Relationship between dynamical entropy and energy

dissipation far from thermodynamic equilibrium

Jason R. Green
a,b
, Anthony B. Costa
b,c
, Bartosz A. Grzybowski
b,c,d
, and Igal Szleifer
b,c,d,e,f,1
a
Department of Chemistry, University of Massachusetts Boston, Boston, MA 02125; and
b
Non-Equilibrium Energy Research Center, Departments of
c
Chemistry
and
e
Biomedical Engineering,
d
Department of Chemical and Biological Engineering, and
f
Chemistry of Life Processes Institute, Northwestern University,
Evanston, IL 60208-3100
Edited

by R. Stephen Berry, The University of Chicago, Chicago, IL, and approved August 29, 2013 (received for review June 26, 2013)
Connections between microscopic dynamical observables and
macroscopic nonequilibrium (NE) properties have been pursued
in statistical physics since Boltzmann, Gibbs, and Maxwell. The
simulations we describe here establish a relationship between
the KolmogorovSinai entropy and the energy dissipated as heat
from a NE system to its environment. First, we show that the
KolmogorovSinai or dynamical entropy can be separated into
system and bath components and that the entropy of the system
h
sys
characterizes the dynamics of energy dissipation. Second, we
nd that the average change in the system dynamical entropy is
linearly related to the average change in the energy dissipated to
the bath. The constant energy and time scales of the bath x the
dynamical relationship between these two quantities. These
results provide a link between microscopic dynamical variables
and the macroscopic energetics of NE processes.
nonequilibrium self-assembly
|
dynamic entropy
|
statistical mechanics
|
Lyapunov exponents
U
ncovering the laws governing nonequilibrium (NE) energy-
dissipating systems is a fundamental challenge, consequen-
tial for the design of new dynamic materials (1, 2) and, ulti-
mately, to the understanding of life (35). From a theoretical
perspective, this challenge is complicated by the inability to use
the methods and simplications familiar from equilibrium ther-
modynamics and equilibrium statistical mechanics. Nevertheless,
there have been numerous important contributions to the un-
derstanding of NE phenomena, the earliest of which applied
thermodynamics to irreversible processes (68) and processes
executed in nite time (9, 10) in the near-equilibrium regime,
often with assumptions of local equilibrium. More recently,
advances have been made in the far-from-equilibrium regime,
including what are now collectively called uctuation theorems
(1114).
Most relevant to the present study is the statistical mechanics
of NE phenomena based on the elements of dynamical systems
theory (15, 16). Missing from this theoretical framework, and
others, is a means of simplifying the description of irreversible
processes, when such a simplication is possible. Most generally
it is not possible to describe the dynamics of NE systems sepa-
rately from the dynamics of their surroundings (e.g., a heat bath)
(17)although progress has been made in reducing the number
of degrees of freedom of NE systems that obey Markovian dy-
namics (18, 19). This difculty is a more general signature of
systems in extreme states of disequilibrium: There is not neces-
sarily a clear physical systembath boundary because all of the
microscopic variables of the bath may be important at one time
or another in the dynamics of mass, energy, or momentum trans-
port through the system. As a consequence, statistical-mechanical
properties of NE systems cannot generally be parameterized by the
static properties of the surroundings.
Central to the dynamical systems approach to NE statistical
mechanics are the KolmogorovSinai (KS) entropy per unit time
(20, 21) and Lyapunov exponents (22). These quantities, for
example, can be used to formulate the deterministic uctuation
theorems that characterize the statistical dynamics of systems far
from equilibrium (12, 13, 23). In the present study, we build on
these two concepts to study the process of NE self-assembly (3, 5,
24). The specic system we analyze is inspired by experiments in
which nanoparticles covered with photoactive ligands are irra-
diated with light, thereby changing their surface properties and
modifying interparticle interactions, resulting in particle assem-
bly (2527). In our simulations, interparticle interactions are
similarly modied, displacing the system from equilibrium and
causing it to evolve under NE conditions until a nal equilibrium
self-assembled state is reached. In both cases, the assembly
process is accompanied by dissipation of energy during the
transient relaxation to an equilibrium state. We analyze this
evolution from our simulations both in terms of the dynamical
(KS) entropy and the energy of (or heat released by) the system.
Model System
Our model is a classical, 3D, periodic 126 Lennard-Jones (LJ)
uid of N particles with a set T of 3N positions and 3N momenta.
The Hamiltonian Hr
ij
; p
k
;
ij
=

3N
k
p
2
k
=2m+

N
i <j
Vr
ij
; t is
used with the interaction potential of particles i and j a distance
r
ij
apart being Vr
ij
; t =4
ij
t
__

rij
_
12

rij
_
6
_
, where
ij
t de-
termines the strength of the interaction at time t and is the
distance at which the attractive and repulsive terms are equal
(i.e., a measure of the particle size). The minority of the particles
(typically, a set S of 7) are switchable and constitute the system to
be assembled into a bound cluster, whereas the majority (typi-
cally, a set B of 143) are the bath particles (note: simulations with
Signicance
The dissipation of energy occurs naturally in systems as diverse
as those in biology and as common as those responsible for the
weather. Systems that dissipate energy while self-assembling
also show promise as a synthetic route to responsive nanoscale
materials. Optimizing the energy efciency of these syntheses
requires reducing the theoretical description to only those
variables that are essential to an accurate prediction of the
energy lost as heat. Here, with computer simulations and
a model of assembly, we establish a methodology for identi-
fying those essential variables. With this minimal description of
the system we also nd a linear relationship between the dy-
namical entropy of the self-assembling system and the energy
dissipated as heat during assembly.
Author contributions: J.R.G., A.B.C., B.A.G., and I.S. designed research; J.R.G. and A.B.C.
performed research; J.R.G., A.B.C., and I.S. analyzed data; and J.R.G., A.B.C., B.A.G., and
I.S. wrote the paper.
The authors declare no conict of interest.

This Direct Submission article had a prearranged editor.

1
To whom correspondence should be addressed. E-mail: igalsz@northwestern.edu.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1312165110/-/DCSupplemental.
www.pnas.org/cgi/doi/10.1073/pnas.1312165110 PNAS | October 8, 2013 | vol. 110 | no. 41 | 1633916343
C
H
E
M
I
S
T
R
Y
other particle numbers yield conclusions identical to those de-
scribed in the following; see Fig. S1 for details).
The method used to simulate NE self-assembly trajectories,
shown schematically in Fig. 1, produces a sample of classical
trajectories. From a given sample of trajectories, average prop-
erties of microscopic dynamical observables can be examined as
they evolve in time. Initial phase points for the trajectories of
interest are selected from the isothermalisobaric ensemble. For
this model, all interparticle interactions
ij
are initially identical.
At a time assigned to be t = 0, the entire collection of particles is
isolated, and the well-depth parameter
ij
for the system particles
is instantaneously changed from a value corresponding to weak
interaction to a new value corresponding to strong attraction.
The effect of this change is to generate an initial NE state from
which system particles self-assemble and concurrently dissipate
their excess kinetic energy to the explicit bath particles, while
conserving the total energy, volume, and number of particles.
The dynamics of all particles is studied until a new equilibrium
state is reached. After the new equilibrium state is reached, the
system is returned to its initial equilibrium ensemble by again
making all interparticle interactions
ij
weakly interacting, and
repeating the entire process.
The above procedure can be performed cyclically to generate
any desired number of trajectories. Equilibration before each NE
trajectory removes the energy dissipated during the previous
assembly trajectory. Data shown here for the assembly trajecto-
ries at constant total energy, volume, and number of particles will
begin at t = 0 and end when assembly is complete. For in-
termediate times, the energy dissipated by the system particles is
entirely accounted for in the energy of the bath. Because the
initial phase points are consistent with the isothermalisobaric
ensemble, the total energy of each trajectory varies from trajectory
to trajectory. In what follows, all distances and times will be
expressed in LJ reduced units, dened in Methods, and the energy
parameter
ij
will be expressed in multiples of a constant (Fig. 1).
One of the methodological advances of this work is the cal-
culation of separate KS entropies per unit time (henceforth,
simply KS entropies) for bath and system components of a
Hamiltonian system. These KS entropies are calculated from
each time step (28) of the set of trajectories characterizing particle
assembly. To compute KS entropies h
KS
=h, orthogonal (Gram
Schmidt) Lyapunov vectors are simulated for the entire collec-
tion of particles; the ith vector has the form of rst variations in
position and momentum q
ji
; p
ji

T
with i; j T evolving under
the linearized Hamiltonian dynamics (29). This basis set for
tangent space is propagated along with the phase point by the
linearized velocity Verlet algorithm, a time step size of 10
3
, and
orthonormalization at every step (for details, see refs. 3032).
At each time step, Lyapunov exponent spectra are estimated
separately for the system and bath particles from the total
comoving set of simulated orthogonal Lyapunov vectors. These
spectra are computed from subvectors identied in the ith Lya-
punov vector q
ji
; p
ji

T
with j S; B and i T. The metric
d
2
i
C; t =

jC
q
ji
t
2
+p
ji
t
2
allows system and bath Lyapunov
spectra to be computed from the associated set of Lyapunov
subvectors. We use C to represent either the set of system S, bath
B, or total T particles. Each set of 2n nite-time Lyapunov
exponents is dened for the total, system, and bath as (29)

i
C; t =jtj
1
ln
d
i
C; t
d
i
C; t
0

C=T; S; B: [1]
For closed dynamical systems, Pesins theorem denes KS entro-
pies as the sum of positive Lyapunov exponents (33). Assuming
the theorem holds for the system and bath separately, their KS
entropies are calculated from the system and the bath Lyapunov
spectra
i
, respectively,
hC; t =

iTj+

i
C; t: [2]
The (nite-time) KS entropies thus dened can be interpreted as
an inverse predictability time scale or uncertainty gain resulting
from the evolution of a trajectory. Put differently, KS entropies
measure the degree of temporal order in dynamics. More
disordered dynamics have a larger h and the evolution of nearby
trajectories is predictable over a shorter time scale. It is empha-
sized that the KS entropy is not the entropy of microstates fa-
miliar from equilibrium thermodynamics: the KS entropy can be
small for a structurally disordered system and large for an
ordered system.
Results and Discussion
With these considerations in mind, and with the aim to bridge
the micro- and macroscale descriptions of our system, we ex-
amined the statistics of KS entropies sampled from ensembles of
classical trajectories.
The distributions of KS entropies f(h), collected from single
time steps in two time segments of the trajectory ensemble, are
shown in Fig. 2 for the assembling system (h
sys
), nonassembling
system (7 randomly chosen particles of the bath, h
nas
), the bath
(h
bath
), and the total (i.e., system and bath particles, h
tot
).
Here, a time segment is dened as an interval of 1,000 numerical
time steps from a single trajectory. For each time segment, we
analyze the statistics of KS entropies over the entire ensemble
of trajectories. Comparison of the distributions in the rst ( = 1)
and last ( = 75) time segments reveals that all KS entropies
distributions shift to larger values during the evolution from the
disassembled to assembled states. Interestingly, f(h
sys
) is skewed
at = 1 but approximately Gaussian at equilibrium, = 75. The
distributions f(h
nas
) for a nonassembling system of 7 randomly
chosen particles of the bath are nearly Gaussian at both times
and their shift is much smaller than that for f(h
sys
). The deviation
of f(h
sys
) from Gaussian at = 1 is attributed to the fact that
A
B
Fig. 1. (A) Scheme illustrating the switching of the interaction potential
between system particle pairs. Equilibrium conguration snapshots are also
shown. The color of system particles corresponds to the pair interaction
potential plotted in B. Before switching, the particles are weakly interacting;
after the switch, they are strongly attractive leading to their assembly. The
time between cycles is long enough to allow relaxation to the new equi-
librium state after the potential switch from blue (red) to red (blue).
16340 | www.pnas.org/cgi/doi/10.1073/pnas.1312165110 Green et al.
self-assembly events of the system particles are being sampled
in the time segment. In contrast with the system distributions,
the bath and total KS entropy distributions in Fig. 2 are nearly
identical to each other and of nearly Gaussian form at all times.
We also note that the bath distributions are narrower than those
of the system. We have found this behavior of the distributions to
be true regardless of the magnitude of the work done by changing

ij
. To summarize, the features of the KS entropy distributions
reect the variety of dynamical processes the system and the bath
undergo after triggering self-assembly.
During the self-assembly process, the KS entropy distribution
of the assembling system shifts more than the other distributions
considered. To better quantify the differences between the sys-
tem and the bath, we consider the percentage increase in sample
average KS entropies h, as shown in Fig. 3. These percentages,
computed from the formula %h =
h h0
h0
100, allow a di-
rect comparison of all average KS entropies simultaneously. All
of the mean KS entropies were smaller at short times (when the
interaction potential was switched on and the system was driven
out of equilibrium) than at long times (when the particles were
equilibrated in the self-assembled structure). Strictly increasing
average KS entropies indicate an increase in the dynamical in-
stability as the system particles are funneled predictably toward
the assembled state. Conversely, they show that the equilibrium
dynamics were less predictable (larger KS entropy) than the out-
of-equilibrium dynamics. The difference in this sensitivity to
initial conditions at and away from equilibrium is clearly most
dramatic for the set of assembling particles. This indicates that
the KS entropies we have dened measure the dynamical in-
volvement of phase space variables in the NE process.
Next, to study the evolution of the system displaced to dif-
ferent distances from equilibrium, we performed simulations in
which the potential parameter of the system particles was
changed from 1 to 5, 10, 15, and 20. In each case, the bath
and the systembath interaction parameters were maintained at
1. Fig. 3 A and B show results for instantaneous switching, at t =
0, of the potential well depth from 1 to 5, and from 1 to 15.
In B, the mean KS entropy h
sys
of the assembling system
increases by nearly 60% during the approach to equilibrium. By
comparison, the mean KS entropies of the bath h
bath
, non-
assembling system h
nas
, and total set of particles h
tot
are not as
responsive to the perturbation from equilibriumincreasing by
only 25% for this switch of the interaction potential. Importantly,
for all perturbations, we observe the monotonic increase of these
KS entropies over time, indicating that the behavior of these
quantities is consistent across a wide range of perturbation
strengths. The more modest increase in these KS entropies, in
contrast with those associated with the assembling system, reects
the change in the dynamics associated with the extra kinetic energy
produced by self-assembly in the bath, and possibly the rearrange-
ment of the bath particles necessary for the self-assembly processes.
The separation of the system KS entropy from the other KS en-
tropies considered suggests it characterizes the entire self-assembly
process and, as we show next, the energy dissipation process.
A
B
Fig. 2. Probability distributions of KS entropies collected from single time
steps in two time segments of the trajectory ensemble, one immediately
after the switch of the potential parameter from 1 to 15 ( = 1) and one
after the particles are assembled ( = 75). Distributions of the assembling and
nonassembling systems are shown in A and those for the bath and the total
set of particles are shown in B.
A
B
Fig. 3. Percentage increase in each h over time for two perturbations from
equilibrium, differing in the depth of the switched potential of system particles:
from 1 to 5 (A) and from 1 to 15 (B). Sample averages were computed from
all trajectories with the single time step values collected in a given time seg-
ment. The KS entropies of the nonassembling system, bath, and total respond
to each switch identically, but that of assembling system is more dramatic. The
increase in all KS entropies depends on the magnitude of perturbation.
Green et al. PNAS | October 8, 2013 | vol. 110 | no. 41 | 16341
C
H
E
M
I
S
T
R
Y
We are now in position to consider the main result of this
work: namely, the relationship between KS entropy and energy
dissipation. To measure the energy dissipated by the system, we
calculated the difference in the (sample mean) total energy of
the system E
sys
over the trajectory ensemble at two consecutive
trajectory segments. We scaled the energy change between times
and +1, E
sys
=E
sys
+1 E
sys
, by the natural energy
scale of the bath, 2 the thermal energy = 1/k
B
T. At the same
time, we scaled the average KS entropy difference of the system
h
sys
=h
sys
+1 h
sys
by the characteristic time scale of
the bath
1
(
1
bath
=

m
2
=48
_
, which for 1, the bath in-
teraction parameter, in LJ time units is 1=

48
p
). This inverse
time scale is determined by the potential of interaction and
eliminates the dimensions both of h
sys
and h
sys
. Fig. 4
summarizes the results by plotting the scaled energy and KS
entropy difference as a function of time. For the two data sets
shown, corresponding to two switches of
ij
, the time de-
pendencies of the two-scaled quantities coincide with one an-
other. Similar agreement holds for other values of
ij
studied (see
Fig. S2 for details), leading to a general relationship of the form

h
sys
_
=2

E
sys
_
: [3]
In writing this relationship, we assumed that for single time step
KS entropies X =X and that the sample mean becomes the
ensemble average h i over all trajectories (in the limit of an inn-
ite number of trajectories). Both sides of this expression are di-
mensionless and decay to zero at equilibrium, when no entropy is
being produced and no energy is being irreversibly dissipated.
The dynamics and energetics of the bath, determining the mech-
anism of energy dissipation, enter only through and .
Further support for Eq. 3 can be found by minor rearrangement
and division by the number of system degrees of freedom 3n. These
operations lead to a quantity = h=3n = 2E=3n.
This , derived from our data, is remarkably similar to a quan-
tity built into the equations of motion often used to study many-
particle systems in NE steady states (34, 35). There, is the sum
over the entire spectrum of Lyapunov exponents (not just those
that are positive). It is interpreted as the (irreversible) entropy
production rate, (viscous) heat production rate, and the phase
space contraction rate, all per degree of freedom. The rst two
interpretations are similarly tting here, given the relation of
to h
sys
and E
sys
. It is also reasonable to interpret our as the
phase space contraction rate of the system compensated for by
the phase space expansion rate of the bath, according to
Liouvilles theorem. The denition of also provides a justi-
cation for the number 2 in Eq. 3. The change in average total
energy of the system is divided by the mean total energy for n
particles
3
2
nk
B
T from the equipartition theorem because of our
choice of initial phase points selected from the isothermal
isobaric ensemble.
It is emphasized that each trajectory in the ensemble sample
evolved under Hamiltonian dynamics, requiring no a priori as-
sumption of thermostatting constraints or NE uxes, as is com-
monly done in atomistic simulations of NE processes (34, 35).
Despite conserving total energy, energy in the system is irre-
versibly dissipated to the explicit bath as heat because of the
choice of initial NE, disassembled states. The system particles
diffuse randomly through the bath, until individual assembly
events cause energy uctuations and a dissipative response in the
bath. This picture of self-assembly raises an interesting connec-
tion. In Langevins theory of Brownian motion, the parameters
and characterize the uctuation, dissipation, and relaxation to
equilibrium. In our work, their product xes the relationship
between the dynamical instability of the self-assembling particles
and the energy they dissipate. Given previous measurements of KS
entropies for Brownian particles (36), experiments should be able
to conrm the validity of Eq. 3 in NE self-assembly processes.
Conclusions
Our results provide direct evidence for the relationship between
the system dynamical entropy and the energy dissipated from the
NE system to a bath. In analogy with equilibrium statistical
mechanics, the role of the bath in the NE process is to act as an
energy sink. As at equilibrium, the temperature determines the
energy scale, whereas the intrinsic relaxation of the bath, mea-
sured by , characterizes the relevant time scale. Because the
relationship in Eq. 3 consists of an entropy and an energy, there
may be an underlying variational principle for a dynamical free
energy. In this context, one difculty with our simulations
results, and many other simulations of NE processes, is that to
obtain the necessary KS entropy statistics we need to calculate
the KS entropy over nite simulation time steps. It is tempting to
assume that Eq. 3 is valid in the limit of an innitesimal time step
and thus it would be valid for the time derivatives. If that were
the case, then the variational free energy would be a constant of
the motion for NE self-assembly processes.
Methods
We performed simulations of a 3D, periodic 612 LJ uid with shifted and
truncated forces (37). Conventional LJ reduced units were used, r* =r= and
t* = =t=

m
2
=
_
. The mass of all particles m was set to 1.0. An ensemble of
constant energy trajectories was generated by sampling transitions between
A
B
Fig. 4. Difference between sample means from consecutive trajectory
segments for the KS entropy and total energy of the system, scaled by ap-
propriate bath parameters, as a function of time. Plotted are dynamical
entropies of the system/bath and the total energy of the system for the
switch from 1 to 5 (A) and1 to 15 (B). All quantities decay to zero at
equilibrium as expected (dashed line), once self-assembly is complete. Similar
results were obtained for switches from 1 to 10 and to 20 (Fig. S2).
1
and are the characteristic time and energy scales of the bath, respectively.
16342 | www.pnas.org/cgi/doi/10.1073/pnas.1312165110 Green et al.
two Hamiltonians [e.g., H1 and H15]. Their initial phase points were
sampled from a trajectory consistent with the isothermalisobaric ensemble
with the temperature T* = 1.0 and pressure P* = 1.0. From each initial
phase point, the potential parameter was instantaneously switched to
that of the second Hamiltonian. Constant energy trajectories were then
propagated forward in time with the velocity Verlet algorithm in steps
of size t* = 10
3
.
A Verlet list with a cutoff of r* = 2.5 and a skin of r* = 0.1 was updated
every 10 time steps. Translational center-of-mass motion was removed every
10 time steps during isothermalisobaric sampling. Combination rules for
interaction parameters were given by the minimum [i.e.,
ij
=min
ii
,
jj
]. In
practice, the switching procedure was performed by cycling between
Hamiltonian and non-Hamiltonian dynamics. Non-Hamiltonian equili-
bration trajectory segments were 25,000 time steps long. The Berendsen
thermostat and barostat maintained constant temperature and pressure;
both used a time constant of * =0:5. Each Hamiltonian, constant-energy
NE trajectory segment was 75,000 time steps long. Phase space volumes
during these segments were then preserved such that the dissipation was
explicitly retained in the degrees of freedom of the total system. During
each constant energy trajectory, tangent space dynamics were also sim-
ulated, which limited the system size that could be studied because of
their high computational cost.
For each constant energy trajectory, the KS entropies were calculated for
each time step from the associated set of Lyapunov exponents. System, bath,
and total system exponents were computed from the same set of orthogonal
Lyapunov basis vectors (GramSchmidt vectors) (29, 38). This tangent space
basis is propagated and comoves with the phase point. Propagation orients
the rst vector parallel to the maximally changing tangent space direction.
Propagation with periodic orthogonalization orients the remaining vectors
in the maximally changing direction of a substance of tangent space and
prevents their collapse onto the maximally changing direction.
The set of GramSchmidt vectors was propagated forward in time with
the linearized velocity Verlet algorithm, a time step size of 10
3
, and
orthonormalization at every step. This algorithm is described in detail else-
where (30, 31). Construction of the linearized velocity Verlet form of the
tangent space propagator required second derivative (Hessian) matrices at
consecutive steps, which used forward differences of the analytical gradients
with a displacement of r* = 10
8
. Initial basis sets were arbitrarily dened as
randomly lled, orthonormal matrices at the initial phase point. All calcu-
lations were performed in double precision.
ACKNOWLEDGMENTS. This material is based upon work supported as part
of the Non-Equilibrium Energy Research Center, an Energy Frontier Research
Center funded by the US Department of Energy, Ofce of Science, Ofce of
Basic Energy Sciences under Award DE-SC0000989.
1. Fialkowski M, et al. (2006) Principles and implementations of dissipative (dynamic)
self-assembly. J Phys Chem B 110(6):24822496.
2. Warren SC, Guney-Altay O, Grzybowski BA (2012) Responsive and nonequilibrium
nanomaterials. J Phys Chem Lett 3(15):21032111.
3. Grzybowski BA, Wilmer CE, Kim J, Browne KP, Bishop KJM (2009) Self-assembly: From
crystals to cells. Soft Matter 5(6):11101128.
4. Soh S, Byrska M, Kandere-Grzybowska K, Grzybowski BA (2010) Reaction-diffusion
systems in intracellular molecular transport and control. Angew Chem Int Ed Engl
49(25):41704198.
5. Whitesides GM, Grzybowski B (2002) Self-assembly at all scales. Science 295(5564):
24182421.
6. Glansdorff P, Prigogine I (1971) Thermodynamic Theory of Structure, Stability and
Fluctuations (Wiley-Interscience, London), p xxiii, 306 pp.
7. Katzir-Katchalsky A, Curran PF (1965) Nonequilibrium Thermodynamics in Biophysics
(Harvard Univ Press, Cambridge, MA), p x, 248 pp.
8. Kubo R, Toda M, Hashitsume N (1991) Statistical Physics (Springer, Berlin), 2nd Ed.
9. Andresen B, Berry RS, Ondrechen MJ, Salamon P (1984) Thermodynamics for processes
in nite-time. Acc Chem Res 17(8):266271.
10. Salamon P, Berry RS (1983) Thermodynamic length and dissipated availability. Phys
Rev Lett 51(13):11271130.
11. Crooks GE (1998) Nonequilibrium measurements of free energy differences for mi-
croscopically reversible Markovian systems. J Stat Phys 90(56):14811487.
12. Evans DJ, Cohen EGD, Morriss GP (1993) Probability of second law violations in
shearing steady states. Phys Rev Lett 71(15):24012404.
13. Evans DJ, Searles DJ (2002) The uctuation theorem. Adv Phys 51(7):15291585.
14. Jarzynski C (1997) Nonequilibrium equality for free energy differences. Phys Rev Lett
78(14):26902693.
15. Dorfman JR (1999) An Introduction to Chaos in Nonequilibrium Statistical Mechanics
(Cambridge Univ Press, Cambridge, UK), p xiv, 287 pp.
16. Gaspard P (1998) Chaos, Scattering, and Statistical Mechanics (Cambridge Univ Press,
Cambridge, UK), p xix, 475 pp.
17. Lebowitz JL (1980) 3rd International Conference on Collective Phenomena - Fore-
word. Ann NY Acad Sci 337(Jun):U1U2.
18. Wang J, Xu L, Wang EK (2008) Potential landscape and ux framework of non-
equilibrium networks: Robustness, dissipation, and coherence of biochemical oscil-
lations. Proc Natl Acad Sci USA 105(34):1227112276.
19. Wang J, Zhang K, Wang E (2010) Kinetic paths, time scale, and underlying landscapes:
A path integral framework to study global natures of nonequilibrium systems and
networks. J Chem Phys 133(12):125103.
20. Kolmogorov AN (1959) Entropy per unit time as a metric invariant of automorphisms.
Dokl Akad Nauk SSSR 124(4):754755.
21. Sinai I (1959) On the concept of entropy for a dynamic system. Dokl Akad Nauk SSSR
124(4):768771.
22. Eckmann JP, Ruelle D (1985) Ergodic-theory of chaos and strange attractors. Rev Mod
Phys 57(3):617656.
23. Gallavotti G, Cohen EGD (1995) Dynamical ensembles in nonequilibrium statistical
mechanicS. Phys Rev Lett 74(14):26942697.
24. Grzybowski BA, Stone HA, Whitesides GM (2000) Dynamic self-assembly of magne-
tized, millimetre-sized objects rotating at a liquid-air interface. Nature 405(6790):
10331036.
25. Klajn R, et al. (2007) Plastic and moldable metals by self-assembly of sticky nano-
particle aggregates. Science 316(5822):261264.
26. Klajn R, Bishop KJM, Grzybowski BA (2007) Light-controlled self-assembly of re-
versible and irreversible nanoparticle suprastructures. Proc Natl Acad Sci USA 104(25):
1030510309.
27. Wei YH, Han SB, Kim J, Soh SL, Grzybowski BA (2010) Photoswitchable catalysis
mediated by dynamic aggregation of nanoparticles. J Am Chem Soc 132(32):
1101811020.
28. Wales DJ, Berry RS (1991) Local interpretation of chaotic dynamics in a many-body
classical Hamiltonian system. J Phys At Mol Opt Phys 24(14):L351L357.
29. Fujisaka H (1983) Statistical dynamics generated by uctuations of local Lyapunov
exponents. Prog Theor Phys 70(5):12641275.
30. Green JR, Jellinek J, Berry RS (2009) Space-time properties of Gram-Schmidt vectors in
classical Hamiltonian evolution. Phys Rev E Stat Nonlin Soft Matter Phys 80(6 Pt 2):
066205.
31. Hinde RJ, Berry RS, Wales DJ (1992) Chaos in small clusters of inert-gas atoms. J Chem
Phys 96(2):13761390.
32. Costa AB, Green JR (2013) Extending the length and time scales of Gram-Schmidt
Lyapunov vector computations. J Comput Phys 246:113122.
33. Pesin YB (1976) Lyapunov characteristic indexes and ergodic properties of smooth
dynamic-systems with invariant measure. Dokl Akad Nauk SSSR 226(4):774777.
34. Evans DJ, Morriss GP (2008) Statistical Mechanics of Nonequilibrium Liquids (Cam-
bridge Univ Press, Cambridge, UK), 2nd Ed, p 314.
35. Hoover WG (1991) Computational Statistical Mechanics (Elsevier, Amsterdam), p xv,
313 pp.
36. Gaspard P, et al. (1998) Experimental evidence for microscopic chaos. Nature
394(6696):865868.
37. Stoddard SD, Ford J (1973) Numerical experiments on the stochastic behavior of
a Lennard-Jones gas system. Phys Rev A 8(3):15041512.
38. Benettin G, Galgani L, Giorgilli A, Strelcyn J-M (1980) Lyapunov characteristic ex-
ponents for smooth dynamical systems and for Hamiltonian systems; a method for
computing all of them. Part 1: Theory. Meccanica 15(1):920.
Green et al. PNAS | October 8, 2013 | vol. 110 | no. 41 | 16343
C
H
E
M
I
S
T
R
Y
Supporting Information
Green et al. 10.1073/pnas.1312165110
Fig. S1. Difference between sample means from consecutive trajectory segments for the KS entropy and energy of the system, scaled by appropriate bath
parameters, as a function of time. Plotted are dynamical entropies of the system and its total energy for the switch from 1 to 5 (A) and 1 to 15 (B). All
quantities decay to zero at equilibrium (dashed line), once self-assembly is complete. These results correspond to n = 200 total particles, 7 of which comprise the
assembling system, whereas the remainder is the explicit bath.
1
bath
and are the characteristic time and energy scales of the bath, respectively.
Green et al. www.pnas.org/cgi/content/short/1312165110 1 of 2
Fig. S2. Difference between sample means from consecutive trajectory segments for the KS entropy and energy of the system, scaled by appropriate bath
parameters, as a function of time. Plotted are dynamical entropies of the system, bath, and the total energy of the system for the switch from 1 to 10 (A) and
1 to 20 (B). All quantities decay to zero at equilibrium (dashed line), once self-assembly is complete.
1
bath
and are the characteristic time and energy scales
of the bath, respectively.
Green et al. www.pnas.org/cgi/content/short/1312165110 2 of 2