Thermalization of Local Observables in the

$$\alpha $$ α -FPUT Chain

Santhosh Ganapa, Amit Apte &

Abhishek Dhar

Journal of Statistical Physics

ISSN 0022-4715
Volume 180
Combined 1-6

J Stat Phys (2020) 180:1010-1030

DOI 10.1007/s10955-020-02576-2

1 23
Your article is protected by copyright and
all rights are held exclusively by Springer
Science+Business Media, LLC, part of
Springer Nature. This e-offprint is for personal
use only and shall not be self-archived in
electronic repositories. If you wish to self-
archive your article, please use the accepted
manuscript version for posting on your own
website. You may further deposit the accepted
manuscript version in any repository,
provided it is only made publicly available 12
months after official publication or later and
provided acknowledgement is given to the
original source of publication and a link is
inserted to the published article on Springer's
website. The link must be accompanied by
the following text: "The final publication is
available at link.springer.com”.

1 23
 Author's personal copy
Journal of Statistical Physics (2020) 180:1010–1030
https://doi.org/10.1007/s10955-020-02576-2

Thermalization of Local Observables in the ˛-FPUT Chain

Santhosh Ganapa1 · Amit Apte1 · Abhishek Dhar1

Received: 9 November 2019 / Accepted: 22 May 2020 / Published online: 9 June 2020
© Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
Most studies on the problem of equilibration of the Fermi–Pasta–Ulam–Tsingou (FPUT)
system have focused on equipartition of energy being attained amongst the normal modes
of the corresponding harmonic system. In the present work, we instead discuss the equili-
bration problem in terms of local variables, and consider initial conditions corresponding to
spatially localized energy. We estimate the time-scales for equipartition of space localized
degrees of freedom and find significant differences with the times scales observed for normal
modes. Measuring thermalization in classical systems necessarily requires some averaging,
and this could involve one over initial conditions or over time or spatial averaging. Here we
consider averaging over initial conditions chosen from a narrow distribution in phase space.
We examine in detail the effect of the width of the initial phase space distribution, and of
integrability and chaos, on the time scales for thermalization. We show how thermalization
properties of the system, quantified by its equilibration time, defined in this work, can be
related to chaos, given by the maximal Lyapunov exponent. Somewhat surprisingly we also
find that the ensemble averaging can lead to thermalization of the integrable Toda chain,
though on much longer time scales.

Keywords Thermalization · Chaos · FPU problem · Toda chain · Equipartition

Dedicated to Joel Lebowitz, to thank him for being a constant source of new ideas, for explaining
Boltzmann, and for his warmth and kindness.

Communicated by David Ruelle.

B Santhosh Ganapa
santhosh.ganapa@icts.res.in
Amit Apte
apte@icts.res.in
Abhishek Dhar
abhishek.dhar@icts.res.in
1 International Centre for Theoretical Sciences, Tata Institute of Fundamental Research, Bengaluru
560089, India

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1011

1 Introduction

The Fermi–Pasta–Ulam–Tsingou (FPUT) problem [1,2] is a classic study with a long history.
The main aim of the FPUT study was to observe thermal equilibration in an isolated nonlinear
system with Hamiltonian dynamics. A chain of oscillators with weak non-linear coupling was
started in a highly atypical initial condition (all energy in a particular mode) and it was asked
if the system evolves at long times to a state with equipartition of energy amongst different
modes. To the surprise of the authors, they did not find such a state in the time scales of their
observations. Instead they found quasi-periodic behaviour and near-recurrences to the initial
state.
Since the original work of FPUT, there have been a number of studies aimed at under-
standing their results, which led to significant developments in statistical physics, nonlinear
dynamics and mathematical physics. We only refer to several of the review articles on the
FPUT problem [3–7]. Some of the recent studies that focus on the specific aspects of the
FPUT problem that are especially relevant to this paper include the role of breathers [8–
12], wave–wave interaction theory [13–15], and the role of breakup of invariant tori and the
stochastic threshold [16–19], and these are described in detail in Sect. 2.
The main aim of the present study is to understand two specific aspects of the equipartition
process: (i) the dependence on initial conditions, and (ii) the role of Lyapunov exponents,
which themselves depend on the initial conditions for a Hamiltonian system. We are mainly
motivated by the statement of the equipartition theorem and hence we focus on the averages
z i ∂ H /∂z i  where z i are phase space coordinates and H (z) is the Hamiltonian of the system.
The plan of the paper is as follows. In Sect. 2 we define the precise model, and describe
in some more detail the results of [13], along with some of the other recent works relevant to
this paper. We motivate and define the methods and objects of our study in Sect. 3. In Sect. 4
we present our numerical results and then in Sect. 5 we discuss the relation between these
results to chaos. We conclude with a discussion in Sect. 6.
In the rest of the introduction, we summarise some of the distinctive features of the
approach we take.

(i) Dependence of equilibration time, τeq , on initial conditions and observed variables—
while most studies look at initial conditions specified in normal mode space, we study
initial conditions with energy localized in real space. Similarly, most studies, beginning
with the first FPUT paper have looked at equipartition of energy among normal modes.
Here we consider equipartition of local variables, which would be especially relevant
when we consider strong nonlinearity, for which the normal mode picture becomes
invalid. Several earlier works have discussed this in the context of harmonic chains
[20–22].
(ii) A necessary step required in studying the question of equilibration is that one needs
some averaging process. Some commonly used protocols involve either an averaging
over initial conditions (chosen from a specified initial distribution) [13] or a temporal
averaging [7,9,23] or a spatial averaging [24,25]. Here we focus (see Sect. 3.1 for details)
on an averaging over initial conditions, and investigate the question of dependence of
the equilibration process on the width of the initial distribution.
(iii) Finally we explore the role of the Lyapunov exponents of the orbits in the process of
equilibration. For a non-linear system, it is expected that a perturbation of the initial
condition will grow exponentially for a non-integrable chaotic system and linearly for
an integrable system. Hence we try to quantify in a precise way how a distribution
of initial conditions expands over the full phase space and at sufficiently long times

123
 Author's personal copy
1012 Santhosh et al.

will reproduce the properties of the microcanonical equilibrium ensemble in not only
chaotic systems but also anharmonic integrable models.

2 Model and Review of Earlier Work

We consider N particles on a ring with displacements and momenta given by {qi , pi }, for
i = 0, 1, . . . , N − 1. The system is governed by the following Hamiltonian:
−1
 

N
pi2 μ(qi+1 − qi )2 α(qi+1 − qi )3
H ( p, q) = + + , (1)
2m 2 3
i=0

where we assume periodic boundary conditions q N ≡ q0 and q−1 ≡ q N −1 , and throughout

the paper we use m = μ = 1. The equations of motion are given by:
 
q̈i = (qi+1 + qi−1 − 2qi ) + α (qi+1 − qi )2 − (qi − qi−1 )2 , (2)

for i = 0, 1, . . . , N − 1. For α = 0, we have a harmonic chain and we can make a linear

change of variables to get an uncoupled set of N oscillators. For periodic boundary conditions,
we define the Fourier modes
N −1 N −1
1  1 
Qk = √ q j e−i2π k j/N , Pk = √ p j e−i2π k j/N . (3)
N j=0 N j=0

The Hamiltonian with α = 0 then takes the form


N −1  
|Pk |2 Ω 2 |Q k |2
H= E k , where E k = + k (4)
2 2
k=0

is the energy of each mode and Ωk = 2 sin(kπ/N ), k = 0, 1, 2, . . . , N − 1. These are the

normal modes of the system and since they do not interact, there is no exchange of energy
between them. The mode k = 0 corresponds to the motion of the centre of mass of the
system and we will always consider the case with E k=0 = 0 (total momentum zero). If the
system is let to evolve from an arbitrary initial condition {q j (0), p j (0)}, the energy of each
mode E k remains constant with time. For α  = 0, the normal modes of the harmonic chain
start interacting with each other and there is sharing of energy between the modes and the
expectation is that this leads to equipartition of energy. For sufficiently small nonlinearity,
the energy contribution from the nonlinear part of the interaction potential should be small
and it is a good approximation to assume that the total energy can still be represented
as a
sum of the energies of the independent oscillators, i.e, the total energy E ≈ E k . In that
case, one check of equipartition would be to see if all the E k (t) converge, at long times, to
the same value e = E/(N − 1) (perhaps with small fluctuations). This was the approach
in [1] (where however the fixed boundary condition case was studied). There, energy was
initially given to the first normal mode (E 1 = 0.08) and with α = 0.25, the time evolution
was studied numerically. Contrary to the expectations, the long-time dynamics appeared to
be almost periodic, with near perfect returns to the initial conditions.
A heuristic estimate of the strength of the nonlinearity can be obtained by comparing the
contribution of the nonlinear interaction part to the total energy. Roughly, if r is the average
spacing between particles, we expect μr 2 ∼ E/(N − 1) = e which gives a length scale

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1013

√
r ∼ e/μ. An estimate of the ratio of the nonlinear and harmonic energies is then given by
the parameter
αr 3 αe1/2
= = 3/2 . (5)
μr 2 μ
This dimensionless number, , and the system size, N , are the only relevant parameters. In
our subsequent discussions we assume that μ = 1 and one can change  = αe1/2 by either
changing the nonlinearity strength α or equivalently, by changing the energy density e. Note
that the cubic potential of the α-FPUT system implies that the system stays bounded only if the
total energy is sufficiently small and the precise condition is E < μ3 /(6α 2 ), corresponding
to all energy contributing to the potential energy of a single particle. In practice this is highly
improbable and one can work with energies slightly higher than this bound.
We summarize some of the attempts to explain the absence of thermalization seen by
FPUT.
1. Continuum limit and closeness to integrable PDEs Kruskal and Zabusky[26] showed that
the continuum limit of the model leads to the KdV equation which is an integrable model.
One might then expect that the long wavelength initial condition used by FPUT remains
close to the continuum description for long times and so one sees non-thermalization as
expected for integrable systems.
2. Stochasticity threshold Izrailev and Chirikov [16] proposed that there exists a stochas-
ticity threshold value of the energy density, ec , such that for e > ec (e = E/N ) one
gets thermalization. This idea was developed using the idea that nonlinearity leads to
broadening of the normal mode frequencies and there is resonance overlap when the
broadening is comparable to the separation between successive modes. The threshold
depends on initial conditions and the system size. For the harmonic chain, for small val-
ues of k(k N ), the separation between successive levels scales as Δω ∼ 1/N , while
for N − k N , Δω ∼ 1/N 2 . Hence the stochasticity threshold is larger for low fre-
quency modes. Since the FPUT study had initial conditions with the lowest mode excited
and the energy density was small, it is plausible that they were below the threshold.
This idea has been studied numerically [17–19] where an attempt was also made to relate
this to chaotic properties. In particular it was pointed out that for generic initial condi-
tions [17], the time evolution of the maximum time-dependent Lyapunov exponent λ(t),
is indistinguishable for the α-FPUT and the Toda systems up to a characteristic time
τtr (called the trapping time) that increase with decreasing energy (at fixed N). Beyond
this time, λ(t) of the α-FPUT system appears to approach a constant Λ, while it keeps
decreasing for the Toda system. This difference was attributed to the untrapping of the
FPUT system from its regular region in phase space and escape to the chaotic compo-
nent of its phase space. The authors in [17] computed τeq , τtr and Λ for system sizes
N = 32, 64, 128 at different energy densities and found the existence of a threshold
ec (N ) such that for e < ec , τeq , τtr seemed to diverge while Λ vanishes. The threshold
decreases with system size as ec (N ) ∼ 1/N 2 . Above ec , power law dependences of the
form τeq ≈ 1/e3 , τtr ≈ 1/e2.5 and λ ≈ e2 were noted. The FPUT parameters correspond
to e ec .
3. Role of breathers Breathers are time-periodic and space-localized solutions of nonlinear
dynamical systems [8]. By nature they are non-thermal and one might expect them to
play a role in preventing thermalization. In a certain sense the idea is similar to the one
relating the presence of solitons in the KdV system to the absence of equilibration in
the FPUT system—the difference being that breathers are stable solutions of the discrete
system while the KdV is a continuum approximation. The unexpected recurrences in

123
 Author's personal copy
1014 Santhosh et al.

the FPUT problem have been linked to the choice of initial conditions used by FPUT,
which are set close to exact coherent time-periodic (or even quasiperiodic) trajectories,
e.g., q-breathers, which show exponential localization of energy in normal mode space
[9–12].
4. Ideas from wave turbulence The FPUT problem has recently been studied [13–15] using
approaches of wave turbulence [27,28]. Based on requirement of resonance between sets
of normal modes, a detailed prediction has been made for the time-scale for equilibration
and it is argued that this is finite for any non-zero strength of non-linearity, for finite sized
systems. In particular, for the case of cubic nonlinearity of strength  in dimensionless
units, the estimated equilibration time scales as  −8 for N = 16, 32, 64. In the thermo-
dynamic limit, this is predicted to change to the form  −4 . The idea of the approach is to
connect the equilibration issue to the presence of high order resonances between dressed
normal modes that appear under repeated canonical transformations. The resonances are
expected to lead to the irreversible transfer of energy and hence thermalization. Their
work[13] showed that resonant triads (three wave resonances) are forbidden, they would
generate a reversible dynamics, which was originally observed by Fermi. They looked for
higher order interactions which are responsible for long term dynamics by a sequence of
canonical transformations. They found that four wave resonant interactions, though they
exist, are isolated from other quartets and cannot spread the energy across the spectrum.
The six wave resonant interactions are interconnected and are the lowest order interac-
tions that lead to an effective irreversible transfer of energy. In order to numerically verify
the predicted equilibration time, the authors in [13] took N = 32 particles on a ring, and
considered two sets of initial conditions. In one set, energy was given only to one normal
mode (E 1 , E 31  = 0), while in the other set, energy was given to five normal modes
(E i  = 0 for i = 1, . . . , 5 and i = 28, . . . , 32). Averages were taken over an ensemble
of 1000 initial conditions by introducing different random phases to each member. From
the time evolution of the ensemble averaged normal mode energies, the time to achieve
equipartition was estimated and for the α-FPUT model it was verified that the equili-
bration time scales as ∼ 1/ 8 . A more recent study [14] of the β-FPUT chain suggests
that wave–wave resonances lead to thermalization at small , where τeq ∼ 1/ 4 , while
at larger  the level broadening mechanism leads to τeq ∼ 1/. It was suggested that no
threshold exists.

In most of the literature, the analysis of equipartition was done for the energy of normal
modes of the corresponding integrable problem, that is, the harmonic chain. The analysis
is in some sense “global,” since normal modes involve all the particles in the chain. The
present work attempts to analyse the problem “locally,” by checking equipartition theorem
at different sites in the chain and tries to compute the time scale the system needs to reach
equipartition. We also investigate the role of the initial ensemble in determining equilibration.

3 Checking Thermalization in an Isolated System

3.1 Averaging Procedures

One can use various methods to check whether a system has reached thermal equilibrium.
It is clear that to check equilibration during the time evolution requires some kind of aver-
aging. The procedure we follow here involves an averaging over initial conditions, with the
expectation that such averages would represent the typical behaviour. We perform an average

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1015

over the initial conditions (q0 , p0 ) which are now chosen from a narrow distribution centred
around a specified point and that are still on the microcanonical surface of constant energy
E, momentum P and number of particles N . Denoting the initial distribution by ρ I (q0 , p0 ),
we then obtain the following average for any observable A = A(q, p):

A(t) = dq0 dp0 A(q(t), p(t))ρ I (q0 , p0 ) . (6)

We then ask if this reproduces the expected equilibrium results which would be obtained
from the equilibrium ensemble corresponding to the macroscopic conserved variables, e.g.
total energy and number of particles. In particular we can ask if equipartition is achieved at
long times, and the time to do so.
In the present work, we focus on this protocol, to explore the issue of equilibration times
in the FPUT problem. Some interesting questions include the dependence of the equilibration
process on the “width” of the initial distribution, and the fluctuations in the measured values.
We explore some of these questions.
Choice of Initial Distribution Here we consider the case where the initial distribution
lies on the constant (E, P, N ) surface in 2N dimensional phase space and has a small spread,
with the size of the spread characterized by a dimensionless number γ . The initial condition
is also chosen to correspond to the initial energy of the chain being localized in a small region.
Thus we set qi = 0, for i = 1, 2, . . . , N and pi = 0 for i = 5, 6 . . . , N − 1, N . The total
energy E is then distributed amongst the four remaining particles in the following way

E 1 = E 2 = (1 − γ )E/4 + νγ E/2, (7)

E 3 = E 4 = (1 − γ )E/4 + (1 − ν)γ E/2, (8)

where 0 < γ < 1 is a number which specifies the “width” of the distribution and ν is a
uniformly distributed random number in the interval (0, 1) which basically gives us some
randomness in the initial conditions. Here we consider initial conditions that have zero
momentum. The first and the second particles are given velocities in opposite directions
as are the third and the fourth. In our simulations we generate R initial configurations from
this distribution and evolve the system with the Hamiltonian dynamics. The ensemble average
of a physical observable A(q, p) is estimated as
R
Ar (t)
A(t) = r =1 ,
R
where the sum is over the R members of the ensemble.
Other averaging protocols commonly used in FPUT studies is temporal coarse graining.
In this case one starts with a fixed initial condition and performs an average over time. In this
case, one can look at microscopic variables, e.g the kinetic energy of individual particles, and
ask whether the time averaged value corresponds to the expected equilibrium value. Starting
the time evolution of the system from an arbitrary initial condition (q0 , p0 ), with energy E,
we can define a time averaged quantity for the observable A = A(q, p) as

1 t
A(t) = ds A(q(s), p(s)) . (9)
t 0
For a non-integrable system we might expect that for generic initial conditions, at long times
we should get thermal equilibration or A(t → ∞) = A E . The time to reach the equilibrium
value should give a measure of equilibration time scales. In some studies an average is taken
over finite large windows of time, centred at different time instances.

123
 Author's personal copy
1016 Santhosh et al.

3.2 Equipartition and Estimation of Thermalization Timescale

A standard test of thermalization would be to check the equipartition theorem expected in

equilibrium systems. Let us recall the precise statement on equipartition. For a system in
thermal equilibrium, the following are true:
   
∂H ∂H
qi = pi = cδi j , (10)
∂q j eq ∂ p j eq

for all i, j = 0, 1, . . . , N − 1 and where ...eq represents an average over an equilibrium

ensemble and c is a constant equal to k B (∂ S/∂ E)−1 for the microcanonical ensemble and
equal to k B T for a canonical ensemble. In the rest of the paper we use the notation
   
1 ∂H ∂H
Ti  = pi , and Vi  = qi . (11)
2 ∂ pi ∂qi
For a weakly nonlinear system, we can neglect the nonlinear terms and then transform to
normal modes coordinates where the system looks like a collection of independent oscillators
as in Eq. (4). In this case the equipartition theorem gives

E k eq = E/N (12)

for the microcanonical ensemble and k B T for the canonical ensemble. We need to remove the
zero mode if we are considering periodic boundary conditions. It is important to remember
that Eq. (12) is an approximate form valid for weak nonlinearity while Eq. (10) is exact.
As a measure of the level of equipartition that is achieved, we define the following function
[19], which has been referred to in the literature as entropy:


N −1
S(t) = − f i (t) ln f i (t) (13)
i=0
−1
where f i (t) = νi (t)/ rN=0 νr (t) for i = 0, 1, . . . , N − 1, and νi could be either Ti  or
Vi  or E i , which correspond to monitoring the equipartition of kinetic energy, qi ∂∂qHi , or the
normal mode energy, respectively.1 The value of S(t) is bounded between 0, corresponding to
the highly nonequilibrium situation with all the energy in a single degree of freedom, and ln N ,
corresponding to the equilibrated system with equipartition between all degrees with { f i }
defining a uniform distribution over the set {0, 1, . . . , N − 1}. Since it can theoretically take
an infinite amount of time to reach ln N , we estimate the equilibration time as the time required
for S(t) to reach a predetermined value of entropy which is close to the equilibration value.
In this work, we consider the following criterion to determine the equilibration timescale:

S(t) − Smax
≤ 0.01. (14)
Smax

The above threshold must be satisfied for two consecutive values of the time that is sampled.
The minimum value of t for which the above criteria is satisfied is termed as equilibration
time (denoted by τeq ).

1 Note that for each fixed time t, the set { f (t)} defines a discrete probability distribution over the set
i
{0, 1, . . . , N − 1}, and then S(t) is just the information entropy.

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1017

   
Fig. 1 α-FPUT chain: Plot shows time evolution of T j (left panel) and V j (right panel) starting from space
localized initial conditions for j = 0, 1, 2, ...N − 1. Parameter values for this plot are N = 32, E = 31,
α = 0.0848 ( = 0.0834), γ = 0.9 and R = 1000. The solid black line corresponds to the equipartition value
T j  = 0.484 and V j  = 0.969. The inset shows the time evolution at the earliest times

4 Numerical Results

As noted in the introduction, our main objectives are to estimate the equilibration time
using observables other than the normal mode energies and to investigate the role of initial
conditions. In the following, we present results on equilibration from the initial ensemble
discussed in Sect. 3.1 corresponding to energy being initially localized in space. We shall
refer to these initial conditions as space localized excitations (SLE) as opposed to normal
mode localized excitations (NMLE) commonly used in most studies. Equipartition   will be
checked by monitoring the entropy S(t) for the local observables T j  and V j which are
defined in Eq. (11). These quantities for the α-FPUT system are compared with those of
the Toda chain and of the harmonic chain, to check the dependence of thermalization on
integrability. This is quantified further by
 the
 analysis
  of the dependence of the entropy S(t)
on time t. In addition, the entropy of V j and T j of the α-FPUT problem are compared
with E k  to check if equilibration depends on whether one is verifying equipartition using
the normal mode or position-momentum coordinates. The scaling laws in the two methods
are compared by plotting the equilibration times as a function of the nonlinear dimensionless
parameter . The dependence of the equilibration time on the width of the initial distribution
in the phase space is also analysed.
In Sect. 5 an attempt is made to relate the thermalization properties of the α-FPUT system
(equilibration time) to the growth of perturbations of initial conditions in the system. For
chaotic systems this growth is exponential and is quantified by the maximal Lyapunov expo-
nent, while for anharmonic integrable systems, the growth is linear. We make comparisons
of the FPUT results with those of the corresponding Toda system.

Simulation Details For most of our numerics, we used a sixth order symplectic integrator
described in [29]. The time-step size was taken as 0.01 and we checked the relative energy
change in the system at the end of the computation to be of the order of 10−11 . Computations
of the Lyapunov exponent were performed by solving the coupled system of 2N + 2N
nonlinear and linearized equations using a fourth order Runge-Kutta integrator. In this case
the time-step size was taken as 0.001 and the relative energy change in the system at the end
of the computation is found to be around 10−9 .

123
 Author's personal copy
1018 Santhosh et al.

Fig. 2 α-FPUT chain: Same as Fig. 1 but with γ = 10−8 . Note that the range of the time t and the averages
T j  and V j  are different from those in Fig. 1

 
Fig. 3 α-FPUT chain: spatial profile of T j at different times for γ = 0.9 (left panel) and γ = 10−8 (right
panel) and other parameters N = 32, E = 31, α = 0.0848 and R = 1000. We see that the initially localized
energy quickly spreads through the chain while equipartition is achieved at much longer time scales

4.1 Evolution of Local Observables from Space Localized Initial Conditions in FPUT
Chain

The space localized excitations (SLE) are parameterized by the variable γ whose magnitude
gives an estimate of the width of the initial distribution in phase space. We note that γ = 0
 γ = 1 corresponds
corresponds to a fixed initial state while  to the broadest distribution. In
Fig. 1 we show the time evolution of T j and V j for γ = 0.9. We see that there is a long
transient period and then we see equipartition at times ∼ 2 × 104 . At the earliest times, the
inset in Fig. 1 shows near-recurrent behaviour. The results for γ = 10−8 are plotted in Fig. 2,
where we now see that equipartition is achieved at somewhat longer times, around t ∼ 105 .
In Fig. 3 we plot the averaged kinetic energy profile at different times. It is seen that the
energy spreads quickly through the entire system, while equipartition is achieved at much
longer time scales.
To demonstrate that the averaging procedure is crucial to the equilibration process in this
set-up, we show the evolution of T j and V j for a single initial condition. In this case, we
see in Fig. 4 that no equilibration is achieved and the oscillatory behaviour persists up to
t = 5 × 105 .

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1019

Fig. 4 α-FPUT chain: Plot shows time evolution of T j (left panel) and V j (right panel) starting from a single
initial condition. Parameter values for this plot are N = 32, E = 31, α = 0.0848 ( = 0.0834). The insets
show zoom-ins of the time evolution at early times and at late times and we see oscillatory behaviour in both
cases. Thus in this case, no equipartition is achieved

 
Fig. 5 α-FPUT chain: Here we examine how the fluctuations seen in T j depend on the number of realizations
 
R. We plot T j for R = 102 (left panel) and R = 104 (right panel). Other parameters were taken as N = 32,
E = 31, α = 0.0848, γ = 0.9. The insets show zoom-ins at short and long times. A vertical line is drawn at
t = 1.2 × 104 , which is the equilibration time. Up to this time, both the plots look nearly the same

An examination of the plots in Figs. 1 and 2 shows that even at late times, the averaged
quantities continue to fluctuate around their equilibrium values. We show in Fig. 5 that these
fluctuations in fact decrease with
 increase in the number of realizations R used to compute
averages. We also see that T j at pre-thermalization times does not depend significantly on
R and the plots are nearly identical for R = 102 and R = 104 (up to the vertical line in
Fig. 5).

4.2 Normal Mode Localized Initial Conditions

We now discuss and compare our results with those obtained in studies on equipartition of
normal mode energies using initial conditions which were investigated in [13]. The authors in
[13] considered initial conditions where the energy was distributed between the modes k = 1
and k = 31 with frequencies Ω1 = Ω31 . Averages were done over 1000 initial conditions
by choosing random phases for the modes. The time evolution of the normal mode energies

123
 Author's personal copy
1020 Santhosh et al.

Fig. 6 α-FPUT chain: left panel shows time evolution of E k  for N = 32, E = 31, α = 0.0848 and
R = 1000. Only modes k = 1, 31 are initially excited. The inset shows quasiperiodic behaviour at short time
scale. Right panel shows the mode energy profile at different times

Fig. 7 α-FPUT chain: plot shows

time evolution of E k starting
from a single initial condition
with energy in two normal
modes. Parameter values for this
plot are N = 32, E = 31,
α = 0.0848. The insets show
zoom-ins of the time evolution at
early times and at late times and
we see oscillatory behaviour in
both cases. Thus in this case, no
equipartition is achieved

was monitored to check for equipartition. Here we reproduce their numerical results and
compare with the results in the previous section. We consider again N = 32 particles with
total energy E = 31. In Fig. 6 we show the time evolution of the energy of all the modes in
the system. Comparing with Figs. 1 and 2 it is clear that equilibration now occurs at a time
scale (∼ 2 × 106 ) that is about an order of magnitude longer. In Fig. 7 we see again that one
does not see any signs of equilibration in the evolution of a single realization.

4.3 Comparison with Temporal Averaging Protocol

As discussed in Sect. 3.1 one can discuss thermalization using a different protocol where one
starts with a single initial condition and then considers a time average of any given observable.
This is given by Eq. (9). In Fig. 8 we show results obtained using this protocol for both the
space-local and normal mode observables. The insets in the figures show that thermalization
time scales are completely different from those obtained by the ensemble averaging protocol.

4.4 Estimation of Equilibration Time from Entropy

To get a more systematic and quantitative estimate of the equilibration time we now look at
the entropy function defined in Eq. (13). This in some sense, performs an average over all the
degrees of freedom and attains its maximum value ln N when all degrees have equilibrated. In

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1021

Fig. 8 α-FPUT chain: plot shows time evolution of running time averages of T j for SLE (left panel) and E k for
NMLE (right panel) starting from a single initial condition. Parameter values for this plot are N = 32, E = 31,
α = 0.0848 ( = 0.0834). The insets show comparison between time averages and ensemble averages and
illustrates that the latter procedure leads to faster thermalization

 
Fig. 9 α-FPUT chain: plot of S S L (t) corresponding to T j for γ = 0.9 (left panel) and γ = 10−8 (right panel)
with other parameters N = 32, E = 31, α = 0.0848. The insets show the convergence of the equilibration
process as the number of realizations is increased

Fig. 9, we plot the evolution of entropy for the two parameter values γ = 0.9 and γ = 10−8 .
The insets show zoom-ins near the equilibrium value ln N , showing the approach to equili-
bration and its dependence on the number of realizations. For higher number of realizations,
the fluctuations in the entropy is lower and also the mean is closer to the equilibrium value.
We use the criterion of Eq. (14) to estimate the equilibration time from the entropy S S L
corresponding to T j  and find τeq ≈ 100300 and 12500 for γ = 10−8 and 0.9 respectively.
In general we find that as the “width” of the distribution γ is increased, thermalization is
found to happen faster. We will discuss this again in Sect. 5. We compute the equilibration
time for different values of the dimensionless parameter  for N = 32. These results are
plotted in Fig. 10 where we find a power-law dependence of equilibration time, τeq , on  of
the form
1
τeq ∝ a , (15)


The value a is found to depend on γ and lies between 4 and 6. This is significantly different
from the form 1/ 8 obtained in [13], by considering equilibration of normal modes. In Fig. 10
we also indicate the relaxation time results for the normal modes which give a ≈ 7.7.

123
 Author's personal copy
1022 Santhosh et al.

Fig. 10 α-FPUT chain:

 graph
showing τeq of T j for
N = 32, E = 31 as a function of
. The slopes of the fitting lines
give a = 5.9, 6.0, 4.0 for
γ = 10−8 , 0.01 and 0.9
respectively. We also show the
equilibration times obtained from
the normal mode entropy
function S N M L for normal mode
localized initial conditions, which
leads to an exponent a ≈ 7.7

Fig. 11 α-FPUT chain: graphs showing the evolution of the entropy corresponding to space-local observables
and normal modes, for γ = 0.9 (left panel) and γ = 10−8 (right panel), with other parameters given by
N = 32, E = 31, α = 0.0848. Results for both space localized initial conditions (SLE) and normal mode
localized initial conditions (NMLE) are shown

It is to be expected that the equilibration time scale should depend not only on the initial
ensemble in which the system is prepared, but also on the observable for which equipartition
is being tested. We investigate this question further by computing the entropy functions S S L
and S N M L for both types of initial conditions, namely space localized (SLE) and normal
mode localized (NMLE). These results have been plotted in Fig. 11. We see clearly that the
relaxation of normal mode coordinates is slower than that of the space localized observables,
irrespective of initial distribution.

4.5 Comparison with Harmonic Chain and Toda Chain

To investigate the role of integrability, we now repeat the above computations in two integrable
models that are related to the FPUT system in the limit of weak nonlinearity. We consider
the harmonic chain which is described by the Hamiltonian in Eq. (1), with α = 0, and the
Toda chain, described by the Hamiltonian

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1023

 
Fig. 12 Harmonic chain: plot of T j as a function of time, for N = 32, E = 31, R = 1000 and γ = 10−8
(left panel) and γ = 0.9 (right panel). In both cases there is no sign of thermalization

 
Fig. 13 Toda chain: plot of T j as a function of time, for N = 32, E = 31, α = 0.0848, b = 2α, g =
b−1 , R = 1000 and γ = 10−8 (left panel) and γ = 0.9 (right panel). Now we observe thermalization in the
right panel

−1
 

N
pi2 g
H ( p, q) = + eb(qi+1 −qi ) . (16)
2 b
i=0

The Toda system is known to be integrable [30,31] and has been much studied as the integrable
limit of the FPUT chain [7,17,32,33]. The parameter choice b = 2α and g = b−1 would
then approximate the α-FPUT potential to leading nonlinearity. Starting with the same space-
localized initial conditions as in the previous sections, we now check equipartition of kinetic
energy Ti . In Fig. 12 we see that no equilibration is achieved for the harmonic chain. On the
other hand, somewhat surprisingly, we see in Fig. 13 that the Toda chain does equilibrate,
provided we start with a wider initial distribution (γ = 0.9).
In Fig. 14 we plot the entropy function S S L and using the criterion in Eq. (14), estimate
the equilibration time and find τeq ≈ 66000 for γ = 0.9. The dependence of the equilibration
time (τeq ) on γ will be discussed quantitatively in the next section.

123
 Author's personal copy
1024 Santhosh et al.

 
Fig. 14 Toda chain: plot of the entropy S S L (t) for T j of the Toda chain with same parameters as Fig. 13 and
−8
γ = 10 (left panel) and γ = 0.9 (right panel). The second case shows clear equilibration

Fig. 15 The time-evolution of the ensemble of initial conditions obtained using Eqs.(7)–(8) is shown for the
FPUT chain with α = 0.0848 (left column), the corresponding Toda chain (middle column), and the harmonic
chain (right column), in the ( p13 , q13 ) and ( p0 , q0 ) planes (top and bottom rows, respectively). The different
lines for harmonic chain are at increased times, with increasing marker size

5 Relation to Chaos

We will now argue that the process of equilibration in the FPUT protocol is intimately
connected to the growth of perturbations of initial condition, and hence to chaos. This idea then
also explains the absence of equilibration in the harmonic chain and the slow equilibration in
the Toda chain. We first quantify the growth of perturbations in the system. Let us consider an
infinitesimal perturbation {δqi (0), δ pi (0)} of an initial condition {qi (0), pi (0)}. We compute
the quantity

N  
δqi2 (t) + δ pi2 (t)
Z (t) = Ni=1  . (17)
i=1 δqi2 (0) + δ pi2 (0)

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1025

We then compute the ensemble averaged time-dependent Lyapunov exponent λ(t) defined
as
1
λ(t) = ln Z (t) , (18)
2t
where ... denotes an average over initial conditions {qi (0), pi (0)} chosen from the distribu-
tion ρ0 (q, p). As described earlier, the numerical integration of {qi (t), pi (t), δqi (t), δ pi (t)}
is done by solving 2N + 2N nonlinear and linearized equations. The largest Lyapunov expo-
nent Λ is then given by Λ = limt→∞ λ(t). For a harmonic chain, transforming to normal
modes shows that Z (t) is bounded. For a Toda chain, a transformation to action angle vari-
ables suggests that Z (t) should grow linearly with time. This is consistent with Λ = 0 for
both the harmonic chain and the Toda chain.
The basic picture that illustrates the difference between the three models is shown in
Fig. 15, where we show the time-evolution of the ensemble of initial conditions obtained
using Eqs. (7)–(8). The FPUT chain shows a fast growth in phase space because of its
positive Lyapunov exponent, while it takes much longer for the Toda chain because of its
integrability (and linear temporal growth of perturbations). There is no spread in the harmonic
chain.
In Fig. 16 we plot log Z (t) for the FPUT chain as well for the corresponding harmonic
chain and Toda chain (with b = 2α, g = b−1 ). We confirm the expected exponential growth
of Z (t) for the FPUT chain at large times, it’s linear growth for the Toda chain and the lack
of growth in the case of harmonic chain. In the inset of the right panel, we also show the line
log Z (t) = 10.3, which corresponds to the equilibration time of the α-FPUT chain. It turns
out that the point of intersection of this horizontal line with the Toda chain is very close to
its equilibration time. This gives us a means to relate the thermalization properties of both
the systems to the growth of their perturbations.
In order to explore a possible relation between the equilibration time τeq and the maximal
Lyapunov exponent Λ, we show, in Fig. 17, the dependence of τeq and Λ−1 on the nonlinearity
parameter  for γ = 0.9 and γ = 10−8 . The slopes are close to each other, suggesting a
possible relation between the equilibration time τeq and the Lyapunov exponent Λ. We
propose that the close relation between the scaling of τeq and Λ−1 with respect to  can
be understood through the following argument which relates the growth of perturbations of
initial conditions to thermalization.
 
N 1/2
Let us define D(t) = i=1 Δqi (t) + Δpi (t)
2 2 to be the distance at time t, of points
that are initially separated by a small but finite separation D(0), due to a finite perturbation
of initial conditions. In particular, we will interpret D(t) as the “spread” of an ensemble of
trajectories, for example, using the ensemble of initial conditions as described in Eqs. (7)–(8).
At early times we expect that the growth can be described by

D(0)eΛt for FPUT,
D(t) ∼ (19)
D(0)t for Toda.
However {Δq(t), Δp(t)} cannot grow forever since {q(t), p(t)} are constrained to be on the
constant energy surface and are themselves bounded. Let us define a √
“phase-space covering”
time scale τph as the time at which each Δq j , Δp j become of order E/N and D becomes
√
of order 2E. Then we have
⎧ √ 
⎨Λ−1 ln 2E for FPUT,
τph ∼  √2E 
D(0)
(20)
⎩ for Toda.
D(0)

123
 Author's personal copy
1026 Santhosh et al.

Fig. 16 Left panel shows Z (t) of α-FPUT chain, Toda chain and the harmonic chain at short times. It can be
seen that the growth of perturbations is bounded for the harmonic chain and linear for the Toda chain. Right
panel shows Z (t) of α-FPUT chain and the Toda chain for long times. The growth of perturbations in the
α-FPUT chain though linear at short times, becomes exponential at long times. It can be seen that the growth
of perturbations is linear for the FPU chain and logarithmic for the Toda chain in the logarithmic scale. A
horizontal line drawn at log Z (t) = 10.3 intersects both the curves. Its significance is explained in the text

Fig. 17 α-FPUT chain: graph

showing equilibration time τeq ,
the phase-space covering time
scale τ ph for γ = 0.9 and
γ = 10−8 and the inverse Λ−1 of
the maximal
 Lyapunov exponent
of T j as a function of α. Other
parameters N = 32, E = 31 and
R = 1000. The slopes of the five
lines (from top to bottom) are
−5.9, −4.7, −4.0, −4.7, −4.7

The initial width D(0) should be proportional to γ , which characterizes the width of our
initial phase-space distribution. The above arguments should work better for smaller γ . One
expects that τeq ∼ τph and we now present some numerical results that support this. From
Eq. (20) we see that τph and hence τeq should scale with γ as ln(γ ) for the FPUT chain and
as 1/γ for the Toda chain. In Fig. 18a we see the logarithmic dependence of τeq on γ for
the FPUT chain, with α = 0.0848. In Fig. 18b we show the dependence of τeq on γ for the
Toda chain, again with α = 0.0848. The slope on a log-log plot is close to 1, supporting the
expectation τeq ∼ 1/γ . Next in Fig. 17 we compare, for the FPUT system, the dependence
of τph and τeq on  for γ = 10−8 , 0.9. For γ = 10−8 we find τeq ∼ 1/ 5.9 and τph ∼ 1/ 4.7 .
Thus we see a reasonable level of agreement though this is not perfect.
The system is translationally invariant. So, the results are identical if we initially distribute
the energy to a different set of four successive particles, while maintaining the order of the
initial distribution. If the energies are distributed in a different permutation we find that, while
the precise equilibration times are different, the dependence of τeq on  and γ are still the
same.
In Fig. 18a, the solid line describes an ensemble of initial conditions described by Eqs. (7)–
(8), referred to as zero volume ensemble, since the latter occupies zero volume in the phase

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1027

 
Fig. 18 Graphs showing the dependence of equilibration time of T j on γ for the α-FPUT chain (left panel)
and the Toda chain for α = 0.0848 (right panel). Other parameters are N = 32, E = 31, and R = 1000. For
the α-FPUT chain, plotted on a linear-log scale, there are two different ensembles. The solid line represents
an ensemble of initial conditions occupying zero volume in the phase space. It is described by Eqs. (7)–(8).
The slope of this line is −4.5 × 103 . The dashed line represents an ensemble of initial conditions occupying
a finite volume in the phase space. The slope of this line is −3.0 × 103 . The slope for the Toda chain, plotted
on a log-log scale, is −0.81

space. The magnitude of the slope of this line on a semilog plot is found to be 4.5 × 103 . Its
inverse is 2.2 × 10−4 , close to the Lyapunov exponent of the system Λ ≈ 4.1 × 10−4 . We
have also studied an ensemble of initial conditions that has randomness in all the degrees of
freedom (while maintaining momentum conservation), thereby occupying a finite volume in
the phase space, quantified by the number γ . The results for this are shown by the dashed
line in Fig. 18a where again we see the logarithmic dependence on γ and in fact find a closer
agreement between the magnitude of the slope (the inverse of which is 3.3 × 10−4 ) and
the Lyapunov exponent, thus making a stronger point for our claim regarding the relation
between the equilibration of local observables and sensitive dependence of the system on
initial conditions.
We believe that the properties of the initial conditions, such as its symmetries and its
vicinity from breather solutions can affect the exponential growth of the perturbations of
the initial conditions, which would lead to a retardation of thermalization. Far away from
breathers we can expect there is no such effect. This needs to be investigated further and is
beyond the scope of this work. Nevertheless, this method gives us a way to link chaos and
thermalization in the α-FPUT system. We have also verified this relation for the β-FPUT
system.

6 Conclusions

We studied the time scale of thermalization of local variables in the α-FPUT chain and its
two limiting integrable versions, namely the harmonic chain and the Toda chain. Considering
systems with N = 32 particles and total energy E = 31, we estimated the thermalization time
τeq by measuring z i ∂ H /∂z i  (z i indicating phase space coordinates) and finding the time to
attain equipartition. The averaging is done over initial conditions chosen from a distribution
whose width is characterized by the parameter 0 ≤ γ ≤ 1, with γ = 1 corresponding to
the broadest distribution and γ = 0 corresponding to a fixed initial condition. The initial
distribution is taken to be one where energy is localized initially in real space instead of normal

123
 Author's personal copy
1028 Santhosh et al.

mode space. The system is described by a single dimensionless parameter  = α(E/N )1/2
characterizing the effective nonlinearity. Some of our main findings are as follows:

(i) For the α-FPUT chain we find τeq ∝ 1/ a with a between 4 and 6, and γ dependent,
in contrast to normal mode equilibration times [13], where one finds a ≈ 8.
(ii) The thermalization time depends on the initial ensemble and we find τeq ∝ ln(γ ),
with the proportionality constant being close to the inverse of the maximal Lyapunov
exponent of the system, thus quantifying the relation between thermalization and chaos
for the α-FPUT system.
(iii) We find that local variables equilibrate at much shorter time scales for the normal mode
localized initial conditions (NMLE) than phase space localized initial conditions (SLE).
(iv) Surprisingly, we find that the Toda chain equilibrated on very long time scales if the
width of the initial distribution is broad enough. In fact we obtain τeq ∼ 1/γ . On the
other hand, the harmonic chain never equilibrates.
(v) We provide a simple geometric understanding of these results—the equilibration time
is simply related to the time it takes for an ensemble of initial conditions in the 2N
dimensional phase space to spread over the microcanonical energy surface. For the
FPUT chain, for energies such that the system is chaotic with a positive Lyapunov
exponent, a fast exponential (in time) spreading occurs. For the Toda chain the growth
is linear and so thermalization takes more time. We provide numerical evidence to
support this picture.

Thermalization in finite Hamiltonian systems has usually been studied by either consider-
ing a time averaging protocol or an ensemble averaging protocol. As we illustrate, they can
lead to very different estimates for the time scale of equilibration. In our example, the time
averaging protocol gives a thermalization time that is several orders of magnitude larger than
that obtained from ensemble averaging. We believe that the ensemble averaging protocol is
relevant for understanding aspects of the classical-quantum correspondence in the context of
thermalization in finite systems. For quantum systems several studies show, e.g [34], that a
finite quantum system prepared in a pure initial state and evolving under unitary dynamics
can exhibit thermalization (without requiring any time averaging). A corresponding state-
ment for the classical system is difficult if (corresponding to the quantum pure state) one
considers a single point in phase space. However if we smear out the point into an initial blob
in phase space, as is done in our study, then we can get an equivalent classical statement.
The smearing out process can be thought of as arising from the uncertainty principle. As an
example we point out recent work [35] on the quantum-classical correspondence in Floquet
systems where such an averaging protocol is essential.
As a concluding remark, we note that another important question is that of thermalization
of macroscopic systems. There we expect that thermalization requires neither a temporal
or an ensemble averaging protocol, but arises from the fact that physical observables are
macro-variables and their measurement typically involves an averaging over many degrees
of freedom [24,25].

Acknowledgements SG would like to thank Hemanta Kumar for help with computations and Varun Dubey
for useful discussions. The numerical simulations were done on Contra, Mario and Tetris computing clusters
of ICTS-TIFR.

123
 Author's personal copy
Thermalization of Local Observables in the α-FPUT Chain 1029

References
1. Fermi, E., Pasta, J., Ulam, S.: Studies of the nonlinear problems, 1955. Los Alamos National Laboratory
Report (LA1940), also in Collected Papers of Enrico Fermi 2
2. Dauxois, T.: Fermi, Pasta, Ulam, and a mysterious lady. Phys. Today 61(1), 55 (2008)
3. Ford, J.: The Fermi-Pasta-Ulam problem: Paradox turns discovery. Phys. Rep. 213(5), 271–310 (1992)
4. Weissert, T.P.: The Genesis of Simulation in Dynamics: Pursuing the Fermi-Pasta-Ulam Problem.
Springer, New York (2012)
5. Berman, G.P., Izrailev, F.M.: The Fermi-Pasta-Ulam problem: fifty years of progress. Chaos 15(1), 15104
(2005)
6. Gallavotti, G.: Introduction to FPU. In: Gallavotti, Giovanni (ed.) The Fermi-Pasta-Ulam Problem: A
Status Report. Springer, Berlin (2008)
7. Benettin, G., Christodoulidi, H., Ponno, A.: The Fermi-Pasta-Ulam problem and its underlying integrable
dynamics. J. Stat. Phys. 152, 195–212 (2013)
8. Marin, J., Aubry, S.: Breathers in nonlinear lattices: numerical calculation from the anticontinuous limit.
Nonlinearity 9, 1501–1528 (1996)
9. Danieli, C., Campbell, D.K., Flach, S.: Intermittent many-body dynamics at equilibrium. Phys. Rev. E
95(6), 060202 (2017)
10. Flach, S., Ivanchenko, M.V., Kanakov, O.I.: q-Breathers and the Fermi-Pasta-Ulam problem. Phys. Rev.
Lett. 95(6), 064102 (2005)
11. Flach, S., Ivanchenko, M.V., Kanakov, O.I.: q-breathers in Fermi-Pasta-Ulam chains: existence, localiza-
tion, and stability. Phys. Rev. E 73(3), 036618 (2006)
12. Christodoulidi, H., Efthymiopoulos, C., Bountis, T.: Energy localization on q-tori, long-term stability,
and the interpretation of Fermi-Pasta-Ulam recurrences. Phys. Rev. E 81(1), 016210 (2010)
13. Onorato, M., Vozella, L., Proment, D., Lvov, Y.V.: Route to thermalization in the α-Fermi–Pasta–Ulam
system. PNAS 112(14), 4208–4213 (2015)
14. Lvov, Y.V., Onorato, M.: Double scaling in the relaxation time in the β-Fermi-Pasta-Ulam-Tsingou model.
Phys. Rev. Lett. 120, 144301 (2018). Apr
15. Pistone, L., Chibbaro, S., Bustamante, M., L’vov, Y., Onorato, M.: Universal route to thermalization in
weakly-nonlinear one-dimensional chains. Math. Eng. 1(4), 672 (2019)
16. Izrailev, F.M., Chirikov, B.V.: Statistical properties of a nonlinear string. Dokl. Akad. Nauk SSSR 166(1),
57–59 (1966)
17. Casetti, L., Cerruti-Sola, M., Pettini, M., Cohen, E.: The Fermi-Pasta-Ulam problem revisited: stochas-
ticity thresholds in nonlinear Hamiltonian systems. Phys. Rev. E 55(6), 6566 (1996)
18. Deluca, J., Lichtenberg, A.J., Ruffo, S.: Energy transitions and time scales to equipartition in the Fermi-
Pasta-Ulam oscillator chain. Phys. Rev. E 51(4), 2877–2885 (1995)
19. Livi, R., Pettini, M., Ruffo, S., Sparpaglione, M., Vulpiani, A.: Equipartition threshold in nonlinear large
Hamiltonian systems: the Fermi-Pasta-Ulam model. Phys. Rev. A 31(2), 1039–1045 (1985)
20. Mazur, P., Montroll, E.: Poincaré cycles, ergodicity, and irreversibility in assemblies of coupled harmonic
oscillators. J. Math. Phys. 1(1), 70–84 (1960)
21. Titulaer, E.M.: Ergodic features of harmonic-oscillator systems. I. Physica 70(2), 257–275 (1973)
22. Titulaer, U.M.: Ergodic features of harmonic-oscillator systems. II. asymptotic ergodicity. Physica 70(2),
276–296 (1973)
23. Benettin, G., Pasquali, S., Ponno, A.: The Fermi-Pasta-Ulam problem and its underlying integrable dynam-
ics: an approach through Lyapunov exponents. J. Stat. Phys. 171(4), 521–542 (2018). Mar
24. Lebowitz, J.L.: Statistical mechanics: a selective review of two central issues. Rev. Mod. Phys. 71(2),
S346–S357 (1999). Mar
25. Goldstein, S., Huse, D.A., Lebowitz, J.L., Tumulka, R.: Macroscopic and microscopic thermal equilib-
rium. Annalen der Physik 529(7), 1600301 (2017). Feb
26. Zabusky, N.J., Kruskal, M.D.: Interaction of “Solitons” in a collisionless plasma and the recurrence of
initial states. Phys. Rev. Lett. 15, 240–243 (1965)
27. Zakharov, V.E., L’vov, V.S., Falkovich, G.: Statistical description of weak wave turbulence. In: Kol-
mogorov Spectra of Turbulence I: Wave Turbulence, pp. 63–82. Springer, Berlin (1992)
28. Nazarenko, S.: Wave Turbulence. Springer, Berlin (2011)
29. Yoshida, H.: Construction of higher order symplectic integrators. Phys. Lett. A 150(5–7), 262–268 (1990)
30. Toda, M.: Vibration of a chain with nonlinear interaction. J. Phys. Soc. Jpn. 22(2), 431–436 (1967)
31. Toda, M.: Studies of a non-linear lattice. Phys. Rep. 18(1), 1–123 (1975)
32. Goldfriend, T., Kurchan, J.: Equilibration of quasi-integrable systems. Phys. Rev. E 99(2), 0022146 (2019)
33. Fu, W., Zhang, Y., Zhao, H.: Universal law of thermalization for one-dimensional perturbed Toda lattices.
N. J. Phys. 21(4), 043009 (2019). Apr

123
 Author's personal copy
1030 Santhosh et al.

34. Rigol, M., Dunjko, V., Olshanii, M.: Thermalization and its mechanism for generic isolated quantum
systems. Nature 452(7189), 854–858 (2008)
35. Howell, O., Weinberg, P., Sels, D., Polkovnikov, A., Bukov, M.: Asymptotic prethermalization in period-
ically driven classical spin chains. Phys. Rev. Lett. 122(1), 010602 (2019)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

123