingly unstable and turbulent. The task of the theory is
to describe this spatio-temporal turbulence and yield quantitative predictions for its
measurable consequences.
Armed with a computer and a great deal of skill, one can obtain a numerical solution
to a nonlinear PDE. The real question is; once a solution is found, what is to be done
with it? The periodic orbit theory is an answer to this question.
Dynamics drives a given spatially extended system through a repertoire of unstable
patterns; as we watch a “turbulent” system evolve, every so often we catch a glimpse
of a familiar pattern:

For any nite spatial resolution, the system follows approximately for a nite time
a pattern belonging to a nite alphabet of admissible patterns, and the long-term dy-
namics can be thought of as a walk through the space of such patterns, just as chaotic
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 65

dynamics with a low-dimensional attractor can be thought of as a succession of nearly

periodic (but unstable) motions. The periodic orbit theory provides the machinery that
converts this intuitive picture into precise calculation scheme that extracts asymptotic
time predictions from the short-time dynamics. For extended systems the theory gives
a description of the asymptotics of partial dierential equations in terms of recurrent
spatio-temporal patterns.
Putkaradze has proposed that the Kuramoto–Sivashinsky system (1) be used as a
laboratory for exploring such ideas. We now summarize the results obtained so far in
this direction by Christiansen et al. [8] and Zoldi and Greenside [9].
The solution u(x; t) = u(x + 2; t) is periodic on the x ∈ [0; 2] interval, so one
(but by no means only) way to solve such equations is to expand u(x; t) in a discrete
spatial Fourier series
+∞
X
u(x; t) = i ak (t)eikx : (2)
k=−∞

Restrict the consideration to the subspace of odd solutions u(x; t)=−u(−x; t) for which
ak are real. Substitution of Eq. (2) into Eq. (1) yields the innite ladder of evolution
equations for the Fourier coecients ak :
∞
X
ȧk = (k 2 − k 4 )ak − k am ak−m : (3)
m=−∞

u(x; t) = 0 is a xed point of Eq. (1), with the k 2  ¡ 1 long wavelength modes
of this xed point linearly unstable, and the short wavelength modes stable. For
 ¿ 1; u(x; t) = 0 is the globally attractive stable xed point; starting with  = 1 the
solutions go through a rich sequence of bifurcations, and myriad unstable periodic
solutions whose number grows exponentially with time.
The essential limitation on the numerical studies undertaken so far have been com-
putational constraints: in truncation of high modes in expansion (3), suciently many
have to be retained to ensure the dynamics is accurately represented. Christiansen
et al. [8] have examined the dynamics for values of the damping parameter close to
the onset of chaos, while Zoldi and Greenside [9] have explored somewhat more tur-
bulent values of . With improvement of numerical codes considerably more turbulent
regimes should become accessible.
One pleasant surprise is that even though one is dealing with (innite dimensional)
PDEs, for these strong dissipation values of parameters the spatio-temporal chaos is
suciently weak that the ow can be visualised as an approximately one-dimensional
Poincare return map s → f(s) from the unstable manifold of the shortest periodic
point onto its neighborhood (see Fig. 1(a)). This representation makes it possible to
systematically determine all nearby periodic solutions up to a given maximal period.
So far some 1000 prime cycles have been determined numerically for various values
of viscosity. The rapid contraction in the nonleading eigendirections is illustrated in

Fig. 1(b) by the plot of the rst 16 eigenvalues of the 1-cycle. As the length of
66 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

Fig. 1. (a) The return map sn+1 = f(sn ) constructed from periodic solutions of the Kuramoto–Sivashinsky
equations (1),  = 0:029910, with s the distance measured along the unstable manifold of the xed point 1. 
Periodic points 0 and 01 are also indicated. (b) Lyapunov exponents k versus k for the periodic orbit 1
compared with the stability eigenvalues of the u(x; t) = 0 stationary solution k 2 − k 4 . k for k¿8 lie below
the numerical accuracy of integration and are not meaningful. From Ref. [8].

Fig. 2. (a) Spatio-temporally periodic solution u0 (x; t) of the Kuramoto–Sivashinsky system, viscosity
parameter  = 0:029910. (b) The dierence between the two shortest period spatio-temporally periodic solu-
tions u0 (x; tT0 ) and u1 (x; tT1 ). From Ref. [8].

the orbit increases, the magnitude of contracting eigenvalues falls o very quickly. In

Fig. 2, we plot u0 (x; t) corresponding to the 0-cycle. The dierence between this solu-
tion and the other shortest period solution is of the order of 50% of a typical variation
in the amplitude of u(x; t), so the chaotic dynamics is already exploring a sizable swath
in the space of possible patterns even so close to the onset of spatio-temporal chaos.
Other solutions, plotted in the conguration space, exhibit the same overall gross struc-
ture. Together they form the repertoire of the recurrent spatio-temporal patterns that is
being explored by the turbulent dynamics.
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 67

2. Periodic orbit theory

Now, we turn to the central issue; qualitatively, these solutions demonstrate that the
recurrent patterns program can be implemented, but how is this information to be used
quantitatively? This is what the periodic orbit theory is about; it oers the machinery
that assembles the topological and the quantitative information about individual solu-
tions into accurate predictions about measurable global averages, such as the Lyapunov
exponents and correlation functions.
Very brie y (for a detailed exposition the reader is referred to Ref. [10]), the task
of any theory that aspires to be a theory of chaotic, turbulent systems is to predict
the value of an “observable” a from the spatial and time averages evaluated along
dynamical trajectories x(t)

Z t
1
hai = lim hAt i; t
A (x) = da(x()) :
t→∞ t 0

The key idea of the periodic orbit theory is to extract this average from the leading
eigenvalue of the evolution operator

t
Lt (x; y) = (y − x(t))eÿA (x)

via the trace formula

X ∞
XX
t Tp (t − rTp )
trL = = erÿAp (4)
p p
|det(1 − Jpr )|
r=1
p

which relates the spectrum of the evolution operator to a sum over prime periodic
solutions p of the dynamical system and their repeats r.
What does this formula mean? Prime cycles partition the dynamical space into neigh-
borhoods, each cycle enclosed by a tube whose volume is the product of its length
Tp and its thickness |det(1 − Jp )|−1 . The trace picks up a periodic orbit contribution
only when the time t equals a prime period or its repeat, a constraint enforced here
by (t − rTp ). Jp is the linear stability of cycle p, so for long cycles |det(1 − Jpr )| ≈
(product of expanding eigenvalues), and the contribution of long and very unstable
cycles are exponentially small compared to the short cycles which dominate trace for-
mulas. The number of contracting directions and the overall dimension of the dynamical
space is immaterial; that is why the theory can also be applied to PDEs. All this infor-
mation is purely geometric, intrinsic to the ow, coordinate reparametrization invariant,
and the same for any average one might wish to compute. The information related to
a specic observable is carried by the weight eÿAp , the periodic orbit estimate of the
t
contribution of eÿA (x) from the p-cycle neighborhood.
68 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

t
The intuitive meaning of a trace formula is that it expresses the average heÿA i as a
discretized integral

over the dynamical space partitioned topologically into a repertoire of spatio-temporal

patterns, each weighted by the likelihood of pattern’s occurrence in the long-time evo-
lution of the system.
Periodic solutions are important because they form the skeleton of the invariant set
of the long-time dynamics, with cycles ordered hierarchically; short cycles give good
approximations to the invariant set, longer cycles renements. Errors due to neglecting
long cycles can be bounded, and for nice hyperbolic systems they fall o exponentially
or even super-exponentially with the cuto cycle length [11]. Short cycles can be
accurately determined and global averages (such as Lyapunov exponents and escape
rates) can be computed from short cycles by means of cycle expansions.
The Kuramoto–Sivashinsky periodic orbit calculations of Lyapunov exponents and
escape rates [8] demonstrate that the periodic orbit theory predicts observable averages
for deterministic but classically chaotic spatio-temporal systems. The main problem
today is not how to compute such averages – periodic orbit theory as well as direct
numerical simulations can handle that – but rather that there is no consensus on what
the sensible experimental observables worth are predicting.
It should be obvious, and it still needs to be said: the spatio-temporally periodic
solutions are not to be thought of as eigenmodes, a good linear basis for expressing
solutions of the equations of motion. Something like a dilute instant approximation
makes no sense at all for strongly nonlinear systems that we are considering here. As
the equations are nonlinear, the periodic solutions are in no sense additive, and their
linear superpositions are not solutions.
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 69

Instead, it is the trace formulas and spectral determinants of the periodic orbit theory
that prescribe how the repertoire of admissible spatio-temporal patterns is to be sys-
tematically explored, and how these solutions are to be put together in order to predict
measurable observables.
Suppose that the above program is successfully carried out for classical solutions
of some eld theory. What are we to make of this information if we are interested
in the quantum behavior of the system? In the semiclassical quantization the classical
solutions are the starting approximation.

3. Stochastic evolution

For the same pragmatic reasons that we found it protable to shy away from fac-
ing the four-dimensional QCD head on in the above exploratory foray into a strongly
nonlinear eld theory, we shall start out by trying to understand the structure of per-
turbative corrections for systems radically simpler than a full- edged quantum eld
theory. First, instead of perturbative corrections to the quantum problem, we shall start
by exploring the perturbative corrections to weakly stochastic ows. Second, instead
of continuous time ows, we shall start by a study of a discrete time process.
For discrete time dynamics a Langevin trajectory in presence of additive noise is
generated by iteration
x n+1 = f(x n ) + n ; (5)
where f(x) is a map, n a random variable, and  parametrizes the noise strength. In
what follows we assume that n are uncorrelated, and that the mapping f(x) is one
dimensional and expanding, but we expect that the form of the results will remain
the same for higher dimensions, including the eld theory example of the preceding
section.
Tracking an individual noisy trajectory does not make much sense; what makes
sense is the Fokker–Planck formulation, where one considers instead evolution of an
ensemble of trajectories. An initial density of trajectories 0 (x) evolves with time as
Z
n+1 (y) = (L ◦ n )(y) = d x L(y; x)n (x) ; (6)

where L is the evolution operator

Z
L(y; x) = (y − f(x) − )P() d = −1 P[−1 (y − f(x))] (7)

and n a random variable with the normalized distribution P(), centered on  = 0.

If the noise is weak, the goal of the theory is to compute the perturbative corrections
to the eigenvalues  of L order by order in the noise strength ,
∞
X m
() = (m) :
m!
m=0
70 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

One way to get at the spectrum of L is to consider the discrete Laplace transform of
Ln , or the resolvent
∞
X X z ∞
zL
z n trLn = tr = (8)
1 − zL 1 − z
n=1 =0

which has a pole at every z = −1  .

The eects of weak noise are of interest in their own right, as any deterministic
evolution that occurs in nature is aected by noise. However, what is most important in
the present context is the fact that the form of perturbative corrections for the stochastic
problem is the same as for the quantum problem, and still the actual calculations are
suciently simple that one can explore many more orders in perturbation theory than
would be possible for a full- edged eld theory, and develop new perturbative methods.
The rst method we try is the standard Feynman–diagrammatic expansion. For semi-
classical quantum mechanics of a classically chaotic system such calculation was rst
carried out by Gaspard [12]. The stochastic version described here, implemented by
Dettmann [13], reveals features not so readily apparent in the quantum calculation.
The Feynman diagram method becomes unwieldy at higher orders. The second
method, introduced by Vattay [14], is based on Rugh’s [11] explicit matrix repre-
sentation of the evolution operator. If one is interested in evaluating numerically many
orders of perturbation theory and many eigenvalues, this method is unsurpassed.
The third approach, the smooth conjugations introduced by Mainieri [15], is per-
haps an altogether new idea in eld theory. In this approach the neighborhood of each
saddlepoint is rectied by an appropriate nonlinear eld transformation, with the focus
shifted from the dynamics in the original eld variables to the properties of the con-
jugacy transformation. The expressions obtained are equivalent to sums of Feynman
diagrams, but are more compact.

4. Feynman diagrammatic expansions

We start our computation of the weak-noise corrections to the spectrum of L by

calculating the trace of the nth iterate of the stochastic √ evolution operator L. A con-
2
venient choice of noise is Gaussian, P() = e− =2 = 2, with the trace given by an
n-dimensional integral on n points along a discrete periodic chain
Z
tr Ln = d x0 : : : d x n−1 L(x0 ; x n−1 ) : : : L(x1 ; x0 )

Z ( )
1 X
= [d x] exp − 2 [xa+1 − f(xa )]2 ;
2 a

n−1
Y dx
x n = x0 ; [d x] = √ a : (9)
a=0 22
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 71

The choice of Gaussian noise is not essential, as the methods that we develop here ap-
ply equally well to other noise distributions, and more generally to the space-dependent
noise distributions P(x; ). As the neighborhood of any trajectory is nonlinearly dis-
torted by the ow, the integrated noise is anyway never Gaussian, but colored.
If the classical dynamics is hyperbolic, periodic solutions of given nite period n
are isolated. Furthermore, if the noise broadening  is suciently weak they remain
distinct, and the dominant contributions come from neighborhoods of periodic points,
the tubes sketched in the trace formula (4). In the saddlepoint approximation the trace
(9) is given by the sum over neighborhoods of periodic points

X X ∞
X
tr Ln → tr Ln |sc = eWc = np n; np r eWpr : (10)
xc ∈ Fixfn p r=1

As traces are cyclic, eWc is the same for all periodic points in a given cycle, independent
of the choice of the starting point xc , and the periodic point sum can be rewritten in
terms of prime cycles p and their repeats. In the deterministic,  → 0 limit this is the
discrete time version of the classical trace formula (4). Eects such as noise-induced
tunnelling are not included in the weak-noise approximation.
We now turn to the evaluation of Wpr , the weight of the rth repeat of prime cycle p.
The contribution of the cycle point xa neighborhood is best expressed in an intrinsic
coordinate system, by centering the coordinate system on the cycle points,

xa → xa + a : (11)

From now on xa will refer to the position of the ath periodic point, a to the deviation
of the noisy trajectory from the deterministic one, fa (a ) to the map (5) centered on
the ath cycle point, and fa(m) to its mth derivative evaluated at the ath cycle point:

fa (a ) = f(xa + a ) − xa+1 ; fa0 = f0 (xa ); fa00 = f00 (xa ); : : : : (12)

Rewriting the trace in vector notation, with x and f(x) n-dimensional column vectors
with components xa and f(xa ), respectively, expanding f in Taylor series around each
of the periodic points in the orbit of xc , separating out the quadratic part and integrating
we obtain
Z Z
−1 0 2 2 2 T 1
eWc = [d] e−(
−V ()) =2 = [d] e−(1=2 )
T
+(···)
c c
Z P 00 k 2 2
= |det
| [d’] e (1=k)tr(
V ()) e−’ =2 : (13)
c

The [n × n] matrix
arises from the quadratic part of the exponent, while all higher
powers of a are collected in V ():
∞
XX m+1
a

−1 0
ab b = −fa a + a+1 ; V () = fa(m) : (14)
a
(m + 1)!
m=2
72 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

The saddlepoint expansion is most conveniently evaluated in terms of Feynman dia-

grams, by drawing
as a directed line
ab = , and the derivatives of V as the
“interaction” vertices

In the jargon of eld theory,
is the “free propagator”. Its determinant

n−1
Y
1
|det
| = ; c = fa0 (15)
|c − 1|
a=0

is the one-dimensional version of the classical stability weight |det(1 − J)|−1 in (4),
with c the stability of the n-cycle going through the periodic point xc .
Standard methods [16] now yield the perturbation expansion in terms of the con-
nected “vacuum bubbles”
∞
X
Wc = −ln |c − 1| + Wc; 2k 2k ; (16)
k=1

;
Wc; 4 = · · · :

In the usual eld-theoretic calculations the Wc; 0 term corresponds to an overall volume
term that cancels out in the expectation values. In contrast, as explained in Section 2,
here the eWc; 0 = |c − 1|−1 term is the classical volume of cycle c. Not only does this
weight not cancel out in the expectation value formulas, it plays the key role both in
classical and semiclassical trace formulas.
In the diagrams sketched above a propagator line connects xa at time a with xb at
later time b by a deterministic trajectory. At time b noise induces a kick whose strength
depends on the local curvature of the ow. A penalty of a factor  is paid, m − 1
deterministic trajectories originate in the neighborhood of xb from vertex V (m) (xb ),
and the process repeats itself, each vertex carrying a penalty of , and higher deriva-
tives of the fb . Summing over all noise kick sequences encoded by a given diagram
and using the periodicity of the trace integral (9) Dettmann [13] obtains expressions
such as
! a−1
r 2r
p −1 rp X 00
fa 2 fa000 Y 0
02 − fd2 : (17)
2 2p − 1 (rp − 1)3 fa fa0
ab d = b+1
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 73

This particular sum is the

Feynman diagram 2 correction to rth repeat of prime cycle p. More algebra leads to
similar contributions from the remaining diagrams. But the overall result is surprising;
the dependence on the repeat number r factorizes, with each diagram yielding the same
prefactor depending only on rp . This remarkable fact will be explained in Section 6.
The result of the Feynman-diagrammatic calculations is the stochastic trace formula
 ∞
zL  XX np tp; k z np 2 (2) 4
tr  = ; tp; k = k
e( =2)wp; k +O( ) ; (18)
1 − zL sc p
1 − tp; k |p |p
k =0

where tp; k is the kth local eigenvalue evaluated on the p cycle. The deterministic,
 = 0 part of this formula is the stochastic equivalent of the Gutzwiller semiclassical
(2)
trace formula [5]. The 2 correction wp; k is the stochastic analogue of Gaspard’s ˝
correction [12]. At the moment the explicit formula is suciently unenlightening that
we postpone writing it down to Section 6.
While the diagrams are standard, the chaotic eld theory calculations are considerably
more demanding than is usually the case in eld theory. Here there is no translational
invariance along the chain, so the vertex strength depends on the position, and the free
propagator is not diagonalized by a Fourier transform. Furthermore, here one is neither
“quantizing” around a trivial vacuum, nor a countable innity of analytically explicit
soliton saddles, but around an innity of nontrivial unstable hyperbolic saddles.
Two aspects of the above perturbative results are a priori far from obvious: (a)
that the structure of the periodic orbit theory should survive introduction of noise, and
(b) a more subtle and surprising result, repeats of prime cycles can be re-summed and
theory reduced to the dynamical zeta functions and spectral determinants of the same
form as for deterministic systems.
Pushing the Feynman–diagrammatic approach to higher orders is laborious, and has
not been attempted for this class of problems. As we shall now see, it is not smart to
keep pushing it, either, as one can compute many more orders of perturbatiion theory
by means of a matrix representation for L.

5. Evolution operator in a matrix representation

An expanding map f(x) takes an initial smooth distribution (x) dened on a subin-
terval, stretches it out and overlays it over a larger interval. Repetition of this process
smoothes the initial distribution (x), so it is natural to concentrate on smooth dis-
tributions n (x), and represent them by their Taylor series. By expanding both n (x)
and n+1 (y) in Eq. (6) in Taylor series Rugh [11] derived a matrix representation of
74 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

the evolution operator

Z
xm X ym
0

d x L(y; x) = Lm0 m ; m; m0 = 0; 1; 2; : : :
m! 0
m0 !
m
m
which maps the x component of the density of trajectories n (x) in Eq. (6) to the
0
ym component of the density n+1 (y) one time step later. The matrix elements follow
0 0
by dierentiating both sides with @m =@ym and evaluating the integral
0 Z

@m xm 
Lm0 m = d xL(y; x)  : (19)
@ym0 m! 
y=0

In Eq. (7), we have written the evolution operator L in terms of the Dirac delta
function in order to emphasize that in the weak-noise limit the stochastic trajectories
are concentrated along the classical trajectory y = f(x). Hence, it is natural to expand
the kernel in a Taylor series [17] in 
X∞ Z
(−)n (n)
L(y; x) = (y − f(x)) +  (y − f(x)) n P() d ; (20)
n!
n=2
(n) n n
where  (y) = (@ =@y )(y). This yields a representation of the evolution operator
centered along the classical trajectory, dominated by the deterministic Perron–Frobenius
operator (y − f(x)), with corrections givenR by derivatives of delta functions weighted
n
by moments of the noise distribution Pn = P() d. We again center the coordinate
system on the cycle points as in Eq. (11), and also introduce a notation for the operator
(7) centered on the xa → xa+1 segment of the classical trajectory
La (a+1 ; a ) = L(xa+1 + a+1 ; xa + a ) :
The weak-noise expansion (20) for the ath segment operator is given by
∞
X (−)n
La (a+1 ; a ) = (a+1 − fa (a )) + Pn (n) (a+1 − fa (a )) : (21)
n!
n=2

As the evolution operator has a simple -function form, the local matrix representation
of La centered on the xa → xa+1 segment of the deterministic trajectory can be
evaluated recursively in terms of derivatives of the map f:
∞
X (−)n
(La )m0 m = Pn (Ba )m0 +n; m ; n = max(m − m0 ; 0)
n
n!
Z
0 m
(Ba ) m0 m = d(m ) (a+1 − fa ())
m!
 m0 m 
1 d 1  
= 0  : (22)
|fa | d fa0 () m! 
=0

0
The matrix elements vanish for m ¡ m, so B is a lower triangular matrix. The diago-
nal and the successive o-diagonal matrix elements are easily evaluated iteratively by
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 75

computer algebra

1 (m + 2)!fa00
(Ba )mm = ; (Ba )m+1; m = − ;::: :
|fa0 |(fa0 )m 2m!|fa0 |(fa0 )m+2

For chaotic systems the map is expanding, |fa0 | ¿ 1. Hence, the diagonal terms drop
o exponentially, as 1=|fa0 |m+1 , the terms below the diagonal fall o even faster, and
truncating La to a nite matrix introduces only exponentially small errors.
The trace formula (8) takes now a matrix form

zL  X z np Lp
tr  = np tr ; (23)
1 − zL SC p
1 − z np Lp

where Lp = Lnp L2 : : : L1 is the contribution of the p-cycle. The subscript SC is a re-

minder that this is a saddlepoint or semiclassical approximation, valid as an asymptotic
series in the limit of weak noise. Vattay [18,19] interprets the local matrix representa-
tion of the evolution operator as follows. The matrix identity log det = tr log together
with the trace formula (23) yields
Y
det(1 − zL)|SC = det(1 − z np Lp ) ; (24)
p

so in the saddlepoint approximation the spectrum of the global evolution operator L

is pieced together from the local spectra computed cycle-by-cycle on neighborhoods of
individual prime cycles with periodic boundary conditions. The meaning of the kth term
in the trace formula (18) is now clear; it is the kth eigenvalue of the local evolution
operator restricted to the pth cycle neighborhood.
Using this matrix representation Palla and SHndergaard [14] were able to compute
corrections to order 12 , a feat simply impossible along the Feynman–diagrammatic
line of attack. In retrospect, the matrix representation method for solving the stochastic
evolution is eminently sensible – after all, that is the way one solves a close relative
to stochastic PDEs, the Schrodinger equation. What is new is that the problem is being
solved locally, periodic orbit by periodic orbit, by translation to coordinates intrinsic
to the periodic orbit. It is this natural local basis that makes the matrix representation
so simple.
Mainieri [15] takes this observation one step further; as the dynamics is nonlin-
ear, why not search for a nonlinear coordinate transformation that makes the intrinsic
coordinates as simple as possible?

6. Smooth conjugacies

This step injects into the eld theory a method standard in the construction of normal
forms for bifurcations [20]. The idea is to perform a smooth nonlinear coordinate
76 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

transformation x = h(y), f(x) = h(g(h−1 (x))) that attens out the vicinity of a xed
point and makes the map linear in an open neighborhood, f(x) → g(y) = J · y.

The key idea of attening the neighborhood of a saddlepoint can be traced back
to Poincare’s celestial mechanics, and is perhaps not something that a eld theorist
would instinctively hark to as a method of computing perturbative corrections. This
local rectication of a map can be implemented only for isolated nondegenerate xed
points (otherwise higher terms are required by the normal form expansion around the
point), and only in nite neighborhoods, as the conjugating functions in general have
nite radia of convergence.
We proceed in two steps. First, substitution of the weak noise perturbative expansion
of the evolution operator (21) into the trace centered on cycle c generates products of
derivatives of -functions:
Z
0
tr Ln |c = · · · + [d]{· · · (m ) (00 − fa (0 ))(m) (0 − fa−1 ()) · · ·} + · · · :

The integrals are evaluated as in Eq. (22), yielding recursive derivative formulas
such as
Z  m 
1 d 1 
(m)
d x (y) = 0 −  ; y = f(x) − x (25)
|y (x)| d x y0 (x) y=0

or n-point integrals, with derivatives distributed over n dierent -functions.

Next, we linearize the neighborhood of the ath cycle point. For a one-dimensional
map f(x) with a xed point f(0) = 0 of stability  = f0 (0), || = 6 1 we search for a
smooth conjugation h(x) such that

f(x) = h(h−1 (x)); h(0) = 0; h0 (0) = 1 : (26)

In higher dimensions,  is replaced by the Jacobian matrix J. For a periodic orbit

each point around the cycle has a dierently distorted neighborhood, with diering
second and higher derivatives, so the conjugation function ha has to be computed
point-by-point

fa () = ha+1 (fa0 h−1

a ()) :

An explicit expression for ha in terms of f is obtained by iterating around the whole

cycle, and using the chain rule (15) for the cycle stability p
n
fap () = ha (p h−1
a ()) ; (27)
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 77

so each ha is given by some combination of fa derivatives along the cycle. Expand

f(x) and h(x)
f(x) = x + x2 f2 + x3 f3 + · · · ; h(y) = y + y2 h2 + y3 h3 + · · · ;
and equate recursively coecients in the functional equation h(y)=f(h(y)) expansion
∞
X
h(u) − h(u) = fm (h(u))m : (28)
n=2

This yields the expansion for the conjugation function h in terms of the mapping f
f2 2f22 + ( − 1)f3
h2 = ; h3 = ;::: : (29)
( − 1) 2 ( − 1)(2 − 1)
The periodic orbit conjugating functions ha are obtained in the same way from
Eq. (27), with proviso that the cycle stability is not marginal, |p | =
6 1.
What is gained by replacing the perturbation expansion in terms of f(m) by still
messier perturbation expansion for the conjugacy function h? Once the neighborhood
of a xed point is linearized, the conjugation formula for the repeats of the map
fr (x) = h(r h−1 (x))
can be used to compute derivatives of a function composed with itself r times. The
expansion for arbitrary number of repeats depends on the conjugacy function h(x)
computed for a single repeat, and all the dependence on the repeat number is carried
by polynomials in r , a result that emerged as a surprise in the Feynman diagrammatic
approach of Section 4. The integrals such as Eq. (25) evaluated on the rth repeat of
prime cycle p
y(x) = fnp r (x) − x (30)
have a simple dependence on the conjugating function h
1 @2 1 r (1 + r ) 2
= (2h2 − h3 ) ;
3! @y2 y0 (0) (r − 1)3

1 @3 1 (r + 1)2 3 52r + 8r + 5

3 0
= −5r r 4
h2 + r h2 h3
4! @y y (0) ( − 1) (r − 1)4

2r + r + 1
− r h4 ;
(r − 1)4

··· = ··· : (31)

The evaluation of n-point integrals is more subtle [15]. The nal result of all these
calculations is that expressions of form (31) depend on the conjugation function de-
termined from the iterated map, with the saddlepoint approximation to the spectral
determinant given by
∞
YY
det(1 − zL )|SC = (1 − tp; k )
p k=0
78 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

in terms of local p-cycle eigenvalues

z np 2 (2) 3 (3) 4 (4) 6

tp; k = e( =2)P2 wp; k +( =3!)P3 wp; k +( =4!)P4 wp; k +O( ) ;
|p |kp
(2)
X (3)
2
wp; k = (k + 1) (2h2a; 2 − ha; 3 ); wp; k = ···;··· :
a

accurate up to order 4 . w(3) , w(4) are also computed by Dettmann, but we desist
from citing them here; the reader is referred to Ref. [15]. What is remarkable about
these results is their simplicity when expressed in terms of the conjugation function
h, as opposed to the Feynman diagram sums, in which each diagram contributes a
sum like the one in Eq. (17), or worse. Furthermore, both the conjugation and the
matrix approaches are easily automatized, as they require only recursive evaluation of
derivatives, as opposed to the handcrafted Feynman diagrammar.
Simple minded as they might seem, discrete stochastic processes are a great lab-
oratory for testing ideas that would otherwise be hard to test. Dettmann, Palla and
SHndergaard have used a one-dimensional repeller of bounded nonlinearity and com-
plete binary symbolic dynamics to check numerically the above results, and computed
the leading eigenvalue of L by no less than ve dierent methods. As anticipated by
Rugh [11], the evolution operator eigenvalues converge super-exponentially with the
cycle length; addition of cycles of period (n + 1) to the set of all cycles up to length
n doubles the number of signicant digits in the perturbative prediction. However, as
the series is asymptotic, for realistic values of the noise strength summations beyond
all orders are needed [21].

7. Summary

The periodic orbit theory approach to turbulence is to visualize turbulence as a se-

quence of near recurrences in a repertoire of unstable spatio-temporal patterns. The
investigations of the Kuramoto–Sivashinsky system discussed above are rst steps
in the direction of implementing this program. So far, existence of a hierarchy of
spatio-temporally periodic solutions of spatially extended nonlinear system has been
demonstrated, and the periodic orbit theory has been tested in evaluation of global
averages for such system. The parameter ranges tested so far probe the weakest non-
trivial “turbulence”, and it is an open question to what extent the approach remains
implementable as the system goes more turbulent.
The most important lesson of this investigation is that the unstable spatio-temporally
periodic solutions do explore systematically the repertoire of admissible spatio-temporal
patterns, with the trace and spectral determinants formulas and their cycle expansions
being the proper tools for extraction of quantitative predictions from the periodic orbits
data.
 P. CvitanoviÃc / Physica A 288 (2000) 61– 80 79

We formulate next a semiclassical perturbation theory for stochastic trace formulas

with support on innitely many chaotic saddles. The central object of the periodic
orbit theory, the trace of the evolution operator, is a discrete path integral, similar
to those found in the eld theory and statistical mechanics. The weak-noise pertur-
bation theory, likewise, resembles perturbative eld theory, and can be cast into the
standard eld-theoretic language of Feynman diagrams. However, we found out that
both the matrix and the nonlinear conjugacy perturbative methods are superior to the
standard approach. In contrast to previous perturbative expansions around vacua and
instanton solutions, the location and local properties of each saddlepoint must be found
numerically.
The key idea in the new formulation of perturbation theory is this: Instead of sepa-
rating the action into quadratic and “interaction” parts, one rst performs a nonlinear
eld transformation which turns the saddle point into an exact quadratic form. The
price one pays for this is the Jacobian of the nonlinear eld transformation – but it
turns out that the perturbation expansion of this Jacobian in terms of the conjugating
function is order-by-order more compact than the Feynman-diagrammatic expansion.

Acknowledgements

I am indebted to my collaborators C.P. Dettmann, G. Vattay, F. Christiansen,

V. Putkaradze, G. Palla, N.SHndergaard, R. Mainieri and H.H. Rugh for co-suering
through all the details omitted in this overview. I am grateful to E.A. Spiegel,
L. Tuckerman and M.J. Feigenbaum for patient instruction. I am not grateful to those
directors and gentlemen of committees who do not nd theoretical physics a vibrant
subject.

References

[1] V.W. Hughes, T. Kinoshita, Rev. Mod. Phys. 71 (1999) S133.

[2] R.S. Van Dyck Jr., P.B. Schwinberg, H.G. Dehmelt, Phys. Rev. Lett. 59 (1987) 26.
[3] K.G. Wilson, Phys. Rev. D 10 (1974) 2445.
[4] A.A. Belavin, A.M. Polyakov, A.S. Swartz, Yu.S. Tyupkin, Phys. Lett. B 59 (1975) 85.
[5] M.C. Gutzwiller, Chaos in Classical and Quantum Mechanics, Springer, New York, 1990.
[6] Y. Kuramoto, T. Tsuzuki, Progr. Theor. Phys. 55 (1976) 365.
[7] G.I. Sivashinsky, Acta Astr. 4 (1977) 1177.
[8] F. Christiansen, P. Cvitanovic, V. Putkaradze, Nonlinearity 10 (1997) 1.
[9] S.M. Zoldi, H.S. Greenside, Phys. Rev. E 57 (1998) R2511.
[10] P. Cvitanovic et al., Classical and Quantum Chaos, Niels Bohr Institute, Copenhagen, 2000;
www.nbi.dk=ChaosBook=.
[11] H.H. Rugh, Nonlinearity 5 (1992) 1237.
[12] P. Gaspard, D. Alonso, Phys. Rev. A 47 (1993) R3468.
[13] P. Cvitanovic, C.P. Dettmann, R. Mainieri, G. Vattay, J. Stat. Phys. 93 (1998) 981; chao-dyn=9807034.
[14] P. Cvitanovic, C.P. Dettmann, G. Palla, N. SHnderga rd, G. Vattay, Phys. Rev. E 60 (1999) 3936;
chao-dyn=9904027.
[15] P. Cvitanovic, C.P. Dettmann, R. Mainieri, G. Vattay, Nonlinearity 12 (1999) 939; chao-dyn=9811003.
[16] P. Cvitanovic, Field theory, Nordita, Copenhagen, 1983;www.nbi.dk=∼predrag=field the=.
80 P. CvitanoviÃc / Physica A 288 (2000) 61– 80

[17] S. Watanabe, Ann. of Prob. 15 (1987) 1.

[18] G. Vattay, P.E. Rosenqvist, Phys. Rev. Lett. 76 (1996) 335, chao-dyn=9509015.
[19] G. Vattay, Phys. Rev. Lett. 76 (1996) 1059.
[20] A. Katok, B. Hasselblatt, Introduction to the Modern Theory of Dynamical Systems, Cambridge
University Press, Cambridge, 1995.
[21] N. SHndergaard, G. Vattay, G. Palla, A. Voros, Asymptotics of high-order noise corrections, J. Stat.
Phys., to appear (chao-dyn=9911016).