Ergodic Properties of Microcanonical Observables
Cristian Giardina
Dipartimento di Fisica Universita di Bologna and INFN Sez. di Bologna,
via Irnerio 46, 40126 Bologna, Italy
giardma@bo.infn.it
Roberto Livi
Dipartimento di Fisica dell'Universitd, INFN Sez. di Firenze and INFM Unitd di Firenze,
Largo E. Fermi 2, 50125 Firenze,
livi@fi.infn.it
Abstract: The problem of the existence of a Strong Stochasticity Threshold in the FPU-/3
2^ ■ model is reconsidered, using suitable microcanonical observables of thermodynamic nature, like
^ , ' the temperature and the specific heat. Explicit expressions for these observables are obtained by
exploiting rigorous methods of difi'erential geometry. Measurements of the corresponding temporal
Oh' autocorrelation functions locate the threshold at a finite value of the energy density, that results to
' be independent of the number of degrees of freedom.
00
X3
><
Keywords: Strong stochasticity threshold, ergodic hypotesis, microcanonical ensemble
. I. INTRODUCTION
^ .
' Models of interest for applications in plasma and condensed matter physics as well as in molecular biology and
, chemistry are described by many degrees of freedom Hamiltonians of the form
o
a^: H = Y^^ + v{{q.}) (1)
1=1
m
'^y where V{{qi}) is a nonlinear interaction potential among the N particles of mass m, located on a regular lattice, whose
sites are labelled by the integer index i. The displacement and momentum canonical coordinates of the i-th particle are
Q ' denoted by qi and Pi , respectively. Apart ramarkable exceptions like the Toda chain , a generic choice for V yields
^2 ■ a chaotic dynamics. One of the main issues in this field is the existence of dilferent levels of chaos that typically occur
when some control parameter (e.g. energy or energy density) is varied. Fermi, Pasta and Ulam in their pioneeering
numerical experiment j^], first observed the strong rigidity of low-A: excitations that prevented equipartion of the
energy among the Fourier modes over exceedingly large time scales. Izrailev and Chirikov provided an explanation
of this fact in terms of the resonance-overlap criterion while more refined numerical experiments showed that, for
^ ' sufficiently high energies, equipartition among the Fourier modes sets in very rapidly l^-Q]. In particular, already
Bocchieri et al. raised the question of the existence of an energy threshold separating a quasi-regular dynamics
from a highly chaotic phase. Nowadays, there is theoretical and numerical evidence that such a threshold does exist
for large values of N. In this respect, it is worth stressing the following points:
• this threshold in general occurs at energy values that prevent any possibility of quantitative explanation in terms
of the canonical perturbation theory;
• in the strongly chaotic phase the time needed to reach equipartition is almost independent of energy and N;
just beyond the threshold it suddenly increases, while the dynamics results to be very weakly chaotic.
These remarks point out that the equipartition threshold (ET) separates definitely different dynamical regimes, in
such a way that for sufficiently small energies the time needed for approaching equilibrium properties may become
so long that the dynamics maintains an ordered structure over any practically available time scale. On the other
hand, it should be stressed that ET does not seem to exhibit analogies with the standard scenario of equilibrium
phase transitions: for instance, it is not at all clear which kind of symmetry breaking mechanism, if any, could be
responsible for the slowing down of energy equipartition among the Fourier modes.
A further crucial problem concerns the persistence of ET in the thermodynamic limit. This is a key point that
recently led various authors to reconsider this problem, by exploiting different approaches and techniques. Let us
1
briefly summarize the state of tlie art. Some investigations have refined the original approach based on the study
of energy equipartition among the Fourier modes: the so-called spectral entropy was first introduced as a global
observable able to quantify the amount of energy equipartition ^,|l^ . The scaling properties of this quantity have
been carefully analyzed in ref.s ^1 12 1, yielding the conclusion that the time needed to reach energy equipartition
scales in a nontrivial way with both the energy and N. Further numerical investigations and theoretical arguments
based on an improved treatment of the Chrikov's resonance overlap criterion |p^-p^ indicate that, at least for low-fc
modes excitations, ET should vanish in the thermodynamic limit.
An alternative approach to the characterization of the different dynamical regimes occurring in models like (|l|)
has been pursued by directly investigating proper geometrical features of the phase space. Concepts and tools taken
from Riemannian geometry were translated in Hamiltonian language in [p5|-p^. By exploiting such techniques it
has been possible to obtain an analytic estimate in the thermodynamic limit of the maximum Lyapunov exponent
as a function of the energy density for the FPU /?- model It exhibits a crossover between a weak and a strong
chaotic regime when the energy density is increased beyond a threshold value, that has been identified as the Strong
Stochasticity Threshold (SST) jl^. Similar conclusions in favour of the existence of SST in the thermodynamic limit
have been drawn by considering other specific geometrical indicators, like generalized curvatures of the phase space,
whose reliability has been tested for various models |20-2^. By the way, theese studies also indicate that in the
strongly chaotic regime these models may still be far from ergodic.
Upon all of these remarks, one might wonder if the existence of a threshold (either ET or SST) may have some
influence on the equilibrium statistical properties of many degrees of freedom hamiltonian models. Beyond the evident
interest of such a question for the foundations of Statistical Mechanics, this is a crucial point for molecular dynamics
applications, that usually assume the validity of the ergodic hypothesis, i.e. the equivalence of time and ensemble
averages. Such a problem would be solved from the very beginning if one could prove ergodicity for a model like (1).
Unfortunately, this is possible only in a few special cases. On the other hand, it is commonly accepted that chaoticity
should be sufficient for reproducing reliable statistical predictions for most of the observables of physical interest.
A partial disproval of this assumption was obtained in psf , where it has been shown that the equivalence between
time and ensemble averages for canonical thermodynamic observables (e.g. internal energy and specific heat) mainly
depend on the model and on the nature of the observable, rather than on the degree of chaoticity of the dynamics.
In this paper we aim to clarify most of these open points by tackling the problem through the direct comparison of
statistical and dynamical properties of microcanonical thermodynamic observables. The general formalism yielding
explicit dynamical expressions for such observables is introduced in Section II, where we also comment on their
geometrical nature. The results obtained by molecular dynamics calculations for the FPU /3-model are reported in
Section III. We show that SST is predicted by the decay rate of the temporal autocorrelation function associated
with the microcanonical temperature and specific heat, while ergodic-like properties are found to depend mainly on
the observable at hand, irrespectively of the degree of chaoticity of the dynamics. Conclusions and perspectives are
contained in Section IV.
II. MICROCANONICAL THERMODYNAMIC QUANTITIES
The equivalence in the thermodynamic limit of the predictions obtained from different statistical ensembles (z^-
canonical, canonical and G-canonical) is a widely accepted and partially proved fact p^ . Explicit expressions for
thermodynamic observables are usually computed by canonical averages. Their relation with the corresponding
microcanonical averages was established in a celebrated paper by Lebowitz, Percus and Verlet ||25[| . Only recently
H.H. Rugh, assuming ergodicity of the phase space, obtained a general rigorous expression for the temperature of
the microcanonical ensemble, that results to be strictly related to the geometrical structure of the phase space [p6| .
Analogously, an explicit formula for the specific heat at constant volume has been also obtained ^t\ .
In what follows we are going to reproduce the expressions of these microcanonical thermodynamic quantities, by
exploiting a different approach, based on standard tools of differential geometry (see the Appendix). Wc find that any
of these quantities is an explicit, even if in some cases complicate, function of the canonical coordinates. As a conse-
quence, this geometrical approach provides a natural bridge between a dynamical and a thermodynamical description
of the microcanonical ensemble. Upon the remarks of the previous Section, we are interested in understanding what
can be inferred about thermodynamics by molecular dynamics experiments performed on hamiltonian models of the
form (1). As we have already stressed, in general such models are known not to be ergodic in a strictly mathematical
sense. In this perspective, it could be hard to find out more appropriate probes for testing ergodicity, than such
microcanonical thermodynamic observables.
2
In analogy with the rigorous approach of Rugh, wc assume ergodicity in order to guarantee that the microcanon-
ical entropy 5 is a well defined quantity associated to some hypervolume in phase space, equipped with a uniform
undecomposable probability measure. In the /^-canonical ensemble S plays the role of a generalized thermodynamic
potential from which any thermodynamic quantity can be obtained by derivations w.r.t. the physical parameters
energy E and volume V:
1 - J- -
T ~ \dEjy Cv ~ \dE
P rdS\ 1 /5T
T \dV J E VuE
(2)
(3)
where a^; is defined as the thermal expansion coefficient in the /U-ensemble. In order to be correctly defined must
fulfill the following properties:
• additivity;
• invariance under adiabatic reversible transformations;
• it must be a non-decreasing function for irreversible adiabatic transformations.
Two choices are possible for the /x-ensemble:
A) S{E, N, V) = \n{cNOj{E, N, V)) = In (^cn J dq dp 6{E - H{q,p))j
da
= In Cat
H=E
VH
B) S{E, N, V) = ln{c^n{E, N, V)) = In ^c^ J dq dp 0{E - H{q,p))j
= In ( c'jv / dV
\ Jh<e y
In these formulae we have set the Boltzmann constant Kb = 1, while cjv and are arbitrary constants to make the
argument of the logarithm dimensionless.
Definition A includes all the microstatcs compatible with the constrain H = E, i.e. those microstates belonging to
the constant energy surface T,e = {{q,p) £ T | H{q,p) = E}. Alternatively, definition B considers as microstates all
those contained inside the hypervolume Ve — {{q,p) & T \ H{q,p) < E} limited by T,e- Both definitions should be
equivalent in the thermodynamic limit.
Let us recall that the phase-space T = E?^ is given the structure of a sympletic manifold by the fundamental
symplectic two-form 0^2 = Y^^=i '^li ^ '^Pi- The Hamiltonian H -.T ^ R generates a vector field / • dH (where I is the
fundamental symplectic matrix) and an associated fiux that preserves the Liouville measure. For smooth potentials
V{{qi}) it seems reasonable to assume that Ve is limited and that is smooth and connected.
Taking into account case A, one can easily obtain an expression for the micro canonical temperature, that from here
on we shall denote with T^, through the relation
dE oj
(4)
Making use of a simple theorem of differential geometry one has
1 Ih=e V ■ ( w) wk ( yg V ...
-^1^ Jh=e wm ^" " ^
The derivation and the full expression of this formula are reported in the Appendix.
Analogously one can obtain an explicit formula for the specific heat per particle at constant volume
3
By applying the differential geometry theorems reported in the Appendix the following expression is obtained
CA
da
E \\VH\
VH
IVHIP
l|VH||
V-
( 1
VH
da \
\\\^H\?)
IIVHPJ
\\^H\\)
(^Ih=E i|VH||)
(V.
[v.(
VH \
VH
\\vHr)
IIVHP
(V.
( VH
))l
\\\vHr
(7)
Similarly, one can treat case B.
The temperature (that from here on we denote Tkim for reasons that will be clear in a moment) can be obtained
by introducing a vector 77 = (0, . . . , 0, x^, . . . , 0), that has the property V • 77 = 1, and using the divergence theorem:
1
dS
Ih^
da
E \\VH\\
Ih.
E \\VH\\
Ih^
da
E wm
This is the usual expression deriving form the virial theorem
rr -I ^\
kin — {^i I a
dxi
that, for Hamiltonian and for rj = 1/N (0, . . . , 0,pi, . . . ,pn), simplifies to
dH da
dx^wm
(8)
(9)
N
i.e. the same expression that is obtained in the canonical ensemble. One could naively conclude that case B should
correspond to the canonical ensemble for any explicit expression of thermodynamic variables. Clearly this is not the
case, as one can easily conclude from the expression of the specific heat per particle
1
dE
1
N
N
OH,
VH
VH
(10)
Notice that cb depends on the ratio between Tkin and T^, that is quite a different expression w.r.t. the corresponding
canonical observable.
III. NUMERICAL ANALYSIS
In this Section we study the dynamical and statistical properties of the microcanonical observables introduced in
Section II. We consider the so-called FPU /3-model, whose Hamiltonian is of the form (0) with
N N
V{{q.}) = 2 - + f Efe+i - (11)
i=0 1=0
We impose fixed boundary conditions, in such a way that the total energy E is the only conserved quantity. All
numerical experiments hereafter reported were performed with /? = 0.1. (Notice that this is not a prejudice of
generality, since rescaling /3 amounts to rescale E). The integration of the equations of motion has been performed by
a bilateral symplectic algorithm in double precision, with a time step At — 0.005, that guarantees the conservation
of the energy at least on the sixth significative figure, for the considered range of energies. The running time average
for an observable / and the corresponding variance, computed over the whole time span, are defined as follows:
{f)t^\jj{p{t').q{t'))dt'
(12)
if) - if?
v{.f) = -^nj)^
Numerical simulations typically have been performed over 0(10®) time steps, after a transient of 10'"' time steps. Since
in the bilateral symplectic algorithm each time step amounts to 2 • At, the total integration time in natural units is
O(IO^), that corresponds approximately to O(IO^) minimum harmonic periods of oscillations. We have considered
different values of the energy density e = E/N and, at fixed e, different values of N . Let us first analyze the strongly
chaotic regime that, according to the results of ref.s ]T^ , pO| -p2t, should extend above e ^ 0(1). For the sake of space
we report here only the running time averages of and Tkm for e = 10 and for N — 2", n = 5, • • • , 9 (see Fig. |^).
Notice that both {Tkin)t and (r^}f approach an asymptotic value depending on N: the former observable exhibits
a fast convergence irrespectively of iV, while the latter is characterized by wild fluctuations that tend to weaken for
increasing values of N . Averaging over many different initial conditions allows one for reducing such wild fluctuations
and to extract more reliable estimates of the asymptotic values attained by {T^)t- Both "temperatures" are found
to converge (from below and from above, respectively) to the same thermodynamic limit value (Too ~ 11.61) with
corrections of 0(1/A^), as it should be expected on the basis of the ergodic theory (see Fig. |^). We have also checked
that the corresponding variances V{Tkin) and V(T^t) scale both like 1/iV thus showing that they are characterized
by robust statistical properties. It is worth stressing that even for such a high value of e the model is far from
ergodic in a strictly mathematical sense ^l| . On the other hand, both temperatures seem to provide equally reliable
thermodynamic predictions. We have verified that the same qualitative behaviour is observed for values of e ranging in
the interval (1, 100). On this basis one can state that, at least for high values of e, the two definitions of temperature
are thermodinamically (even if not dynamically) equivalent, while remarking that measurements of Tkin are much
more efficient for any practical purpose.
A very different scenario is obtained for e <^ 0(1), i.e. in the weakly chaotic regime. As an example, the running
time averages of and Tkin for e — 0.01 and N = 128 are shown in Fig. 3, for three different initial conditions. Tkin
still fastly relaxes to an asymptotic value independently of the initial condition. Conversely, exhibits a dramatic
dependence on the initial condition, despite its fluctuations are much less wild than in the strongly chaotic regime.
The comparison between the two temperatures is quite illuminating. Tkin appears as a "good" thermodynamic
observable that converges to its equilibrium value, irrespectively of ergodicity. In the perspective of the weak ergodic
theorem by Khinchin one would expect that at least the temporal autocorrelation function of Tkin decays to zero.
The behaviour of T^ clearly excludes this possibility, since its dependence on the initial conditions shows that the
trajectories over which both quantities have been measured remain "trapped" for extremely long integration times in
different regions of the phase space. As an illustration, in Fig.s 4 a,b we show the normalized temporal autocorrelation
functions C^{t) and Ckin{t) of the dynamical observables defining the two temperatures, for e = 0.01 and N — 128.
The corresponding power spectra clearly indicate that the dynamics amounts to a quasi-periodic motion with a few
dominating harmonic components, whose frequencies and amplitudes depend on the initial conditions. If one looks at
the explicit expressions for and Tkin this result appears much less mysterious: the former observable is sensitive
to the geometry of the phase space (mainly depending on V) through its explicit dipendence on the g^'s, while the
latter is sampled over the submanifold of the momenta pi 's that makes it insensitive to any geometrical intricacy of
the phase space ( a straightforward calculation shows that averaging Tkin over a quasiperiodic orbit yields the same
prediction of equilibrium statistical mechanics).
The measurement of C'^(t) and Ckin{t) in the strongly chaotic regime provides additional elements of information,
that confirm the previous conclusions and allow one for obtaining a quantitative characterization of the typical
relaxation times of both temperatures to their equilibrium values. In all numerical simulations hereafter reported
C^{t) and Ckin{t) have been averaged over 5 • 10^ initial conditions. Typical results are shown in Fig.s 4 c,d for
£ = 10 and N = 256. Both autocorrelation functions exhibit a fast exponential decay modulated by an oscillation
that is rapidly damped for large times. At variance with Ckin{t), C^{t) presents also a typical " hydrodynamic tail"
that superposes to the exponential decay at large values of t. This shows that, even in the strongly chaotic regime,
the two temperatures still exhibit quite different dynamical features that originate from their different geometrical
nature, making them sensitive or insensitive to the geometrical structure of the phase space.
A systematic inspection in a wide range of e in the strongly chaotic regime and for 7V = 2",n = 5,---,9 yields the
following conclusions. The hull of the autocorrelation functions can be confidently fitted by the following laws
Ck^n{t) = e-*/^i (14)
5
I- A
(15)
1 + t/T2
where ti is the same for both quantities and results to be independent of TV, as well as T2, and A is just a normalization
constant. The main outcome of this analysis is that, when varying e in the strongly chaotic regime, ti is found to
obey the following remarkable scaling law
with Ec ~ 0.8 (see Fig. 5). It is worth stressing that this result allows to locate unambigously the SST of the FPU
/?-model at a specific value of the energy density that agrees with all the estimates obtained by the methods mentioned
in the Introduction. We want to observe that also T2 is found to diverge at Ec, altough the poor numerical accuracy
in fitting the " hydrodynamic" decay of Cfj.{t) does not allow, in this case, to extract a clean scaling law.
The same analysis has been performed for the observables ca, cb and for the canonical specific heat per particle
obtained in as a function of the microcanonical fluctuations of the total kinetic energy K
The running time averages of ca and Ccan exhibit behaviours very similar to those reported in Fig. 1 for and Tkm,
respectively. For the sake of space, in Fig. 6 we simply report the scaling with N of the asymptotic values attained
by CA and Ccan, that are found to converge to the same thermodynamic limit value Coo- We have also looked at the
temporal autocorrelation function of those observables, whose averages determine the expressions of the specific heats
per particle. In practice, the only new interesting quantity, with respect to those already considered in the study
of the temperatures, is the numerator of the fraction in eq.(0). We have verified that its autocorrelation functions
coincide with those of for each value of e. All these results indicate that these specific heats provide a scenario
fully consistent with the temperatures analysis.
On the contrary, the running time averages of cb are characterized by extremely wild fluctuations, that do not show
any tendency to a smooth approach to the expected thermodynamic limit value over available time spans. This fact
definitely enforce the conclusion that relaxation mechanisms may strongly depend on the very nature of the observable
at hand.
Upon assuming ergodicity, thermodynamic variables in the microcanonical ensemble can be explicitly written as
dynamical quantities, by exploiting standard methods of differential geometry. In the thermodynamic limit they can
be shown to coincide with the corresponding quantities obtained in the canonical and gran-canonical ensembles (this
can be easily checked at least for some basic quantities like temperature, specific heat etc.).
On the other hand, many degrees of freedom hamiltonian systems are known not to be ergodic even at high
energies In particular, for sufficiently small energy densities such models exhibit quasi-periodic evolution along
trajectories, that remain trapped inside some submanifold of the phase space for extremely long integration times.
This peculiar dynamical feature is at the basis of the long-standing problem of "ergodicity threshold" raised by the
celebrated paper of FPU . An up to date summary of this problem is contained in the introduction.
In this paper we have pointed our attention on the FPU /3-model, providing an unambiguous description of the
Strong Stochasticity Threshold. This was obtained using the previously mentioned microcanonical observables as
probes of ergodicity. In particular, this analysis allowed us for locating the SST at a finite value of the energy density
Ec — 0.8, that results to be independent of N. It is worth stressing that this value agrees with rougher estimates still
originating from the study of "geometrical" observables [l^p^ .
Moreover, the comparison with the corresponding canonical observables allows us for concluding that the crucial
point about the ergodic hypothesis concerns the very nature of the observable and not just the features of the dynamics.
Such a point of view was already raised by Khinchin p9[ | , although we have shown that it may exist thermodynamic
observables (like Tkin and Ccan) that converge to the predictions of equilibrium statistical mechanics even when time
averages are performed over highly non-chaotic evolutions, where the "weak ergodic theorem" by Khinchin does not
certainly apply.
Let us finally observe that one can obtain other different dynamical expression for microcanonical observables
by proper choices of the coordinates space representation. The only constraint for such a choice is to maintain
n oc (e - e,)-i/2
(16)
(17)
IV. CONCLUSIONS AND PERSPECTIVES
6
invariant the flux through the constant energy surface. For instance, in case B, one could choose the vector
Tj = l/2N{qi, - ■ ■ ,qN,pi, - ■ ■ ,pn), in such a way that the temperature depends also on the qi coordinates. This
arbitrariness, obviously, does not affect thermodynamic limit properties, since all expressions for the same observables
coincide, while, dynamically, the scenario should not change significantly with respect to the one described in Section
III.
Nonetheless a detailed investigation of these different choices can provide a deeper insight on the way "statistical
fluctuations" can be consistently introduced in the microcanonical ensemble. This point goes beyond the aims of this
paper and will be discussed elsewhere.
ACKNOWLEDGMENTS
We are indebted with H.H. Rugh, M. Pettini, S. Lepri, L. Casetti, R. Franzosi, A. Franchi, M. Peyrard and T.
Dauxois for useful discussions and comments. We want to thank I.S.I, in Torino for the kind hospitality during the
workshop of the EU HC&M Network ERB-CHRX-CT940546 on "Complexity and Chaos", where part of this work
was performed.
APPENDIX A:
In this appendix we report a simple theorem of differential geometry that allows to express the derivatives of the
entropy w.r.t. the energy E. Moreover, we give the explicit expressions for the temperature and of the quantities
defining the specific heat of standard Hamiltonian .
Theorem: Consider S^; as a (2A^ — l)-dimensional hypersurface of R^^ parametrized by H{x) — E, where
H{x) : Ris the Hamiltonian fmiction in the F-space. Given a function f{x) : i?^^ ^ i?, that is C'^^{R^^),
we let
fix) da
(Al)
H=E
where da denotes the (2A^ — l)-dimensional measure. Suppose that it exists a constant c > such that ||Vif||
(ViJ, V-ff)^/^ > c. Then the following formula holds for the n-th derivative of ljj{E):
d"LU
dE^
A''f{x)da
H=E
where A" denotes the n-th iterate of the operator A, defined by
Afix) - V • fix)
iViJl
(A2)
(A3)
Proof: It clearly sufficies to prove the result for n — 1. We assume, without loss of generality, that VH points
towards the inside of the hypersurfaces parametrized by H{x) = E. Consider the difference quotient for uj{E)
u>{E + h) - uj{E)
Exploiting the identity
1 =
f{x)da- / f{x)da
E+h
(A4)
and, remembering that n(x) = n^^n is the unit inner normal to Tie at point x, one can write
ring that n{x) = yvh\\
ui{E + h)-oj{E) 1
E
t(x)- -ripda
E lVi/||
(A5)
7
where Ue is the unit normal pointing toward the exteriors of the annular region
{x G R^^ : E < H{x) < E + h}. We may now applay the divergence theorem tt
Lu{E + h)~Lu{E) _1 r^+'' ^ V7U \
h h Je
Finally, using the co-area formula (sec ) , wc have
and, taking the limit for h
to obtain
dE'
Je Jh=e'
• 0, we conclude that
V.(/(.)^l..
VH
llViJl
da
VH
J{E) = f V
JH=E
1
If we consider the theorem with
VH
\\VH\\ J \\VH
7 da
q.e.d.
(A6)
(A7)
(A8)
l|Vi?|l
and the cases n = 1, 2 we obtain the expressions that enter in the formula for temperature and specific heat
the previous Sections. In particular:
'dE
JH=E
V
VH
da
VH IP ; II VH
VH \ VH
vH\\y ]rv^"2
da
dE^ J^^E [ U^HW^J \\VH\\^\\\VH\\
In the case of standard Hamiltonian (|^), for which one has dH/dpi = dV/dqi — --Fi and dH/dpi = pi, the
dynamical expressions are:
given m
(A9)
(AlO)
explicit
VH
VH IP
AH
WW
1
ivffip
■VH
i=l Pi ^
Pi
V •
VH
VH
VH IP / II VH IP
^ WW
{El,pi+F?y
^.VH + A
\VH¥
1
\VHP
(All)
(A12)
{AHr (^-eLII
iViJiP
EliPf + Fi
AH
WW
VH =
+ 4
{EIiPI + F?^
: ^ Q
A
iViJll
(El^Pf + Fi
= —2 • = —
(yZ^pI + f?]
J2k=iPl + J2k=i ^»fft) ^
iEl^pf + pf"
8
Figure Captions
FIG. 1. Running time averages for the temperatures Tkin (solid lines) and (dot-dashed lines) for e = 10 and
N = 2^ , n = 5,---,9. Increasing A'^, Tmu (T)i) better and better approximates the equilibrium expectation from below
(above).
FIG. 2. The inverse reduced temperature IRT=|T(iV) - Tod"^ versus N for T^{N) (circles) and Tki„{N) (triangles). Solid
lines are best fits.
FIG. 3. Running time averages for the temperatures Tkin (solid lines) and (dot-dashed lines) for s = 10~^, N = 128 and
for three different initial conditions.
FIG. 4. The temporal autocorrelation functions of Tkin and T^ for N = 256 and e = 10~^ ( a and b, respectively) and for
e = 10 (c and d , respectively)
FIG. 5. Log-log plot of Tj ^ versus (e — Ec). The solid line is the best fit obtained assuming the otimal exstimate £c = 0.79 .
FIG. 6. The inverse reduced specific heat IRC = \c{N) — Coo|~^ versus A'^ for CAiN) (circles) and Ccan{N) (triangles). Solid
lines are best fits.
9
References
M. Toda, Phys. Rep. 18 : 1 (1975).
E. Fermi, J. Pasta and S. Ulam, in Collected papers of E. Fermi (University of Chicago Press, Chicago, 1965), Vol.2, p. 78 .
F. Izrailev and B. Chirikov, Sov. Phys. Doklady 11 : 30 (1966).
B. Chirikov, F. Izrailev and V. Tayurskij, Comp. Phys. Comm. 5 : 11 (1973).
P. Bocchieri, A. Scotti, B. Bearzi and A. Loinger, Phys. Rev. A 2 : 2013 (1970).
M. Casartelli, G. Casati, E. Diana, L. Galgani and A. Scotti, Theor. Math. Phys. 29 : 205 (1976).
G. Benettin and A. Tenenbaum, Phys. Rev. A 28 : 3020 (1983).
L. Casetti, R. Livi, A. Macchi and M. Pettini, Europhys. Lett. 32 : 549 (1995).
R.Livi, M. Pettini, S. Ruffo, M. Sparpaglione and A. Vulpiani, Phys. Rev; A 31 : 1039 (1985).
R.Livi, M. Pettini, S. Ruffo and A. Vulpiani, Phys. Rev; A 31 : 2740 (1985).
H. Kantz, R.Livi and S. Ruffo, J. Stat. Phys. 76 ; 627 (1994).
J. De Luca, A.J. Lichtenberg and S. Ruffo, Phys. Rev. E 51, 2877 (1995).
J. De Luca, A.J. Lichtenberg and M.A. Lieberman, CHAOS 5, 283 (1995).
D. Shepelyansky, preprint, November 1996.
M. Pettini, Phys. Rev. E 4.7 : 828 (1993).
L. Casetti and M. Pettini, Phys. Rev. E 48 : 4320 (1993).
M. Cerruti-Sola, M. Pettini, Phys. Rev E 53 : 179 (1995)
L. Casetti, R. Livi and M. Pettini, Phys. Rev. Lett. 74 : 375 (1995).
M. Pettini, M. Cerruti Sola, Phys. Rev. A 44 : 975 (1991)
C. Alabiso, M. Casartelh and P. Marenzoni, Phys. Lett. A 183 : 305 (1993).
C. Alabiso, M. Casartelh and P. Marenzoni, J. Stat. Phys. 79 : 451 (1995).
C. Alabiso, N. Besagni, M. Casartelh and P. Marenzoni, J. Phys. A 29 : 3733 (1996).
R. Livi, M. Pettini, S. Ruffo and A. Vulpiani J. Stat. Phys. 48 : 539 (1987).
D. Ruelle, Statistical mechanics: Rigorous Results, Addison- Wesley (1989).
J. Lebowitz, Percus and Verlet, Phys. Rev. 153 : 250 (1967).
H.H. Rugh, Phys. Rev. Let t. 78 : 772 (1997)
H.H. Rugh, preprint 1997, chao-dyn/97030i;
L. Casetti, Physica Scripta 51 : 29 (1995).
A.I. Khinchin, Mathematical Foundations of Statistical Mechanics (Dover, New York 1949).
10
12.1 4 /
11.9
1^
11.7
11.5
u
I
i / —
Li/ --^
1
2e+05
4e+05 6e+05
t
8e+05 1e+06
Q
S.
CO
1.08
1.06 -'
1.04 -
1.02 -
1.00
0.98 -
0.96
2e+05 4e+05 6e+05 8e+05 1e+06
t
Q.
<"
CQ
CO
