arXiv:1102.0103v1 [cond-mat.

stat-mech] 1 Feb 2011

Non-equilibrium Dynamics, Thermalization and

Entropy Production

Haye Hinrichsen1 , Christian Gogolin2 , and Peter Janotta1

1
Universität Würzburg, Fakultät für Physik und Astronomie 97074 Würzburg, Germany
2
Institute for Physics and Astronomy, Potsdam University, 14476 Potsdam, Germany

E-mail: [email protected]
STATPHYS-KOLKATA VII Conference Proceedings

Abstract. This paper addresses fundamental aspects of statistical mechanics such as the
motivation of a classical state space with spontaneous transitions, the meaning of non-
equilibrium in the context of thermalization, and the justification of these concepts from the
quantum-mechanical point of view. After an introductory part we focus on the problem of
entropy production in non-equilibrium systems. In particular, the generally accepted formula
for entropy production in the environment is analyzed from a critical perspective. It is shown
that this formula is only valid in the limit of separated time scales of the system’s and the
environmental degrees of freedom. Finally, we present an alternative simple proof of the
fluctuation theorem.

1. Introduction
Classical non-relativistic statistical physics is based on a certain set of postulates. Starting point
is a physical entity, called system, which is characterized by a set Ωsys of possible configurations.
Although this configuration space could be continuous, it is useful to think of it as a countable set
of discrete microstates s ∈ Ωsys . Being classical means that the actual configuration of the system
is a matter of objective reality, i.e. at any time t the system is in a well-defined configuration
s(t). The system is assumed to evolve in time by spontaneous transitions s → s′ which
occur randomly with certain transition rates wss′ > 0. This commonly accepted framework
is believed to subsume the emerging classical behavior of a complex quantum system subjected
to decoherence.
Although in a classical system the trajectory of microscopic configurations is in principle
measurable, an external observer is usually not able to access this information in detail. The
observer would instead express his/her partial knowledge in terms of the probability Ps (t) to
find the system at time t in the configuration s. In contrast to the unpredictable trajectory s(t),
the probability
P distribution Ps (t) evolves deterministically according to the master equation
Ṗs (t) = s′ ∈Ωsys (Ps′ (t)ws′ s − Ps (t)wss′ ).
Figure 1. Cartoon of a complex statistical system as a space Ωsys of configurations (red dots). At any time
the system is in one of the configurations and evolves by spontaneous transitions (indicated by arrows) selected
randomly according to specific transition rates.

A system is said to equilibrate if the probability distribution Ps (t) becomes stationary in

the limit t → ∞. In addition, if the probability currents Jss′ = Ps (t)wss′ and Js′ s =
Ps′ (t)ws′ s between all pairs of configurations s, s′ individually cancel each other in the stationary
state, the system is said to thermalize, obeying detailed balance. The predictive power of
equilibrium statistical mechanics relies on the fact that the stationary probability distribution
of a thermalized system is universal and can be classified into a small number of thermodynamic
ensembles. In particular, an isolated system thermalizes in such a way that each available
configuration is visited with the same probability. This fundamental equal a priori probability
postulate is at the core of equilibrium statistical mechanics, from which all other thermodynamic
ensembles can be derived.
Classical thermodynamics is concerned with time-dependent phenomena close to equilibrium.
Introducing notions such as thermodynamic forces, potentials and currents it makes statements
how such systems relax towards thermal equilibrium. Contrarily, non-equilibrium statistical
physics deals with systems that do not thermalize, meaning that their probability currents do
not vanish even in the stationary state. Typically such systems need to be driven from the
outside in order to prevent them from thermalizing.
The basic framework of statistical physics and thermodynamics sketched above is often taken
for granted. However, one should be aware that the underlying postulates are highly non-
trivial. In this paper we address some of these issues from a critical perspective. Where does
the cartoon of spontaneous hopping between configurations come from? How can we be sure
that the system will equilibrate into a stationary state? What is known about thermalization
and the justification of the equal a priori probability postulate? What is the meaning of non-
equilibrium in an environment that thermalizes? What is behind the commonly accepted formula
for the entropy production of non-equilibrium systems? In the following we make an attempt to
discuss some of these questions in a common context, giving partial answers and pointing out
open questions. Moreover, we would like to draw the readers attention to the fascinating links
between classical statistical physics and recent developments in quantum information science.
The paper is organized as follows. In the following section we summarize existing textbook
knowledge on stochastic system in a compressed form suitable for beginners in the field. Sect. 3
deals with the question how we can justify the basic assumptions and postulates of classical
statistical physics from the quantum perspective. In Sect. 4 we will discuss the problem of
entropy production in non-equilibrium systems, analyzing the conditions under which commonly
accepted formulas are valid. A discussion of the fluctuation theorem together with an alternative
compact proof is presented in Sect. 5. The paper closes with concluding remarks in Sect. 6.
2. Setup of classical statistical physics

Configuration space and dynamics

As outlined in the introduction, classical statistical physics is mainly concerned with models of
complex systems having the following properties:

(i) The system is characterized by a certain set Ωsys of configurations s ∈ Ωsys , also called
microstates. Usually the configuration space is implicitly specified by the definition of
a model. For example, in a reaction-diffusion model this space is the set of all possible
particle configurations while in a growth model the microstates are identified with the
possible configurations of an interface.
(ii) The states are classical, i.e., at any time the system is in one particular configuration s(t).
(iii) The system evolves randomly by instantaneous transitions s → s′ occurring spontaneously
with certain transition rates wss′ ≥ 0. In numerical simulations, this dynamics is
approximated by random-sequential update algorithms.

Starting with an initial configuration s0 the system evolves randomly through an unpredictable
sequence of configurations s0 → s1 → s2 → . . . by instantaneous transitions. These transition
take place at certain points of time t1 , t2 , . . . which are distributed according to a Poisson
distribution like shot noise. Such a sequence of transitions is called a stochastic path.
Although the actual stochastic path of the system is unpredictable, the probability Ps (t) to
find the system in configuration s at time t evolves deterministically according to the master
equation
d X 
Ps (t) = Js s (t) − Jss (t) ,
′ ′ (1)
dt ′ s ∈Ω

where
Jss′ (t) = Ps (t)wss′ (2)
is the probability current flowing from configuration s to configuration s′ . The system is said

• to be stationary or equilibrated if the probability distribution Ps (t) is time-independent,

meaning that for a given configuration s all incoming and outgoing probability currents
cancel.
• to equilibrate if the master equation evolves into a stationary distribution in the limit t → ∞,
denoted as Ps = Ps (∞).

For simplicity we will assume that this stationary state is unique and independent of the initial
state. Note that systems with a finite configuration space always relax into a stationary state
while for systems with an infinite or continuous configuration space a stationary state may not
exist. Moreover, we will assume that the dynamics of the system under consideration is ergodic,
i.e., the network of transitions is connected so that each configuration can be reached.
Detailed balance
A stationary system is said to thermalize if it evolves into a stationary state which obeys detailed
balance. This means that the probability currents between all pairs of configurations cancel, i.e.

Jss′ = Js′ s ∀s, s′ , (3)

or equivalently
wss′ Ps′
= ∀s, s′ , (4)
ws′ s Ps
where Ps is the stationary probability distribution. It is worth being noted that detailed balance
can be defined in an alternative way without knowing the stationary probability distribution.
To see this let us consider a closed loop of three transitions s1 → s2 → s3 → s1 . For these
transitions Eq. (4) provides a system of three equations. By multiplying all equations one can
eliminate the probabilities Psi , arriving at the condition ws1 s2 ws2 s3 ws3 s1 = ws1 s3 ws3 s2 ws2 s1 . A
similar result is obtained for any closed loop of transitions, hence the condition of detailed
balance can be recast as Y Y
wsi si+1 = wsi+1 si (5)
i i

for all closed loops in Ωsys . A system with this property is said to have a balanced dynamics.
Note that in a system with balanced dynamics we may rescale a pair of opposite rates
by wss′ → Λwss′ and ws′ s → Λws′ s without breaking the detailed balance condition. This
intervention changes the dynamics of the model and therewith its relaxation, but the stationary
state (if existent and unique) remains the same. This statement holds even if pairs of opposite
rates are set to zero as long as this manipulation does not break ergodicity.

Equal a priori postulate

Classical statistical physics is based on the so-called equal a priori postulate. This postulate
states that an isolated system, which does not interact physically with the outside world,
will thermalize into a stationary state where all accessible configurations occur with the same
probability, i.e.
Ps = const = 1/|Ωsys | . ∀s ∈ Ωsys (6)
It is assumed that this state obeys detailed balance since otherwise one could use the non-
vanishing probability currents to construct a perpetuum mobile. The immediate consequence
would be that all transitions are reversible and that opposite rates coincide, i.e. wss′ = ws′ s .

Entropy
Entropy is probably the most fundamental concept of statistical physics. From the information-
theoretic point of view, the entropy of a system is defined as the amount of information (measured
in bits) which is necessary to describe the configuration of the system. Since the description of
a highly ordered configuration requires less information than a disordered one, entropy can be
viewed as a measure of disorder.
The amount of information which is necessary to describe a configuration depends on the
already existing partial knowledge of the observer at a given time. For example, deterministic
systems with a given initial configuration have no entropy because the observer can compute
the entire trajectory in advance, having complete knowledge of the configuration as a function
of time even without measuring it. Contrarily, in stochastic systems the observer has only a
partial knowledge about the system expressed in terms of the probability distribution Ps (t). In
this situation the amount of information needed to specify a particular configuration s ∈ Ωsys is
− log2 Ps (t) bits, meaning that rare configurations have more entropy than frequent ones.
Different scientific communities define entropy with different prefactors. In information
science one uses the logarithm to base 2 so that entropy is directly measured in bits.
Mathematicians instead prefer a natural logarithm while physicists are accustomed to put an
historically motivated prefactor kB in front, giving entropy the unit of an energy. In what follows
we set kB = 1, defining the entropy of an individual configuration s as

Ssys (t, s) = − ln Ps (t) . (7)

Since this entropy depends on the actual configuration s, it will fluctuate along the stochastic
path. However, its expectation value, expressing the observer’s average lack of information,
evolves deterministically and is given by
X
Ssys (t) = hSsys (t, s)is = − Ps (t) ln Ps (t) , (8)
s∈Ωsys

where h. . .i denotes the ensemble average over independent realizations of randomness. Apart
from the prefactor, this is just the usual definition of Shannon’s entropy [1, 2].
Up to this point entropy is just an information-theoretic concept for the description of
configurations. The point where entropy takes on a physical meaning is the equal a priori
postulate, stating that an isolated system thermalizes in such a way that the entropy takes the
largest possible value Ssys = ln |Ωsys |. As is well-known, all other thermodynamic ensembles can
be derived from this postulate.
The numerical determination of entropies is a nontrivial task because of the highly non-linear
influence of the logarithm. To measure an entropy numerically, one first has to estimate the
probabilities Ps (t). The resulting symmetrically distributed sampling errors in finite data sets
are amplified by the logarithm, leading to a considerable systematic bias in entropy estimates.
Various methods have been suggested to reduce this bias on the expense of the statistical error,
see e.g. [3, 4].

Subsystems
In most physical situations the system under consideration is not isolated, instead it interacts
with the environment. In this case the usual approach of statistical physics is to consider the
system combined with the environment as a composite system. This superordinate total system
is then assumed to be isolated, following the same rules as outlined above. To distinguish the
total system from its parts, we will use the suffixes ‘tot’ for the total system while ‘sys’ and ’env’
refer to the embedded subsystem and its environment, respectively.
The total system is characterized by a certain space Ωtot of classical configurations c ∈ Ωtot
(not to be confused with system configurations s ∈ Ωsys ). The number of these configurations
may be enormous and they are usually not accessible in experiments, but in principle there
should be a corresponding probability distribution Pc (t) evolving by a master equation

d X  
Pc (t) = Jc′ c (t) − Jcc′ (t) , Jcc′ (t) = Pc (t)wcc′ (9)
dt
c′ ∈Ωtot

with certain time-independent transition rates wcc′ ≥ 0.

Figure 2. A subsystem is defined by a projection π which maps each state s ∈ Ωtot of the total system (left)
onto a particular state c ∈ Ωsys of the subsystem (right), dividing the state space of the total system into sectors.
The figure shows a stochastic path in the total system together with the corresponding stochastic path in the
subsystem.

Let us now consider an embedded subsystem. Obviously, for every classical configuration
c ∈ Ωtot of the total system we will find the subsystem in a well-defined unique configuration
s ∈ Ωsys . Conversely, for a given configuration of the subsystem s ∈ Ωsys the environment (and
therewith the total system) can be in many different states. This relationship can be expressed
in terms of a surjective map π : Ωtot → Ωsys which projects every configuration c of the total
system onto the corresponding configuration s of the subsystem, as sketched in schematically
Fig. 2.
The projection π divides the space Ωtot into sectors π −1 (c) ⊂ Ωtot which consist of all
configurations which are mapped onto the same s. Therefore, the probability to find the
subsystem in configuration s ∈ Ωsys is the sum over all probabilities in the corresponding sector,
i.e. X
Ps (t) = Pc (t) , (10)
c(s)

where the sum runs over all configurations c ∈ Ωtot with π(c) = s. Likewise, the projected
probability current Jss′ in the subsystem flowing from configuration s to configuration s′ is the
sum of all corresponding probability currents in the total system:
XX X X
Jss′ (t) = Jcc′ (t) = Pc (t) wcc′ . (11)
c(s) c′ (s′ ) c(s) c′ (s′ )

This allows us to define effective transition rates in the subsystem by

P P
Jss′ (t) c(s) Pc (t) c′ (s′ ) wcc′
wss′ (t) = = P . (12)
Ps (t) c(s) Pc (t)

In contrast to the transition rates of the total system, which are usually assumed to be constant,
the effective transition rates in the subsystem may depend on time. With these time-dependent
rates the subsystem evolves according to the master equation

d X  
Ps (t) = Js′ s (t) − Jss′ (t) Jss′ (t) = Ps (t)wss′ (t) . (13)
dt
s′ ∈Ωsys

From the subsystems point of view this time dependence reflects the unknown dynamics in the
environment. Moreover, ergodicity plays a subtle role: Even if the dynamics of the total system
was ergodic, the dynamics within the sectors π −1 (c) is generally non-ergodic and may decompose
into several ergodic subsectors. As we will see in Sect. 4, this allows the environmental entropy
to increase even if the subsystem is stationary.

Systems far from thermal equilibrium

In Nature many systems are not thermalized but out of equilibrium. For this reason the study
of non-equilibrium systems plays an increasingly important role in statistical physics. It should
be noted that different communities use term ‘non-equilibrium’ in a different way. In the
context of thermodynamics, non-equilibrium refers to non-stationary situations close to thermal
equilibrium, while statistical physicists use this term for systems violating detailed balance. In
the following we use the nomenclature that a system is

• in thermal equilibrium if its probability distribution is stationary obeying detailed balance.

• thermalizing if it relaxes towards thermal equilibrium with balanced rates obeying Eq. (5).
• in a non-thermal equilibrium if its probability distribution is in a stationary state without
satisfying detailed balance.
• in out of thermal equilibrium if it is non-stationary violating Eq. (5).

Since in Nature isolated systems are expected to thermalize, we can conclude that conversely
a non-thermalizing system must always interact with the environment. This means that an
external drive is needed to prevent the system from thermalizing, maintaining its non-vanishing
probability currents. On the other hand, the total system composed of laboratory system and
environment should thermalize. This raises the question how a thermalizing total system can
contain a non-thermalizing subsystem?
The answer to this question is given in Eq. (12). Even if the total system was predetermined
to thermalize, meaning that the rates wcc′ obey Eq. (5), one can easily show that the effective
rates wss′ of transitions in the subsystem are generally not balanced. Therefore, a thermalizing
‘Universe’ may in fact contain subsystem out of thermal equilibrium. The apparent contradiction
is resolved by the observation that the projected rates wss′ (t) depend on time: Although these
rates may initially violate detailed balance, they will slowly change as the ‘Universe’ continues to
thermalize, eventually converging to values where they do obey detailed balance. This process
reflects our everyday experience that any non-thermalizing system will eventually thermalize
when the external drive runs out of power.

3. Justification from the quantum perspective

The standard setup of statistical mechanics as described in section 2 is amazingly successful
in explaining a wide range of physical processes. In stark contrast to this strong justification
by corroboration the question of whether and how it can be justified microscopically is still
open to a great extent. Over the course of the last century many famous physicists, such as
Ludwig Boltzmann, have tried to derive statistical physics from Newtonian mechanics. Despite
major efforts no fully convincing and commonly accepted microscopic foundation for statistical
physics could be found [5]. Nowadays, quantum mechanics is the commonly accepted microscopic
theory but a complete justification of statistical mechanics from quantum mechanics is yet to
be achieved. This is surprising as already the founding fathers of quantum theory have worked
on this problem [6].
Recently, stimulated by new experiments [7,8] and novel methods from quantum information
theory, a renewed interest in such fundamental questions ignited a flurry of activity in this field
with significant new insights, leading to a reconsideration of the axiomatic system of statistical
mechanics. As we shall see below, quantum mechanics and statistical physics apparently
contradict each other but, very surprisingly, at the same time some genuine features of quantum
mechanics can be used to justify essential assumptions of statistical mechanics. In this Section
we discuss some of these recent developments, which we think are of interest to a wider audience,
in a non-technical fashion. Readers not interested in the relationship of quantum mechanics and
statistical physics can safely skip this section.

Emergence of the classical state space

One aspect of the quantum mechanical foundations of statistical mechanics is to explain why
macroscopic systems can be well described without considering the quantum mechanical nature
at the microscopic level. In the following we discuss a general mechanism that explains how
classical ensembles can emerge from the microscopic quantum dynamics.
As an example for an intrinsically quantum mechanical process that can be successfully
described within the framework of statistical mechanics, let us consider the emission and
absorption of light (photons) by atoms. This example includes both equilibrium situations,
like the interaction of atoms with black body ration, and extreme nonequilibrium situation, like
the pumping of a laser. The statement that emission and absorption of light can be described
within the framework of statistical mechanics is to be understood in the following sense: quantum
mechanics enters the description only in so far as it defines the energy levels of the atom. The
relevant energy eigenstates of the quantum Hamiltonian are then taken to be the configurations s
that constitute the classical state space Ωsys of the atom while the dynamics is entirely described
in terms of transition rates between these levels, corresponding to absorption, spontaneous
emission, and stimulated emission. Although this description turned out to be useful, it is
not at all clear why this simplified, classical treatment of-light matter interaction is eligible.
As an essential feature, quantum mechanics allows systems to be in coherent superpositions of
energy eigenstates rather than probabilistic superpositions. Why is it possible to simply ignore
this fundamental feature of quantum mechanics and work with a classical description that only
includes incoherent, probabilistic superpositions of energy eigenstates and transitions between
them?
The first step in resolving this problem is to realize that it is in general impossible to
completely isolate a system from its environment. By tracing out the degrees of freedom in
the environment it is easy to show that the time evolution of such an interacting system is not
unitary anymore. In particular, the interaction with the environment can suppress coherent
superpositions in the system, a process called decoherence [9–11]. The insights obtained in the
field of decoherence theory have proved to be extremely valuable, for both applications and from
a fundamental perspective. However, most of the results are restricted to specific models or rely
on assumptions such as a special form of the interaction or on approximations that are hard to
control. Given the broad applicability of statistical mechanics and thermodynamics one would
rather want to have an explanation for the emergence of classical state spaces from quantum
theory that does not rely on such details.
Recently, building on earlier works [12, 13], a result in this spirit was obtained in Ref. [14].
It can be summarized informally as follows. Whenever a system interacts weakly with an
environment its quantum state (density operator) is close to being in a purely probabilistic
superposition of energy eigenstates for most times. The result shows that quantum mechanics
implies a generic mechanism leading to a suppression of coherent superpositions in such a way
that an observer measuring the system at an arbitrary point in time is most likely to find the
system in a state where only classical probabilistic superpositions contribute significantly. In
other words, the coherent quantum dynamics of system plus environment suppresses coherence
 Time reversal objection (Loschmidt):

t t
−→ ~v → −~v −→

Recurrence objection (Ponicaré):

t t t t
−→ . . . −→ −→ . . . −→

Figure 3. Schematic depiction of the two objections against irreversibility in systems based on Newtonian
mechanics. According to Loschmidt’s time reversal objection, for every process that increases the entropy of a
system, say the spreading of particles in a container after removal of a wall, there exists a corresponding time-
reversed process, which would reduce the entropy of the system. According to Ponicaré’s recurrence objection,
whenever the entropy of a system increased, the recurrent nature of Newtonian evolution implies that the entropy
has to decrease after a sufficiently long period of time. The two objections illustrate that the Second Law in its
usual form is incompatible with Newtonian mechanics. Equivalent objections apply to quantum mechanics which
is also time reversal invariant and recurrent.

in the subsystem. The result can be expressed as a rigorous inequality bounding the off-diagonal
entries in the density matrix of the subsystem. Its proof does not rely on any special properties
of the interaction, it rather follows directly from the full unitary evolution of the total system
without any approximations. The inequality is meaningful only if the interaction with the
environment is weak, meaning that it must be smaller (or at least not significantly larger) than
the gaps between the energy levels of the system.
Can this result explain why the energy eigenbasis is the correct choice for the state space
in the classical description of an atom absorbing and emitting light? Thinking of the atom
plus the surrounding light field as a quantum system subjected to decoherence by interaction
with other atoms of the gas it does. In fact, the energy gaps of a few eV are several orders of
magnitude larger than the thermal energy of the atoms (about 25 meV at room temperature),
which sets the relevant energy scale for the coupling to other atoms of the gas. The weak
perturbation caused by thermal scattering in the gas naturally leads to a decoherence into the
energy eigenbasis, meaning that the classical description in terms of a probability Ps (t) to be in
the energy eigenstate s at time t describes the state of the system almost completely.

Equilibration and apparent irreversibility

Another important issue is an apparent incompatibility between quantum mechanics and
statistical physics: While the Second Law, a corner stone of statistical physics, postulates the
irreversible convergence of Ps (t) to a stable equilibrium distribution in the limit t → ∞, the
unitary dynamics of isolated quantum systems is time reversal invariant and recurrent; meaning
that for every initial state the system will eventually return to an almost indistinguishable
state after a certain (possibly extremely large) recurrence time. The famous objections brought
forward by Loschmidt and Ponicaré against Boltzmann’s H-Theorem, which was intended to
be a proof of the Second Law for Newtonian mechanics, the so called time reversal objection
and the recurrence objection (see Fig. 3) equally apply if quantum mechanics is taken to be the
underlying microscopic theory.
These two objections show that true equilibration in the usual sense of statistical mechanics
is impossible in quantum theory. The origin of the tendency to evolve towards equilibration,
which constitutes a crucial part of the framework of statistical physics, on first sight, appears
to be miraculous from the quantum perspective. However, recent studies show a way how it
is possible to resolve this apparent contradiction. All we can hope for is to find equilibration
in a weaker sense, meaning that the system is almost equilibrated for most times. That is, the
expectation values of a set of relevant observables could evolve towards and then stay close to a
certain equilibrium value for most times during the evolution, even if the system started far from
equilibrium. The notion of equilibration in classical statistical mechanics that comes closest to
the quantum version of equilibrium seems to be what is called a non-thermal equilibrium above.
Of particular interest are observables acting on a small part of a larger composite system.
Surprisingly, as shown in a series of recent papers [12, 15, 16], it is possible to rigorously
prove equilibration for almost all times for such observables from the unitary, time reversal
invariant, and recurrent time evolution of quantum mechanics. More specifically, whenever the
dimension of the Hilbert space of a subsystem is much smaller than a quantity called the effective
dimension of the initial state, it is shown that the unitary dynamics of the full system is such
that for most times the states of all small subsystems are practically indistinguishable from
apparent equilibrium states. The proof requires that the Hamiltonian is fully interactive and
non-degenerate, i.e., it does not decompose into non-interacting subsystems, – a very weak and
quite natural assumption.
The effective dimension, which is defined as deff (ρ) = 1/Trρ2 , measures how many eigenstates
of the Hamiltonian contribute significantly to the initial state. Can we expect it to be large in
realistic situations? In realistic many particle systems energy is approximately extensive. Even
if we assume it to grow at most polynomially with the number of constituents, then an energy
interval of fixed width will usually contain exponentially many energy eigenstates. This implies
that even states with a very small energy uncertainty will usually consist of exponentially many
energy eigenstates and thus it is safe to assume that the effective dimension will be very large
in realistic situations. Thus we can conclude that equilibration of small subsystems in large,
interacting quantum systems is a generic property. Similar results can be obtained for certain
sets of global observables [15, 16].

Maximum entropy principle

In his seminal papers [1, 2] E. T. Jaynes suggested a maximum entropy principle as a possible
foundation for statistical mechanics. In short, Jaynes argues that the correct method to calculate
expectation values of observables that give only limited knowledge about a system is to take
the state with maximum entropy among the configurations that are compatible with our partial
knowledge. His argument is based on information theoretic considerations of the method of
statistical inference.
Surprisingly, the unitary dynamics of quantum mechanics does also naturally imply a
maximum entropy principle that is quite similar in style. Recently it was shown in Ref. [17]
that whenever an expectation value of some observable equilibrates in the sense defined above,
it equilibrates towards the expectation value it would have in the state that maximizes the von
Neumann entropy among the states that have the same expectation values for all conserved
quantities, i.e., of all observables that commute with the Hamiltonian. Again the result is
derived without making any special assumptions and without approximations. In contrast to
the common approach in which the maximum entropy principle appears as an axiom, this
result follows directly from first principles when we chose a microscopic quantum description.
What makes the maximum entropy principle from quantum dynamics different form the usual
Jaynes’ principle is that the number of conserved quantities of a quantum system grows with the
dimension of the Hilbert space and thus exponentially with the system size. Contrarily, Jaynes
formulated his maximum entropy principle having situations with partial knowledge about a
handful of natural physical observables in mind. How exactly these two maximum entropy
principles are related is yet to be explored in more detail.

Thermalization
One of the most important applications of equilibrium statistical physics is to calculate the
properties of systems at a well defined temperature.The standard assumption going into these
calculations is that the state of such a system is described by a Gibbs state. The Gibbs state
and the canonical ensemble can be derived from the equal a priory probability postulate under
certain assumptions about the density of states of a bath with which the system can exchange
energy. Alternatively it is possible to justify the Gibbs state by using Jaynes’ maximum entropy
principle, showing that the Gibbs state is the state that maximizes the conditional entropy given
a fixed energy expectation value. However, it remains unclear how, and under which conditions,
subsystems of quantum systems actually thermalize, by which, in this section, we mean that it
equilibrates towards a Gibbs states with a well defined temperature. Note that it is not easy to
relate the notion of thermalization we use in this section to the detailed balance condition used
throughout the rest of this article.
Earlier works attempting to solve this problem [18–20] either rely on certain unproven
hypotheses such as the so-called eigenstate thermalization hypothesis, or they are restricted
to quite special situations such as coupling Hamiltonians of a special form, or they merely prove
typicality arguments instead of dynamical relaxation towards a Gibbs state. Although the results
obtained in these papers are very useful and have significantly improved our understanding of
the process of thermalization, they do not yet draw a complete and coherent picture.
An attempt to settle the question of thermalization will be made in a forthcoming paper [21].
As discussed above we already know conditions under which we can rigorously guarantee
equilibration [12, 15, 16]. What remains to be done is to identify a set of conditions under
which one can guarantee that the equilibrium state of a subsystem is close to a Gibbs state.
By using a novel perturbation theory argument and carefully bounding all the errors in an
approximation similar to that of [20] one can indeed identify such a set of sufficient (and more
or less necessary) conditions, that can be summarized in a non-technical way as follows:

(i) The energy content and the Hilbert space dimension of the bath must be much larger than
the respective quantities of the system.
(ii) The coupling between them must be strong enough, in particular much stronger than the
gaps of the decoupled Hamiltonian. This ensures that the eigenbasis of the full Hamiltonian
is sufficiently entangled (a lack of entanglement provably prevents thermalization [17]).
(iii) At the same time, the coupling must be weak in the sense that it is much smaller than the
energy uncertainty of the initial state. This is a natural counterpart to the weak coupling
assumption known from classical statistical mechanics. item[(iv)] The energy uncertainty
of the initial state must be small compared to the energy content and at the same time
large compared to the level spacing. Moreover, the energy distribution must satisfy certain
technical smoothness conditions.
(v) The spectrum of the bath must be well approximable by an exponential on the scale of
the energy uncertainty and the density of states must grow faster than exponential. This
property of the bath is ultimately the reason for the exponential form of the Gibbs state and
is also required in the classical derivation of the canonical ensemble. Most natural many
particle systems have this property.
Figure 4. Entropy production: A non-thermalizing system cannot exist on its own but must be driven from
the outside. The external drive, that keeps the system away from thermal equilibrium, inevitably increases the
entropy in the environment.

In summary, one can say that more or less the same conditions that are used in the classical
derivation of the canonical ensemble appear naturally in the proof of dynamical thermalization.

Time scales
The most important open problem for the approach described above is that rigorous bound on
the time scales for decoherence/equilibration/thermalization are not yet known. The results
derived in [12–17] only tell us that decoherence/equilibration/thermalization must eventually
happen under the given conditions, but they do not tell us how long it takes. In general this
seems to be tough question, but for exactly solvable models the time scales can be derived [22].

4. Entropy production
Returning to the classical framework, let us now study the problem of entropy production. As
outlined in Sect. 2, thermalizing systems (i.e. systems with balanced rates relaxing into thermal
equilibrium) can contain subsystems which are out of thermal equilibrium in the sense that
the transition rates wss′ do not obey detailed balance. The apparent contradiction is resolved
by observing that the effective rates in the subsystem are generally time-dependent and will
eventually adjust in such a way that the subsystem thermalizes as well. However, for a limited
time it is possible to keep them constant in such a way that they violate detailed balance. This
is exactly what happens in experiments far from equilibrium – typically they rely on external
power and will quickly thermalize as soon as power is turned off.
The external drive which is necessary to keep a subsystem away from thermal equilibrium
will on average increase the entropy in the environment, as sketched in Fig. 4. In the following
we discuss various attempts to quantify this entropy production.

Entropy changes
For a subsystem embedded in an environment we distinguish three types of configurational
entropies, namely, the configurational entropy of the total system (‘Universe’), the entropy of
the subsystem (experiment) and the entropy in its environment:

Stot (t, c) = − ln Pc (t) , (14)

Ssys (t, s) = − ln Ps (t) , (15)
Senv (t, c) = Stot (t, c) − Ssys (t, π(c)) . (16)
Averaging over many realizations the corresponding mean entropies are given by
X
S̄tot (t) = hStot (t, c)ic = − Pc (t) ln Pc (t) , (17)
c∈Ωtot
X
S̄sys (t) = hSsys (t, s)is = − Ps (t) ln Ps (t) , (18)
s∈Ωsys

S̄env (t) = hSenv (t, c)is = Stot (t) − Ssys (t) . (19)

The time derivative of these averages is given by

d X Pc (t)
S̄tot (t) = Jcc′ (t) ln , (20)
dt Pc′ (t)
c,c′ ∈Ωtot
d X Ps (t)
S̄sys (t) = Jss′ (t) ln , (21)
dt Ps′ (t)
s,s′ ∈Ωsys

where we used the master equations (13) and (9).

Let us now consider a temporal regime in which the rates of the subsystem can be considered
as constant. In this case the subsystem, which is often small compared to the environment, may
quickly reach a non-thermalized stationary state, which is often referred to as non-equilibrium
steady state (NESS) in the literature. In this case the average entropy of the subsystem will
saturate whereas the total system thermalizes according to the Second Law, meaning that the
average entropy in the environment increases. This environmental entropy production is the
price Nature has to pay for keeping a subsystem away from thermal equilibrium.
To be more specific, let us now assume that the total system follows a particular stochastic
path Γ : t 7→ c(t)

Γ : c0 → c1 → c2 → . . . at times t0 , t1 , t2 , . . . (22)

Whenever π(ci ) 6= π(ci+1 ) a transition in the total system implies a transition in the subsystem,
as sketched in Fig. 2. Denoting the corresponding transition times by tni , the projected stochastic
path of the subsystem γ = π[Γ] reads

γ : s0 → s1 → s2 → . . . at times tn0 , tn1 , tn2 , . . . (23)

where si = π(cni ). Along their respective stochastic paths the configurational entropies of the
total system and the subsystem are given by
Γ
Stot (t) = − ln Pc(t) (t) , (24)
γ
Ssys (t) = − ln Ps(t) (t) . (25)

How do these quantities change with time? Following Ref. [23] the temporal evolution of the
configurational entropy is made up of a continuous contribution caused by the deterministic
evolution of the master equation and a discontinuous contribution occurring whenever the system
hops to a different configuration. This means that the time derivative of the systems entropy is
given by
d γ Ṗs(t) (t) X Psj (t)
Ssys (t) = − − δ(t − tnj ) ln . (26)
dt Ps(t) (t) Psj−1 (t)
j
Similarly, the total entropy of the ‘Universe’ is expected to change as

d Γ Ṗc(t) (t) X Pcn (t)

Stot (t) = − − δ(t − tn ) ln (27)
dt Pc(t) (t) n
Pcn−1 (t)

so that the environmental entropy production is given by their difference:

d Γ d γ d Γ
Senv (t) = Ssys (t) − Stot (t) . (28)
dt dt dt
This formula is exact but useless from a practical point of view because the actual stochastic
trajectory Γ of the total system (the whole ‘Universe’) is generally not known.

Effective environmental entropy production

Based on previous work by Andrieux and Gaspard [24], Seifert suggested a very compact formula
for the effective entropy production in the environment caused by the embedded subsystem [23]:

d γ X ws s (t)
Senv (t) = − δ(t − tnj ) ln j j+1 . (29)
dt wsj+1 sj (t)
j

This formula tells us that each transition s → s′ in the subsystem causes an instantaneous
change of the environmental entropy by the log ratio of the forward rate wss′ divided by the
backward rate ws′ s . Together with Eq. (26) this formula would imply that the total entropy
changes according to

d γ Ṗs(t) (t) X Psj (t)wsj sj+1 (t)

Stot (t) = − − δ(t − tnj ) ln . (30)
dt Ps(t) (t) Psj−1 (t)wsj+1 sj (t)
j

This expression differs significantly from the exact formula (27) so that it can be only meaningful
in an effective sense under certain conditions or in a particular limit.
Before discussing these underlying assumptions in detail, we like to note that Eq. (29) is
indeed very elegant. It does not require any knowledge about the nature of the environment,
instead it depends exclusively on the stochastic path of the subsystem and the corresponding
transition rates. Moreover, this quantity can be computed very easily in numerical simulations:
Whenever the program selects the move s → s′ , all what has to be done is to increase the
environmental entropy variable by ln(wss′ /ws′ s )1 . Note that the logarithmic ratio of the rates
requires each transition to be reversible.
To motivate formula (29) heuristically, Seifert argues the corresponding averages of the
entropy production reproduce a well-known result in the literature. More specifically, he shows
that Eq. (29) averaged over many possible paths gives the expression.
d X wss′ (t)
S̄env (t) = Jss′ (t) ln . (31)
dt ws′ s (t)
s,s′ ∈Ω sys

Combined with Eq. (21) one obtains the average entropy production in the total system
d X Ps (t)wss′ (t)
S̄tot (t) = Jss′ (t) ln . (32)
dt Ps′ (t)ws′ s (t)
s,s′ ∈Ω sys

1
In order to avoid unnecessary floating point operations, it is useful to store all possible log ratios of the rates
in an array.
This formula was first introduced by Schnakenberg [25] and has been frequently used in chemistry
and physics [26]. It is in fact very interesting to see how Schnakenberg derived this formula. As
described in detail in Appendix A, he considered a fictitious chemical system of homogenized
interacting substances which resemble the dynamics of the master equation in terms of particle
concentrations. Applying standard methods of thermodynamics, he was able to prove Eq. (32).
The rational behind this derivation is to assume that the environment is always close to thermal
equilibrium.

Limit of fast thermalization in the environment

In the following we show that the entropy production formula is correct in the limit where
the environment equilibrates immediately whenever a transition occurs in the subsystem. This
requires a separation of time scales of the internal dynamics of the subsystem on the one hand
and the relaxation in the environment on the other.
As shown in Sect. 2, the projection π : Ωtot → Ωsys divides the configuration space of
the total system into sectors of configurations c mapped onto the same s (see Fig. 2). It is
important to note that even if the dynamics of the total system was ergodic, the dynamics
within these sectors is generally non-ergodic, meaning that they may split up into various
ergodic subsectors. For example, if the subsystem starting in a certain configuration returns
to the same configuration after some time, the corresponding subsector may have changed,
reflecting the change of entropy in the environment. For a given stochastic path of the small
s(t)
system, we will denote the corresponding subsector as Ωtot ⊂ Ωtot .
Let us now assume that the environmental degrees of freedom thermalize almost instantan-
eously, reaching maximal entropy within the actual subsector. This means that the system
quickly reaches a state where the probabilities
s(t)

Pc (t) = Ps (t)/Ns (t) if c ∈ Ωtot (33)
0 otherwise.
s(t)
are constant on Ωtot . This implies that in the formula (28) for the environmental entropy
production, namely

d Γ Ṗc(t) (t) Ṗs(t) (t) X Psj (t) X Pcn (t)

Senv (t) = − + δ(t − tnj ) ln − δ(t − tn ) ln , (34)
dt Pc(t) (t) Ps(t) (t) Psj−1 (t) n
Pcn−1 (t)
j

the first two terms cancel. Therefore, this expression reduces to

d Γ X Psj (t) X Pπ(cn ) (t)Nπ(cn−1 ) (t)

Senv (t) = δ(t − tnj ) ln − δ(t − tn ) ln . (35)
dt Psj−1 (t) n
Pπ(cn−1 ) (t)Nπ(cn ) (t)
j

Obviously, only those terms in the second sum will contribute where π(cn−1 ) 6= π(cn ), i.e. where
n = nj , hence the sum can be reorganized as

Psj (t) Psj (t)Nsj−1 (t) Nsj (t)

 
d γ X X
Senv (t) = δ(t − tnj ) − ln = δ(t − tnj ) ln (36)
dt Psj−1 (t) Psj−1 (t)Nsj (t) Nsj−1 (t)
j j

which now depends only on the stochastic path γ of the subsystem. This result, saying that
the entropy increase is given by the logarithmic ratio of the number of available configurations,
is very plausible under the assumption of instantaneous thermalization of the environment. It
Figure 5. Configurational entropy of an isolated system, its temporal derivative and the corresponding
probability distribution of entropy differences.

remains to be shown that this ratio is related to the ratio of the effective rates. In fact, inserting
(33) into (12) we obtain
P P P P
c(s) Pc (t) c′ (s′ ) wcc′ c(s) c′ (s′ ) wcc′
wss′ (t) = P = , (37)
c(s) Pc (t) Ns (t)

hence wss′ /ws′ s = Ns′ /Ns . Inserting this relationship into Eq. (36) we arrive at the formula for
the effective entropy production (29). This proves that this formula is valid under the assumption
that the environmental degrees of freedom thermalize instantaneously after each transition of
the subsystem.

Conjecture for systems in a non-thermalized environment

What happens if the environment does not thermalize immediately after each transition of
the subsystem? To answer this question we performed numerical simulations of a three-state
system in an environment with 50 configurations and randomly chosen transition rates. The
results (not shown here) suggest that the entropy production predicted by formula (29) deviates
from the true entropy production in both directions. However, averaging over many independent
stochastic paths while keeping the rates fixed, the formula seems to systematically underestimate
the actual entropy production. This leads us to the conjecture that formula (29) may serve as
a lower bound for the expectation value of the entropy production.

5. Fluctuation theorem revisited

The Second Law of thermodynamics tells us that the entropy Stot of an isolated system
increases on average during thermalization, approaching the maximal value Stot = ln |Ωtot |.
The fluctuation theorem generalizes this statement by studying the actual fluctuations of the
total entropy, including the Second Law as a special case [27–35]. In this section we want to
suggest an alternative particularly transparent proof of the fluctuation theorem.
The generic situation is sketched in Fig. 5. Although the average entropy of an isolated system
will increase, the actual configurational entropy will not grow monotonously, instead it jumps
discontinuously by finite differences ∆Stot . As the average entropy increases, these differences
will be preferentially positive, but sometimes fluctuations in opposite direction may occur.
Depending on the system under consideration, these differences will be distributed according
to a certain asymmetric distribution. The fluctuation theorem states that this distribution is
constrained by the condition
P (∆Stot )
= e∆Stot . (38)
P (−∆Stot )
It relates pairs of values at opposite locations on the abscissa, as marked by the red arrows in
Fig. 5. In other words, given one half of the distribution, the fluctuation theorem predicts the
other half.
The fluctuation theorem holds for any stochastic system and is usually proved by comparing
the entropy production along a given stochastic path with the entropy production along the
reverse path. Here we suggest an alternative simple proof. Starting point is the observation
that the fluctuation relation is invariant under convolution, i.e. if two function f, g satisfy the
property f (x) = ex f (−x) and g(x) = ex g(−x) their convolution product will satisfy the same
property:
Z Z
(f ∗ g)(x) = dy f (y)g(x − y) = dy ey f (−y)ex−y g(−x + y) (39)
Z Z
= ex dy f (−y)g(y − x) = ex dy f (y)g(−y + x) = ex (f ∗ g)(−x) .

This means that the sum of random variables obeying the fluctuation relation will again obey
the fluctuation relation. The remaining proof consists of two steps:

(i) First we prove the fluctuation theorem for a single transition. According to Eq. (27), an
individual transition c → c′ changes the configurational entropy of an isolated system by
∆Stot = ln Jcc′ − ln Jc′ c . Since transition occurs with frequency Pc wcc′ = Jcc′ , we have
P (∆Stot ) = Jcc′ and similarly P (−∆Stot ) = Jc′ c for the reverse transition. Therefore the
fluctuation relation  
∆Stot = ln P (∆Stot )/P (−∆Stot ) (40)
holds trivially for a single transition.
(ii) The entropy change ∆S over a finite time is the sum of entropy changes caused by individual
transition. Summing random variables means to convolve their probability distributions.
Since the fluctuation theorem is invariant under convolution, it follows that this sum will
automatically obey the fluctuation relation as well.

As well-known, the fluctuation theorem implies the nonequilibrium partition identity

hexp(−∆Stot )i = 1 (41)
as well as the Second Law of thermodynamics
h∆Stot i ≥ 0. (42)
Note that the fluctuation theorem holds exactly only in isolated systems. However, it may also
hold approximately in the limit t → ∞ for the environmental entropy under certain assumptions
if the subsystem is stationary.

6. Concluding remarks
In this paper we have addressed several aspects of classical non-equilibrium statistical physics,
describing its general setup and its justification from the quantum perspective. In particular,
we have focused on the problem of entropy production. As we have pointed out, the commonly
accepted formula for entropy production in the environment ∆Senv = ln(wss′ /ws′ s ) holds only
in situations where the environment thermalizes almost immediately after each transition of the
subsystem. Whether this separation of time scales is valid in realistic situations remains to be
seen. Moreover, we have suggested a conjecture that this formula gives a lower bound to the
average entropy production in the environment.
Appendix A. Tracing the historical route to entropy production
It is instructive to retrace how the formula for entropy production was derived by Schnakenberg
in 1976 [25]. To quantify entropy production, Schnakenberg considers a fictitious chemical
system that mimics the dynamics of the master equation. This fictitious system is based on the
following assumptions:

• Each configuration c of the original system is associated with a chemical species Xc in an

ideal homogeneous mixture of molecules.
• The molecules react under isothermal and isochoric conditions by Xc ⇋ Xc′ in such a way
that their concentrations [Xc ] = Nc /V evolve in the same way as the probabilities in the
master equation, i.e.
d X 
[Xc ] = [Xc′ ]wc′ c − [Xc ]wcc′ . (A.1)
dt ′ c
• The reactions are so slow that standard methods of thermodynamics can be applied.

Under isothermal and isochoric conditions the chemical reactions change the particle numbers
Ni in such a way that the Helmholtz free energy F is maximized. In chemistry the corresponding
thermodynamic current is called the extent of reaction ξcc′ , which is defined as the expectation
value of the accumulated number of forward reactions Xc → Xc′ minus the number of backward
reactions Xc ← Xc′ . Note that ξcc′ does not account for fluctuations, instead it is understood
as a macroscopic deterministic quantity that grows continuously as
˙ ′ = Nc wcc′ − Nc′ wc′ c .
ξcc (A.2)

According to conventional thermodynamics, a thermodynamic flux is caused by a conjugate

thermodynamic force which is the partial derivative of the thermodynamic potential with respect
˙ ′ is the
to the flux. In chemistry the thermodynamic force conjugate to the extent of reaction ξcc
so-called chemical affinity 
∂F 
Acc′ = . (A.3)
∂ξcc′ V,T
With this definition the temporal change of the free energy is given by

˙′.
X
Ḟ = Acc′ ξcc (A.4)
cc′

The affinity is related to the chemical potential of the involved substances as follows. On the one
hand, the reaction changes the particle number by Ṅc = −ξP ˙ ˙
cc′ and Nc′ =P+ξcc˙ ′ . On the other
∂F
hand, the change of the free energy can be expressed as Ḟ = c ∂Nc Ṅc = c µc Ṅc . Comparing
this expression with Eq. (A.4) the affinity can be expressed as

Acc′ = µc′ − µc . (A.5)

For an ideal mixture the Helmholtz free energy is given by

X
F = Nc (qc + kB T ln Nc ) , (A.6)
c

where qc is a temperature-dependent constant. The chemical potential of species Xc is

∂F
µc = = µ0c + kB T ln Nc (A.7)
∂Nc
with µ0c = qc + kB T , so that the affinity is given by

Nc′
Acc′ = µ0c − µ0c′ + kB T ln . (A.8)
Nc

The fictitious chemical system relaxes towards an equilibrium state that corresponds to the
stationary state of the original master equation. In this state the particle numbers Nc attain
certain stationary equilibrium values Nceq . Moreover, the thermodynamic flux and its conjugate
force vanish in equilibrium:
Aeq eq
cc′ = ξcc′ = 0 . (A.9)
Because of Aeq
cc′ = 0 we have
Nceq
µ0c − µ0c′ = kB T ln
′
, (A.10)
Nceq
which allows one to express the affinity as

Nc′ Nceq
Acc′ = kB T ln . (A.11)
Nc Nceq
′

eq
On the other hand, ξcc ′ = 0 implies that

Nceq wc′ c
eq = . (A.12)
Nc ′ wcc′

Inserting this relation into Eqs. (A.11) and (A.4) the change of the free energy (caused by all
reactions Xc ⇋ Xc′ ) is given by
X Nc′ wc′ c
Ḟ = kB T ξcc′ ln . (A.13)
Nc wcc′
cc′

Since temperature T and internal energy U of the mixture remain constant the variation of the
free energy F = U − T S is fully absorbed in a change of the entropy, i.e. Ḟ = −T Ṡ. This allows
one to derive a formula for the entropy production
X [Xc ]wcc′
Ṡ = −kB ξcc′ ln , (A.14)
[Xc′ ]wc′ c
cc′

where we used the fact that Nc /Nc′ = [Xc ]/[Xc′ ].

Since the extent of reaction ξcc′ just counts the number of reactions from c to c′ , this formula is
just the continuum limit of Eq. (30), proving the formula for entropy production (29). However,
it is important to recall the underlying assumptions. Each component of the fictitious chemical
systems was assumed to be internally equilibrated and methods of ordinary thermodynamics
were used to derive Eq. (A.14). This means that the formula will only be meaningful in the
limit where the environment equilibrates almost immediately.
Acknowledgments
This work was initiated at conference STATPHYS-VII in Kolkata, India. HH would like to
thank the organizers for the warm hospitality and for the lively and productive atmosphere. CG
and PJ thank the Centre for Quantum Technologies, Singapore, where part of this work was
done.

References
[1] Jaynes E, 1957, Information theory and statistical mechanics I, Phys. Rev. 106 620.
[2] Jaynes E, 1957, Information theory and statistical mechanics II, Phys. Rev. 108 171.
[3] Schürmann T and Grassberger P, 1996 Chaos 6 414.
[4] Bonachela JA, Hinrichsen H and Muñoz MA, 2008 J. Phys. A: Math. Theor. 41 202001.
[5] Uffink J, Compendium of the foundations of classical statistical physics,
http://philsci-archive.pitt.edu/2691/.
[6] Schrödinger E, 1927 Ann. Phys. 388, 956; von Neumann J, 1929 Z. Phys. 57, 30.
[7] Bloch I, Dalibard J and Zwerger W, 2008 Rev. Mod. Phys. 80 885.
[8] Trotzky S, Chen Y-A, Flesch A, McCulloch IP, Schollwöck U, Eisert J, and Bloch I, 2011 preprint
arXiv:1101.2659.
[9] Zurek WH, 2003 Rev. Mod. Phys. 715.
[10] Joos E, Zeh H, Kiefer C, Giulini D, Kupsch J, Stamatescu IO, 1996 Decoherence and the Appearance of a
Classical World in Quantum Theory (Springer, Berlin).
[11] Schlosshauer M, 2007 Decoherence and the Quantum to Classical Transition (Springer, Berlin).
[12] Linden N, Popescu S, Short AJ, and Winter A, 2009 Phys. Rev. E 79, 061103.
[13] Linden N, Popescu S, Short AJ, and Winter A, 2010 New J. Phys. 12, 055021.
[14] Gogolin C, 2010 Phys. Rev. E 81, 051127.
[15] Reimann P, 2008 Phys. Rev. Lett. 101, 190403.
[16] Short AJ, 2010 preprint arXiv:1012.4622.
[17] Gogolin C, Müller MP, and Eisert J, 2011 Phys. Rev. Lett. 106, 040401.
[18] Srednicki M, 1994 Phys. Rev. E 50 888. .
[19] Tasaki H, 1998 Phys. Rev. Lett 80 1373.
[20] Goldstein S, 2006 Phys. Rev. Lett. 96 050403.
[21] Riera A, Gogolin C, and Eisert J, in preparation.
[22] Cramer M, Dawson CM, Eisert J, and Osborne TJ, 2008 Phys. Rev. Lett. 100, 030602; Cramer M and Eisert
J, 2010 New J. Phys. 12, 055020.
[23] Seifert U 2005 Phys. Rev. Lett 95, 040602.
[24] Andrieux D and Gaspard P 2004 J. Chem. Phys. 121 6167
[25] Schnakenberg J 1976 Rev. Mod. Phys. 48 571
[26] Jiu-li L, Van den Broeck C, and Nicolis G, 1984 Z. Phys. B: Condens. Matter 56, 165.
[27] Evans D J, Cohen E G D and Morriss GP, 1993 Phys. Rev. Lett. 71 2401.
[28] Evans D J and Searles D J, 1994 Phys. Rev. E 50 1645.
[29] Gallavotti G and Cohen E G D, 1995 Phys. Rev. Lett. 74 2694.
[30] Kurchan J, 1998 J. Phys. A: Math. Gen. 31 3719.
[31] Lebowitz J L and Spohn H, 1999 J. Stat. Phys. 95 333.
[32] Maes C, 1999 J. Stat. Phys. 95 367.
[33] Jiang D-Q, Qian M and Qian M-P, 2004 Mathematical Theory of Nonequilibrium Steady States (Berlin:
Springer)
[34] Harris R J and Schütz G M, 2007 J. Stat. Mech. P07020.
[35] Kurchan J, 2007 J. Stat. Mech. P07005.