Onsager theory: overview

Pearu Peterson
December 18, 2006
1 Introduction
Our aim is to study matter that consists of large number of molecules. A
complete mechanical description of such a system is practically impossible
due to inability (i) to solve the system of equations with enormous size that
describes the motion of molecules; (ii) to determine the initial state of the
system. So, in the following the statistical description of matter will be in-
troduced that should resolve the issues with handling the system of equation
size as well as dealing with irreversible character of matter.
1.1 Mechanical ensemble theory
To overcome the diculty of specifying initial conditions, Gibbs introduced
the concept of an ensemble that consists of a large number of systems be-
ing equivalent in a macroscopic sense. Each such a system is described as a
distribution of points in phase space (a space of positions and momenta of
molecules). Two systems are considered equivalent when their macroscopic
properties are equal within the same macroscopic specication of the ensem-
ble. The distribution of phase points is used to assign probabilities to points
in phase space based on their frequency in the ensemble. The evolution of
this probability cloud is described by the Liouville equation. According to
Gibbs, an estimate of the value of any measurable property in the system is
obtained by averaging over the same property in the ensemble. To describe
a system using Gibbs ensemble approach, two questions arise:
When is the specication of the ensemble good enough so that the
resulting average will give good estimates for any system in ensemble
as well as will reliably predict the future? To answer the question, it is
necessary to use empirical information.
1
For such a good specication, what is the corresponding distribution
in phase space? There is no answer to this within the connes of the
mechanical theory and so the answer must be postulated.
The basic empirical guide to good ensemble specications is given by
equilibrium thermodynamics. In thermodynamics the state of a large system
(but far from critical points) is determined by its entropy as a function of
extensive variables: the internal energy E, the volume V , and the number
of molecules N. This is the specication of micro-canonical ensemble. The
canonical ensemble specication uses temperature T instead of internal en-
ergy. The limit where N and V become innite with N/V xed, is called the
thermodynamic limit.
In mechanical ensemble theory we have to introduce postulates about the
nature of initial distribution functions which cannot themselfs be measured.
For a non-equilibrium ensemble, this is especially a serious problem.
1.2 Physical ensemble theory
Measurements on macroscopic systems are restricted to a small number of
variables. If a given set of variables will lead to a self-contained description
of a large molecular system then it constitutes a contracted description of a
macroscopic system. Similar to Gibbs ensemble theory, a question of what
variables will provide a good contracted description of a macroscopic system
in the sense that the future behaviour of the system will be well predicted
from the knowledge of initial values? The answer is given by empirical ob-
servations. For example:
For an isothermal solution undergoing slow chemical reactions at con-
stant pressure the concentrations of the various reacting species give a
good contracted description.
For a uid ow that is not too rapid, a hydrodynamic-level description
involving mass, energy, and momentum densities as functions of spatial
positions provides a good contracted description.
For a matter close to equilibrium the thermodynamic-level description
involving extensive variables E, V , and N provides a good contracted
description.
In the contracted description of matter at equilibrium two systems which
are at equilibrium and poses the same values of the extensive variables are
identical, for the purposes of most measurements. But in the molecular state
or location in phase space, they are not the same. To measure the dierences
2
which appear in these systems that have the same macroscopic description,
molecular uctuations are used. These uctuations must be small.
The need for uctuations demonstrate our limited ability to make mea-
surements and so contracted descriptions must be inherently statistical. A
physical ensemble consists of a collection of systems, identically prepared
with respect to the contracted description. Dierent from Gibbs ensemble
where one requires a knowledge of the statistical distribution in phase space,
in a contracted description for physical ensembles the statistical properties
of ensemble, that are variables which one has been chosen to use in the con-
tracted description, are described. Mathematically physical ensembles are
treated by the theory of stochastic processes.
2 Stochastic Processes
A vector-valued stochastic process is denoted by n(t) = (n
1
(t), . . . , n
k
(t))
where n
i
(t) is random variable for xed t. It is characterised by the joint
distribution for all possible nite specications of times. If
Prob[n(t
1
) n
1
, . . . , n(t
m
) n
m
]
denes the joint probability that the variables are less than the indicated
values n
1
, . . . , n
m
then the join distribution of n(t
1
), . . . , n(t
m
) is
W
m
(n
1
, t
1
; . . . ; n
m
, t
m
)dn
1
dn
m
= Prob[n
1
n(t
1
) n
1
+dn
1
, . . . , n
m
n(t
m
) n
m
+dn
m
]
We have
W
m1
(n
1
, t
1
; . . . ; n
m1
, t
m1
) =

W
m
(n
1
, t
1
; . . . ; n
m
, t
m
)dn
m
.
A knowledge of all the joint probability densities provides a complete statis-
tical description for measurements of n made at a nite number of dierent
times.
Two-time averages of the scalar functions f and g are dened by
< f(n(t))g(n(t

)) >
2
=

f(n(t))g(n(t

))W
2
(n, t; n

, t

) dndn

.
Fluctuations about the average is dened as
n(t) = n(t) < n(t) > .
Two-time conditional (transition) probability density is deed as follows
Prob[n(t
0
) = n
0
, n n(t) n +dn] = P
2
(n
0
, t
0
|n, t) dn.
3
We have
W
2
(n
0
, t
0
; n, t) = W
1
(n
0
, t
0
)P
2
(n
0
, t
0
|n, t),
.
.
.
W
m
(n
1
, t
1
; . . . ; n
m
, t
m
) = W
m1
(n
1
, t
1
; . . . ; n
m1
, t
m1
)P
m
(n
1
, t
1
; . . . ; n
m1
, t
m1
|n
m
, t
m
).
The conditional average of a function f(n(t)) is dened by
< f(n(t))) >
0
=

f(n)P
2
(n
0
, t
0
|n, t) dn.
2.1 Markov processes
A Markov process (without memory) has a property that conditions prior
the most recent one in a conditional ensemble do not aect the subsequent
probability distribution:
P
m
(n
1
, t
1
; . . . ; n
m1
, t
m1
|n
m
, t
m
) = P
2
(n
m1
, t
m1
|n
m
, t
m
).
Markov process is completely dened by W
1
(n, t) and P
2
(n
0
, t
0
|n, t).
2.2 Stationary processes
A stationary stochastic process is the one which is invariant to time transla-
tions by the same interval , that is, n(t
1
+), . . . , n(t
m
+) are statistically
indistinguishable from the untranslated n(t
0
), . . . , n(t
m
) for all m and .
For a stationary process all single-time averages are constants, the single-
time probability density, W
1
, is independent of time, the joint probability
densities depend only on pairwise time dierences:
W
2
(n
1
, t
1
; n
2
, t
2
) = W
2
(n
1
, 0; n
2
, t
2
t
1
),
P
2
(n
0
, t
0
; n, t) = P
2
(n
0
, 0|n, t t
0
).
Ensembles at thermodynamic equilibrium or at non-equilibrium steady state
are stationary, for example.
2.3 Gaussian processes
A stochastic process is called Gaussian if all the joint and conditional prob-
ability densities have a Gaussian form:
G(n) =
exp((n n)
T

1
(n n)/2)

(2)
m
det
,
4
where is covariance (nonsingular, symmetric, positive denite) matrix and
n is average.
For a stationary Gaussian process, the lowest-order joint density will be
time independent and the average values and covariance will be constant.
For conditional Gaussian ensembles, the average values and covariance of
n(t) will be time dependent.
2.4 Ornstein-Uhlenbeck processes
An Ornstein-Uhlenbeck process is generated by a linear stochastic dierential
equation which includes a white noise term, i.e. a Langevin-like equation:
da
dt
= Ha +

f,
where H is negative semi-dened matrix, < a >= 0,

f is white noise with
the strength matrix = (H +H
T
), = lim
t
< a(t)a
T
(t) >
0
.
3 The Onsager picture
Lars Onsager developed a consistent statistical theory of irreversible pro-
cesses that is able to relate measurable quantities (transport coecients,
thermodynamic derivatives) to the results of experimental measurements.
This theory is linear and limited in validity to a neighbourhood of equilib-
rium. However, the basic structure of the Onsager theory contains many
features that are needed for more general situations. In fact, the Onsager
theory is a special case of equilibrium for the general theory described in this
course.
The Onsager theory describes the properties of equilibrium ensembles.
For measurable physical quantities one must specify some subset of exten-
sive variables, collected in n, that characterise the system. The aim is to
characterise the dynamics of n in the equilibrium ensemble, that is, how the
initial value n
0
evolve to some possible value of n at a later time t. These
transitions depend only on the elapsed time t since an equilibrium ensemble
is stationary. According to the Onsager principle, the rate of the change
in the conditional averages < n(t) >
0
are linearly related to the average
deviations of the conjugate intensive variables from their equilibrium values
where the coupling matrix is symmetric and non-negative denite. This is
the linear law for irreversible processes in an equilibrium ensemble. For ease
of describing the linear relaxations of the averages to their equilibrium val-
ues, the deviations of n from their equilibrium values n
e
as a = n n
e
will
5
be used. The conditional average values of a will be denoted as a =< a >
0
.
The thermodynamically conjugate variable to n
i
is denoted by F
i
=
S
n
i
and
is called as thermodynamic force. The variable conjugate to a
i
is then
X
i
=
S
n
i

S
n
i

e
.
The Onsager principle reads now
J
i

da
i
dt
=

j
L
ij
X
j
,
where J
i
is the average thermodynamic ux and L
ij
satises the following
constraints:
The coupling matrix L must satisfy the Onsager-Casmir reciprocal sym-
metry relationships:
L
ij
(B) =
i

j
L
ji
(B),
where
i
= 1 when n
i
is odd variable, +1 otherwise. Recall that
variable n
i
is called odd when its sign changes on time reversal. In
general, time reversal changes the sign of time, velocities, and magnetic
eld. The reciprocal relations follow from microscopic reversibility that
itself follows from the two general symmetry properties of the two-time
correlation function:
C() < a
0
a
T
>
1
= E
1
exp(H
T
)
where W
1
(a) = G(0, E, a) and E = S/k
B
.
The coupling matrix L must be positive semi-denite. This property
is closely related to the Second Law of thermodynamics from which
follows that the entropy must be maximised at the equilibrium state
and has the following form if the system is displaced by a small amount
of a from the equilibrium
S(a) = S(0) +
1
2

i,j
S
ij
a
i
a
j
,
where S
ij
=

2
S
n
i
n
j

e
is negative semi-denite matrix. On the other
hand for small a we have
X
i
=
S
n
i
(0) +

j
S
ij
a
j
(
S
n
i
)
e
6
and S(a) = S(0) +
1
2

i
X
i
a
i
. Since the Second Law holds only on the
average, then
dS(a)
dt
=

i
X
i
da
i
dt
=

i,j
L
ij
X
i
X
j
0,
which is implies that L is a positive semi-denite matrix. Quadratic
function is called the Rayleigh-Onsager dissipation function. Only
the symmetric part of L, L
s
, causes dissipation.
The Onsager principle can be written in the following form
da
dt
= Ha,
where H = LS. Its solution reads as a(a
0
, t) = exp(Ht)a
0
, where a
0
is the
vector of initial deviations from equilibrium in the conditional ensemble.
A given system in the ensemble will not of course exhibit this exponential
relaxation due to molecular uctuations a(t) = a(t) a(a
0
, t). According to
Onsager, these small uctuations are similar to an impressed macroscopic de-
viation, only they appear spontaneously. So, they should satisfy a dynamical
equation similar to of a. This is called the Onsagers regression hypothesis
for uctuations:
da
dt
= Ha +

f,
where

f is a random term. From this follows Langevin type of equation for
a(t):
da
dt
= Ha +

f.
So a(t) must be a stationary, Gaussian, Markov process and regression hy-
pothesis is equivalent to assumption that uctuations in the equilibrium en-
semble are Gaussian and Markovian. It also means that the random term

f
is a multivariate white noise and a is an Ornstein-Uhlenbeck process.
For white noise we have
<

f(t) > = 0,
<

f(t

f
T
(t) > = (t t

).
where satises the uctuation-dissipation theorem: H +H
T
= with
the equilibrium covariance matrix =< aa
T
>
e
= E
1
= k
B
S
1
. We
have also H = LS so that = (LS(k
B
)S
1
(k
B
)S
1
SL
T
) = 2k
B
L
s
.
So, the strength of the random terms is proportional to the symmetric part
of L.
7
In summary, the Onsager theory requires choosing proper extensive vari-
ables describing the macroscopic properties of ensemble, nding the dynamic
coupling matrix L, determining from the local equilibrium entropy the ther-
modynamic coupling matrix S and then it follows that the deviations of the
extensive variables around their equilibrium values are stationary, Gaussian,
Markov processes. The single-time probability density of the deviations is a
Gaussian centred at zero with the covariance = k
B
S
1
. As a function
in time, the deviations satisfy the Langevin type equation with white noise
term as random force. The strength of the random force is dened by the
symmetric part of the dynamic coupling matrix L which is also responsible
for the dissipation function to be non-negative.
8