July 1, 2009 11:54 WSPC/102-IDAQPRT 00360

Innite Dimensional Analysis, Quantum Probability

and Related Topics
Vol. 12, No. 2 (2009) 213229
c World Scientic Publishing Company
A MATHEMATICAL APPROACH TO JARZYNSKIS
IDENTITY IN NONEQUILIBRIUM STATISTICAL
MECHANICS
EVELINA SHAMAROVA
Grupo de Fsica-Matem atica da Universidade de Lisboa
Portugal
[email protected]
Received 18 April 2008
Communicated by F. Fagnola
We develop a mathematical approach to the nonequilibrium work theorem which is
traditionally referred to in statistical mechanics as Jarzynskis identity. We suggest a
mathematically rigorous formulation and proof of the identity.
Keywords: Nonequilibrium work theorem; Jarzynskis identity; nonequilibrium statistical
mechanics; probability measures on phase space paths.
AMS Subject Classication: 60J25, 60J35, 82C05
1. Introduction
The nonequilibrium work theorem is an equation in statistical mechanics that
relates the free energy dierence F to the work W carried out on a system dur-
ing a nonequilibrium transformation. The identity appeared in dierent, but as we
show below, equivalent formulations in Refs. 1 and 2 in 1997, and in the series of
papers, Refs. 47 in 19771981.
In physics literature, the identity is usually written in the form
e
W
= e
F
, (1)
where the average is taken over all possible system trajectories in the phase space,
and is an inverse temperature. The identity rst appeared in this form in Refs. 1
and 2. Traditional equilibrium thermodynamics tells us that W F while the
transformation of the system is innitely slow. The identity (1) is a stronger state-
ment, and in addition to this, it is valid for arbitrary transformations of the system.
The identity is used eectively in computer simulations, as well as in experimen-
tal physics, to calculate the free energy dierence between two states of the system
by running many trajectories and taking the average value of e
W
(see Refs. 812,
1416 and references therein).
213
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
214 E. Shamarova
The paper Ref. 3, discusses the connection between two dierent versions of the
identity, and shows that Refs. 47 use a dierent denition of work. The identity
obtained in Refs. 47 (referred to below as BochkovKuzovlevs identity) reads:
e
W0
= 1,
where W
0
is the work (in BochkovKuzovlevs sense) performed on the system, and
the angle brackets have the same meaning as in (1). The present paper shows that
Jarzynskis and BochkovKuzovlevs identities easily follow from each other.
Since the identities involve taking an average over trajectories, it is natural
to interpret this average as the expectation relative to a probability measure on
trajectories, while assuming that the system evolves stochastically. In terms of
expectations, the identities can be represented by the formulas
E[e
W
] = e
F
and E[e
W0
] = 1,
where E is the expectation relative to a probability measure on phase space paths.
For this probability measure, some analytical assumptions under which the identi-
ties hold are found.
2. Notation and Assumptions
Let us assume that the evolution of our system is described by a Markov process

t
(), t [0, T], given through its transition density function. Let X = R
2d
be the
phase space for our system, i.e. the set of values of
t
. We assume that at time t = 0
the distribution on the phase space X is given by the following density function:
q
0
(x) =
e
H(x,0)
_
X
e
H(x

,0)
dx

=
1
Z
0
e
H(x,0)
,
where H( , ) : X R is a Hamiltonian parametrized by an externally controlled
parameter , R
l
is an open set; =
1
kBT
, k
B
is the Boltzmann constant, T
is the temperature of the system, Z
0
=
_
X
e
H(x,0)
dx is the partition function.
We assume that for all
0
,
_
X
e
H(x,0)
dx < . We consider the situation
when the external parameter is a function of time [0, T] , i.e. we actually
consider a time-dependent Hamiltonian H(x, (t)). Let E
0
denote the expecta-
tion relative to the measure q
0
(x)dx. Below we assume that the changing in time
external parameter belongs to the space
V [0, T] = { : [0, T] , =
c
+
step
;
c
C
V
,
step
L
step
},
where C
V
= C
V
[0, T] is the space continuous function of bounded variation on
[0, T], and
L
step
= L
step
[0, T] =
_
() =
n1

i=0

i
I
[ti,ti+1)
() +
n
I
{T}
()
_
is the space of right continuous step functions corresponding to dierent partitions
P = {0 = t
0
< < t
n
= T} of [0, T] and dierent nite sets of values {
i
}.
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 215
To emphasize the fact that the function V [0, T] gives rise to the process

t
, we will use the notation

t
. Let p

(s, x, t, y), s, t [0, T], s < t, x, y X,

be the transition density function for

t
. Let X
[0,T]
denote the space of all paths
[0, T] X. In terms of p

we can construct a probability measure L

on X
[0,T]
by
means of the nite-dimensional distributions
_
X
[0,T]
f((t
0
), (t
1
), . . . , (t
n
)) L

(d)
=
_
X
dx
0
q
(t0)
(x
0
)
_
X
dx
1
p

(t
0
, x
0
, t
1
, x
1
)

_
X
dx
n
p

(t
n1
, x
n1
, t
n
, x
n
)f(x
0
, x
1
, . . . , x
n
), (2)
where {0 = t
0
< < t
n
= T} is a partition of the interval [0, T], and f : X
n+1
R
is a bounded and measurable function. By Kolmogorovs extension theorem, the
right-hand side of this equality denes a probability measure on the minimal -
algebra of X
[0,T]
generated by all cylindrical sets. We denote this -algebra by

c
(X
[0,T]
).
Assumption 1. The measure L

is concentrated on the right continuous trajec-

tories without discontinuities of the second kind.
Assumption 1 is fullled, for example, when the function p

satises one of
the conditions given in Ref. 17 (Chap. 2, paragraph 1). Also, we assume that the
-algebra
c
(X
[0,T]
) is augmented with all subsets of L

-null sets.
We take X
[0,T]
as the probability space, i.e. we set = X
[0,T]
, and L

as the
probability measure on it. Then, L

-a.s.,

t
() = (t).
For each xed

, we introduce another transition density function
p(s, x, t, y,

) which represents the situation when the system evolves being con-
trolled by a constant in time parameter. The following Assumption 2 is the key
assumption under which the nonequilibrium work theorem holds.
Assumption 2. If |
[s,t]


, then p

(s, x, t, y) = p(s, x, t, y,

), where
p(s, x, t, y,

) conserves the canonical distribution on the phase space X. Speci-

cally, it satises the identity
_
X
q

(x) p(s, x, t, y,

) dx = q

(y). (3)
We make further assumptions:
Assumption 3. If is constant on [s, t), and discontinuous at the point t, then
for Lebesgue almost all x X,
_
X
dy p

(s, x, t, y)f(y) =
_
X
dy p(s, x, t, y, (s))f(y),
where f : X R is bounded and Borel measurable.
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
216 E. Shamarova
Assumption 4. For all t [0, T), and for all compacts K X,
lim
0+
sup
xK
__
X
p(t, x, t +, y, (t))f(y)dy f(x)
_
= 0,
where f : X R is bounded and continuous.
Lemma 1. Let P = {0 = t
0
< t
1
< < t
n
= T} be a partition of [0, T], and let
() =

n1
i=0

i
I
[ti,ti+1)
( ) +
n
I
{T}
( ). We assume that if |
[s,t]


, where


is constant, then p

(s, x, t, y) = p(s, x, t, y,

). Further we assume that the function

p

satises Assumption 3. Then, for all s, t, 0 s < t T, for all x X, and for
all bounded and Borel measurable functions f : X R,
_
X
dy p

(s, x, t, y) f(y)
=
_
X
dx
1
p(s, x,
1
, x
1
, (s))
_
X
dx
2
p(
1
, x
1
,
2
, x
2
, (
1
))

_
X
dy p(
k
, x
k
, t, y, (
k
)) f(y), (4)
where {s <
1
< <
k
< t} = (P {s, t}) [s, t].
Proof. The proof follows from Assumption 3.
3. Jarzynskis Identity
Let =
c
+
step
be the decomposition of V [0, T] into a sum of a
c
C
V
[0, T]
and a
step
L
step
[0, T]. Further let

H : X R
l
denote the partial derivative
with respect to the second argument (i.e. with respect to the control parameter).
We assume that the partial derivative

H exists on X([0, T]), and is bounded.

Also, we assume that for each xed

([0, T]), H( ,

) is bounded and Borel

measurable. Everywhere below, the probability space is the space X
[0,T]
. We
dene the work W

: R performed on the system by the formula

W

() =
_
T
0

H(

t
(), (t)) , d
c
(t)
R
l
+
n

i=1
(H(

ti
(), (t
i
)) H(

ti
(), (t
i
0))), (5)
where {0 = t
0
< t
1
< < t
n
= T} are discontinuity points of , ,
R
l is the
scalar product in R
l
, and the integral on the right-hand side is the Lebesgue
Stieltjes integral, i.e. the sum of the LebesgueStieltjes integrals with respect to the
components of
c
. In the following, we skip the sign of the scalar product in the
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 217
rst term of (5), and simply write
W

() =
_
T
0

H(

t
(), (t)) d
c
(t)
+
n

i=1
(H(

ti
(), (t
i
)) H(

ti
(), (t
i
0))).
Let F

=
1

ln Z

(free energy of the system), and let F = F

(T)
F
(0)
(free
energy dierence). Let E
L

denote the expectation relative to the measure L

.
3.1. Jarzynskis identity, case L
step
[0, T]
Clearly, if =

n1
i=0

i
I
[ti,ti+1)
+
n
I
{T}
L
step
[0, T],
W

() =
n

i=1
(H(

ti
(),
i
) H(

ti
(),
i1
)).
Theorem 1. (Jarzynskis identity: case L
step
[0, T]) Let L
step
[0, T], and let
the transition density function p

satisfy Assumptions 2 and 3. Further let H( ,

)
be bounded and Borel measurable on X for each xed

([0, T]). Then the func-
tion e
W

is L

-integrable, and
E
L

[e
W

] = e
F
.
Proof. L

-a.s.,
W

() =
n

i=1
(H((t
i
),
i
) H((t
i
),
i1
)).
Note that W

() is a cylinder function. By Lemma 1,

E
L

[e
W

] =
_
X
[0,T]
e

P
n
i=1
(H((ti),i)H((ti),i1))
L

(d)
=
_
X
dx
0
q
0
(x
0
)
_
X
dx
1
p(t
0
, x
0
, t
1
, x
1
,
0
)

_
X
dx
n1
p(t
n2
, x
n2
, t
n1
, x
n1
,
n2
)

_
X
dx
n
p(t
n1
, x
n1
, t
n
, x
n
,
n1
) e

P
n
i=1
(H(xi,i)H(xi,i1))
.
(6)
Assumption 2 implies that for all s < t, for all y X, and

,
_
X
dx
e
H(x,

)
e
H(y,

)
p(s, x, t, y,

) = 1. (7)
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
218 E. Shamarova
Taking into account this, and changing the order of integration in (6), we obtain
E
L

[e
W

] =
1
Z
0
_
X
dx
n
e
H(xn,n)

_
X
dx
n1
e
H(xn1,n1)
e
H(xn,n1)
p(t
n1
, x
n1
, t
n
, x
n
,
n1
)

_
X
dx
0
e
H(x0,0)
e
H(x1,0)
p(t
0
, x
0
, t
1
, x
1
,
0
). (8)
Starting from the end, we replace each integral in (8) with 1, which is valid by the
relation (7), until we reach the very rst integral (taken with respect to x
n
), which
we replace with Z
n
. Noticing that
0
= (0) and
n
= (T), we obtain
E
L

[e
W

] =
Z
(T)
Z
(0)
.
The theorem is proved.
3.2. Jarzynskis identity, case C
V
[0, T]
Theorem 2. (Jarzynskis identity: case C
V
[0, T]) Let C
V
[0, T], and let the
transition density function p

satisfy Assumptions 2, 3, and 4. Let the probability

distribution L

of

t
be given by (2), and satisfy Assumption 1. In addition, let the
following assumptions be fullled:
5. If
n
L
step
[0, T], and as n ,
n
uniformly on [0, T], then L

n L

weakly relative to the family of bounded continuous cylinder functions;

6. The function H( ,

) is bounded and Borel measurable on X for each xed

([0, T]);
7. The partial derivative

H exists at all points of X ([0, T]), and is bounded;

8. The functions

H(x, ) are equicontinuous as a family of functions parametrized

by x X;
9. The function

H( ,

) is continuous for each xed

.
Then, the function e
W

is L

-integrable, and
E
L

[e
W

] = e
F
. (9)
Lemma 2. Let V [0, T]. Then, under Assumptions 1 and 79, the function
W

() given by (5) is
c
-measurable.
Proof. Below, for an arbitrary small we construct a
c
-measurable function
F : R such that
sup

|F() W

()| < . (10)

July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 219
We nd a function
step
=

n1
i=0
(t
i
)I
[ti,ti+1)
+ (t
n
)I
{tn}
L
step
[0, T] such that
for all , for all t [0, T],
|

H(

t
(), (t))

H(

t
(),
step
(t))| < .
This is possible by Assumption 8. Thus, it suces to nd a measurable function
F verifying (10) for W

of the form
_
r
s

H(

t
(),

) d
c
(t), where

([0, T]) is
xed, and
c
is not a constant on [s, r].
By Assumption 1, L

-a.s., the paths of

t
are right continuous and without
discontinuities of the second kind. This implies that the map

: [0, T]
X is (B([0, T])
c
, B(X))-measurable (where B([0, T]) and B(X) are Borel -
algebras) which follows from Ref. 17 (Chap. 2, Theorem 11). By Assumption 7,

H( , ) is bounded, say by a constant M. We divide the ball of radius M in

R
l
with the center in the origin into a nite number of sets O
i
whose diameter is
smaller than

V
T
0
[c]
, where V
T
0
[
c
] denotes the variation of
c
on [0, T]. Further let
A
i
=

H( ,

)
1
(O
i
), and C
i
= (

)
1
(A
i
) [0, T] . The sets A
i
are open
by Assumption 9. The sets C
i
are B([0, T])
c
-measurable by the (B([0, T])

c
, B(X))-measurability of the map

. Fix x
i
A
i
, and consider the function
(, t) =

i

H(x
i
,

)I
Ci
(, t). Clearly, sup
t[0,T],
|

H((t),

) (, t)| <

V
T
0
[c]
. We dene
F() =
_
r
s
(, t)(dt) =

H(x
i
,

({t : (, t) C
i
}),
where

is the LebesgueStieltjes measure on [0, T] corresponding to the function

. Fubinis theorem implies that the function

({t : (, t) C
i
}) is
c
-
measurable. The inequality (10) is obviously satised.
Proof of Theorem 2. We take a sequence of partitions P
n
= {0 = t
n
0
< t
n
1
<
< t
n
n
= T}, and consider the functions

n
(t) =
n1

i=0
(t
n
i
)I
[t
n
i
,t
n
i+1
)
(t) +(T)I
{T}
(t).
Clearly,
n
on [0, T], and V
T
0
[
n
] < V
T
0
[], where V
T
0
is the variation on [0, T].
We set
n
i
= (t
n
i
), and dene the functions

n
() = e

P
n
i=1
(H((t
n
i
),
n
i
)H((t
n
i
),
n
i1
))
.
By Theorem 1,
_

n
() L

n(d) =
Z
(T)
Z
(0)
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
220 E. Shamarova
for all n. We denote () = e
W

()
, and prove that
lim
n
_

(
n
() ())L

n(d) = 0, (11)
lim
n
_

(
n
() ())L

(d) = 0, (12)
lim
n
lim
m
_

(
n
() ())L

m(d) = 0. (13)
For this, we rst replace the functions
n
in (11)(13) with more suitable functions

n
such that (
n
()
n
()) 0 as n , uniformly on . We have:
n

i=1
(H((t
n
i
),
n
i
) H((t
n
i
),
n
i1
))
_
T
0

H((t),
n
(t)) d
n
(t)
=
n

i=1

H((t
n
i
),

n
i
)(
n
i

n
i1
)
n

i=1

H((t
n
i
),
n
i
)(
n
i

n
i1
), (14)
where in the rst term in (14) we applied the mean value theorem to each summand,
and chose

n
i
[
n
i
,
n
i+1
]. Since by assumption

H(x, ) is equicontinuous, the

absolute value of the dierence in (14) does not exceed V
T
0
[] where is chosen so
that |

H(x,

1
)

H(x,

2
)| < whenever |

2
| < , and is chosen by the
equicontinuity argument. The relation (14) shows that if we prove (11)(13) with

n
() = e

R
T
0

H((t),
n
(t)) d
n
(t)
substituted for
n
, then we prove (11)(13).
We dene the functions:

n
(t, ) =
n1

i=0
(t
n
i
)I
[t
n
i
,t
n
i+1
)
(t) +(T)I
{T}
(t);

n
(t, ) = (0)I
{0}
(t) +
n

i=1
(t
n
i
)I
(t
n
i1
,t
n
i
]
(t); (15)

n
(t) = (0)I
{0}
(t) +
n

i=1
(t
n
i
)I
(t
n
i1
,t
n
i
]
(t).
With the help of these functions, the second term in (14) can be represented as
_
T
0

H((t),
n
(t))d
n
(t) =
_
T
0

H(
n
(t, ),
n
(t))d
n
(t)
=
_
T
0

H(
n
(t, ),

n
(t))d(t). (16)
By Assumption 8, we choose an > 0 so that |

H(
n
(t, ),

n
(t))

H(
n
(t, ), (t))| < whenever sup
t[0,T]
|

n
(t) (t)| < , and is chosen
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 221
by the equicontinuity argument. This means that the relations (11), (12), and (13)
are equivalent to
lim
n
_

(
n
() ())L

n(d) = 0,
lim
n
_

(
n
() ())L

(d) = 0,
lim
n
lim
m
_

(
n
() ())L

m(d) = 0,
(17)
where
n
() = e

R
T
0

H(
n
(t,),(t))d(t)
, and
n
(t, ) is given by (15). We show
that the relations (17) hold. Since

H( , ) is bounded on X([0, T]) by Assump-

tion 7, is a function of bounded variation on [0, T], and the exponent is Lipschitz
on bounded domains, for all m we obtain the estimate

(
n
() ())L

m(d)

< K
L
_

m(d)
_
T
0
||(dt)|

H(
n
(t, ), (t))

H((t), (t))|
= K
L
_
T
0
||(dt)
_

m(d)|

H(
n
(t, ), (t))

H((t), (t))|, (18)

where K
L
is the Lipschitz constant for the exponent, ||(dt) is the Lebesgue
Stieltjes measure corresponding to the total variation function ||(t), and Fubinis
theorem has been applied to pass to the last integral. The same estimate holds for
L

. Namely,

(
n
() ())L

(d)

= K
L
_
T
0
||(dt)
_

(d)|

H(
n
(t, ), (t))

H((t), (t))|. (19)

We would like to show that
lim
n
_

n(d)|

H(
n
(t, ), (t))

H((t), (t))| = 0, (20)

lim
n
_

(d)|

H(
n
(t, ), (t))

H((t), (t))| = 0, (21)

lim
n
lim
m
_

m(d)|

H(
n
(t, ), (t))

H((t), (t))| = 0. (22)

July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
222 E. Shamarova
Let t (t
n
i1
, t
n
i
], and let
n
= t
n
i
t. Then
n
(t, ) = (t +
n
). Let us show (20).
We have:
_

n(d) |

H((t +
n
), (t))

H((t), (t))|
=
_
X
q
0
(x
0
)dx
0
_
X
dxp

n(t
0
, x
0
, t, x)

_
X
dx
i
p(t, x, t +
n
, x
i
, (t)) |

H(x, (t))

H(x
i
, (t))|.
Note that
F
n
(x) =
_
X
dx
i
p(t, x, t +
n
, x
i
, (t)) |

H(x
i
, (t))

H(x, (t))|
converges to zero uniformly in x running over compacts in X which follows from
Assumption 4. Indeed,
|F
n
(x)|
2

_
X
dx
i
p(t, x, t +
n
, x
i
, (t)) |

H(x
i
, (t))

H(x, (t))|
2
.
The right-hand side of this inequality converges to zero uniformly in x K X,
where K is an arbitrary compact. This easily follows from Assumption 4 after we
separate in the right-hand side the terms depending on x and on x
i
. Dene the
measures:

n(A) =
_
X
q
0
(x
0
) dx
0
_
A
p

n(t
0
, x
0
, t, x) dx,

(A) =
_
X
q
0
(x
0
) dx
0
_
A
p

(t
0
, x
0
, t, x) dx.
By Assumption 5, L

n converges weakly to L

relative to the family of bounded

continuous cylinder function, whenever
n
. This implies that

weakly
(relative to the family of bounded continuous functions), as
n
. By Prokhorovs
theorem, the family {

n,

} of probability measures on X is tight. We x an

arbitrary > 0 and nd a compact K

such that

n(XK

) < for all n. Since

F
n
0 on K

, and all F
n
are bounded on X by a constant, say M, not depending
on n, we can nd an N N such that |F
n
(x)| < on K

for all n > N. We obtain

the estimate:

_
X

n(dx)F
n
(x)

sup
xK
|F
n
(x)| +M

n(XK

) +M.
This proves (20). The relation (21) follows from the right continuity of , Assump-
tion 9, and Lebesgues theorem. The relation (22) follows from Assumption 5 and
(21). Application of Lebesgues theorem to the integral (18) (taken over [0, T] with
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 223
respect to ||(dt)) implies that as n , the integral (18) converges to zero. This
implies (17), and thus (11), (12) and (13) are proved. By (11), we obtain:
0 = lim
n
_
(
n
L

n L

n) = lim
n
_

n
L

n lim
n
_
L

n
=
Z
(T)
Z
(0)
lim
n
_
L

n.
On the other hand, (12) and (13) imply:
0 = lim
n
lim
m
_
(

nL

m L

m)
= lim
n
lim
m
_

n
L

m lim
m
_
L

m
= lim
n
_

n
L

lim
m
_
L

m =
_
L

lim
m
_
L

m.
Comparing the last two relations gives:
_

()L

(d) =
Z
(T)
Z
(0)
.
The theorem is proved for the case C
V
[0, T].
3.3. Jarzynskis identity for V [0, T] and its corollaries
Corollary 1. (Corollary of Theorem 2) Let the assumptions of Theorem 2 be
fullled, and let f : X R be bounded Borel measurable. Then,
E
L

[(f
T
) e
W

] = E
(T)
[f] E
L

[e
W

], (23)
where
t
: X
[0,T]
X, (t) is the evaluation mapping, E
(T)
is the expectation
relative to the measure
1
Z
(T)
e
H(x,(T))
dx.
Proof. Assuming that L
step
[0, T], we repeat the argument of (6) and (8), while
using the relation (7). Specically, we obtain:
E
L

[(f
T
) e
W

]
=
1
Z
(t0)
_
X
e
H(xn,(tn))
f(x
n
) dx
n

_
X
e
H(xn1,(tn1))
e
H(xn,(tn1))
p(t
n1
, x
n1
, t
n
, x
n
, (t
n1
)) dx
n1

_
X
e
H(x0,(t0))
e
H(x1,(t0))
p(t
0
, x
0
, t
1
, x
1
, (t
0
)) dx
0
= E
(T)
[f]
Z
(tn)
Z
(t0)
= E
(T)
[f] E
L

[e
W

].
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
224 E. Shamarova
Now let C
V
[0, T], and let the constant M
f
be such that sup
xX
|f(x)| < M
f
.
As in the proof of Theorem 2, we nd a sequence of step functions
n
converging
to uniformly on [0, T]. The identity (23) holds for each
n
. The argument of
passing to the limit as n is similar to the argument used in the proof of
Theorem 2. We dene the function F : R, F() = f((T)), and repeat the
argument of Theorem 2 until the inequality (19) with the following replacements:
F,
n
F
n
,
n
F
n
,
n
F
n
. The right-hand sides of (18) and
(19) will be the same as in the proof of Theorem 2 but the Lipschitz constant K
L
will be replaced with K
L
M
f
. The part of the proof of Theorem 2 following after
the inequality (19) remains unchanged until the last two arguments of passing to
the limit. Those arguments now will be:
0 = lim
n
_
(F
n
L

n FL

n) = lim
n
_
F
n
L

n lim
n
_
FL

n
= E
(T)
[f] E
L

[e
W

] lim
n
_
FL

n.
On the other hand,
0 = lim
n
lim
m
_
(F

nL

m FL

m)
= lim
n
lim
m
_
F
n
L

m lim
m
_
FL

m
=
_
FL

lim
m
_
FL

m. (24)
This implies the identity (23).
Corollary 2. (Corollary of Theorem 2) Let the assumptions of Corollary 1 except
Assumption 5 of Theorem 2 be fullled with respect to the function C
V
[0, T].
Let a , and let
a
= I
[0,T)
+a I
{T}
. Let us assume that
If
n
L
step
[0, T], and
n

a
on [0, T], then L

n L

a weakly relative to
the family of bounded continuous cylinder functions.
Then,
E
L

a
[(f
T
) e
W

] = E
(T)
[f]
Z
(T)
Z
(0)
. (25)
Proof. If L
step
[0, T], then
a
=

n1
i=1

i
I
[ti,ti+1)
+a I
{T}
. The representation
(4) shows that E
L

a
[(f
T
) e
W

] is the same for the functions

a
with dierent
a. Hence,
E
L

a
[(f
T
) e
W

] = E
(T)
[f]
Z
(T)
Z
(0)
.
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 225
Now let C
V
[0, T]. Note that in the relations (18) and (19) we integrate
with respect to ||(dt) where || is continuous. Hence, by Lebesgues theorem, the
investigation of convergence at the point T in (20)(22) is not necessary to prove
the convergence in (18) and (19). Finally we note that by assumption, and by the
relations (19) and (21), with
a
substituted for , for the rst term in the identity
(24) we obtain:
lim
n
lim
m
_
F
n
L

m =
_
FL

a .
This proves (25).
Let [s, t] [0, T] be an arbitrary subinterval, X
[s,t]
be the space of all functions
on [s, t]. We introduce further notation: Let C
V
[s, t] be the space of continuous
functions of bounded variation on [s, t], L
step
[s, t] be the space of right continuous
step functions on [s, t] taking a nite number of values. Below, on X
[s,t]
we dene
the distributions L
x
;s,t
and L
;s,t
, V [0, T], x X. Let s <
1
< <
k1
< t
be a partition of [s, t], f : X
k+1
R be bounded and measurable. The nite-
dimensional distributions of L
x
;s,t
we dene by the formula:
_
X
[s,t]
f((s), (
1
), . . . , (
k1
), (t)) L
x
;s,t
(d)
=
_
X
dx
1
p

(s, x,
1
, x
1
))
_
X
dx
k
p

(
k1
, x
k1
, t, x
k
) f(x, x
1
, . . . , x
k
).
The nite-dimensional distributions of L
;s,t
we dene by:
L
;s,t
[f] = E
(s)
[L

;s,t
[f]].
We extend the measures L
x
;s,t
and L
;s,t
to
c
(X
[s,t]
) by Kolmogorovs extension
theorem. We assume that
c
(X
[s,t]
) is augmented with all subsets of L
;s,t
-null sets.
Theorem 3. (Jarzynskis identity, case V [0, T]) Let V [0, T], and let
{t
1
< < t
k
= T} be the set of its discontinuity points. Further let the transi-
tion density function p

of the Markov process

t
satisfy Assumptions 2, 3 and 4,
and the probability distribution L

be given by (2), and satisfy Assumption 1. Let

Assumptions 6, 7, 8 and 9 of Theorem 2 be fullled. Additionally, we assume that
10. If
m
L
step
[0, T], and
m
on [0, T], as m , then for each i, 0 i <
n, L

m
;ti,ti+1
L
;ti,ti+1
weakly relative to the family of bounded continuous
cylinder functions on X
[ti,ti+1]
;
11. For all bounded and continuous functions f : X
[ti,ti+1]
R, the function X
R, x L
x
;ti,ti+1
[f] is Borel measurable.
Then, the function e
W

is L

-integrable, and
E
L

[e
W

] = e
F
. (26)
The continuity of functions X
[ti,ti+1]
R is understood with respect to the
topology of pointwise convergence.
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
226 E. Shamarova
Proof of Theorem 3, case V [0, T]. Let =
c
+
step
, where
c
C
V
[0, T],
and
step
=

n1
i=0

i
I
[ti,ti+1)
+
n
I
{T}
L
step
[0, T]. On each interval [t
i1
, t
i
] we
dene the function
i
= |
[ti1,ti]
, and identify each X
[0,T]
with the sequence
(
1
, . . . ,
n
). Since

t
is a Markov process, we obtain:
_
X
[0,T]
e
W

()
L

(d)
=
_
X
dx
0
q
(t0)
(x
0
)
_
X
[t
0
,t
1
]
L
x0
;t0,t1
(d
1
)

_
X
[t
1
,t
2
]
L
1(t1)
;t1,t2
(d
2
)

_
X
[t
n1
,tn]
L
n1(tn1)
;tn1,tn
(d
n
)e
W

(1,...,n)
,
where
W

(
1
, . . . ,
n
) =
n

i=1
_
ti
ti1

H(
i
(t), (t))d
c
(t)
+
n

i=1
(H(
i
(t
i
), (t
i
)) H(
i
(t
i
), (t
i
0))).
On each interval [t
i1
, t
i
], we dene continuous functions
(i)
(t) = (t), t [t
i1
, t
i
),
and
(i)
(t
i
) = (t
i
0), 1 i n. Let
W

(i)
(
i
) =
_
ti
ti1

H(
i
(t), (t))d
(i)
(t).
We obtain:
_
X
[0,T]
e
W

()
L

(d)
=
_
X
dx
0
q
(t0)
(x
0
)

_
X
[t
0
,t
1
]
e
W

(1)
(1)
e
(H(1(t1),(t1))H(1(t1),(t10)))
L
x0
;t0,t1
(d
1
)

_
X
[t
1
,t
2
]
e
W

(2)
(2)
e
(H(2(t2),(t2))H(2(t2),(t20)))
L
1(t1)
;t1,t2
(d
2
)

_
X
[t
n1
,tn]
e
W

(n)
(n)
e
(H(n(tn),(tn))H(n(tn),(tn0)))
L
n1(tn1)
;tn1,tn
(d
n
).
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 227
Let E

denote the expectation relative to the measure q

(x)dx. We can rewrite the

above relation in terms of expectations:
E
L

[e
W

]
= E
(t0)
[E
L

;t
0
,t
1
[e
W

(1)
(1)
e
(H(1(t1),(t1))H(1(t1),(t10)))
E
L

1
(t
1
)
;t
1
,t
2
[ E
L

n1
(t
n1
)
;t
n1
,tn
[e
W

(n)
(n)
e
(H(n(tn),(tn))H(n(tn),(tn0)))
] ]]].
(27)
Note that in (27) we always deal with a function of the form:
X
[ti1,ti]
R, e
W

(i)
()
f((t
i
)), (28)
where f is bounded and Borel measurable by Assumptions 9 and 11. Applying
Corollary 2, we obtain:
E
L
;t
i1
,t
i
[e
W

(i)
()
f((t
i
))] =
Z
(ti0)
Z
(ti1)
E
(ti0)
[f]. (29)
Also, note that the following relation holds:
E
(ti0)
[e
(H(,(ti))H(,(ti0))
E
L

;t
i
,t
i+1
[F]]
=
Z
(ti)
Z
(ti0)
E
(ti)
[E
L

;t
i
,t
i+1
[F]] =
Z
(ti)
Z
(ti0)
E
L
;t
i
,t
i+1
[F]. (30)
Applying the identities (30) and (29) to (27), we obtain:
E
L

[e
W

] =
Z
(t10)
Z
(t0)
Z
(t1)
Z
(t10)

Z
(tn0)
Z
(tn1)
Z
(tn)
Z
(tn0)
=
Z
(T)
Z
(0)
. (31)
Corollary 3. (Corollary of Theorem 3) Let the assumptions of Theorem 3 be
fullled, and let f : X R be bounded and Borel measurable. Then,
E
L

[(f
T
) e
W

] = E
(T)
[f] E
L

[e
W

]. (32)
Proof. We repeat the argument that we used in the proof of Theorem 3 to obtain
(27) in connection to the expression E
L

[(f
T
) e
W

]. Instead of the very last

expectation in (27), we obtain
E
L

n1
(t
n1
)
;t
n1
,tn
[e
W

(n)
(n)
f(
n
(t
n
)) e
(H(n(tn),(tn))H(n(tn),(tn0)))
].
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
228 E. Shamarova
Applying Corollary 2, we obtain:
E
L
;t
n1
,tn
[e
W

(n)
()
f(
n
(t
n
)) e
(H(n(tn),(tn))H(n(tn),(tn0)))
]
=
Z
(tn0)
Z
(tn1)
E
(tn0)
[e
(H( ,(tn))H( ,(tn0)))
f]
=
Z
(tn0)
Z
(tn1)
Z
(tn)
Z
(tn0)
E
(tn)
[f]. (33)
Instead of the multiplier
Z
(tn0)
Z
(t
n1
)
Z
(tn)
Z
(tn0)
in the relation (31) we obtain
Z
(tn0)
Z
(t
n1
)
Z
(tn)
Z
(tn0)
E
(tn)
[f] according to (33). This gives (32).
4. BochkovKuzovlevs Identity
Here we give a rigorous mathematical proof of the identity announced in Refs. 47
where the authors used a dierent denition of work to that in Refs. 1 and 2. The
dierence between the two denitions of work was analyzed in Ref. 3, and it was
found that
W() W
0
() = H((T), (T)) H((T), (0)),
where W
0
is the work in the BochkovKuzovlev sense. We use this equality as the
denition of the new work W
0
, and will prove the following theorem:
Theorem 4. Let the assumptions of Theorem 3 be fullled. Then,
E
L

[e
W0
] = 1. (34)
Moreover, the identities (34) and (26) are equivalent.
Proof. Applying Corollary 3 and the identity (26) we obtain:
E
L

[e
W0
] = E
L

[e
W
] E
(T)
[e
(H((T),(T))H((T),(0)))
]
= E
L

[e
W
]
Z
(0)
Z
(T)
=
Z
(T)
Z
(0)
Z
(0)
Z
(T)
= 1.
This relation also shows that the identities (34) and (26) are equivalent.
Acknowledgments
The author thanks C. Dellago for attracting her attention to Jarzynskis iden-
tity, and O. G. Smolyanov for useful discussions. This work was supported by
the research grant of the Erwin Schr odinger Institute for Mathematical Physics
(Vienna), by the Austrian Science Fund (FWF) under START-prize-grant Y328,
and by the Portuguese Foundation for Science and Technology (FCT) under the
project PTDC/MAT/69635/06.
July 1, 2009 11:54 WSPC/102-IDAQPRT 00360
A Mathematical Approach to Jarzynskis Identity 229
References
1. C. Jarzynski, Nonequilibrium equality for free energy dierences, Phys. Rev. Lett. 78
2690 (1997) 26902693.
2. C. Jarzynski, Equilibrium free energy dierences from nonequilibrium measurements:
A master-equation approach, Phys. Rev. E. 56 (1997) 50185035.
3. C. Jarzynski, Comparison of far-from-equilibrium work relations, Comptes Rend.
Phys. 8 (2007) 495506.
4. G. N. Bochkov and Yu. E. Kuzovlev, General theory of thermal uctuations in non-
linear systems, Zh. Eksp. Teor. Fiz. 72 (1977) 238243 [JETP 45 (1977) 125130].
5. G. N. Bochkov and Yu. E. Kuzovlev, Fluctuation-dissipation relations for nonequilib-
rium processes in open systems, Zh. Eksp. Teor. Fiz. 76 (1979) 10711079 [JETP 49
(1979) 543551].
6. G. N. Bochkov and Yu. E. Kuzovlev, Nonlinear uctuation-dissipation relations and
stochastic models in nonequilibrium thermodynamics, I. Generalized uctuation-
dissipation theorem, Physica 106A (1981) 443479.
7. G. N. Bochkov and Yu. E. Kuzovlev, Nonlinear uctuation-dissipation relations and
stochastic models in nonequilibrium thermodynamics, II. Kinetic potential and vari-
ational principles for nonlinear irreversible processes, Physica 106A (1981) 480.
8. G. E. Crooks, Path-ensemble averages in systems driven far from equilibrium, Phys.
Rev. E. 61 (2000) 23612366.
9. G. E. Crooks, Nonequilibrium measurements of free energy. Dierences for microscop-
ically reversible Markovian systems, J. Statist. Phys. 90 (1998) 14811487.
10. G. E. Crooks, Entropy production uctuation theorem and the nonequilibrium work
relation for free energy dierences, Phys. Rev. E 60 (1999) 27212726.
11. C. Dellago and E. Sch oll-Paschinger, A proof of Jarzynskis nonequilibrium work
theorem for dynamical systems that conserve the canonical distribution, J. Chem.
Phys. 125 (2006) 054105.
12. C. Dellago, P. L. Geissler, W. Lechner and H. Oberhofer, Equilibrium free energies
from fast-switching simulations with large time steps, J. Chem. Phys. 124 (2006)
044113.
13. E. B. Dynkin, Markov Processes (Academic Press, 1965).
14. G. Hummer and A. Szabo, Free energy reconstruction from nonequilibrium single-
molecule pulling experiments, Proc. Nat. Acad. Sci. 98 (2001) 3658.
15. G. Hummer and A. Szabo, Free energy surfaces from single-molecule force spec-
troscopy, Acc. Chem. Res. 38 (2005) 504513.
16. J. Liphardt et al., Equilibrium information from nonequilibrium measurements in an
experimental test of Jarzynskis equality, Science 296 (2002) 1832.
17. N. I. Portenko, A. V. Skorokhod and B. M. Shurenkov, Markov processes, Itogi nauki
i techniki. Sovremennye problemy matematiki, Fund. Talnye Napravleniya 46 (1989)
5248 (in Russian).