Journal of Statistical Mechanics:

Theory and Experiment

Related content
- Entropy production and the arrow of time
Non-equilibrium statistical mechanics: partition J M R Parrondo, C Van den Broeck and R
Kawai
functions and steepest entropy increase - Maximum entropy production rate in
quantum thermodynamics
Gian Paolo Beretta
To cite this article: Sergio Bordel J. Stat. Mech. (2011) P05013
- Non-equilibrium Dynamics, Thermalization
and Entropy Production
Haye Hinrichsen, Christian Gogolin and
Peter Janotta
View the article online for updates and enhancements.

This content was downloaded from IP address 157.88.41.117 on 06/11/2018 at 15:23

J ournal of Statistical Mechanics: Theory and Experiment

Non-equilibrium statistical mechanics:

partition functions and steepest entropy
increase

J. Stat. Mech. (2011) P05013

Sergio Bordel
Department of Chemical and Biological Engineering, Chalmers University of
Technology, Kemivägen 10, SE-412 96 Göteborg, Sweden
E-mail: velasco@chalmers.se

Received 21 February 2011

Accepted 15 April 2011
Published 12 May 2011

Online at stacks.iop.org/JSTAT/2011/P05013
doi:10.1088/1742-5468/2011/05/P05013

Abstract. On the basis of just the microscopic definition of thermodynamic

entropy and the definition of the rate of entropy increase as the sum of
products of thermodynamic fluxes and their conjugated forces, we have derived
a general expression for non-equilibrium partition functions, which has the
same form as the partition function previously obtained by other authors using
different assumptions. Secondly we show that Onsager’s reciprocity relations are
equivalent to the assumption of steepest entropy ascent, independently of the
choice of metric for the space of probability distributions. Finally we show that
the Fisher–Rao metric for the space of probability distributions is the only one
that guarantees that dissipative systems are what we call constantly describable
(describable in terms of the same set of macroscopic observables during their
entire trajectory of evolution towards equilibrium). The Fisher–Rao metric is
fundamental to Beretta’s dissipative quantum mechanics; therefore our last result
provides a further justification for Beretta’s theory.

Keywords: large deviations in non-equilibrium systems, Boltzmann equation

2011
c IOP Publishing Ltd and SISSA 1742-5468/11/P05013+12$33.00
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

Contents

1. Introduction 2
2. The non-equilibrium partition function 3
3. The steepest entropy ascent and Onsager’s relations are equivalent 6
4. The Fisher–Rao metric is the natural metric 9

J. Stat. Mech. (2011) P05013

4.1. The definition of constantly describable systems . . . . . . . . . . . . . . . 9
4.2. Time evolution of the probability distribution of constantly describable
systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
5. Conclusions 11
References 11

1. Introduction

Non-equilibrium thermodynamics is still a field of research open to many different

interpretations [1]. One of the reasons for this lack of consensus could be the absence
of an accepted microscopic statistical description analogous to Gibbs’ formalism for
equilibrium thermodynamics. This statistical description would enable us to derive the
relationships between different macroscopic variables from a partition function. Several
authors [2]–[5] have derived expressions for non-equilibrium statistical distributions using
different methods, such as the assumption of maximization of the Shannon information
entropy. In the second section of this work, we derived a non-equilibrium partition
function that coincides with the one obtained by previous authors, in particular Zubarev
and Kalashnikov [2]; however, in our case the only assumption that is introduced is the
definition of the rate of entropy increase as the sum of the products of the thermodynamic
forces and fluxes and the microscopic definition of the thermodynamic entropy [6]. No
extra assumptions about the maximization of Shannon’s information entropy [2] or about
the form of the Liouville operator that describes the time evolution of the system [7] were
made to derive our expression.
One of the reasons that made the development of a microscopic description of
non-equilibrium processes difficult is the fact that the classical and quantum Liouville
operators describing the microscopic evolution of an isolated system do not explain
the increase of entropy consistently with the second law of thermodynamics. In order
to overcome this contradiction, some authors have introduced dissipative terms in the
Liouville operator that are explained as interactions with a reservoir [7, 8], an assumption
that seems in principle to be in contradiction with the definition of an isolated system.
Beretta [9, 10] developed a remarkable theory that explains thermodynamic irreversibility
at a fundamental level without assuming interactions with a reservoir. Beretta’s theory
predicts that the evolution of the density operator (in a quantum representation) follows
the direction of the steepest entropy increase. In order to define a direction of steepest

doi:10.1088/1742-5468/2011/05/P05013 2
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

increase, a metric in the space of density operators has to be introduced. Beretta chose a
Fisher–Rao metric [10, 11].
In a recent work [12] we argued, using information-theoretical concepts, that the
most probable direction of evolution of a system is the direction of maximal entropy
ascent and showed that Onsager’s reciprocity relations appear as a consequence of this
assumption. However we did not pay attention to how the choice of metric for the space
of probability distributions would affect these conclusions. In section 3 of this work
we make a stronger case for the steepest entropy ascent and show that independently
of the choice of metric, the maximal entropy ascent is a condition both necessary and
sufficient for the satisfaction of Onsager’s reciprocity relations. This result shows a clear

J. Stat. Mech. (2011) P05013

interrelation between Onsager’s formalism (which is universally accepted for systems close
to equilibrium) and Ziegler’s principle or the maximal entropy production principle [13]–
[15], which is still subject to controversy and is treated with scepticism by many scientists.
As a result of the absence of a microscopic statistical description for systems far
away from equilibrium, the study of dissipative systems has relied so far on the use
of systems of phenomenological differential equations describing the evolution of a set
of macroscopic observables that characterize the system [16, 17]. Phenomena such as
non-equilibrium phase transitions are then described using methods such as stability
analysis and bifurcation theory. This treatment of dissipative systems carries an implicit
assumption, namely that the same set of macroscopic observables is able to describe the
state and the dynamics of evolution of the system along its entire trajectory towards
equilibrium. We have named the systems satisfying this property constantly describable.
In the fourth section of this work we show that with the steepest entropy increase
assumption, the only metric that guarantees that a system is constantly describable is
the Fisher–Rao metric.
Beretta [10, 11] justifies his choice of the Fisher–Rao metric as such that the rate of
entropy production does not go to infinity when the probability of a microstate (in the
classical approach) or an eigenvalue of the density matrix (in the quantum approach) is
equal to zero. The argument that we give in this work could help to strengthen this choice,
which is central to Beretta’s quantum dissipative mechanics [9, 10].

2. The non-equilibrium partition function

In this section we adopt again most of the arguments developed in a previous work [12] in
order to derive a probability distribution for states far from equilibrium. We consider an
isolated system far away from equilibrium and analyse the time evolution of the system’s
thermodynamic entropy towards its maximum, taking in account the fact that in an
isolated system, energy and matter are conserved.
Given the microscopic definition of entropy [6], the rate of entropy increase in an
isolated system, can be defined as follows:

dS  dpi  dpi
= −kB ln(pi ) − kB . (1)
dt i
dt i
dt

doi:10.1088/1742-5468/2011/05/P05013 3
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

The sum of the probabilities of all the accessible microstates must remain constant
and equal to 1; therefore its time derivative is equal to 0:
 dpi
= 0. (2)
i
dt

The rate of the entropy increase becomes

dS  dpi
= −kB ln(pi ). (3)
dt i
dt

J. Stat. Mech. (2011) P05013

A number of macroscopic variables (energy, numbers of certain chemical species etc)
are conserved during the evolution of the system. Therefore the following constraints
must be satisfied:
 dpi
Cir = 0. (4)
i
dt

The variable C r indicates each of the conserved quantities.

The previous conditions, together with the normalization condition, impose
→˙
limitations on the values that the vector p can take. Using matrix notation we can
express these limitations in the following way:
˙
C p = 0.
→
(5)
The matrix C has as many columns as possible states of the system and as many
rows as constraints. The first row contains 1s and the following rows the values of Cir
˙
→
If the system has n possible states, the vector p is an element of Rn . Equation (5)
→˙
constrains the possible values of p to the kernel of the matrix C. The dimension of ker(C)
is the n-rank (C). If {ek } is a basis of ker(C), any path followed by the system can be
expressed as

p˙ =
→
kek . (6)
k

The rate of entropy increase can be rewritten as follows:

dS 
= −kB k eki ln(pi ). (7)
dt ki

Or, in a more compact form,

dS 
=− k πk (8)
dt
k

πk = kB eki ln(pi ). (9)
i

Note that thermodynamic equilibrium implies that πk is equal to zero for every k, or
in other words, that the vector of components ln(pi ) is perpendicular to the kernel of the

doi:10.1088/1742-5468/2011/05/P05013 4
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

matrix C and therefore belongs to the vector space generated by its rows. This condition
implies the following relation:

ln(pi ) = λ + λC r C r . (10)
r

The previous expression is identical to the grand canonical distribution obtained from
Gibbs’ formalism.
As we argued in a previous work [12], the thermodynamic fluxes can be defined as
the rates of change of non-conserved quantities:

Jf = φf i ṗi .

J. Stat. Mech. (2011) P05013

(11)
i

Expressing the rates of change of the probabilities according to equation (6) we obtain
the following expression:
  
Jf = k eki φf i = k akf . (12)
k i k

The rate of entropy increase is normally defined in terms of thermodynamic fluxes

and forces:
dS   
= J f Xf = k akf Xf . (13)
dt f k f

Comparing equations (8) and (13) and considering that when all the thermodynamic
forces become zero (in equilibrium), all the elements πk should also be zero, the following
relation between thermodynamic forces and the elements πk arises:

πk = − akf Xf . (14)
f

By combining equations (9) and (14) we obtain the following relation between the
thermodynamic forces and the non-equilibrium probability distribution:
 
kB eki ln(pi ) = − akf Xf ∀k. (15)
i f
  
kB eki ln(pi ) = − eki φ f i Xf ∀k. (16)
i i f

If we rewrite the previous equation using matrix formulation, with E being the matrix
formed by the row vectors {ek }, we obtain the following relation:
 
−−→  →
kB E ln(p) = −E Xf φ f . (17)
f

The following relation arises immediately from the previous equation:

−−→  →
kB ln(p) = − 
Xf φf + ψ. (18)
f

doi:10.1088/1742-5468/2011/05/P05013 5
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

The vector ψ  can be any vector orthogonal to the subspace generated by {ek };
therefore it can be generated as a linear combination of the rows of the matrix C:
 
kB ln(pi ) = λ + λC r Cir − Xf φ f i . (19)
r f

This equation allows defining a non-equilibrium partition function

    
λ  λC r  Xf
Z(λC r , Xf ) = exp − = exp Cir − φf i . (20)
kB i r
k B
f
k B

J. Stat. Mech. (2011) P05013

This partition function allows defining the same kinds of relationships between
variables as the usual equilibrium partition function. For example, the entropy of a
system far away from equilibrium can be obtained as follows:
 
S = kB ln Z − λC r C r  + Xf φf . (21)
r f

The average values of the non-conserved variables can be obtained from the partial
derivatives of the partition function with respect to their associated thermodynamic forces:

∂ ln Z
φf  = −kB (22)
∂Xf
and the average values of the conserved quantities (energy or chemical species):
∂ ln Z
C r  = kB . (23)
∂λC r
However, the Lagrange multipliers corresponding to the conserved quantities can no
longer be directly identified with the inverse of the temperature (for the energy) or with the
negative chemical potentials divided by the temperature (for conserved chemical species),
as these variables are difficult to define in a system far away from equilibrium that could
be inhomogeneous.
A similar non-equilibrium partition function was already derived using different
methods by Zubarev and Kalashnikov [2], who also identified the parameters equivalent
to Xf in their expression as thermodynamic forces. The most interesting feature of the
non-equilibrium probability distribution obtained here is that it corresponds to the same
distribution as would be obtained by maximizing the entropy of the system with the values
φf  constrained to be constant. However this condition was not used in our derivation
and the expression arises only on defining the rate of entropy increase as the sum of the
products of thermodynamic fluxes with their respective forces.

3. The steepest entropy ascent and Onsager’s relations are equivalent

In a previous work [12] we argued from geometric considerations that the most probable
path of evolution of a system towards equilibrium is the direction of steepest entropy ascent
and showed that Onsager’s reciprocity relations can be deduced from this consideration
and hold also far from equilibrium. However, in order to define a gradient and a direction

doi:10.1088/1742-5468/2011/05/P05013 6
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

of steepest entropy ascent, it is necessary to define a distance in the space of vectors p.
In our previous work we implicitly assumed the distance to be defined as follows:

d 2 = dp2i . (24)
i

This decision appears to be arbitrary. Beretta [10, 11] argued that the natural metric
of a space of discrete probability distributions is the Fisher–Rao metric. The same metric
was used by Crooks [18] to define what he called the thermodynamic distance between
two states, although he only applied it to equilibrium states:
1 1 2

J. Stat. Mech. (2011) P05013

d 2 = dp . (25)
4 i pi i

In this section we aim to prove that independently of the choice of metric, Onsager’s
relations can be deduced from the steepest entropy ascent and hold even far away from
equilibrium. Given an arbitrary metric for our space of probability distributions,

d 2 = gij (
p) dpi dpj . (26)
ij

The bilinear form with components gij is positive definite, and locally dependent on
the probability distribution.
By choosing the basis {ek } in such a way that it is orthonormal with respect to
the scalar product defined by the bilinear form gij , the metric with respect to this basis
becomes

d 2 = d2k . (27)
k

Looking at equation (8) we can see that the direction of maximal entropy ascent
corresponds to
1
 = − π . (28)
τ
The parameter τ has units of time and has been chosen (instead of β = 1/τ in our
previous work [12]) in order to follow Beretta’s notation [10, 11]. By substituting the
last expression in equation (12) we obtain the following equation for the thermodynamic
fluxes:
1
Jf = − πk akf (p). (29)
τ k

In this case the basis {ek } is chosen to be orthonormal with respect to the bilinear
form gij (
p); therefore, its components are functions of the probability distribution and so
are the components akf ( p):

akf (
p) = eki (
p)φf i . (30)
i

doi:10.1088/1742-5468/2011/05/P05013 7
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

If we combine equations (14) and (29) we arrive at the following relationship between
the fluxes and forces:
1 
Jf = akg (
p)akf (p)Xg . (31)
τ k g

Equation (31) can be rewritten in an Onsager-like form with coefficients satisfying

Onsager’s relations:
1
Jf = lf g (
p)Xg . (32)
τ g

J. Stat. Mech. (2011) P05013

Each element lf g is the scalar product of two rows of the matrix akf ; therefore
Onsager’s relations are automatically satisfied by them:

lf g (
p) = akg (
p)akf (
p) = lgf (
p) (33)
k

p)2 ≤ lf f (p)lgg (
lf g ( p). (34)

We can conclude that Onsager’s relations can be obtained from the condition of
maximal entropy ascent independently of the metric chosen for the space of probability
distributions and hold also far away from equilibrium. Beretta also obtained the same
result using the Fisher–Rao metric [10, 11].
Now we will show that the condition of maximal entropy ascent is not only consistent
with Onsager’s relations but also necessary. Let us assume that the vectors  and π are
not anti-parallel but are related by the following equation:
1
 = − Uπ . (35)
τ
The matrix U is unitary and different from the identity matrix (and therefore it is
asymmetric). Equations (29) and (31) get transformed in the following way:
1 
Jf = − ukq πq akf (
p) (36)
τ q k

1 
Jf = ukq aqg (
p)akf (
p)Xg . (37)
τ k g q

Therefore Onsager’s reciprocity relations are broken:

 
lgf (
p) = ukq aqg (
p)akf (
p) = ukq aqf (
p)akg (
p) = lf g (p). (38)
kq kq

The reciprocity relations are satisfied only if the unitary matrix U becomes the identity
matrix; therefore Onsager’s reciprocity relations are satisfied only if the system evolves
in the direction of steepest entropy increase for any particular metric of the space of
probability distributions.

doi:10.1088/1742-5468/2011/05/P05013 8
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

4. The Fisher–Rao metric is the natural metric

4.1. The definition of constantly describable systems

We define a system as being fully describable with the set of variables {bs } if given the
values of all the variables in the set, the state of the system (and its dynamics) becomes
fully determined. According to the non-equilibrium partition function defined in the
second section, the probability distribution of a system far from equilibrium is describable
with the set of variables {C r , φf }.
We define a system far from equilibrium as being constantly describable with the set
of variables {bs } if this set of variables is sufficient for describing the system at every

J. Stat. Mech. (2011) P05013

point of its trajectory towards equilibrium. All the dissipative systems that are described
in the literature in terms of autonomous systems of differential equations [16, 17] are
implicitly being assumed to be constantly describable. If they were not, new variables
and their corresponding differential equations would need to be introduced during the
time evolution of the system towards equilibrium.
db
= F (b). (39)
dt
This definition can seem superfluous, as most of the real systems seem to satisfy it, but
its utility will become apparent in section 4.2.

4.2. Time evolution of the probability distribution of constantly describable systems

Consider a system far from equilibrium with the following probability distribution:
−−→  → 
kB ln(p) = λ1 − Xf φ f + λC r C r . (40)
f r

It is constantly describable if all the probability distributions that it reaches during

its future evolution are also describable with the same set of conserved and non-conserved
−−→
quantities or in other words, if the vector ln(p) remains in the subspace as is generated
→ −−→
by the vectors {1, C r , φf }. If during the evolution of the system ln(p) leaves the
aforementioned subspace, new thermodynamic forces and their conjugated variables would
appear and these conjugated variables would be necessary for describing the system;
therefore it would not be constantly describable. This appearance of new variables seems
to be counterintuitive for dissipative systems. And it is indeed non-consistent with the
usual treatment of dissipative systems [16, 17]. Therefore we postulate that the real
systems are constantly describable and that any condition necessary for a system to be
constantly describable should be satisfied by a real system.
The reasons given by Beretta [10, 11] for the choice of a Fisher–Rao metric are related
to the fact that for a usual metric, the rate of entropy production becomes infinite
whenever the probability of a microstate tends to zero (or for a quantum representation,
one eigenvalue of the quantum density operator). Therefore it could seem that the choice
is just an arbitrary calculus artefact and this could be one of the reasons why Beretta’s
remarkable theory has not yet been given the credit and attention that it deserves. Here
we aim to provide an extra justification for the choice of a Fisher–Rao metric, namely its
compatibility with constantly describable systems.

doi:10.1088/1742-5468/2011/05/P05013 9
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase
−−→ →
For the vector ln(p) to remain in the vector subspace generated by {1, C r , φf }, its
time derivatives must also belong to this subspace. Rewriting equation (6) in matrix
notation we obtain
p˙ = E t.
→
(41)
Now we write the time derivatives of the probabilities as a function of the time
derivatives of their logarithms
−→˙
P ln(p) = E t. (42)
The matrix P is a diagonal matrix that contains the values of the probabilities and E

J. Stat. Mech. (2011) P05013

has the elements of the basis {ek } as rows. Note that this basis depends now on the
probability distribution, as we argued in section 3.
Let us call G the matrix of the elements gij ( p) defining the metric of the space
of probability distributions and remember that the basis {ek } has been chosen to be
orthonormal with respect to the scalar product defined by the symmetric bilinear form
G. If we multiply both sides of the equation (42) by EG, we obtain
−→˙
EGP ln(p) = I. (43)
Now we will apply the steepest entropy ascent condition, which as we showed in
section 3 holds for any metric:
−→˙ 1
EGP ln(p) = − π . (44)
τ
By writing equation (14) in matrix notation we obtain
−→˙ 1 
EGP ln(p) = − AX. (45)
τ
A is the matrix of components akf (p) introduced in section 3. The matrix A can be
rewritten using equation (30) in matrix notation:

A = EΦ (46)
−→
˙ 1 
EGP ln(p) = − EΦX. (47)
τ
→  belongs
The matrix Φ contains the vectors {φf } in its columns; therefore the vector ΦX
→
to the subspace generated by {φf }.
Following the same argument as in equation (17) we can see that
−→˙ 1  
GP ln(p) = − ΦX + ψ. (48)
τ
 belongs to the subspace generated by {1, C r }. Therefore, for an arbitrary
The vector ψ
−→˙
metric defined by G we have concluded that the vector GP ln(p) belongs to the subspace
→ −→˙
generated by {1, C r , φf }. Now, if we want to assure that ln(p) also belongs to the same
subspace we need the following condition to be satisfied:
G ∝ P −1 . (49)

doi:10.1088/1742-5468/2011/05/P05013 10
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

This means that for a system to be constantly describable, the metric of the space
of probability distributions has to be of the Fisher–Rao type, exactly as Beretta had
proposed.
An interesting characteristic of Beretta’s solution of the problem of thermodynamic
irreversibility (which makes it different from other approaches to solving the same
problem) is that the density operator is defined both for positive and negative times [10].
Therefore, previous states of the system under study can be deduced from the present
state. A system being constantly describable implies that the same set of macroscopic
variables can describe the state of the system at all the points of its trajectory towards
thermodynamic equilibrium. There may be an epistemological relationship between these

J. Stat. Mech. (2011) P05013

two characteristics. If a system were not constantly describable, a completely different
(and unknown) set of observables could have defined its state in the past, which would
make the past states also unknown and therefore not defined for negative times.

5. Conclusions

We have derived a general expression for a non-equilibrium partition function, which is

equivalent to the one proposed previously by other authors [2]. The only assumptions
involved in its derivation are the microscopic definition of entropy and the definition of
the entropy production rate as the sum of products of thermodynamic fluxes and forces.
We have shown the equivalence between steepest entropy increase and Onsager’s
reciprocity relations, as well as the fact of this equivalence being independent of the
choice of metric for the space of probability distribution. This implies that accepting
the validity of Onsager’s reciprocity relations implies accepting the steepest entropy
production principle. Onsager’s relations are universally accepted for situations close to
equilibrium and it has also been argued [19] that they hold also far away from equilibrium.
If this were the case, the steepest entropy ascent would be a universal characteristic of
the evolution of systems towards equilibrium, as Beretta’s theory implies [9, 10].
Finally we defined what we call constantly describable systems and showed that for
a system to be constantly describable, the metric introduced in the space of probability
distributions should be the Fisher–Rao metric, which lies at the heart of Beretta’s theory of
quantum dissipative systems. We expect this finding to bring more attention to Beretta’s
elegant conciliation of thermodynamics and quantum mechanics.
References
[1] Lebon G, Jou D and Casas-Vázquez J, 2008 Understanding Non-equilibrium Thermodynamics (Berlin:
Springer)
[2] Zubarev D N and Kalashnikov V P, 1970 Physica 46 550
[3] Algarte A C, Vasconcellos A R, Luzzi R and Sampaio A J, 1985 Rev. Bras. Phys. 15 106
[4] Robin W A, 1990 J. Phys. A: Math. Gen. 23 2065
[5] Attard P, 2006 Phys. Chem. Chem. Phys. 8 3585
[6] Jaynes E T, 1957 Phys. Rev. 106 620
[7] McLennan J A, 1961 Phys. Fluids 4 1319
[8] Kossakowski A, 1972 Rep. Math. Phys. 3 247
[9] Beretta G P, 1987 Found. Phys. 17 365
[10] Beretta G P, 2009 Rep. Math. Phys. 64 139
[11] Beretta G P, 2008 Entropy 10 160
[12] Bordel S, 2010 Physica A 389 4564
[13] Martyushev L M and Seleznev V D, 2006 Phys. Rep. 426 1

doi:10.1088/1742-5468/2011/05/P05013 11
 Non-equilibrium statistical mechanics: partition functions and steepest entropy increase

[14] Dewar R C, 2003 J. Phys. A: Math. Gen. 36 631

[15] Niven R K, 2009 Phys. Rev. E 80 021113
[16] Nicolis G, 1986 Rep. Prog. Phys. 49 873
[17] Haken H, 1977 Synergetics (Berlin: Springer)
[18] Crooks G E, Phys. Rev. Lett. 99 100602
[19] Hurley J and Garrod C, 1982 Phys. Rev. Lett. 23 1575

J. Stat. Mech. (2011) P05013

doi:10.1088/1742-5468/2011/05/P05013 12