(Sekimoto) Stochastic Energetics
(Sekimoto) Stochastic Energetics
(Sekimoto) Stochastic Energetics
Editorial Board
R. Beig, Vienna, Austria
W. Beiglbock, Heidelberg, Germany
W. Domcke, Garching, Germany
B.-G. Englert, Singapore
U. Frisch, Nice, France
F. Guinea, Madrid, Spain
P. Hanggi, Augsburg, Germany
W. Hillebrandt, Garching, Germany
R. L. Jaffe, Cambridge, MA, USA
W. Janke, Leipzig, Germany
H. v. Lohneysen, Karlsruhe, Germany
M. Mangano, Geneva, Switzerland
J.-M. Raimond, Paris, France
M. Salmhofer, Heidelberg, Germany
D. Sornette, Zurich, Switzerland
S. Theisen, Potsdam, Germany
D. Vollhardt, Augsburg, Germany
W. Weise, Garching, Germany
J. Zittartz, Koln, Germany
The Lecture Notes in Physics
The series Lecture Notes in Physics (LNP), founded in 1969, reports new developments
in physics research and teaching quickly and informally, but with a high quality and
the explicit aim to summarize and communicate current knowledge in an accessible way.
Books published in this series are conceived as bridging material between advanced grad-
uate textbooks and the forefront of research and to serve three purposes:
to be a compact and modern up-to-date source of reference on a well-defined topic
to serve as an accessible introduction to the field to postgraduate students and
nonspecialist researchers from related areas
to be a source of advanced teaching material for specialized seminars, courses and
schools
Both monographs and multi-author volumes will be considered for publication. Edited
volumes should, however, consist of a very limited number of contributions only. Pro-
ceedings will not be considered for LNP.
Volumes published in LNP are disseminated both in print and in electronic formats, the
electronic archive being available at springerlink.com. The series content is indexed, ab-
stracted and referenced by many abstracting and information services, bibliographic net-
works, subscription agencies, library networks, and consortia.
Proposals should be sent to a member of the Editorial Board, or directly to the managing
editor at Springer:
Christian Caron
Springer Heidelberg
Physics Editorial Department I
Tiergartenstrasse 17
69121 Heidelberg / Germany
christian.caron@springer.com
K. Sekimoto
Stochastic Energetics
ABC
Ken Sekimoto
Matieres et Systemes Complexes
CNRS-UMR7057
Universite Paris 7
France
and
Gulliver
CNRS-UMR7083
ESPCI ParisTech
Paris
France
ken.sekimoto@espci.fr
http://www.pct.espci.fr/sekimoto/
Sekimoto, K.: Stochastic Energetics, Lect. Notes Phys. 799 (Springer, Berlin Heidelberg
2010), DOI 10.1007/978-3-642-05411-2
What is the Stochastic energetics? In brief, this is the framework that connects
the missing link between the stochastic dynamics and the thermodynamics. See
Fig. 1. Since the nineteenth century, the thermodynamics has been established
through the efforts of Carnot, Mayer, Clausius, and many others. Later on the statis-
tical mechanics is developed by Maxwell, Boltzmann, Gibbs, Onsager, Einstein, and
others. The latter discipline relates the thermodynamics to the microscopic mechan-
ics. The study of the stochastic dynamics, or the phenomena of thermal fluctuations,
has also a long history since Robert Brown reported the so-called Brownian motion
of a fine particle of pollen (Brownian motion, 1827). The modern framework of
stochastic process is established by Einstein [1], Smoluchowski [2], Perrin [3],
Langevin [4], Fokker and Planck [5, 6], Kramers [7], Ito [8], and others. More
recently this framework has been justified from micro mechanics with the aid of
the so-called projection methods by Zwanzig [9], Mori [10], Kawasaki [11], and
others. It is, therefore, natural that there is a link between the stochastic dynamics
and the thermodynamics, as Fig. 1 suggests. This book is the first lecture notes in
English on this linkage.1
The Stochastic energetics adapts to the modern development of nanotechnolo-
gies. The vast improvement of the spatiotemporal resolution in these technologies
has enabled the access to the individual thermal random processes at sub-micron
(1m = 104 cm) and sub-millisecond scales. While the entropy, which charac-
terizes ensemble of fluctuating states and processes, is a core concept of the ther-
modynamics and statistical mechanics, a complementary approach to deal with the
individual realizations is now needed also. Those what were mere thought experi-
ments (Gedankenexperiment) in the mid-twentieth century are now testable models
such as so-called Feynman ratchet and pawl [13] and many models related to the
Maxwells demon [14].
This book is, therefore, for those people who study and work on the scale of
thermal fluctuations, such as microfluidics, nano machines, nano sensing devices,
1 This is a highly enlarged and revised version of the Japanese book published before by the author
[12]. The present version has about 1.5 times of pages with many new figures and a new chapter
(Chap. 6) with respect to the Japanese version.
vii
viii Preface
Stochastic
Stochastic energetics Thermo
dynamics dynamics
Projection
methods
Statistical
Micro
mechanics
mechanics
Fig. 1 The Stochastic energetics as completion of the missing link
nanobiology, nanoscopic chemical engineering, etc. In other words, those who are
interested in the following questions will find the answer, or at least some clues, in
this book:
What is the heat associated to the thermal random (Brownian) motion of a meso-
scopic particle?
What work do we need for the operation and observation of small system?
How much is the work to operate an ideal Carnot engine? Is it reversible?
Can we cool a drop of water by agitating a nanoparticle immersed therein?
How does the heat flow if a particle undergoing Brownian motion pulls a polymer
chain?
Is the energy conserved during an individual realization of Brownian motion?
Is the projection methods, which eliminates rapid microscopic motions, compat-
ible with reversible or quasiequilibrium process?
Can we measure the free energy of the system by a single realization of stochastic
process?
Are there quantum mechanics-like uncertainty or irreversibility upon the mea-
surement of thermal random process?
Is the definition of the heat unique? Is the thermodynamics unique for any partic-
ular system ?
Does a particle carry the chemical potential when it enters into an open system
from the environment?
Why does the chemical potential of a molecule depend on its density even if the
molecule does not interact with other molecules?
Do we need an irreversible work to make a copy of the information in a bit
memory?
Can we detect reversibly the arrival of a Brownian particle with 100% of sureness
at finite temperature?
Do molecular motors need to stock a large energy in order to do a large work?
Preface ix
The readers of this book are assumed to have the very basics of thermodynamics
and Newtonian mechanics, as well as elementary analysis and the ordinary differ-
ential equation. No advanced knowledge at the level of physics graduate courses is
required. Examples in the main text are all simple.
Limitation of the scope of this book: First, we do not deal with the quantum
fluctuations. We assume that the temperature of the environment is high enough
that the quantum interferences are negligible. The mesoscopic quantum systems
are already an established field (see, for example, an inspiring book [15]). Sec-
ond, we discuss very little about the subjects of nonequilibrium statistical mechan-
ics around the fluctuation theorem (FT) and the (Jarzynski) nonequilibrium work
relation. These subjects have been rapidly developed between the late 1990s and
early 2000s, when Stochastic energetics has also been formulated. These sub-
jects deal with the ensemble of stochastic processes, where the former frame-
work has been used. I hope that some comprehensive books will be written by
the people who initiated these vast subjects. Last, as the schema of Fig. 1 shows,
the Stochastic energetics works where the mesoscopic scale is more or less well
defined as distinguished both from micro level and from macro level. It, therefore,
does not apply to the phenomena where the interference among these scales is
strong [16].
Organization of the book: This book consists of three parts. See Fig. 2. Glance
at the table of Contents will give you more detailed composition of each chap-
ters. In the three chapters of Part I we will prepare the basics of the stochas-
tic dynamics (mainly the Langevin equation), the thermodynamics, and the reac-
tion dynamics (including the master equation). The following three chapters in
Part II will introduce the basic concept of the heat on the mesoscopic scale and
describe its consequences. In the last two chapters belonging to Part III we will
see, through simple examples, various strategies and constraints in the fluctuating
world. The asterisk at the head of sections or subsections indicates the core part
of the book. The technically advanced descriptions are given in the Appendix at
the end of the book. There are already good textbooks of stochastic dynamics such
as [1719, 8], where the thermodynamic equilibrium and the fluctuationresponse
relations are well described in the context of stochastic dynamics. Complementarily
to these textbooks, we consider in this book the stochastic dynamics whose param-
eter(s) are changed in time from outside or by other subsystems. This generaliza-
tion opens a fertile field of physics and enables to fill up the missing link men-
tioned above. I hope that the readers enjoy the richness of physics in the Stochastic
energetics.
About the references: The papers cited in the main text are those to which I
owe directly their ideas, methods, and perspectives. Therefore, the references in this
book do not cover the whole activities of the communities related to the stochastic
dynamics. I apologize those people of whom I overlooked important papers either
because of my limited capacity of understanding or bibliographical search.
I would like to take this opportunity to express my deep thanks to my coauthors and
the others to whom the contents of this book owes a lot:
(The coauthors on related subjects) T. Hondou, Y. Miyamoto, E. Muneyuki,
E. Muto, T. Ooshida, C. Ribrault, S.I. Sasa, K. Sato, T. Shibata, F. Takagi, K.Tawada,
and A. Triller. (The other reserchers) A. Allahverdyan, H. Asai, C. Bagshaw,
M. Bazant, H. Hasegawa, S. Ishiwata, C. Jarzynski, T. Kodama, R. Kanada,
R. Kamiya, K. Kawasaki, T. Komatsu, A. Libchaber, S. Leibler, N. Miyamoto,
M. Mizuno, T. Munakata, N. Nakagawa, H. Noji, F. Oosawa, J. Parrondo, L. Peliti,
K. Sasaki, F. Sasaki, T. Sasada, M. Schindler, U. Seifert, A. Shimizu, K. Sutoh, Y.
Tanaka, H. Tasaki, T. Tsuzuki, M. Tokunaga, C. Van den Broeck, A. Yamada, T.
Yanagida, R. Yasuda, T. Yuge, and S. Yukawa.
I also thank the colleagues of Groupe de Physico-Chimie Theorique of E.S.P.C.I.
and the Laboratory of Matieres et Systemes Complexes of Paris 7th University, who
provided me with intellectual and comfortable research environment.
Concerning the proposition of this book to Springer, I gratefully acknowl-
edge those who have given invaluable encouragement and concrete help; M. Doi,
R. Kawai, J. Kurchan, E. Muto, S. Nakamura, I. Ojima, and S. Ramaswamy.
I especially acknowledge R. Kawai, Y. Oono, S. Sasa, and the anonymous review-
ers for reading through the draft and giving precious comments. Also R. Kawai
kindly provided me with Fig. 4.10 related to their paper [20]. R. Adhikari helped
me by pre-proofreading the beta version as a native English speaker and a physicist.
Also I appreciate P. Marcq, who kindly prepared the translation of the first few chap-
ters. Although all the comments by the above mentioned people have been taken into
account as far as I can, the responsibility of all the eventual errors, ambiguities, or
incompleteness is to the author.
I thank the senior editor of the Springer publishing, C. Caron, and the senior
editor of Iwanami Books Ltd., N. Miyabe, for their professional editorship.
References
1. A. Einstein, Investigation on the theory of the Brownian Movement, (original edition, 1926)
ed. (Dover Pub. Inc., New York, 1956), chap.V-1, pp. 6875 vii
2. M. von Smoluchowski, Ann. der Phys. 21, 756 (1906) vii
xi
xii Acknowledgment
xiii
xiv Contents
Appendix A . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
A.1 Appendix to Chap. 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
A.1.1 Examples of Probability Distribution: Appendix
to Sect. 1.1.2.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
A.1.2 A Particular Aspect of Gaussian Distribution:
FluctuationResponse: Appendix to Sect. 1.1.2.1 . . . . . . . . . 282
A.1.3 Sketch of Derivation of (1.10), (1.11), and (1.12):
Appendix to Sect. 1.1.2.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
A.1.4 Derivation of (1.23): Appendix to Sect. 1.1.3.2 . . . . . . . . . . . . 285
A.1.5 Langevin Equation Obtained by the Method of Projection
Operators: Appendix to Sect. 1.2.1.5 . . . . . . . . . . . . . . . . . . . 285
A.1.6 The Distinction Between Different Types of Calculus:
Appendix to Sect. 1.2.2.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
A.1.7 Conversion of Itos Lemma (1.58) into Stratonovich Form:
Appendix to Sect. 1.2.2.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
A.1.8 Derivation of FokkerPlanck Equation (1.73) and Kramers
Equation (1.74): Appendix to Sect. 1.2.3.1 . . . . . . . . . . . . . . 289
A.1.9 Jacobian to Transform ( ) into x( ):
Appendix to Sect. 1.3.1.1 . . . . . . . . . . . . . . . . . . . . . . . . . . 290
A.1.10 Derivation of FluctuationDissipation (FD) Relation:
Appendix to Sect. 1.3.1.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
A.1.11 Derivation of (1.93) and (1.94): Appendix to Sect. 1.3.2.1 . . 291
A.1.12 Derivation of (1.108): Appendix to Sect. 1.3.3.2 . . . . . . . . . . 293
A.1.13 Derivation of the Mean First Passage Time: Appendix
to Sect. 1.3.3.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
A.2 Appendix to Chap. 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
A.2.1 Maxwell Relation in the Fundamental Relation Assures
the Existence of Thermodynamic Function: Appendix
to Sect. 2.1.3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
A.2.2 Invariance of Thermodynamic Relations and the Choice
of Reference Energy and Entropy: Appendix to Sect. 2.1.3.4 296
A.3 Appendix to Chap. 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
A.3.1 Derivation of (3.27): Appendix to Sect. 3.3.1.4 . . . . . . . . . . . 297
A.3.2 Derivation of (3.72) and (3.73): Appendix to Sect. 3.3.3.4 . . 298
A.4 Appendix to Chap. 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
A.4.1 Derivation of (4.13): Appendix to Sect. 4.1.2.2 . . . . . . . . . . . 299
A.4.2 Error in the Euler Scheme: Appendix to Sect. 4.1.2.5 . . . . . . 300
xviii Contents
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
Chapter 1
Physics of Langevin Equation
Sekimoto, K.: Physics of Langevin Equation. Lect. Notes Phys. 799, 366 (2010)
DOI 10.1007/978-3-642-05411-2 1 c Springer-Verlag Berlin Heidelberg 2010
4 1 Physics of Langevin Equation
a cellular organelle waiting for the arrival of a molecular signal from the surrounding
cytoplasm. Or, suppose that a protein motor, say, myosin is waiting for the arrival of
an ATP molecule. The arrival of a signal molecule or an ATP molecule is at random
in the sense that there is no regularity except for its average frequency, which serves
for nothing as a signal. However, this random event can become a useful signal or
a source of mechanical work because the receptor or the protein motor are designed
so that they can respond to such event of arrival by changing their internal state,
called the conformational change [5]. The unbinding of those molecules can be
made difficult by this conformational change, and the receptor or the protein motor
will start a cascade of functional process.1
1 For the receptors or the motor proteins to function, we also need the global nonequilibrium of
the environment, which promotes the process in one direction against the reversed one.
1.1 Random Events or Fluctuations 7
m = i Pr ob[m = i].
i
In the case of the proper die, we have the average of n or m n 2 , i.e., n = 7/2
and n 2 = 91/6, or the variance, [n n]2 = 35/12. The random variable bears
both the statistical nature and the uncertainty, and therefore fits with our notion of
the variable which undergoes fluctuations.
What if we want to know the average f (n) in knowing Pr ob[n = i]? The
answer is
f (n) = f (i) Pr ob[n = i].
i
Two advanced examples of the discrete random variables (binomial distribution and
Poisson distribution) are given in Appendix A.1.1.
Continuous random variable: We may consider a random variable y that can take
any real value, for example, on the interval [0, 1]. If we chose completely at ran-
dom a point y on this interval, every value of y is, by definition, equally likely.
But the probability to realize a particular real value y is 0. Still we can discuss the
probability of finding y within a segment [a, b] with 0 a < b 1. In this case
Pr ob[ y [a, b]] = b a. In general, if a random variable y takes a value on
the real axis R, we define the probability associated to y in terms of the probability
density, p(y), a nonnegative-valued real function such that
dy dy
Pr ob y y ,y+ = p(y)dy.
2 2
2 Note that f (y)(y y)dy = f (y)(y y)dy = f ( y).
3 Those who are not interested or familiar to the analysis may skip the details of the characteristic
functions and also its applications to the Gaussian random variables which appear later on.
8 1 Physics of Langevin Equation
+
y () = eiy p(y)dy. (1.1)
d 2 i y
For example, for g(y) = y 2 , we have ( 1i d ) e = y 2 ei y |=0 = y 2 .
=0
Gaussian random variables: Among the continuous random variables, the those
called the Gaussian random variables are of particular importance for the reasons
which we will see below in different occasions. The probability density for this type
of random variables (Gaussian distribution, for short) is of the following form:
1 (yy0 )2
p (G) (y) = e 2 2 . (1.3)
2 2
Here y0 is the average, y0 = y, and 2 is the variance, 2 = [ y y]2 . The char-
acteristic function y () for the Gaussian random variable takes especially simple
form,
2 2
y(G) () = eiy0 e 2 .
y 2n y 2 n
= n . (1.4)
(2n)! 2 n!
The reader might verify the identity y f (y) = y 2 f (y), which is a simple
case of so-called Novikovs theorem.
A particular aspect of Gaussian distribution is explored in the context of the
relation between fluctuation and response [6]. In Appendix A.1.2 we describe its
outline.
Mathematicians view of random variables: The mathematicians view of probabil-
ity is apparently somehow different from the physicists view, though the essential
notions are the same. The former view assumes a fundamental random variable,
, with the probabilities (or measure) P associated to each value of in the
domain . With some technical precision for the admissible ensemble (set) of those
1.1 Random Events or Fluctuations 9
values,4 called family of sets, F, they define the probability space (, P, F). If
is a continuous set, one may imagine a monochrome canvas as where the local
brightness corresponds to the density of probability.
In this setup, any other random variable, say X , should be the function of the
fundamental one, i.e., X = f X (). Each random variable X has its own domain
X and is associated by its own probabilities PX as well as the family of sets
F X . That is, a new probability space ( X , PX , F X ) is produced. For the physi-
cists view, this change from to X could be interpreted as a change of resolution
of observation or measurement of a particular aspect of fluctuations. The ultimate
fundamental random variable should describe everything, leaving no room for the
fluctuation.
1
e 2 ( y y0 ) M( y y0 ) ,
1 t
p (G) ( y) = (1.5)
(2)d/2 det(M)
i y0 1 t 1
y(G) () =e exp M . (1.6)
2
4 Though we do not go into the detail of the technical precision, this is an essential point concern-
ing the continuous random variables. The situation is somewhat analogous and related to the fact
that the rational numbers are dense in the real number, but is very rare compared with irrational
numbers. See, p. 23 of [1].
10 1 Physics of Langevin Equation
The meaning of the matrix M is made clear from y(G) (), which tells immediately
that (yk yk,0 )(yl yl,0 ) = (M1 )kl , or M is the inverse of the covariance matrix,
( y y0 ) ( y y0 )t .
Stochastic process as a random variable: Especially important case of multicompo-
nent random variable is the case where the number of components is infinite and the
index to distinguish those components of the variable represents the time.5 We call
this random variable the stochastic process. Each realization of a stochastic process
is, therefore, a function of time. The time index, say t, of a stochastic process, say
, is represented either as the suffix like t or as the argument of a function like (t).
In order to remind that a realization of stochastic process is a random variable, we
will sometimes call it the path and denote as () where the generic index, t, is
suppressed.
When we talk about a function of a stochastic process, say f [ ], in fact f [y]
is the functional which takes a function () as variable. So is the probability den-
sity, p[y]. Then its average, f [ ], is obtained by performing infinitely multiple
integrals called the functional integral of the product f [y] p[y]. Each integral is
for each variable of integration, (t), with the index, t. In this book we will not go
into details of the functional integral approach to the stochastic process. However,
we should always keep in mind that any average f [ ] is not the average over the
value of (t) with particular index t, but over the whole paths (). It is only when
each component of obeys the mutually uncorrelated fluctuations that the average
(t) reduces to the average over a single random variable, (t).
Characteristic functional for the stochastic process: As an extension of the charac-
teristic function y () for the single random variable, y, we introduce the charac-
teristic functional for the stochastic process, [7]:
[] exp i (t) (t)dt , (1.7)
where (t) is any real function of time with proper smoothness. The brackets have
the same meaning as before, that is, we should take the sum over all the possible
realizations of .
The Gaussian stochastic process is a straightforward generalization of Gaussian
random variable of multiple components. Later we will assert that the thermal ran-
dom force is a special type of stochastic process. We define this process by its char-
acteristic functional, []. For simplicity, we consider the case where (t) = 0.6
Instead of the double sum in the bilinear form t M1 in (1.6), we need the double
integral with respect to the index of :
As any order correlation y n was given in terms of the variance and the average of y
for a single-component Gaussian random variable, any N -point correlation function
for (t1 ), , (t N ) of a Gaussian stochastic process is represented in terms of the
covariances, (t j ) (tk ). Taking twice the functional derivative of (1.8) with respect
to (t) and (t ), and then letting (t) = 0 for all t, we see that the kernel K (t, t )
yields the correlation function of , that is,
As another example of the usage of the characteristic function, let us calculate the
probability density ( (t)) for (t) at a time t. By choosing (t) = z(t), (1.8)
and (1.9) yield [z(t)] = ei z (t) = e (t) z /2 . We then have7
2 2
+ dz
2
i z i z (t) e 2 (t)2
( (t)) = e e = 1
.
2 2 (t)2 2
We assume further that these random variables share the same statistical properties.
For example, let us suppose that there is a huge container filled with a gas and
that we observe, in this container, N nonoverlapping domains of volume v. We
define the random variable ai as the total kinetic energy of the gas particles found at
a time within the ith domain. We assume that the sum of the volumes we observe,
N v, is much smaller than the total volume of the container and that each domain
is well separated from the other domains. Then, unless the gas is extremely dilute,
we expect that the random variables {ai } are mutually almost independent and that
+
ei z dz = 2 ( ) as well as eax d x =
2
7 We use /a.
8 We use the notation ai instead of yi simply to emphasize the aspect that {ai } are different random
variables, rather than the components of a multicomponent random variable y.
12 1 Physics of Langevin Equation
their probability densities are almost identical. In terms of the probability density,
we can express the independence as
p(a1 , . . . a ) = pi (ai ), {ai }: independent.
i
N
1
A N N ai .
i=1
Unlike the average of a random variable, the empirical average is still a random vari-
able. It is easy to verify that A N = ai ( a ). Our interest is how the statistical
properties of A N change with N .9
We may intuitively expect that the probability for A N to take the value away
from a decreases with N . In fact this is true but there are three different levels of
expression for this result, which we summarize briefly and intuitively here and leave
the detailed explanations in Appendix A.1.3.
9 In composing the part below the author has referred to the lecture note by Y. Oono at the Univer-
sity of Illinois at Urbana Champaign together with his private communications.
1.1 Random Events or Fluctuations 13
Roughly speaking, the error regarding the empirical average A N as the true
average a becomes 0.
Central limit theorem:
c w2
| A N a | e 2
Pr ob <c dw for N , (1.11)
c 2
N
The result (1.11) first indicates that, usually, the fluctuations of the macro-
variable A N ( N ) do not fluctuate visibly because its typical deviations, A N a ,
are of order of N 2 . Second, the typical deviations, A N a , for large N can be
1
10 Sections 1.1.3.4 and 1.2.1.5 deal with the more than one dimensions.
14 1 Physics of Langevin Equation
As an example, let us consider the simple Brownian motion, i.e., the random
motion of a fine particle (Brownian particle) freely suspended in the fluid at rest
with a homogeneous temperature. The fluid molecules undergo the collisions with
the Brownian particle. We are interested in the character of the force from the fluid
molecules onto the Brownian particle. It would, however, be too difficult to describe
those forces with the temporal resolution of the microscopic characteristic time m ,
which is related to the collisions of individual fluid molecules upon the Brownian
particle. Only the full molecular dynamic simulations can do it. But we are not
interested in the detailed information of the movement of the fluid molecules. We,
therefore, focus on the following aspects:
Single process: We look for a model of stochastic process of the Brownian particle,
i.e., the generator of individual realizations of the motion of the Brown-
ian particle. It means that we do not model the ensemble average of the
force at a given instant of time. If we do the average of some quantity, it
should be related to the temporal accumulation or empirical average (see
Sect. 1.1.2.3) in a single realization.
Mesoscipic object: We are not interested in the detailed motion of the water
molecules either causing the motion of the Brownian particle or respond-
ing to the Brownian particle. We only look at the Brownian particle and
need a good description on the mesoscopic level.
Mesoscopic timescale: We want to know the coarse-grained force along the time
axis, i.e., the rate of momentum transfer, on the Brownian particle, being
averaged over a timescale which is much larger than the microscopic
time, m .
Statistical reproducibility: It is sufficient to have a good dynamical model which
reproduces the statistical characteristics of the actual force or actual motion,
and we do not need to predict the random motion of the Brownian particle
from a particular initial condition.11 In other words, we look for a model
of the stochastic process which generates the ensemble of paths approxi-
mately identical to the ensemble of observed processes.
We should immediately say that it is not evident to be able to find the closed descrip-
tion for the force that satisfies the above conditions. If the environment of the Brow-
nian particle has a significant longtime memory, we may not separate the motion of
the Brownian particle from the dynamics of the environment. In this book we will
not study what microscopic conditions are necessary to realize such an environment.
When the environment allows the descriptions under the above constraints, we call
such environment the thermal environment.
11 In fact, any high-resolution setup of the initial condition cannot control the chaotic molecular
motion for the time
m .
1.1 Random Events or Fluctuations 15
Next we shall consider the collision force due to the fluid molecules on the Brow-
nian particle over the duration of observation t.12
(1) First we consider a hypothetical situation where the Brownian particle is fixed
in space. For t much smaller compared with m , the force of collision by the
fluid molecules from the left or from the right will be at most a single spike
(Fig. 1.1a, top).
(2) For t
m , the number of collisions by water molecules, n, will be numerous
with various amplitudes of force (Fig. 1.1a, bottom). The cumulated momentum
transfer during t, or t tot (t)dt, may obey approximately the Gaussian dis-
tribution (Fig. 1.1b).13
From the symmetry of the setup, the average of this distribution must be
0. Moreover, since the Brownian particle undergoes many random collisions
with water molecules during t, the cumulated momentum transfer over this
timescale should be sharply peaked around this average, according to the law of
large
number. In the same context, the fluctuation around the average, which is
t tot (t)dt itself, should behave as a Gaussian random variable, except for the
very rare large deviations.
force
time
t t+t
(t: small)
force
time
(t: large)
(a) (b)
Fig. 1.1 (a) Schematic time series of the force on a Brownian particle during t, which is small
(top) or large (bottom) as compared with microscopic time. (b) The distribution of the cumulated
momentum transfer during a large t
If the water does not have a long-term memory as we suppose, the random
variable t tot (t)dt for a time interval t and that for another time interval
t just next to the former should be independent. If t is small enough as
compared with the characteristic timescale of the Brownian movement itself,
we may find such an N
1 that N t is still small compared with the latter
timescale. We then have a model of random force as i.i.d. random variables.
(3) Next we suppose that the Brownian particle moves forward with a constant
velocity V relative to the fluid environment, which we define to be at rest.
The dominant collisions then come from the front of the moving particle.
For the velocity
V not being very large, we may assume a linear relationship
between t tot (t)dt and V . We will characterize this relationship by a coeffi-
cient (> 0) as follows:
tot (t)dt = V t, (1.13)
t
where the minus sign on the right-hand side indicates that the average force
should act so as to resist the movement of the Brownian particle. By this physi-
cal interpretation, we call the viscous friction
coefficient.
(4) What about the remaining fluctuations of t rem (t)dt t tot (t)dt (
V t)? By definition the average of t rem (t)dt is zero. Whether V = 0 or not,
there underlie many independent degrees of freedom. By the same reason as in
(2) above, the fluctuation must, therefore, be well approximated by a Gaussian
random
variable.14 We further approximate that as Gaussian random variable
t rem (t)dt for V = 0 is of the same character as t tot (t)dt at rest (V = 0).
14 The above argument is adopted from the lecture note by Y. Oono, UIUC.
1.1 Random Events or Fluctuations 17
Remark 1 Generally the Gaussian stochastic process with the property of (1.14) is
called white Gaussian (stochastic) process, since its frequency spectrum does not
depend on the frequency like white light.15 As mathematical object, is singular
and cannot be formulated as it is. The Wiener process and the stochastic calculus
which we will describe later have been developed to circumvent the difficulty.
Remark 2 As model of physical random force on the Brownian particle, however, we
should remember that in the context of (1.14) t = t always implies |t t |
m
and that the equality t = t has to be understood as |t t | t, where t
m .
In this context, we may replace (t) by any sharply peaked but smooth function
+
t that mimics (t), keeping the identity tdt = 1. The formula about the -
function such as t (t) = (t) may no more hold for (t) since t (t) has double
peak. However, as long as the peak is narrower than t the replacement, (t)
(t), makes no difference on the calculation of quantities of the time resolution
(smoothness) coarser than m .
Remark 3 With (1.8) and (1.14), the thermal random force as white Gaussian
stochastic process is characterized by the characteristic functional:
(t) (t)dt
ei = eb 2 (t)dt
. (1.15)
15 The remarks on the mathematical treatment of (t) will be given later in Sect. 1.2.3.1.
18 1 Physics of Langevin Equation
However, the microscopic origin of the free Brownian motion is the deterministic
mechanics. Therefore, we seek to formulate the free Brownian motion from the
viewpoint of Newtonian mechanics while we keep the timescale of interest much
larger than m . To this end, we take into account all the mechanical forces that the
Brownian particle receives from the thermal environment. From the argument in
the Sect.1.1.2.4, first of all there is a random thermal force, (t), that is, a white
Gaussian process (1.14) with the average 0, (t) = 0. Another force is the viscous
friction force, v(t), where v(t) is the velocity of the Brownian particle at time
t with respect to the thermal environment. (See (1.13) in the previous subsection.)
The viscous friction force can be written either as p/m or as d x/dt, where
m, p, and x are, respectively, the mass, momentum, and position of the Brownian
particle:16 We, therefore, have the following Newtonian-like equations of motion of
free Brownian motion:17
dp p dx p
= + (t), = . (1.16)
dt m dt m
See Fig. 1.2a for the trajectory of x(t) obtained by solving (1.16).18
Einstein introduced the basic concepts of the Brownian motion [9]. We refer
the readers to Sect. 1.2.1 of [1] for his original reasoning. Einsteins paper finally
convinced people that the heat is molecular motions [4]. The Brownian particle was
the testimony of the thermal motion. In this book we will see how the heat in the
thermal environment turns into the energy of the Brownian particle and vice versa.
(a) (b)
Fig. 1.2 Two-dimensional trajectories of a free Brownian particle. (a) A solution x(t) of (1.16),
(b) A solution x(t) of (1.19). In (a), the velocity of the Brownian particle ( vth ) is approximately
maintained along the segment for vth (m/ ). In (b) the trajectory shows fictive fine structures
even below the length vth (m/ )
16 We ignore the renormalization of the mass m due to the entrainment of the fluid molecules. See
the remark after (1.16).
17 Hereafter, we will often omit the hat, , which signifies the stochastic process. In fact (), p,
and x in (1.16) are stochastic processes.
18 The figures are drawn by simple program: ( pn+1 pn )/t = ( pn+1 + pn )/(2m) + n ,
(xn+1 xn )/t = ( pn+1 + pn )/(2m), where m = t = 1 and n are i.i.d. Gaussian random
variables with zero average. In (a) = 0.1 and in (b) = 2.
1.1 Random Events or Fluctuations 19
Einstein Relation
The compatibility requirement between (1.16) and the canonical equilibrium dis-
tribution (Maxwell distribution) of the particle velocity leads to the relation named
after Einstein:
b = kB T, (1.17)
where b appeared in (1.14). Therefore, the thermal random force (t) as the Gaus-
sian white noise is finally characterized by the following relations:
By substituting
into this integral a particular solution of the first equation of (1.16),
say, p(t ) = 0 e( /m)u (t u)du, we have
t
p2 e m (u 1 +u 2 ) 1
= lim (t u 1 ) (t u 2 )dt du 1 du 2 .
2m t 0 0 2m t 0
Since the thermal random force is the stochastic process having effectively no
time correlation (see (1.14)), the longtime average (i.e., the empirical average) in
the angular brackets with large t is the average over many independent random vari-
ables. By the law of large numbers (see Sect. 1.1.2.3), such average should converge
in the limit of t to the (ensemble) average, (t u 1 ) (t u 2 ) = 2b(u 1 u 2 )
(see (1.14)).19 By performing the remaining integrals with respect to u 1 and u 2 , we
find
p2 b
= .
2m 2
Now that we apply the same logic of the law of large numbers directly to the left-
hand side, i.e., the kinetic energy of the Brownian particle. If the temporal fluctua-
tion of the kinetic energy has only a finite correlation time, which we expect to be
of order of m/ , its longtime average (i.e., the empirical average) must converge to
the (probabilistic) average. The latter can be calculated using the canonical partition
p 2
function at temperature T . The result is 2m = kB2T , being known as the law of
20
equipartition [4]. Equating the last two estimates of the kinetic energy, we obtain
b = kB T .
dx
0 = + (t). (1.19)
dt
See Fig. 1.2b for the trajectory of x(t) obtained by solving (1.19).
The derivation is as follows: We substitute the solution of the first equation of
(1.16), i.e., p(t ) = 0 e( /m)u (t u)du, into the second equation of (1.16). The
result writes
19Note that we have not used the ensemble average, but used the convergence of the empirical
average, (t u 1 ) (t u 2 ), to the ensemble average, (t u 1 ) (t u 2 ).
20 We did all the calculations in one dimension. The generalization to three dimensions does not
change the result.
1.1 Random Events or Fluctuations 21
u
0 = d x(t)/dt + (t), (t) e m (t u)du. (1.20)
0 m
The newly defined thermal random force, , is also the Gaussian process with zero
average, (t) = 0. The temporal correlation of is21
(t) (t ) = 2 kB T e m |tt | 2 kB T (t t ).
2m
The last equation is justified as long as the temporal resolution t satisfies t
m/ . By redefining as we arrive at (1.19).
Remark 1 We have derived (1.19) without taking the statistical average of (1.16).
That is, a particular realization of free Brownian motion looks differently under
different time resolutions.22 When an evolution equation bears a time resolution t,
the solutions of the equation contain reliable information only in their frequency
components with the frequencies smaller than (t)1 . In other words, if we calcu-
late the time integral of the solution with another function of time, the latter must
vary on the timescale larger than t. Otherwise the result may lead to unphysical
conclusion. For example, if we solve (1.19) (t) using a path of Gaussian white
noise, the solution x(t) behaves like that of (1.16) only on the timescales coarser than
m/ . This explains the difference in the trajectories in Fig. 1.2. If we were to calcu-
late the velocity d x/dt using the (1.19), it is unboundedly large! 23 In short, (1.19)
is useless for the study of the momentum, p = m d x/dt, while (1.16) describes a
finite p.
Remark 2 The Eq. (1.19) contains no characteristic timescales. It implies that this
result is universal for all the time resolutions well beyond m/ . In other words, the
further coarse graining of this equation leaves the form of equation invariant.
1.1.3.2 * Diffusion by the Free Brownian Motion and the Einstein Relation
The Eq. (1.16) implies that, due to the inertia effect, a Brownian particle keeps its
velocity and orientation of movement during the time m/ , while the velocities
at the times separated more than m/ are no more correlated among each other.
From (1.16) we can calculate the mean square displacement (MSD), which yields24
where
kB T
D= . (1.22)
One can obtain the same formulas from (1.19), formally for any |t|, but |t|
m/
is always assumed (see the Remark 1 of Sect. 1.1.3.1).
The result (1.21) implies
that the net displacement of a Brownian particle during
the time t is typically 2Dt. This type of behavior is called diffusion with the
diffusion coefficient, D. The formula (1.22) which relates D with the viscous fric-
tion coefficient is also called the Einstein relation. In Sect. 3.3.1.6 we will see that
the Einstein relation is required for the detailed balance condition.
The relation (1.22) can be understood qualitatively by recalling the central limit
theorem.25 We approximate the trajectories
N of the Brownian particle as the piece-
wise linear steps, x(t) x(0) = i=1 ai . Each linear segment ai corresponds to
the time (m/ ) during which the Brownian particle moves at a constant velocity.
This velocity is typically of thermal velocity, vth = kB T /m, derived from the
law of equipartition mentioned above. Then the distance of each segment |ai | is
vth (m/ ) and the number of segments N is therefore N = t/(m/ ). Every steps
are independent, and the MSD after the N steps is N vth (m/ ). 26 Identifying the
MSD with 2Dt we recover (1.22) up to a constant factor.
25 Originally Einstein has derived this relation by considering a stationary density distribution of
suspending particles in fluid under the gravitation [14]. He required the compatibility between the
thermodynamic argument, The stationary state is the balance between the gravitational force and
the osmotic pressure, and the kinetic argument, The stationary state is the balance between the
sedimentation
and the diffusion.
N 2
N
26 a
i=1 i = i=1 (ai )2 = N (ai )2 , where we used ai a j = 0 for i = j because of the
independence.
27 The fluid viscosity is usually a decreasing function of temperature, and and are propor-
tional in many cases.
28 For a spherical particle in an incompressible simple viscous fluid (Newtonian fluid), is linearly
proportional to the radius of the particle [15].
1.1 Random Events or Fluctuations 23
(see (1.18)).29 However, a larger particle receives also stronger frictional forces .
Since the latter effect is more important for large , the thermal random force has
little influence on the motion of the macroscopic objects.
Which to measure, D or ?: For some experiments the spontaneous movements
are easier to measure than the response under external driving. For some theories
the situation could be reversed. One can study either D or , thanks to the Einstein
relation. For example, experimentalists [12] measured the 1D diffusion coefficient
D of a rod-like filament (microtubule) on a substrate coated with inactivated molec-
ular motors (dyneins). A theory [13] explained their result through the estimation of
when the filament is driven at a constant velocity.30
1 X2
P(X, t) = e 4Dt . (1.23)
4 Dt
That it takes the form of Gaussian distribution is the general consequence of the lin-
ear combination of Gaussian random variables. The derivation is given in Appendix
A.1.4.
From (1.23) we can verify that the distribution function P(X, t) obeys the fol-
lowing partial differential equation called the diffusion equation:
P 2 P
= D 2. (1.24)
t X
d
= 2D (t), (1.25)
dt
dx
= 2D u ut + 2D u ut (t), (1.26)
dt
where D = kB T / , D = kB T / , and D = kB T / are the diffusion coeffi-
cients determined from the shape of the object, and ut (t) (cos (t), sin (t)) and
ut (t) ( sin (t), cos (t)) are the transpose of u (t) and u (t), respectively. We
have also introduced the (normalized) independent thermal random forces, (t) and
t (t) = (1 (t), 2 (t)), such that
32 Experimental realization of 2D Brownian motion requires attentions both by technical [21] and
fundamental [22] reasons.
1.1 Random Events or Fluctuations 25
It takes a finite time to reorient the anisotropic object by the rotational diffusion.
During this time, the object diffuses spatially in the direction for which the diffusion
constant is large. By this reason, the spatial trajectories of the anisotropic object
show some memory, even if we ignore the inertia. See Fig. 1.3 for the comparison
between the diffusion of an isotropic molecule and that of an anisotropic one. If D
is much larger than D , the trajectory looks to have the characteristic length scale
1/2
D /D over which the preferential local axis of the diffusing object is kept
approximately constant.
While Eq. (1.26) takes a simple form, the form is not similar to the previous
Eq. (1.19), which resembles more to the Newtonian equation. In fact (1.26) can be
rendered to the form more analogous to (1.19) [23]:
with
The latter form makes explicit that the actual motion of x is the superposition of
(instantaneous) 1D Brownian motion along u and u . The transformation between
the two representations is given by the formula:
= 2 kB T u (t), = 2 kB T u (t).
Even if the orientation of the particle is not observable, the spatial trajectory
presents some informations on the rotational Brownian motion. If D = D , the
data of x(t) are in principle enough to find the values of these parameters and that
1 0 1 0
1 1
1 1 0 0
0 0
1 1
1 0 1 0
Fig. 1.3 Numerically generated trajectories of isotropic Brownian particle (thick dot: left) and of
anisotropic Brownian particle (thick bar: right) (Adapted from Fig. 2 in [19])
26 1 Physics of Langevin Equation
of D [21, 19].33 The most important point is that the diffusion phenomena does
not reflect the head and tail of the diffusing object. That is, even if the shape of
the object does not have any mirror symmetry, the linear relation (1.26) between ddtx
and (t) preserves the invariance under u u and u u . A consequence
of this fact [24, 25] will be discussed in Sect. 1.3.4.2.
dp U (x, a) p dx p
= + (t), = , (1.31)
dt x m dt m
33 In two dimension, the equations are explicitly solvable by integration, since the angular part
can first be solved separately. In three dimension, it is not the case because of the non-Abelian
character of rotation [27], as well as the possible couplings between rotation and translation for
chiral particles.
34 What then corresponds to the Newtons third law? This is the subject the chapters from Chap. 4.
35 English translation of the original paper of Langevin is found in [8].
1.2 Construction of Langevin Equations 27
where m, p, x, , and (t) are defined as before, and a is the parameter by which the
potential energy U can be changed through an external operation.36 For a moment
we assume that the parameter a does not depend on the time, t.
If the time resolution of our interest, t, is such that t
m/ , and if and T
are constant in time, we can replace (1.31) approximately by the following equation
without inertia term (cf. (b) in Sect. 1.1.3.1).
U (x, a) dx
0= + (t). (1.32)
x dt
The derivation from (1.31) to (1.32) will be given later (Sect. 1.3.2). The Eqs. (1.31)
or (1.32) are called the Langevin equations. In both cases we recall that (t) is the
white and Gaussian stochastic process satisfying (1.18). The equations of motion
for the free Brownian particle is the special cases of Langevin equation.
36 In the literatures, the control parameters are often denoted by the greek characters, , , t , etc.,
and are distinguished from the roman characters for the fluctuating variables, x, etc. In this book,
however, we will often use the roman characters also for the control parameters, like a. Although
such choice makes no essential differences, our standpoint is that the roles of x and a are relative
in certain cases, especially when we are interested in different scales of descriptions.
28 1 Physics of Langevin Equation
resolution, t
m/ ) and K x 2 /2 = kB T /2, and also with the Einsteins rela-
tion, = kB T /D.37 Finally the artificial thermal random force is characterized by
(t) = 0 and (t) (t ) = 2 kB T (t t ).
P2 p j 2
N
H = + + U (X, {xk })
2 j=1
2m
N 2
m j 2 j
U (X, {xk }) = U0 (X ) + xj X .
j=1
2 m j 2
We have assumed that the mass of the Brownian particle is unity while the micro-
scopic molecules have the mass m. The equations of motion (Hamilton equations)
for the Brownian particle and for the microscopic molecules then write
N
dX dP U0 (X ) j
= P, = + j xj X (1.34)
dt dt X j=1
m j 2
dx j pj dp j
= , = m j 2 x j + j X. (1.35)
dt m dt
Schematically, the Brownian particle is under the influence of its proper poten-
tial energy, U0 (X ), together with the harmonic potential energies through which
37 D can be measured from the MSD [x(t) x(0)]2 = 2Dt of the free diffusion (K = 0).
38Historically, this solvable model has been proposed in the course of the development of the
general framework described in Sect. 1.2.1.5.
1.2 Construction of Langevin Equations 29
the surrounding microscopic molecules exert forces to the Brownian particle (see
Fig. 1.4). The last equation for dp j /dt indicates that the spring constant of the
harmonic potential is m j 2 for the jth microscopic molecule and that j represents
the strength of the coupling between this molecule and the Brownian particle.39 The
object is to eliminate all of {x j , p j } to find the equation of motion for X and P.
Below we sketch the derivation. The points are the appearance of temperature and
the change of the property of time-reversal symmetry.
Since the Equations (1.35) are linear in {x j , p j }, we can integrate these equations
with respect to {x j , p j } assuming that X is known as a function of time:
j t
j
d X (t )
x j (t) = X (t) cos j (t t ) dt
m j 2 0 m j 2 dt
j sin( j t)
+ x j (0) X (0) cos( j t) + x j (0) , (1.36)
m j 2 j
where x j (0) = p j (0)/m is the initial velocity of the jth microscopic molecule.
Substitution of this solution into (1.34) yields the following concise form:
dX
=P
dt
dP U0 (X ) t
d X (t )
= (t t ) dt + (t). (1.37)
dt X 0 dt
Here the friction coefficient (with memory), (t), and the thermal random force
(t) are defined as follows:
N
j2
(t) cos( j t), (1.38)
j=1
m j 2
N
j sin j t
(t) j x j (0) X (0) cos j t + x(0) . (1.39)
j=1
m j 2 j
The memory term including in (1.37) represents the dynamical response of the
microscopic molecules at time t to the motion of the Brownian particle at the past
time t (< t). Although Equations (1.37) take a form similar to (1.31), the former are
purely mechanical in the following two aspects:
(i) There is no information on the temperature of the environment, that is, of the
ensemble of the microscopic molecules.
(ii) The time-reversal symmetry is strictly observed.40
(i) We assume that the positions and velocities of the microscopic molecules at
the initial time are distributed so that its probability density corresponds to the
canonical equilibrium at temperature T of the statistical mechanics:
2
N
m x j (0)2 m j 2 j
Pr ob {x j (0), x j (0)} exp + x j (0) X (0) ,
2 2 m j 2
j=1
40 The time-reversal symmetry requires that the equations of motion are unchanged upon this oper-
ation. This operation is the replacement of P P, p j p j , t t, and d/dt d/dt,
therefore, x j x j . Intuitively, this operation is like the playing back of movies. Concomitantly,
the initial conditions are regarded as the terminal conditions.
41 Any linear combination of random variables obeying Gaussian probability distributions also
obeys Gaussian probability distribution.
1.2 Construction of Langevin Equations 31
comes from our assumption that only X and P are the slowly varying degrees
of freedom. Once these assumptions and approximations are introduced, we do
not use (1.39) anymore. Therefore, the characters of the microscopic molecules
are not directly reflected anywhere in (t). Instead, we generate (t) as a white
and Gaussian stochastic process satisfying (1.40) with (t) 2 (t).
With these two operations which are not the consequence of the mechanics, the
equation of motion (1.37) for X and P is converted to the form of the Langevin
equation:42
dX
= P,
dt
dP U0 (X ) d X (t)
= + (t). (1.41)
dt X dt
This is the outline of the derivation of the Langevin equation from a simple and
purely mechanical model.
What do the two steps (i) and (ii) change the physical properties of the equation
of motion for X and P?
(a) Once the canonical distribution is adapted, the translational invariance and,
therefore, the conservation relation of the total momentum are lost.
(b) Once the microscopic construction of (t) (see (1.39)) is abandoned, the oper-
ation of the time reversal, i.e., P P, t t and d/dt d/dt, does
not leave invariant the equation of motion. In order to bring the time-reversal
symmetry into (1.41) we need to define the new time-reversal operation for the
sum, d X (t)/dt + (t).
t
42 We should evaluate the integral including the delta function with applying the rule, 0 (t
t )dt = 12 , since its original form, (t t ), is even function of its argument.
43 In Sect. 1.1.3.1 we have already used the Markov approximation to derive (1.19) without men-
tioning it.
32 1 Physics of Langevin Equation
is not the only example of the Markovian processes. The master equation, which
we will discuss in Chap. 3, also describes a Markovian process. The Hamiltonian
equations of motion also describe Markovian processes without stochastic nature.44
Remark The loss of the mechanical time-reversal symmetry ((a) above) related
to the Markovian approximation does not imply the incompatibility between the
Langevin equation and the equilibrium thermodynamics. We should recall that the
equation of free Brownian motion, (1.16), supplemented with the relation (1.18) can
reproduce the property of equipartition in equilibrium. This implies that we need to
distinguish the mechanical reversibility from the thermodynamic reversibility. In
Chap. 5 we will see that not only the equilibrium states but the reversible thermody-
namic processes can also be described on the basis of the Langevin equation.
44 The Schrodinger equation also describes deterministic Markovian process, where the wave func-
tions represent the states. The classical or quantum Liouville equation describes also Marcovian
process.
45 That is, the interactions are not quadratic in their variables.
46 The original idea of this type of projection operator is due to [33], where it was applied to the
probability density, complementarily to the observable quantities.
1.2 Construction of Langevin Equations 33
The last relation assures that the distribution of A in equilibrium obeys the canonical
one:
Ueq (A(t))
P eq (A(t)) e kB T . (1.44)
The potential energy function Ueq (a) for the slow variable A is the Helmholtz free
energy at the temperature T with the value of A being constrained at a:
Remark 1 The fact that the potential energy Ueq (a) is a (constrained) free energy
is understandable since the projection treats many microscopic degrees of freedom
as environment. This fact implies also the following thing: If one can find a ther-
modynamic relation, like the first law and so on, from the Langevin equation, then
the thermodynamic variables thus derived should not necessarily be the same as
those of the conventional thermodynamics. The conventional thermodynamic vari-
ables can be related, through the classical statistical mechanics, to the microscopic
Hamiltonian with bare potential energy functions. In other words, the thermody-
namic framework derived from the Langevin equation can describe or even find the
thermodynamic relations at the levels of description different from the macroscopic
one. From experimental point of view it is crucial to know the relation between
the thermodynamics of different levels, the issue which will be addressed later in
Chap. 6.
Remark 2 We can consider a system which interacts with two heat baths of different
temperatures. Experimentally, it will be possible to realize such a system by using a
fine charged particle which is immersed in fluid at a temperature T on the one hand
and at the same time subject under an equilibrated electromagnetic field (i.e., a black
body radiation) at another temperature T . (We assume that the radiation field does
not practically interact with the fluid.) In such setup the two thermal environments
will interact via the motion of a Brownian particle. This is a mechanical model
of heat conduction since the interaction transports the energy from the fluid to the
radiation field and vice versa. We can model such a system as coupled Langevin
equations. In a later chapter (Sect. 4.2.1.2) we will discuss this phenomenon on the
basis of the energetics of the Langevin equation. Such equations, however, will not
34 1 Physics of Langevin Equation
The projection technique applied above eliminates the slowly varying component
from the fluctuating force (t) in the Langevin equations of the form (1.31) or
(1.32). If the correlation time in (t) is negligibly small as compared with the time
resolution of our interest, the Markov approximation is justified and we replaced
finally the time correlation of (t) by (1.14), i.e., (t) (t ) = 2 kB T (t t ). Such
idealization involving the delta function has a merit of simplifying description and
calculation of represented quantities. However, as (t) or (t) are not mathematically
ordinary functions, a care must be taken when we do the calculations involving
these objects. Otherwise, the results could be widely, even qualitatively, different
from what one should have by the calculation without this idealization. (The same
remark may apply also to the description of a point charge as a delta function,
which can actually represent,
for example, an ionized molecule of finite size.) For
example, an integral x(t) (t)dt for the variable x(t) obeying d x/dt = (t) does
not have a finite average for the idealized white Gaussian process (t).
A naive solution to avoid the ambiguity related to the -function and other singu-
lar objects would be to come back to the description with a finite correlation time in
(t) and to do all the integrals using such smoothened variables. But such option
would require a case-by-case treatment of problem where complicated estimation of
errors and justification of the (re-)limiting procedure are needed. It will, therefore,
be more efficient if there are systematic rules to deal with these singular mathemati-
cal objects as they are. For this purpose, we need therefore (i) to know the situations
where the rules of ordinary calculus of the analysis are not applicable to (t) or
(t), (ii) to learn the rules of calculus specific to these singular objects, and (iii) to
recognize the results of new calculus as a limit of ordinary calculus.
It is Ito [34] who formulated as mathematical framework the Langevin equation
and any other dynamical processes involving (t) obeying (1.14). The formulation
is principally summarized by what is called Itos lemma (Sect. 1.2.2.3 below). Just
as the axiomatization of Dirac delta function by Schwartz has opened the field of
distribution in mathematics, the mathematical foundation of the (equivalent of)
Langevin equation through Itos lemma has opened the new field of analysis called
the stochastic differential equation or the stochastic calculus.
Below we describe briefly the main concepts and tools of the stochastic calcu-
lus which we will use later in this book.47 We will give mathematical meaning to
47 The author referred to the lecture note by Y. Oono at the University of Illinois at Urbana Cham-
paign in composing the present section.
1.2 Construction of Langevin Equations 35
(t) (Sect. 1.2.2.1), then to the integral with (t) (Sect. 1.2.2.2), and to the func-
tions containing (t) (Sect. 1.2.2.3), before reformulating the Langevin equation
(Sect. 1.2.2.4). Those who are not interested in the mathematical details of the cal-
culus may skip to Sect. 1.2.2.5.
or
t+dt
(s)ds = 2 kB T d Bt . (1.47)
t
where (t) is as introduced in Sect. 1.1.2.4 and in Sect. 1.1.3.1, and we have defined
d Bt Bt+dt Bt , (1.48)
d Bt = 0. (1.49)
48 We adopt the convention in the field of stochastic calculus which often represents the time
argument t as a suffix of the stochastic process, e.g., Bt for B(t).
49 We recall that, means the average over the paths (realizations of stochastic process).
36 1 Physics of Langevin Equation
d Bt d Bt = 0, if t = t ,
where t = t implies that the two intervals, [t, t + dt] and [t , t + dt ], have no
overlap.
2. Since (s) with s between t and t obeys the Gaussian distribution, its linear
functional, Bt Bt , does also.
3. Thetotal variation of Bs between s = 0 and s = t is defined by the n limit
of nk=1 |B nk t B k1 t |. This limit is divergent with probability 1 for the Wiener
n
process. It is then said that the Wiener process is not of bounded variation.50
4. The following limit holds with probability 1 (t > 0):
n
2
B nk t B k1
n t
t (n ). (1.50)
k=1
Since each segment of time, nk t k1n
t , brings an independent contribution,
the above result allows the following replacement:
(d Bt )2 = dt, (1.51)
almost everywhere (a.e.), that is, with probability 1. Notice that in (1.51) there is
no average over the paths.
As for the last two properties, 3 and 4, we might have the following intuitive expla-
nations:51 Equations (1.47) and (1.18) yield |B nk t B k1 t |2 = t/n, which then
n
implies |B nk t B k1 t | t/n. If we add up such contribution for n intervals, we
n
would have the sum nt. The last quantity diverges in the limit of infinitely fine
division of the interval. Since B nk t B k1 t with different values of k are mutually
n
independent, nk=1 |B nk t B k1 t |2 has the average, nk=1 |B nk t B k1 t |2 = t, and
n
2
n
n
k=1 |B n t B n t | t = 2t /n. Therefore, the difference
52 2 2
the variance, k k1
n
between k=1 |B nk t B k1 t | and t becomes effectively 0 in the limit of fine seg-
2
n
mentation, n . When several such calculus are combined, we should keep only
50A real-valued function f (x) is said to be of bounded variation on the interval [a, b] if, with
any partitioning, a = x0 < x1 < . . . < x N = b, and with any natural number N , the sum
N 1
k=0 | f (x k+1 ) f (x k )| is less than a positive number, M.
51 For more rigorous proofs the reader should consult a book on stochastic process, for example,
Sect. 4.2.5 of [1].
52 We recall the formula (1.4), i.e., y 2 p /(2 p)! = y 2 p /(2 p p!) with integer p 0 for y obeying
a Gaussian distribution with the average 0. This formula can be derived by expanding the special
case of (1.15), i.e., ei y = e 2 y .
1 2 2
1.2 Construction of Langevin Equations 37
the terms up to the order of dtand neglect higher order ones, where d Bt is regarded
as a quantity of the order of dt. In summary, the following prescriptions for the
infinitesimal quantities should be applied:
dt 2 0, d Bt dt 0, (d Bt )2 dt. (1.52)
As mentioned at the end of Sect. 1.1.3.1, the solution of d x/dt = (t) or the Wiener
process has no intrinsic length or timescale (scale invariance or similarity law).
Other aspects of Wiener process are described in [35].
Other interpretations are also possible. If Bt were of bounded variation, these two
would have the same limit when n . However, withBt being the Wiener
process, the difference between the above two, 2 1 , is nk=1 |B nk t B k1 t |2 +
n
O(t/n). As we have seen above, the sum on the right-hand side tends to t in the
limit of n .
To avoid such ambiguity
of the integral, we introduce detailed definitions of the
integral of the type, f (s)d Bs , together with new notations. Below, s denotes a
positive finite interval of time. Calculus of Ito type:
f (s + s) + f (s)
[Bs+s Bs ] f (s) d Bs . (1.54)
2
Notice the different symbols between f (s) and d Bs . As the essential difference is
only of the lowest order in s, the latter definition can also be regarded as the
38 1 Physics of Langevin Equation
To make clear the implication of this equation, let us temporarily introduce the
notation xt to denote the average over the processes for [t, t +dt] with a given ini-
tial condition xt . Such average can be defined because xt is a Markovian process.
Using d Bt = 0 and the nonanticipating property of the Ito-type product, we have
$ %
d x t xt = a(xt , t)dt, (d x t )2 xt
= [b(xt , t)]2 dt. (1.57)
In (1.56) the term a(x t , t)dt, which is of order dt, cannot be neglected despite the
dominance of b(xt , t) d Bt , since the average of the latter disappears due to the
nonanticipating property of b(xt , t): 53 b(xt , t) d Bt = 0.
Next we will consider any function f (xt ) including the xt as its argument and
define its increment through infinitesimal time dt(> 0):
Then d f (xt ) can be rewritten by the following formula, called Itos lemma:
b(xt , t)2
d f (xt ) = a(xt , t) f (xt ) + f (xt ) dt + b(xt , t) f (xt ) d Bt , (1.58)
2
f (xt )
f (xt+dt ) = f (xt ) + f (xt )d xt + (d xt )2 + . . . . (1.59)
2!
All the differences from the real analysis come out from the dt property of d Bt ,
which is herited by d x$t through % (1.56) and then by d f (xt ). We have kept the second
order in d xt because (d x t )2 xt = [b(xt , t)]2 dt (see above) indicates that (d xt )2 is
also of order dt, instead of (dt)2 . Explicitly, the only term of O(dt) in (d x t )2 is
b(xt , t)2 dt, and we arrive at (1.58). Itos lemma gives how the time evolution of
f (xt ) is generated by that of the Wiener process, d Bt . If we can invert the relation
f t = f (xt ), then (1.58) gives the time evolution of f t directly without solving (1.56).
The SDE (1.56) can be rewritten in terms of the Stratonovich-type calculus. The
result writes as follows:
b(xt , t) b(xt , t)
d xt = a(xt , t) dt + b(xt , t) d Bt . (1.60)
2 xt
d f (xt ) = f (xt ) d xt .
U (x, a) p p
dp = dt dt + 2 kB T d Bt . dx = dt, (1.61)
x m m
dz t = i z t (0 dt + 2kB T d Bt ). (1.62)
Even without solving (1.62), one can verify the relation, d(z t z t ) = z t dz t + z t
dz t = 0. We can rewrite (1.62) in the Ito type, and the result is
dz t = i z t (0 dt + d Bt ) kB T z t dt
= z t (i0 kB T )dt + i z t d Bt . (1.63)
Although (1.62) and (1.63) are mathematically equivalent, the former is more
appealing than the latter from the viewpoint of a single realization of stochastic
process. The Ito-type representation is, on the other hand, more adapted for the
estimation of averaged properties: In order to understand the spurious damping
term, kB T z t dt in (1.63), let us take the path average of (1.63). The result writes
dz t = z t (i0 kB T )dt. The solution z t = z 0 ei0 t ekB T t shows the damp-
ing of the amplitude of z t due to the desynchronization, i.e., the phase diffusion
among the rotators of different realizations.
We note that, if we interpreted dz = i(0 + (t))z as of the Ito type, i.e., as the
dt
SDE: dz t = i(0 dt + 2kB T d Bt ) z t , then z t z t would increased exponentially,
obeying d(z t z t ) = 2kB T (z t z t )dt.
A modified model is the amplitude modulation of the growth rate of a real
quantity z t : The model
writes dz t = (0 dt + d Bt ) z t . It has a solution,
z t = z 0 exp(0 t + (Bt B0 )), that is, log(z t ) at a given t is the Gaussian random
variable with the average of log(z 0 ) + t and the variance of t. The probability
density of z t is then said to obey the log-normal distribution.
This example will show the utility of Ito-type calculus. As seen in (1.63)
above, a practical advantage of Ito-type representation is that the average over the
path ensemble is easily taken due to its nonanticipating property. Here is another
example.
Let us consider the overdamped Langevin equation (1.32), that is,
U (x, a) dx
0= + (t).
x dt
where we have used (A.14). Since U (x, a)/ x on the second line is nonanticipating
with respect to d Bt , the path average over the Wiener stochastic processes of this
product vanishes. The path average of the above equation then yields60
2
U (x, a) U (x, a)
(t) = kB T (overdamped). (1.65)
x x2
Physical reason for the different result is that, in the case with inertia, the fluctuation
1
of U/ x, or essentially that of d xt is smooth and, therefore, O(dt), not O((dt) 2 ).
where is a constant and (t) the normalized white Gaussian process satisfying
(t) = 0 and (t) (t ) = (t t ). We will show below that first-order schemes
of SDE can leave the errors of the 1.5th order.
The simplest first-order (explicit) Euler scheme generates x(h) at time t = h
shortly after t = 0 as follows:
1& '
x(h) = x(h) + x(0) + F(x(h))h + r (h) . (1.70)
2
This algorithm assures the second-order convergence in h for simple additive noises
as above. In the case of multiplicative noise, however, this method assures only
first-order convergence. We then should use improved methods (see, for example,
[4042]). The Langevin equation on manifolds (Sect. A.4.7.3) needs the multiplica-
tive noise.
where
44 1 Physics of Langevin Equation
1/2
1 U 1 2kB T
a(x) = + kB T , b(x) = . (1.72)
x x
Using the Itos lemma, (1.58), applied to (X xt ), we obtain the following equation
called the FokkerPlanck equation:61
P(X, t) 1 U
= + kB T (X ) P(X, t). (1.73)
t X (X ) X X
This equation is called the Kramers equation [43], derived around 1940. The gener-
alization to the higher dimensionality of the (X, P)-space is straightforward.
Throughout this book, we use the term FokkerPlanck equation to mention
generically both the FokkerPlanck equation (1.73) and the Kramers equation
(1.74).
where G(X, t|X 0 , t0 ) is the solution (called Green function) of the linear
equation (1.73) satisfying the initial condition, P(X, t0 ) = (X X 0 ). The
same principle applies to the case where the effect of inertia is taken into
account.
Equation of continuity for the probability: The FokkerPlanck equation (1.73),
takes the form of equation of continuity,
61 It was first written down by Einstein [9]. See, for example, the descriptions in [1].
1.2 Construction of Langevin Equations 45
P Jx
= , (1.76)
t X
1 U
Jx + kB T P(X, t), (1.77)
X X
where the probability flux, Jx , is the probability that flows per unit of time
through the position X toward the +X direction.62 This continuity (1.76)
reflects the conservation of the probability and the fact that in each path xt
is continuous in time. For the Kramers equation, (1.74), the equation of conti-
nuity and the definition of the probability flux are constructed on the 2D phase
space, (X, P):
P Jx Jp
= , (1.78)
t X P
P
Jx P(X, P, t),
m
U P
Jp P(X, P, t) [ kB T P(X, P, t)]. (1.79)
X m P
62 Here the X also represents the state where the particle is at the space position, X .
63 With the detailed balance, (i) the forward driving of x by the potential gradient is counterbal-
anced by the more frequent backward displacements by diffusion and (ii) the diffusive displace-
ments of x into high-friction regions (i.e., large regions) are as often discouraged as do the
diffusive displacements out of such regions.
46 1 Physics of Langevin Equation
P2
e( 2m +U (X,a))/kB T
P eq (X, P, a; T ) = . (1.80)
2mkB T eU (X ,a)/kB T d X
This entropy is nonnegative and vanishes if and only if P(t) coincides with
P eq .65 By the uses of (1.73) one can verify that D(P(t) P eq ) decreases
strictly:
d D(P(t) P eq )
< 0, (P P eq ). (1.82)
dt
It thus implies the convergence of P(t) to P eq in the t limit. The func-
tional of P(t) having the property of (1.82) is referred to as Lyapnov functional
for P(t). In general, a theorem proving the approach to the equilibrium density
by way of strictly nondecreasing quantity (entropy) is called the H-theorem.
The FokkerPlanck equation and the Kramers equation are mathematically equiv-
alent to the Langevin equations, respectively, without or with the effect of inertia
in the sense that the Langevin equations can be reconstructed if we know the for-
mer ones for the probability densities. However, the Langevin equations are more
fundamental than the former equations since the particular realization of stochastic
process is the basis of its statistical distribution.
64 This distance D(P(t) P eq ) does not fulfill standard axioms of the distance, in particular,
the reflection law. D(P(t) P eq ) is also called the relative entropy or P with respect to Peq . The
KullbackLeibler distance is closely related to the LDP. See [44].
65 To verify this, we can use the inequality, p log( p/q) q + p, where p and q are arbitrary
nonnegative quantities. The above inequality can, in turn, be obtained by multiplying by p the
simple inequality, log( p/q) = log(q/ p) q/ p + 1.
1.3 Physical Implications of Langevin Equations 47
1
P[ ] exp (s)2 ds , (1.83)
4 kB T 0
except for the normalization factor. Substituting (t) by ddtx + dUd x(x) and taking care
of the change of variable, the same probability is rewritten in terms of x( ) (apart
from the initial probability distribution of x):
( t ) * +
1 dx dU 2 2U
P[x] exp + 2kB T 2 ds , (1.84)
4 kB T 0 dt dx x
with (t) (s) = 2 kB T (t s). For the process x( ) a sporadic increment in (t0 )
and a small force at the time t0 are not distinguishable. The effect of former on
x( ) is characterized by the velocity correlation in the absence of external force,
while the effect of weak external force is characterized by the linear response func-
tion R(u) for the velocity (x)
+
x(t) = R(t s) f (s)ds. (1.87)
By the causality requirement, R(u) satisfies R(u) = (u) R(u), where (u) is the
step function; (u) = 1 for u > 0, (u) = 0 for u < 0, and (0) = 1/2.
In the case (1.85), we have
2 2kB T
R(t s) = (t s) (t s), C(t s) = (t s), (1.88)
therefore
The last form of the relation between the correlation function and the response
function is called the fluctuationdissipation (FD) relation (of the first kind). This
relation (1.89) is closely related to $ the Einstein
% relation because the double time
integral of C(t) gives D = limt [ x(t)2 /(2t)] = kB T / .
By contrast, the correlation of the random force is related to the friction constant
through
The latter type of relation is called the FD relation of the second kind.
The relation (1.89) holds very generally for equilibrium fluctuations, even if we
add the potential force dU (x)/d x and also the inertial term m x to the equation of
1.3 Physical Implications of Langevin Equations 49
free Brownian motion. In other words, the FD relation for the free Brownian motion,
which is essentially that of ( ), is reserved upon the transformation of stochastic
process from ( ) to x( ) through the Langevin equation (see Sects. 1.3.1.1). This is
one of the main result of the linear response theory, including the quantum case.
See, for example, [4, 2] for the general argument. In Appendix A.1.10 we describe
the essence of the FD relation. We will come back to the FD relation in more general
case in Part II, in the context of energetics.
1 (t)1 (t ) = 21 kB T (t t ), 2 (t)2 (t ) = 22 kB T (t t ),
and 1 (t)2 (t ) = 0. We will solve these equations under the initial conditions
x1 (0) = x2 (0) = 0 at time t = 0, which has no bias toward +x or x directions.
The direct calculation shows that x(t) is correlated with the relative displacement
w(t) x1 (t) x2 (t), i.e., w(t)x1 (t) > 0 with w(t) = x1 (t) = 0. This implies
that, statistically, the motor head can detect the direction of its own displacement by
monitoring the conformation of the molecule, w(t) = x1 (t) x2 (t), without directly
referring to the coordinate system (Fig. 1.5). The Eq. (1.90) can be separated by
introducing the coordinate of the diffusion center, X = (1 x1 + 2 x2 )/(1 + 2 )
and w = x1 x2 :
2 2 k B T
2K
t
x1 (t)w(t) = x2 (t)w(t) = 1e > 0, (1.91)
1 1 + 2 K
We have seen that the Langevin equation is obtained as a result of elimination of the
rapidly varying degrees of freedom in a more microscopic model. It is, therefore,
necessary to fix a finite time resolution, t, of the description by the Langevin
equation. Concomitantly there is a finite resolution in the variables (e.g., x or p).
However, once a Langevin equation is written down, the t appears nowhere in the
equation, and its solutions (i.e., the paths or the realizations) contain arbitrarily fine
details as function of time. In general, the fine details of its solutions at timescales
below t have nothing to do with the original microscopic model.67
A logical consequence of the above fact is that there can be more than one
Langevin equations to describe the same phenomenon, one at a time resolution of
t1 and the other at t2 (> t1 ), etc. The solutions of the former Langevin equation
contain supplemental informations about the timescales between t1 and t2 . It
should, therefore, be possible to derive the Langevin equation with the time resolu-
tion t2 from the one with better time resolution t1 (< t2 ). Below we will see
two examples of the different levels of Langevin equations.
66 More precisely, we should study the probability density for w(t) with a given value of x (t),
1
where the nonzero constrained average, wx1 (t) , is an indication of the memory. Using the
probability density of w(t) (see (1.33) with x w and x0 = 0), we can show wx1 (t) =
x1 (1 + 2 /1 )/((2K t/ ) + 2 /1 ) > 0, where (s) s/(1 es ). For large 2 /1 , the memory
is maintained up to t 2 /K .
67 One might compare this situation with the fact that the touches of a painting of a landscape on
canvas do not reproduce the details of the original landscape.
1.3 Physical Implications of Langevin Equations 51
U (x, a) p p
dp = dt dt + 2 kB T (x) d Bt , dx = dt. (1.92)
x m m
The time resolution, t1 , is finer than the damping time of the inertia effect, that is,
t1 m/ .
Our aim is to convert this equation to a new Langevin equation which is (only)
valid for the time resolution t
m/ .70 In other words, we will eliminate the
momentum, p, in (1.61) as a fast variable. This problem has long been discussed
since the landmark paper by Kramers [43], and there is generally a consensus about
the result [4749]. It is expressed as SDEs, either of Ito type or of Stratonovich type,
as follows:
U (x, a)
dx = 2 kB T (x) d Bt dt, (1.93)
x
kB T (x)
d x = 2 kB T (x) d Bt U (x, a) + dt. (1.94)
x 2
The derivation has often been done by way of the FokkerPlanck equation (see
Sect. 1.2.3.1). In Appendix A.1.11 we present a derivation that uses only the coarse
graining in time along a single realization [49].
Below are the comments on the physical aspects of the above result:
1. The term T (x) in (1.94) represents the effect of thermophoresis, that is, the
[direct] thermal random force drives a Brownian particle toward the cooler region
from the hotter region. To understand precisely what it means, suppose that there
is no potential force (U (x, a) =const.). From (1.93) we find
$ %
d x t xt = 0, (d xt )2 = 2(kB T (x)/ )dt,
where xt is the conditional path average under a given value of xt . The for-
mer equation tells that there is equal probabilistic weights on d x t > 0 and on
d x t < 0 up to time t + dt, starting from a given position, xt , at time t. But if
These two equations are equivalent. Using (1.95) we can obtain the FokkerPlanck
equation corresponding to these SDEs. It writes
P 1 U
= P+ (T (x)P) . (1.97)
t x (x) x x
potential a0 x + V (x).
Um (x) V (x) 1
x = + a0 + [2 kB T ] 2 m (t), (1.98)
x x
where m (t) is a Gaussian white noise with zero average and the variance,
m (t)m (t ) = (t t ). The spatial variation of V (x) is supposed to be much
smoother than that of Um (x), while a0 need not be smaller than the magnitude of
|Um / x|, see Fig. 1.6. The spatial resolution of (1.98), which we denote by x1 ,
is finer than the period of Um , that is, x1 .
Our aim is to convert this equation to a new Langevin equation which is (only)
valid for the spatial resolution x
. (x should, however, be small compared
with the characteristic length scale of the change of V (x)/ x.)
The first step is to analyze the case with V (x) 0. The spatiotemporal coarse
graining of this case has been studied analytically [51, 52], and the result is written
in the form of free Brownian motion under drift:
1
x = vs (a0 ) + [2Deff (a0 )] 2 (t), (1.99)
where x(t) is the position of the particle on this coarse-grained description, and
vs (a0 ) x a0 is the coarse-grained steady mean velocity,72 and Deff (a0 ) is the
coarse-grained diffusion coefficient such that
The new Langevin equation (1.99) is valid for the spatial resolution x satis-
fying x
. The quantity in the limt approaches to Deff (a0 ) only for
t
2 /Deff (a0 ). A very rough estimate of Deff (a0 ) is kBT eUm /kB T , where Um is
the amplitude of the variation of Um (x), i.e., its barrier height (see Sect. 7.1.1.4).
The second step is to take into account of the spatially slowly varying potential,
V (x). Since | V / x| is small, it can be treated as a local variation of the force
parameter a as a0 a0 + a with identifying a with Vx . In the context of
Appendix A.1.2, the effect of small change of the parameter on the mean velocity is
V(x)
a0x
Um(x)
Fig. 1.6 The three potential
energies in (1.98) x
1 x a0 +a x a dvs (a)
lim = . (1.100)
(a0 ) a0 a da a=a0
1
= kB (a0 )Deff (a0 ), (1.101)
(a0 )
kB (a0 ) plays the role of u a(1) in (A.2).73 The linear perturbation, vs (a) vs (a0 ) +
1
(a0 )
a, results in [54]
1 V 1
x vs (a0 ) + [2Deff (a0 )] 2 m (t).
(a0 ) x
1 V
0 = (a0 )[x vs (a0 )] + [2 (a0 )kB (a0 )] 2 m (t) . (1.102)
x
73 Generally, v (a) is a nonlinear and complicated function of a and is not antisymmetric with
s
respect to a.
74 We use X for the spatial coordinate.
1.3 Physical Implications of Langevin Equations 55
t
1
P[0,t]
res
(x; X ) (X x(t )) dt . (1.103)
t 0
P[0,t]
res
(x; X ) P eq (X, a; T ), as t , (1.104)
1 U (X,a)/kB T
P eq (X, a; T ) e . (1.105)
Zx
Here Z x eU (X ,a)/kB T d X is the configurational canonical partition function
of the statistical mechanics.75
The convergence problem of (1.104) is an example of LDP discussed in
Sect. 1.1.2.3 along the time axis. But it is understandable intuitively that the rate
of convergence depends on the form of U (X, a)/kB T : For example, let us compare
the case of a single-well potential U (X, a) = U2 (X ) 4kB T X 2 and the double-well
potential U (X, a) = U4 (X ) 100kB T (X 2 1)2 . The latter potential has a barrier
of 100kB T around X = 0 and the valleys of 10 times narrower than the quadratic
potential U2 (X ). If we take as unit of time the convergence time of P[0,t]
res
(x; X ) for
the U2 potential, then the convergence time for the U4 potential should be much
larger than unity because of the rare chance to cross the energy barrier, i.e., about
e100 /10 1042 .76 If the convergence time for the U2 potential is a nano second
(109 s), the convergence time for U4 would be 1024 years, which is practically
unattainable.
This result implies that, in between nano second and 1024 years, there is a well-
defined time range where almost all paths of x( ) remain in one of the valleys of
the U4 potential. Within this time range, the empirical distribution P[0,t]
res
(x; X ) must
converge to P ceq
(X, a; T, x(0)):
1 U4 (X )/kB T
Z x+
e X/x(0) > 0
P ceq (X, a; T, x(0)) (1.106)
0 X/x(0) < 0
of a function of x, e.g., f (x(t)) = f (X )(X
75 An important corollary: The longtime average
x(t))d X , converges to its canonical average, f (X )P eq (X, a; T )d X .
76 The factor 101 is from the narrowness of the valley bottom, but is not essential here.
56 1 Physics of Langevin Equation
if x(0) > 0. Here Z x+ 0 eU4 (X )/kB T d X is the constrained equilibrium partition
function in the right valley. See Fig. 1.7.
Remarks
1. The time for a Brownian particle to escape (for the first time) from a valley does
not depend on the profile of the potential energy beyond the potential barrier.
Therefore, that profile can be replaced by a sharp and infinitely deep valley.
Kramers [43] used this fact to derive the efficient formula of the escape time,
called the first passage time.
2. There could be a theoretical standpoint where one is interested only in the
infinitely large time, t, to find the unique probability density, P eq (X, a; T ). In
such a case, however, we implicitly assume that the potential function is realized
by some material whose characteristic relaxation time is larger than this time, t.
Therefore, the word infinitely large has only a limited meaning.
3. In Sect. 7.1.2 we will discuss the case where a system with double minimum
potential represents a single bit of memory device. In that case the constrained
density P ceq corresponds to a well-defined bit, and the full equilibrium den-
sity P eq represents the state where the initial memory is lost.
Fig. 1.7 Brownian particle in an symmetric double-well potential (left: curve) and the initial con-
dition x(0) (left: dot). At high temperature, the residence time distribution reproduces the canonical
equilibrium distribution P eq (center), while at low temperature it reproduces the constrained equi-
librium distribution P ceq , which depends on the initial condition (right)
The result can be understood intuitively: For each realization x( ), there is enough
time for the particle to explore the instantaneous profile of the potential U ( , a) for
each value of a along the protocol. By the convergence theorem of the empirical res-
idence time distribution (1.104), the temporal integral around each a can, therefore,
be replaced by the statistical average, (, a)eq . This is the notion of quasistatic
process, and the above formula will be fully used in Sect. 5.2.1.2.
It is noteworthy that the rate of the convergence of I depends on the function
(X, a). If (X, a) takes appreciable values for those X which are rarely realized
in P eq (X, a; T ), we will need extremely slow operation (i.e., extremely large top ) for
the convergence.
78 op for operation.
79There is a related problem called the exit problem, where we study the location on the border of
at FPT.
58 1 Physics of Langevin Equation
(i) Solve the static differential elliptic equation L (x) = 1 for x under the
[Dirichlet] boundary condition, (x) = 0 for x .
(ii) is given as the value of (x0 ).
U
Here L x (x)1
x
+ x kB T (x) is the operator of Fokkerlanck equation, and
1 80
L is its adjoint operator, i.e., L = U x
+ kB T (x) x (x) x
.
For example, when we study the Brownian particle in the double-well potential
energy U4 (x) (see Sect. 1.3.3.1) we can use the MFPT to estimate the waiting time
until the Brownian particle crosses over the potential barrier. This waiting time is
the MFPT with the domain = [0, ], for example. In such case the FPT depends
little on the initial position x0 as long as it is near the bottom of the valley.
The first passage problem is applied also to the free Brownian motion but with
various shape of domain . For example, the Brownian particle is a macro-ion and
is an absorbing surface.
1.3.4.1 Background
Models of autonomous heat engine on the mesoscopic scale have long been pro-
posed: The Feynman ratchet (Feynmans pawl and ratchet wheel) [55] and the
ButtikerLandauer ratchet [56, 57] are among the most studied. In the former model,
the system has simultaneous access to the two thermal environments of temperatures
T and T , while in the latter model, the system has alternative access to one of these
environments.
The aim of this subsection is to describe their models and show how it is mod-
eled using the Langevin equations. These models have played important roles in the
physics of stochastic phenomena. Parrondo [58] was the first who added a caveat
on the analysis by Feynman on his ratchet model.81 Dereny, Bier, and Astumian
80 L is also called thebackward Fokkerlanck operator. In general when the operators L and L
satisfy g(L f )d x = (L g) f d x for any f and g satisfying a common homogeneous boundary
condition, L and L are called adjoint.
81 Soon later [60] also pointed out it independently.
1.3 Physical Implications of Langevin Equations 59
[59] were among the first who argued that the overdamped limit of the Buttiker
Landauer ratchet is a singular limit. Stochastic energetics has first been applied to
the Feynmans model [60]. We will come back to these models in Sect. 4.2.2.2 for
the analysis from the energetics viewpoint.
dx U
0 = + 2 kB T (t),
dt x
dy U
0 = + 2 kB T (t), (1.110)
dt y
T T
Fig. 1.9 Feynman pawl and ratchet wheel [55]. The left heat bath (temperature T ) moves the
ratchet wheel, while the right heat bath (temperature T ) agitates the pawl. In between the elements
of the potential energy U (x, y) are presented, from the left to the right, as f x for the load,
U1 (x (y)) for the pawl-ratchet wheel coupling, and U2 (y) for the pawl
60 1 Physics of Langevin Equation
where U (x, y) = U1 (x (y)) + U2 (y) f x, and (t) and (t) are mutu-
ally independent white Gaussian random noises with zero mean and (t) (t ) =
(t) (t ) = (t t ). Modified models have also been studied recently [6163].
From Fig. 1.9 we expect that a light load can be lifted up if T > T , because
the rotation of the wheel axis in the forward (lifting-up) direction is most likely
registered by the cool pawl, while the rotation in the opposite direction is more
likely hindered by the cool pawl. For 0 < T < T , the wheel has a tendency to
rotate in the opposite direction to the case of T > T .
The hot pawl moves rather by thermal random force than by the potential force
of U2 (y). A weak back-and-forth Brownian motion of the ratchet wheel can cause
the repositioning of the pawl to the neighboring spaces between the ratchet tooth.
Such events occur in an asymmetric manner with respect to x reflection, due to the
asymmetry of the profile (x). These results are verified numerically. It is possible
to take into account of the inertia of the pawl and the ratchet wheel, and the essential
behavior of the system is the same as long as T > 0 [62].
dp U (x) p dx p
= f + (t), = , (1.111)
dt x m dt m
where the thermal random force (t) is the white Gaussian noise with (t) (t ) =
2 kB T (x)(tt ).82 f is a load on the particle toward the (x) direction. If we could
justify to take the overdamping limit (see Section 1.3.2.1), the resulting Langevin
equation is (1.112), that is,
kB T (x)
dx = 2 kB T (x) d Bt U (x, a) + f x + dt. (1.112)
x 2
U(x)
x
Fig. 1.10 Potential energy U (x) of Buttiker and Landauer ratchet (adapted from [59]). The tem-
perature T (x) takes the value T on the thick solid line and T on the dashed line
82 Recall that, in the presence of the inertia, the variation of x(t) during dt is O(dt), not O(dt 1/2 ).
The multiplicative character of the (t) here does not cause ambiguities [74].
1.3 Physical Implications of Langevin Equations 61
In fact, however, the neglect of the inertia completely changes the energetic aspect
of the model. See Sect. 4.2.2.2.
The intuitive explanation for the mechanism of this heat engine is as follows: Let
us assume that T > T and that the temperature profile is as described in Fig. 1.10
and its caption. In the cool region of T (the dashed line in the rightward downhill),
the thermal activation to climb up the potential is less probable than the other side
of the potential peak. The probability of barrier crossing is then less frequent from
right to left than the inverse. In [59] an alternative explanation is given in terms
of the effective potential, Ueff (x) = kkBBTT(x)
0
U (x), immersed in a fictive homogeneous
thermal environment of the temperature T0 . In any case, the particle migrates toward
the right on the average if T > T .
83 There are ratchet models that can work without the thermal random force [80].
62 1 Physics of Langevin Equation
U(x): on
x
U(x): off
Fig. 1.11 Flushing ratchet model [64]: Two alternative potential profiles of the flushing ratchet. In
the period when the sawtooth potential U (x) is on, the state x is driven toward one of its valleys.
While U (x) is off (horizontal dashed line), x diffuses symmetrically. The average displacement is
positive, d x/dt > 0
U (x) dx
0 = (t) + (t), (1.113)
x dt
where and (t) are as usual (see (1.32)), and (t) temporarily switches between 0
and 1.84
Fig. 1.12 Model of the heat engine. Wedge-shaped object and rectangle-shaped object are con-
strained to move together horizontally. These objects are immersed in the gas reservoirs at different
temperatures. If the environment for the wedge-shaped object is hotter than that for the rectangle-
shaped object, the combined objects move to the left and vice versa (Figure adapted from Fig. 1
of [25])
Remarks
The directed transport by spontaneous symmetry breaking is not covered by the
Curie principle. Such case is demonstrated in the collective ratchet model [70].
On the fluctuating level, temporal directional transport in a particular realization
is possible under the detailed balance condition, as long as the spatial x symmetry
is broken [6163].87
Another limitation of the Curie principle is that it does not tell anything about the
sign and amplitude of the cross-coupling effect. Whether or not an active transport
against external load occurs depends on concrete structure of the model.
Langevin equation is not the only way to model heat engines. A heat engine is
modeled using the nonlinear kinetic couplings between the system and its environ-
ments [24, 25]. Figure 1.12 illustrates their model. Since the spatial symmetry is
broken by this kinetic coupling, the model is not directly reducible to the Langevin
equations. The latter have only linear kinetic coupling through the friction coeffi-
cients ( etc.).
where the integral is done over all phase space (i.e., (x, y)-plane for Feynman
ratchet, and x-axis for Buttiker and Landauer ratchet) and J is the probability flux
that appears in the FokkerPlanck equations, tP = J.
A simple derivation is to compare this equation and the identity
(x x(t)) d x
= (x x(t)) ,
t dt
1.4 Discussion
References
1. C.W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry for Natural Sciences,
3rd edn. (Springer, 2004) 3, 9, 18, 36, 40, 44
2. H. Risken, T. Frank, The Fokker-Planck Equation: Methods of Solutions and Applications,
2nd edn. (Springer, 1996) 3, 49
3. R. Zwanzig, J. Stat. Phys. 9, 215 (1973) 3, 26, 28, 29
4. R. Kubo, M. Toda, N. Hashitsume, Statistical Physics II: Nonequilibrium Statistical Mechan-
ics, 2nd edn. (Springer-Verlag, Berlin, 1991) 4, 18, 20, 49
88 In the latter approach the probability density of a single counterion is assumed to be proportional
to the charge density of counterion.
References 65
5. B. Alberts et al., Essential Cell Biology, 3rd edn. (Garland Pub. Inc., New York & London,
2009) 6
6. K. Sato, Y. Ito, T. Yomo, K. Kaneko, PNAS 100, 14086 (2003) 8
7. N.G. van Kampen, Stochastic Processes in Physics and Chemistry, Revised ed. (Elsevier Sci-
ence, 2001) 10
8. D.S. Lemons, A. Gythiel, Am. J. Phys. 65, 1079 (1997) 16, 26
9. A. Einstein, Ann. Phys. (Leipzig) 17, 549 (1905) 18, 44
10. A. Rahman, Phys. Rev. 136, A405 (1964) 19
11. B.J. Alder, T.E. Wainwright, Phys. Rev. Lett. 18, 988 (1967) 19
12. R.D. Vale, D.R. Soll, I.R. Gibbons, Cell 59, 915 (1989) 23
13. K. Tawada, K. Sekimoto, J. Theor. Biol. 150, 193 (1991) 23
14. A. Einstein, Investigation on the theory of the Brownian Movement, (original edition, 1926)
ed. (Dover Pub. Inc., New York, 1956), chap.V-1, pp. 6875 22
15. L.D. Landau, E.M. Lifshitz, Fluid Mechanics (Course of Theoretical Physics, Volume 6), 2nd
edn. (Reed Educational and Professional Publishing Ltd, Oxford, 2000) 22
16. E. Entchev, A. Schwabedissen, M. Gonzalez-Gaitan, Cell 103, 981 (2000) 24
17. F. Perrin, Le Journal de Physique et Le Radium 5, 497 (1934) 24
18. F. Perrin, Le Journal de Physique et Le Radium 7, 1 (1936) 24
19. C. Ribrault, A. Triller, K. Sekimoto, Phys. Rev. E 75, 021112 (2007) 24, 25, 26
20. S.C. Harvey, J. Garcia de la Torre, Macromolecules 13, 960 (1980) 24
21. Y. Han et al., Science 314, 626 (2006) 24, 26
22. A.J. Levine, T.B. Liverpool, F.C. MacKintosh, Phys. Rev. E 69, 021503 (2004) 24
23. M.X. Fernandes, J. Garcia de la Torre, Biophys. J. 83, 3039 (2002) 25
24. C. Van den Broeck, R. Kawai, P. Meurs, Phys. Rev. Lett. 93, 090601 (2004) 26, 63
25. C. Van den Broeck, R. Kawai, Phys. Rev. Lett. 96, 210601 (2006) 26, 63
26. G. Ford, M. Kac, P. Mazur, J. Math. Phys. 6, 504 (1965) 26
27. D.A. Beard, T. Schlick, Biophys. J. 85, 2973 (2003) 26
28. G. Goupier, M. Saint Jean, C. Guthmann, Phys. Rev. E 73, 031112 (2006) 27
29. G.E. Uhlenbeck, L.S. Ornstein, Phys. Rev. 23, 823 (1930) 27
30. N.G. Van Kampen, Stochastic Processes in Physics and Chemistry (North-Holland, Amster-
dam, 2007) 27
31. H. Mori, Prog. Theor. Phys. 33, 423 (1965) 32
32. K. Kawasaki, J. Phys. A 6, 1289 (1973) 32
33. R. Zwanzig, J. Chem. Phys. 33, 1338 (1960) 32
34. K. Ito, Stochastic Processes (Lecture Notes from Aarhus University 1969) (Springer Verlag,
New York, 2004) 34
35. G. Gallavotti, Statistical Mechanics, A Short Treatise, 1st edn. (Springer-Verlag, New York,
1999) 37
36. N.G. van Kampen, J. Stat. Phys. 24, 175 (1981) 40
37. E. Wong, M. Zakai, Ann. Math. Statist. 36, 1560 (1965) 40
38. R. Kupferman, G. Pavliotis, A. Stuart, Phys. Rev. E 70, 036120 (2004) 40
39. R. Mannella, Int. J. Mod. Phys. C13, 1177 (2002) 43
40. P.E. Kloeden, E. Platen, H. Schurz, Numerical Solution of SDE Through Computer Experi-
ments (Springer, New York, 1994) 43
41. P.E. Kloeden, E. Platen, Numerical Solution of Stochastic Differential Equations (Springer,
New York, 1992) 43
42. J. Qiang, S. Habib, Phys. Rev. E 62, 7430 (2000) 43
43. H.A. Kramers, Physica 7, 284 (1940) 44, 51, 56
44. Y. Oono, Progr. Theor. Phys. Suppl. 99, 165 (1989) 46
45. T. Hondou, J. Phys. Soc. Jpn. 63, 2014 (1994) 48, 63
46. T. Hondou, Y. Sawada, Phys. Rev. Lett. 75, 3269 (1995) (Erratum; T. Hondou, Y. Sawada,
Phys. Rev. Lett. 76 1005 (1996)) 48, 63
47. J.M. Sancho, M.S. Miguel, D. Durr, J. Stat. Phys. 28, 291 (1982) 51, 52
66 1 Physics of Langevin Equation
48. M. Doi, S.F. Edwards, The Theory of Polymer Dynamics, 1st edn. (Oxford Science Pub,
Oxford, 1986), Sect. 3.3 51, 52
49. K. Sekimoto, J. Phys. Soc. Jpn. 68, 1448 (1999) 51
50. R. Landauer, Phys. Rev. A 12, 636 (1975) 52
51. G. Costantini, F. Marchesoni, Europhys. Lett. 48, 491 (1999) 53
52. P. Reimann et al., Phys. Rev. Lett. 87, 010602 (2001) 53
53. K. Sekimoto, A. Triller, Phys. Rev. E 79, 031905 (2009) 52
54. K. Hayashi, S. Sasa, Phys. Rev. E 69, 066119 (2004) 54
55. R.P. Feynman, R.B. Leighton, M. Sands, The Feynman Lectures on Physics vol.1 (Addison
Wesley, Reading, Massachusetts, 1963), Sects.46.146.9 58, 59
56. M. Buttiker, Zeitschrift der Physik B 68, 161 (1987) 58, 60
57. R. Landauer, J. Stat. Phys. 53, 233 (1988) 58, 60
58. J. Parrondo, P. Espanol, Am. J. Phys. 64, 1125 (1996) 58
59. I. Derenyi, M. Bier, R.D. Astumian, Phys. Rev. Lett. 83, 903 (1999) 59, 60, 61
60. K. Sekimoto, J. Phys. Soc. Jpn. 66, 1234 (1997) 58, 59, 61
61. N. Nakagawa, T. Komatsu, Progr. Theor. Phys. Suppl. 161, 290 (2006) 60, 63
62. N. Nakagawa, T. Komatsu, Physica A 361, 216 (2006) 60, 63
63. N. Nakagawa, T. Komatsu, J. Phys. Soc. Jpn. 74, 1653 (2005) 60, 63
64. A. Ajdari, J. Prost, C. R. Acad. Sci. Paris. Ser. II 315, 1635 (1992) 61, 62
65. M.O. Magnasco, Phys. Rev. Lett. 71, 1477 (1993) 61
66. J. Prost, J. Chauwin, L. Peliti, A. Ajdari, Phys. Rev. Lett. 72, 2652 (94) 61, 62
67. S. Leibler, D.A. Huse, J. Chem. Biol. 121, 1357 (1993) 61
68. M.O. Magnasco, Phys. Rev. Lett. 72, 2656 (1994) 61
69. R.D. Astumian, M. Bier, Phys. Rev. Lett. 72, 1766 (1994) 61
70. F. Julicher, J. Prost, Phys. Rev. Lett. 75, 2618 (1995) 61, 63
71. H. Zhou, Y. Chen, Phys. Rev. Lett. 77, 194 (1996) 61
72. P. Reimann, Phys. Rep. 361, 57 (2002) 61, 63
73. J.M.R. Parrondo, B.J. De Cisneros, Appl. Phys. A75, 179 (2002) 61
74. M. Matsuo, S. Sasa, Physica A 276, 188 (2000) 60
75. J. Rousselet, L. Salome, A. Ajdari, J. Prost, Nature 370, 446 (1994) 62
76. L.P. Faucheux, L.S. Bourdieu, P.D. Kaplan, A.J. Libchaber, Phys. Rev. Lett. 74, 1504 (1995) 62
77. L.D. Landau, E.M. Lifshitz, L.P. Pitaevskii, Electrodynamics of Continuous Media, 2nd edn.
(Elsevier Butterworth-Heinemann, Oxford, 2004) 62
78. P. Curie, J. Phys. (Paris) 3eme serie 3, 393 (1894) 62
79. F. Julicher, A. Ajdari, J. Prost, Rev. Mod. Phys. 69, 1269 (1997) 62
80. M. Porto, M. Urbakh, J. Klafter, Phys. Rev. Lett. 85, 491 (2000) 61
81. M. Bernstein, L.S. Brown, Phys. Rev. Lett. 52, 1933 (1984) 63
82. R. Kanada, K. Sasaki, J. Phys. Soc. Jpn. 68, 3579 (1999) 63
83. I. Nishio, S.-T. Sun, G. Swislow, T. Tanaka, Nature 281, 208 (1979) 64
84. R. Messina, C. Holm, K. Kremer, Phys. Rev. E 64, 021405 (2001) 64
Chapter 2
Structure of Macroscopic Thermodynamics
System: Of the whole world, a part which is properly cut out is called the sys-
tem. In this chapter we suppose that a system contains a macroscopic number
(say, 1023 ) of constituting elements of the same kind like particles, spins,
etc. If a system is divided into more than one part by some criterion, each
Sekimoto, K.: Structure of Macroscopic Thermodynamics. Lect. Notes Phys. 799, 6792 (2010)
DOI 10.1007/978-3-642-05411-2 2 c Springer-Verlag Berlin Heidelberg 2010
68 2 Structure of Macroscopic Thermodynamics
part is called a subsystem. A system having no contact with any other system
is called isolated system.
Internal energy of a system: If one can access the microscopic Hamiltonian of
the system, or if one can observe or calculate the mechanical energy, either
classical or quantum, then we naturally identify this energy as the internal
energy of the macroscopic thermodynamics.
Extensive variables: When a system and its identical copy are combined to
make a new system, certain variables take the values twice as large as the
original ones before combining with its copy. Such variables are called exten-
sive variables. For a system consisting of single-component particles, the
internal energy, E, the volume, V , the number of particles, N (as well as the
entropy, S, in the equilibrium states, see below), are extensive variables.
The other types of variables which do not change their values are called
intensive variables of the system, but the latter refers to the equilibrium states
(see below).
Environment: Those background systems to which the system of our interest
is connected are called the environments, reservoirs, or baths. According
to the entity that is exchanged with the system, the environment is called
thermal environment or heat reservoir/bath (for energy exchange), particle
environment or particle/chemical reservoir (for mass exchange) or pressure
environment (for volume exchange).1
As an environment forms the background of a system and not the main
object of study, they are considered to satisfy several simplifying assump-
tions.
(1) We ignore the interaction energy between the system and its environ-
ments. (In case that it is not realistic assumption, we try to define the
interface layer of interaction as a new system.)
(2) Environments are big in the sense that the conservation laws for the
total energy, mass, volume, or momentum have no effect on the state of
the environment.
(3) Environments return instantaneously to their equilibrium states (see
below), and keep no memories of the systems action in the past. In
other words, any action on the environments is supposed to be qua-
sistatic from macroscopic point of view.2
Because of these simplifications, environments are characterized only by
their intensive parameters, that is, temperature, chemical potential (or the
density of particles), and pressure.
Heat exchange: It is the form of exchange of energy not through work, or
involving mass exchange. This process occurs either between a system and
its environment, or among different subsystems.
1 In this book we will use the words heat bath and thermal environment interchangeably.
2 In other words, we do not discuss the irreversible entropy production proper to the environments.
2.1 Basic Concepts of Thermodynamics 69
Some aspects of macroscopic thermodynamics are peculiar to the fact that the sys-
tem is macroscopic.
1. When the system consists of several macroscopic subsystems, a subsystem can
do work on the other subsystems as an external system.
2. The presence of extensive variables and the negligence of interaction energy
between the system and environments suppose that the surface contributions
3 More about the macroscopic chemical thermodynamics will be described in the next chapter.
4 For a review, see [3]. cf. Sect. 4.2.1.2
5 The author acknowledges Izumi Ojima for his comment on this point.
6 Note that certain textbooks of thermodynamics adopt the opposite sign convention.
70 2 Structure of Macroscopic Thermodynamics
to those variables are negligible compared with the bulk contributions. If the
surface-to-volume ratio is larger than some inverse characteristic length (e.g., the
interaction range of particles), this hypothesis is no more valid.
3. The extensive character implies the absence of fluctuations. It is related to
the law of large numbers. It is often called the self-averaging property of
macroscopic system. Entropy and temperature which are ensemble-based con-
cept can be assigned to a single macroscopic system. Systems at critical point
of phase transition, or with long-range interactions, cannot be described by
usual thermodynamics.
Zeroth law: If a system is left isolated for sufficiently long time from any
environment and from any external system, it will reach a state with no
further macroscopic changes.7 Such a state is said to be in an equilibrium
state. For a single-component gas isolated in a single compartment, for
example, the set of energy, volume, number of gas particles, (E, V, N ), is
sufficient to characterize its equilibrium state, and the entropy S is then
a function of these variables. These variables are called thermodynamic
variables, and the functions which relate the thermodynamic variables in
an equilibrium system are called thermodynamic functions. If any process
occurs slowly enough so that the system remains almost in equilibrium
at each instant, the process is said to be quasistatic (or quasiequilibrium)
process.
First law: We cannot realize a perpetual machine of the first kind, that is, there
is no autonomous system, isolated from the environment, that produces work
(i.e., W < 0) through a cyclic process. The inverse process is also not pos-
sible. More concretely, a balance of energy should always be established
among the changes in the internal energy of the system, E, the work, W,
and a heat, Q, during any process:
E = W + Q. (2.1)
7 Strictly speaking, there may remain global motions due to conserved total momentum or total
angular momentum of the whole system.
2.1 Basic Concepts of Thermodynamics 71
Second law: One cannot realize a perpetual machine of the second kind, that
is, being in contact with a single thermal environment at constant tempera-
ture, there is no autonomous cyclic engine which converts heat into work.
The inverse process from work to heat is possible. More precisely, there is
a thermodynamic function of equilibrium systems called entropy (S) which
never decreases during processes of an isolated system. Entropy is an exten-
sive variable.
Third law: This law, also called the NernstPlancks theorem, concerns the limit
to absolute zero temperature. At absolute zero, the quantum interference is
believed to impose a unique ground state at least in most cases, and the exten-
sivity of the macroscopic thermodynamics should be carefully tested [4]. As
quantum fluctuations are beyond the scope of this book, we will not go into
the detail of this law.
Fourth law 8 : Any thermodynamic variable characterizing an equilibrium
state is either extensive one or intensive one. For the system whose exten-
sive variables are (E, V, N ) and S, the intensive variables are temperature T ,
pressure p, and chemical potential . The intensive variables are the homo-
geneous function of the extensive variables of zeroth order. The intensive
variables characterize the equivalence relationship of equilibrium. If an
equilibrium system A and another one B satisfy A B, it means that the
contact of these two systems cause no changes in their extensive variables.
The system B can be an environment with which the system A is in contact.
The transitive property of this equivalence relationship, ( A C)
(B C) A B, allows to compare, for example, the temperature T
of two systems A and B using the thermometer C. The reflective property,
A A, implies that the intensive variables remain unchanged upon combin-
ing the two identical systems A. The symmetric property, A B B A,
implies that the assignment of the measuring system and the measured sys-
tem is relative in the context of macroscopic equilibrium. Further properties
of the extensive and intensive variables will be described below.
If the system is not macroscopic, the intensive variables can still be used to char-
acterize the environment. But they are not the state variables of the small system.
In this section we will describe several basic consequences of the laws of thermo-
dynamics.
One important aspect of the thermodynamics, especially from the zeroth and
fourth laws, is the presence of thermodynamic function by which various
8 There is no general consensus on which law should be put as the fourth law. Here we take up a
version.
72 2 Structure of Macroscopic Thermodynamics
1 p
d E = T dS p dV + d N, or dS = d E + dV d N, (2.2)
T T T
where differentials, (d S, d E, d V, d N ), are the local variables along (S, E, V, N ),
with the origin relocated to the tangent point, and the coefficients (1/T, p/T, /T )
are independent of these differentials. Thus we can have essentially all relations
between the extensive variables and the intensive ones, such as T1 = ES V,N for these
equations. We call the relations (2.2) the fundamental relations. Those thermody-
namic functions that generate the fundamental relations, such as S(E, V, N ) or
E(S, V, N ), are called the complete thermodynamic functions. The pairs of ther-
modynamic derivatives associated with the partial derivative like T (E, V, N )1 =
S
E V,N
are called conjugate variables. (T, S), ( p, V ), and (, N ) are pairs of con-
jugate variables.9
Complete thermodynamic functions are not unique, and therefore, the funda-
mental relations are not either. However, all the fundamental relations of a system
represent the same interdependencies. To obtain a new complete thermodynamic
function, we must apply the Legendre transformation to the original complete
thermodynamic function. For example, we can obtain the complete thermody-
namic function, Helmholtz free energy, F(T, V, N ) of a simple gas system by
the Legendre transformation of E(S, V, N ): We solve T = E(S, V, N )/ S to find
(formally) S(T, V, N ). (This function, S(T, V, N ), is not a complete thermody-
namic function.) Then we substitute this into E(S, V, N ) T (S, V, N )S to obtain
F(T, V, N ) = E(S(T, V, N ), V, N ) T (S(T, V, N ), V, N ) S(T, V, N ). The vari-
able S is now found through ( F/ T )V,N = S, instead of ( E/ S)V,N = T .
The exchange of the conjugate pair, (S, T ), as independentdependent variables
9 Incomplete thermodynamic functions can also be made. For example, we can solve
T (E, V, N )1 = ES V,N for V and substitute the result in S(E, V, N ). Then we have a func-
tion S = S(E, V (E, T, N ), N ). Such function, although correct, cannot generate all the other
thermodynamic variables; there remains an undetermined additive constant.
2.1 Basic Concepts of Thermodynamics 73
requires the subtraction of the term T S from the original complete thermodynamic
function.10
Physically, the subtraction of a product of conjugate thermodynamic variables
(e.g., T S) in the transform (e.g., E F = E T S) implies the attachment or
detachment of an environment. For example, suppose that an isolated simple gas
system has the complete thermodynamic function, S(E, V, N ). We now attach a
thermal environment to this system. Since the combined system is a new isolated
system, we may use the entropy of the whole system S(E, V, N )+ Sres (E tot E) as a
new thermodynamic function. Since the system can exchange energy with this envi-
ronment satisfying the law of energy conservation, we control no more the energy
E, but control the energy of the whole system, E tot . Since the reservoir retains a
constant temperature Tres by definition, we choose as a new independent thermody-
namic variable Tres rather than E tot , where Tres is given by Sres / E tot = 1/Tres .11
Then we have
E
S(E, V, N ) + Sres (E tot E) S(E, V, N ) + Sres (E tot )
Tres
1
= (E Tres S) + const. (2.3)
Tres
The repartitioning of the energy into E and E tot E is self-adjusted so that the
temperatures are equilibrated, T = Tres . The new complete thermodynamic func-
tion F = E T S then describes the combined system (the system and the thermal
environment) with {T, V, N } as independent variables.
As a consequence of the second law of thermodynamics, the work W done to the
system through arbitrary isothermal system is bounded below by the change in the
Helmholtz free energy, F = E T S 12 :
W F.
2 E ( p)
= = .
V S,N V N N S,V
These types of equations are called the Maxwell relation. Inversely, if the coeffi-
cients of the Eq. (2.2) satisfy the Maxwell relation, these local equations can be
integrated to define a surface. See Appendix A.2.1 for a brief proof.
T S pV + N = E. (2.5)
A
A
x
It is easily derived from either one of the relations in (2.4) by differentiating with
respect to and then putting = 1. This type of identity is called the Eulers
theorem.14,15
Eulers theorem applies also to the expressions of the intensive variables, e.g.,
= E/ N = (S, V, N ). Since the value of should remain the same when the
system and its copy is combined, we have
(S, V, N ) = (S, V, N ).
S +V +N =0 (2.6)
S V N
and similar identities.
SdT + T d S ( pd V + V dp) + d N + N d = d E.
SdT V dp + N d = 0. (2.7)
An open system exchanges not only heat but also particles with its environment.
See Fig. 2.2.16 Thermodynamics of open system is an important issue in biology
since protein molecular motors are usually open systems which consume source
molecules like ATP in the environment. The problem of organic or inorganic adsorp-
tion and diffusional transport across membranes are also the subject of open system.
In this section and in Sects. 2.3.2 and 2.3.3, we deal explicitly with an open system.
For a closed system, we can identify the system with the particles, etc., that are
therein. An open system, however, is rather a container of particles, and a particle
belongs to the system while it is in the container.17
RT
S
R
Fig. 2.2 An open system containing gas particles, S, in contact with a thermal environment, RT ,
and a particle environment, R . The thick lines and filled rectangle are isolating walls and piston,
respectively, and the thin walls are thermally conducting walls. A thick dashed line is the wall
permeable for the gas particle
16In quantum physics, often the same word means the systems in contact with (only) a thermal
environment.
17 In the problem of adsorption, the container is the ensemble of adsorption sites. Those particles
on the adsorption sites belong to the open system.
2.1 Basic Concepts of Thermodynamics 77
Let us consider a gas system in which work is done through the change of volume,
d W = pd V . The first law of thermodynamics reads
d E = pd V + d Q tot , (2.8)
where d Q tot denotes all the energy entering into the system except for the
work pd V .18 We suppose that the system is in contact with a thermal environ-
ment of temperature T and also with a particle environment of chemical potential
. Therefore, if a small number of particles, d N , have entered into the open system
in keeping the equilibrium, and the volume is increased by d V , then the change in
the internal energy is given by the fundamental relation, d E=T d S pd V + d N .
The substitution of this expression for d E in (2.8) yields
d Q tot = T d S + d N . (2.9)
In general, both d N and d S are nonzero when a particle enters or leaves an open
system. In other words, the chemical potential is not the (only) energy carried by
a particle upon its migration across the systems border:19 When the only change
is the immigration of a single particle, (d N )1 = 1, we will denote by (d S)1 the
associated change of entropy. The heat associated with this immigration is then
the sum, (d Q tot )1 = T (d S)1 + . When the particle is an ideal gas particle, such
energy changes should not involve the concentration of the particle in the parti-
cle environment simply by the definition of ideal particle. On the other hand, the
particle concentration of the particle environment c enters in in entropic form,
kB T log c. In order that the (d Q tot )1 is independent of the particle concentration
of the environment, the last entropic term should be cancelled by a similar term in
T (d S)1 .
18Hereafter we will use d to denote the infinitesimal transfer of work, d W , and of (generalized)
heat, d Q to distinguish from d which is related to the thermodynamic variables.
19 Statistical mechanics shows that the chemical potential is related to the fugacity, i.e., the proba-
bility to escape, of anonymous particles. This is why depends on the concentration of the particles
even if the particles do not interact among each other.
78 2 Structure of Macroscopic Thermodynamics
The chemical reaction is similar to the open system in the context that discrete
atomic masses are transferred between one state to the other. For an open system
the transfer is between a system and its particle environment, while for a chemical
reaction it is between a molecular species and the another. These two processes can
coexist. Typical examples are surface catalytic reactions and motor proteins20 or the
transport of molecules across a membrane channel. In the latter case, we can regard
the transport as unimolecular reaction from one species to another species, such
as cytoplasm nucleus. If both sides of a membrane are macroscopic, the
channels on the membrane can be also described as an open system that exchanges
particles with two particle environments.21
20 Different species of molecules can share the same space as distinct particle environments if the
spontaneous reaction between these species is negligible.
21 In the master equation formalism and related discrete Langevin equation discussed in the next
chapter, the change in closed systems and the migration of particles in open systems are both
described as a state transition of the whole system.
2.2 Free Energy as an Effective Potential Energy for the External System 79
system
bath
Fig. 2.3 Thermodynamic system viewed from the external system. Under quasistatic operations of
the piston (top figure), the black box (dashed rectangle) which consists of a gas in the cylinder
and the heat environment behaves like a spring (bottom). The temperature, instead of the force of
spring, is equilibrated between the system and the thermal environment (heat bath)
tells us that the associated work W done onto the system is partitioned into the
increment of the energy of the system, E, and the heat released to the thermal
environment, (Q).22 For the external system, the system and the thermal environ-
ment, therefore, look like a single black box. The external system measures only the
relation between the work W and the displacement of a piston.
To characterize this black box, let us denote by a the position of the piston and by
W (a1 ; a2 ) the work done through the displacement of the piston from a1 to a2 . If this
work can be described by a potential, (a), such that W (a1 ; a2 ) = (a2 ) (a1 ),
then one might regard this black box as an elastic spring whose potential energy
is (a) up to an additive constant (see Fig. 2.3 (bottom)). Macroscopic thermody-
namics tells that (1) such (a) exists if the parameter a is changed quasistatically
and (2) this potential energy, (a), is the Helmholtz free energy, F = E T S,
up to an arbitrary additive constant.23 Thus the word free energy for F is
justifiable because it behaves like a potential energy (for energy), and it is avail-
able as work (for free). A similar interpretation can be done for the free energies
G = E T S + pV (= N ) or J = E T S N (= pV ) when the system
undergoes a process keeping in contact with a pressure environment (for G) or with
a particle environment (for J ), respectively.
Let us summarize how the quasistatic operation of the piston in the Fig. 2.3 can
be described differently according to the standpoint of the observer: the Ext who
sees the system only through the parameter a, the Sys who has access to both E and
22 Note the sign convention about the work and heat, see Sect. 2.1.1.
23 d E = T d S pd V + d N under the constraints, dT = d N = 0, yields d(E T S) =
d F = pd V = d W .
80 2 Structure of Macroscopic Thermodynamics
S, and the Tot who regards the totality of the system and the thermal environment
as an isolated system.
Ext: The work W is apparently stored in the system as the increment of the
potential energy, F (= E T S).24 If the data of F are accumulated
at different temperatures, the change of the (internal) energy E can be
obtained by the following formula called the vant Hoff equation (see,
e.g., [5]):
F
E = F T , (2.13)
T a
where the suffix a indicates that the initial and final values of a should be
fixed upon taking the derivative.25 Equation (2.13) comes from the relation:
E = F + T S = F T F/ T.
Sys: The work W is partly stored in the system as E, while the rest is trans-
ferred to the thermal environment as heat, T S.26 The actual partitioning
of the work into E and (T S) depends on the details of the system.
For example, F should be E for a system made of metallic spring but it
should be (T S) for a system of ideal gas or ideal polymer network.
Tot: The work W is stored in the isolated system as the increment of (true)
energy, E + T Senv , where the entropy of the thermal environment, Senv ,
changes so that S + Senv = 0.
For nonquasistatic processes the second law of thermodynamics tells us
that the entropy of the isolated system is nondecreasing: S + Senv 0.
This inequality relates the work done to the system, W = E (T Senv ),
27
and the change of the Helmholtz free energy, F = E T S, by the
inequality,
W F. (2.14)
In other words, the work obtainable from the system (W ) is limited by the
decrease of the free energy of the system (F):
(W ) (F). (2.15)
24 The quantities denoted like M implies the increment of M through the change of the
parameter a.
25 Remember that the partial derivative can give different results depending on what variables
are fixed: For example, p = (U/ V ) S is rewritten as p = (U/ V )T + T ( p/ T )V .
Generally, given the fundamental relation, f (x1 , . . . , xn ) and R j f / x j , we have a formula,
( f / xi ) R j = ( f / xi )x j R j ( Ri / R j )xi .
26 Remember the sign convention: the heat, Q = T S, is positive if the energy is transferred from
the thermal environment to the system.
27Note that the thermal environment is supposed to be in equilibrium and, therefore,
Q = T Senv holds always.
2.3 * Free-Energy Conversion 81
T T
We here discuss reaction heat and the work produced by a heat engine.
A thermal environment is supposed to be an inexhaustible buffer of energy.
Because of this character, the isothermal processes exhibit sometimes
28 If the (almost) quasistatic process takes an extremely large time t, diffusive exchange of heat
1
between the two thermal environments via the system will be of order t 2 . We neglect this as
compared to the systematic part of t.
29 It is called local equilibrium.
82 2 Structure of Macroscopic Thermodynamics
counterintuitive behaviors about their energetics. The rubber band mentioned above
is an example: little heat is released from an ideal rubber band when it freely shrinks
from a stretched state, while a quasistatic shrinkage can absorb heat. Free expansion
of an ideal gas behaves similarly: although a quasistatic expansion from the volume
Vini to Vfin can do the reversible work of kB T log(Vfin /Vini ), the work-free expansion
(JouleThomson process) absorbs no heat.
In the biochemical context, the hydrolysis reaction of an ATP (adenosine triphos-
phate) with a water molecule into an ADP (adenosine diphosphate) and an inorganic
phosphate (Pi), ATP + H2 O ADP + Pi , can do about 20kB T of work under
usual physiological conditions. It does not mean that we will observe this amount
of dissipative heat when no work is extracted through this reaction. (See below.)
We might also recall the existence of endothermic chemical reaction while the
reaction proceeds with consuming the relevant free energy (see below), the system
absorbs heat, instead of absorbing heat. The point is that, for open systems, the heat
released to the thermal environment, (Q), is not the total increase of entropy of
the environment Senv , that is Q = T Senv : The total change Senv generally
includes a purely combinatorial part (mixing entropy change).
The chemomechanical coupling is a mechanism which converts a chemical free
energy into work under constant temperature and pressure. We will see the inter-
play of heat and work in this mechanism. Figure 2.5 shows a schematic setup of
chemomechanical coupling. In this setup, the material before the reaction (called
the substrate) and the one after the reaction (called the product) are separated as sub-
systems in order to avoid a work-free reaction. As we are interested in the process
under constant pressure p, the energy of the pressure environment is included in the
energy of the system. That is, we use enthalpy, H E + pV , instead of E, where
V is the volume. By adopting enthalpy as the energy of the system, the work due to
volume change is not counted in W . Finally there is an engine that works between
the substrate and the product subsystems. The two subsystems and the engine are
immersed in a thermal environment (Fig. 2.5). The combined system, therefore,
consists of four subsystems, i.e., the two particle subsystems, the pressure reservoir,
and the chemical engine.
pV
substrate product
Fig. 2.5 Schematic setup of (W)
chemomechanical coupling.
The system consists of the
substrate subsystem, the
product subsystem, the
ambient pressure (piston),
and the chemical engine
(Q)
(circle), all immersed in a
thermal environment
2.3 * Free-Energy Conversion 83
In the case of the quasistatic process, the analysis is done by the aid of a thought
experiment called a vant Hoff reaction box (see e.g., [5]). Here we include the
general case of a nonquasistatic process and consider what occurs upon one cycle
of operation of the engine. Let us denote by (Q) the heat released from the
combined system to the thermal environment and by (W ) the work done by
the combined system to the external system. Let us also denote by S, V , and
H (= E + pV ), the changes in the entropy, volume, and the enthalpy of the
combined system.30
Application of the first law of thermodynamics: In the present context this law gives
E = W + ( pV ) + Q, or by noting that p is constant,
(H ) = (W ) + (Q). (2.16)
The minus signs are incorporated in the above equation so that we can use the
extracted work, (W ), and the released heat, (Q). There are two extreme cases:
Case where no work is extracted: We substitute W = 0 in (2.16) to have (Q) =
(H ), which we shall call the heat of JouleThomson, (Q)JT . That is,
(H ) = (Q)JT , W = 0. (2.17)
In case where the particle environments are treated as environment, the chemical
potentials should be the independent variables. See Sect. 2.1.4.2.
General case: In the general case, the total entropy is nondecreasing, S + Senv
0, by the second law of thermodynamics. Therefore,
(W ) (W )rev
(Q) (Q)rev . (2.19)
(W) (Q)
( H) >0 :
0 ( G) ( H
(W)
( H)=0 :
0 (Q) ( G)
( H)
(Q)
( H)<0 :
( H) 0 (W) ( G)
Fig. 2.6 The partition of the extracted work, (W ), (filled arrows) and the released heat, (Q),
(open arrows) under constant temperature and pressure. They obey the first law; (H ) =
(W ) + (Q). The reaction is either exothermic ((H ) > 0, top), athermal ((H ) = 0,
middle), or endothermic ((H ) < 0, bottom). The figures show only the case of engine,
i.e., (G) > 0
Figure 2.6 summarizes the above relations for the cases with (H ) > 0, (H ) =
0, and (H ) < 0, respectively.31
Remark: We should avoid confusions like the following: If the extracted work
(W ) is less than its maximum available value, (G), then the difference (G)
(W ) should be measured as the released heat. The correct statement is . . . , then
we have the irreversible heat, (G) (W ) 0, in addition to the reversible one,
(H ) (G).
Chemical pump: In the reverse case, that is, when the work is done to the system in
order to drive the reaction or transport against the natural tendency, the Gibbs free
energy increases, G > 0.
The equations and inequalities above are always valid with appropriate reinter-
pretation. For example, (W ) (W )rev of (2.19) should be read as W Wrev ,
that is, the work done to achieve a cycle of engine is no less than the reversible one,
Wrev , which in turn is G according to (2.18).
Finally we note that the model of autonomous chemomechanical coupling is
beyond the scope of macroscopic thermodynamics because it concerns the dynamics
of the system.
31 If the volume, instead of the pressure, of the system is fixed, all the enthalpy H in the above
formulas should simply be replaced by the energy E.
2.3 * Free-Energy Conversion 85
FUEL
Fh Fl
Ll Lh
LOAD
Fig. 2.7 Two reservoirs of the fuel (F) particles and two others of the load (L) particles with
the high (h) and low () chemical potentials. The passive diffusion of the fuel particles along
their chemical potential gradient (two arrows in the top layer) is coupled by the chemical engine
(hexagon at the center) to the active transport of the load particles against their chemical potential
gradient (two arrows in the bottom layer)
two species of particles, the fuel (F) and the load (L) particles. For each species
of particles, we prepare two particle reservoirs with high and low densities, which
we distinguish by the suffixes, h and , respectively (see Fig. 2.7). For example, Fh
denotes the high-density reservoir of fuel particles. We will denote the chemical
potential of each reservoir by Fh , etc. By definition, Fh > F and Lh > L . A
chemical engine (denoted by a hexagon at the center of Fig. 2.7) enables the active
transport of the load particles from L to Lh at the expense of the passive transport
of the fuel particles from Fh to F .
Gibbs free-energy surface: As in the previous section, we regard the four particle
reservoirs and the chemical engine as a combined system working at a constant
temperature and pressure. The relevant thermodynamic potential is then the Gibbs
free energy,
where NFh [NLh ] are the number of the fuel [load] particles in their respective high-
density reservoirs, and we have used the fact that the total number of the fuel [load]
particles are constant.32 Equation (2.20) defines an inclined plane in the 3D space
of (NLh , NFh , G tot ). See Fig. 2.8. We shall call this plane the Gibbs free-energy
surface. The thermodynamically allowed processes are those which decrease this
free energy, G tot . Suppose that the reactions L Lh and Fh F are coupled
at the ratio of n L :n F particle transport on the average. This imposes a condition on
(NLh , NFh ) that
32 After each cycle of the chemical engine, the number of the particles in the machinery returns to
the same value. We then ignore the cyclically varying part of G tot .
86 2 Structure of Macroscopic Thermodynamics
NLh NLh
0
nL
= , (2.21)
NFh NFh
0 nF
where (NLh 0
, NFh
0
) are constants. While the chemical engine consumes the n F fuel
particles, decreasing the free energy by G F = (Fh F )(n F ) < 0, it pumps
up the n L load particles, increasing the free energy by G L = (Lh L )n L > 0.
Figure 2.8 shows how the active transport, G L > 0, is realized while satisfying
the second law of thermodynamics, G tot G F + G L < 0.
In most cases, the phrase efficiency of a heat engine would refer to the Carnot
cycle. When Carnot discussed his famous cycle, people had not yet recognized that
heat can be transferred without transporting any material, nor had they discovered
the second law of thermodynamics. It was, therefore, natural that the efficiency was
defined so that it is unity when all the heat from the high-temperature environment
is turned into a useful work. After thermodynamics was established, the efficiency
of a heat engine was redefined in reference to the Carnots theoretical maximum,
allowed by the second law of thermodynamics.
For a single thermodynamic process, it is easy to introduce the efficiency of
energy conversion in reference to the second law of thermodynamics. For example,
for an isothermal process of changing the parameter a from a1 to a2 , we can define
the efficiency as
2.3 * Free-Energy Conversion 87
(W )
= (Isothermal process), (2.22)
(F)
where (W ) is the actually extracted work and (F) = F(a1 ) F(a2 ) is the
decrement in the Helmholtz free energy, that is, the maximally available work of
this process (Sect. 2.1.2). By definition, satisfies (0 ) 1.
When a heat engine converts energy indefinitely, it must undergo a cycle of
thermodynamic processes. The ideal cycle requires generally four quasistatic pro-
cesses, as is the case with the ideal Carnot cycle: two processes in contact with
the two thermal environments, respectively, one to absorb heat and the other to
discard heat, and the other two processes for adjusting the system between these
environments. The Carnot cycle is in this sense a minimal cyclic process to avoid
irreversible losses. Below we derive the maximal work available from one cycle of
Carnot heat engine. Later we will discuss the Carnot cycle on the fluctuating scale
(Sect. 8.1.1).
When the cycle is ideal, each of the above mentioned thermodynamic pro-
cesses must be reversible. Moreover, the attachment and detachment between the
consecutive processes should cause no irreversibility. This requires that, at the
end of each adiabatic process, the temperature of the system should be equal
to the temperature of the thermal environment with which the system will be
attached.33 Let us denote the temperatures of the hot [cool] thermal environment
by Th [T ], respectively. For the isothermal process in contact with the hot envi-
ronment the extracted work, (Wh ), is the change of Helmholtz free energy of the
system, Fh = E h Th Sh , and similarly for the isothermal process with cool
environment, i.e.,
(Wh ) = (Fh ) = (E h ) + Th Sh ,
(W ) = (F ) = (E ) + T S .
During adiabatic processes the extracted work is equal to the change of the energy
of the system:
where the suffixes ad:h and ad:h indicate, respectively, the adiabatic cooling and
heating processes. Because of the cyclicity of the processes, we impose
(E h ) + (E ad:h ) + (E ) + (E ad:h ) = 0.
33 We can assume the continuity of the temperature of the system from an isothermal process to
the following adiabatic process.
88 2 Structure of Macroscopic Thermodynamics
Reversibility implies
(Sh ) + (S ) = 0.
The maximal available work per a cycle of the ideal heat engine, (W tot )max , is
then
(W tot ) (W tot )
= 1. (2.24)
(W tot )max (Th T )Sh
This result can be rewritten in terms of the total Gibbs free energy G tot defined by
G tot = h Nhres + Nres + G engine , where Nhres , etc., denote the number of particles
34 In the present context the substrate and product, Fig. 2.5, should be reread as high-density reser-
voir and low-density reservoir, respectively.
35 etc. implies the change of suffix, h .
2.3 * Free-Energy Conversion 89
in a particle reservoir and G engine is that of the chemical engine. Since a balance
of the number of particles transferred between a reservoir and the chemical engine
imposes Nhres +Nh = 0 and Nres +N = 0, and since the system undergoes a
cycle, G engine = 0, we arrive at (h )Nh = G tot . Therefore, the relevant
definition of the efficiency of the chemical engine is
(W tot )
= . (2.26)
(G tot )
Remark: Any arbitrary positive quantity whose upper bound is unity does not
necessarily deserve to be called efficiency: if we have an inequality in the form of
A B C 0 about the cycle of an energy converter, we might ask ourselves
which of A/(B + C), ( A B)/C, or (A C)/B is the most appropriate definition
of the efficiency of energy conversion.
General remark: It is only the difference that the thermodynamic functions
appeared in the thermodynamic efficiency or in the Carnot efficiency . It was
so also in the heat and work of reversible reactions (2.18). It is because the ther-
modynamic functions have the arbitrariness of an additive constant and the experi-
mentally observable results do not depend on that constant. So as to be compatible
with the extensive character of these functions, this arbitrariness is reduced to the
arbitrary additive constants in the molar [specific] energy and entropy. All the ther-
modynamic laws and relations are invariant under the change of these constants. See
Appendix A.2.2.
The analysis of a Carnot cycle by conventional macroscopic thermodynamics
is tautological, since macroscopic thermodynamics is constructed to explain the
Carnot cycle. The above review will be still useful as a reference frame when a
similar cycle is studied on the scale of thermal fluctuations.
The cost of the operations, i.e., of connecting or disconnecting with the environ-
ments, macroscopic thermodynamics is supposed to be negligible. For mesoscopic
systems, irreversibilities associated to the operation on the system require attention.
This issue will be discussed later (Chap. 7).
36 The generalization to many component systems is known as the Gibbs phase rule.
90 2 Structure of Macroscopic Thermodynamics
G TS II
CII
I
C
CI
II I
0 T 0 H
Tc Tc
Fig. 2.9 (Left) Schematic representation of Gibbs free energy of two phases vs. T . From the gra-
dient G/ T the entropy S = T G/ T is estimated, while from the intersection at T = 0
of the tangent lines, the enthalpy, H = G T G/ T is obtained. The smaller of G I and G II is
realized; i.e., G I for T < Tc and G II for T > Tc . (Right) The jump of H and of T S across the
phase transition CI CII on the (H, T S)-plane. The gradient of the jump CI CII (a diagonal of
the dashed square) is 45 because of the H S compensation (2.28)
where we distinguish the two phases by I and II, see Fig. 2.9 (left), where the pres-
sure p is fixed. Exactly at the transition point, either one of the states is realized or
the two phases coexist macroscopically in the system.
Gibbs free energy consists of enthalpy (i.e., the internal energy of the system plus
the pressure environment) H and the entropic term, TS, that is G = H TS. In the
case of the liquidvapor transition of water, the liquid phase (I) gains the enthalpic
part of G I by its cohesive energy, while the vapor phase (II) gains the entropic term
(TS) by having a large specific volume per molecule. As a result, the enthalpy of
these two phases satisfies HII HI > 0, while their entropy obeys T (SII SI ) > 0. At
the phase transition (2.27) tells that the differences of these two terms must balance
( Enthalpyentropy (H S) compensation):
ing water molecules.37 Biological molecules are thought to use such transition-like
phenomena. Some soft materials often undergo the transition under physiological
conditions.
The H S compensation has sometimes been an object of controversy because
it can also arise as an experimental artifact [10]. In fact, if the term T G/ T is
predominant in the formula H = G T G/ T over the magnitude of G, we have
TS = T G/ T H. Therefore, we would have H T S for any change
of parameters, including T , within a relatively narrow range. The estimation of this
artifact may be a good exercise for error analysis problems.38
37 Upon a principal conformation change of a myosin molecule (called the isomerization), about
120 water molecules leave from the surface, according to the analysis based on dielectric measure-
ments [9]. There, the enthalpy changes of about H = +51(kJ/mol) is mostly compensated, leaving
little Gibbs energy difference between the conformations. (cf. kB T = 2.4(kJ/mol).)
38 [29] is a good introduction of error analysis.
92 2 Structure of Macroscopic Thermodynamics
without the extensivity of the system behavior. In the theory of glass [transition],
the lack of quasistatic processes gives rise to many interesting phenomena, such
as memory effect, aging, and plasticity. The related issue will be addressed in
Sects. 7.1.2 and A.7.2. The so-called effective temperature [28] of glassy states
indicates the (decaying) memory of the initial preparation.
References
1. Y. Oono, M. Paniconi, Prog. Theor. Phys. Suppl. 130, 29 (1998) 67, 70, 91
2. E. Lieb, J. Yngvanson, Phys. Rep. 310, 1 (1999) 67, 70, 91
3. H.S. Leff, A.F. Rex, Maxwells Demon: Information, Entropy, Computing (A Hilger (Europe)
and Princeton U.P. (USA), 1990) 69
4. M. Toda, R. Kubo, N. Saito, Statistical Physics I: Equilibrium Statistical Mechanics, 2nd edn.
(Springer-Verlag, Berlin, 1992) 71
5. E. Fermi, Thermodynamics, 1st edn. (Dover Publications, New York, 1956) 80, 83
6. S.R. De Groot, P. Mazur, Non-Equilibrium Thermodynamics (Dover Publications, New York,
1984) 81, 91
7. T.L. Hill, Free Energy Transduction and Biochemical Cycle Kinetics, dummy edn. (Springer-
Verlag, New York, 1989) 81, 91
8. E. Grunwald, L.L. Comeford, in Protein-Solvent Interactions, ed. by R.G. Gregory
(Marcel Dekker, Inc., New York, 1995), Chap.10, p.421 90
9. M. Suzuki et al., Biophys. J. 72, 18 (1997) 91
10. A. Cornish-Bowden, J. Biosci. 27, 121 (2002) 91
11. F. Weinhold, J. Chem. Phys. 63, 2479 (1975) 91
12. F. Weinhold, J. Chem. Phys. 63, 2488 (1975) 91
13. F. Weinhold, J. Chem. Phys. 63, 2495 (1975) 91
14. G. Oster, A. Perelson, A. Katchalsky, Nature 234, 393 (1971) 91
15. L. Onsager, Phys. Rev. 37, 405 (1931) 91
16. L. Onsager, Phys. Rev. 38, 2265 (1931) 91
17. L.D. Landau, E.M. Lifshitz, Fluid Mechanics (Course of Theoretical Physics, Volume 6), 2nd
edn. (Reed Educational and Professional Publishing Ltd, Oxford, 2000) 91
18. M. Paniconi, Y. Oono, Phys. Rev. E 55, 176 (1997) 91
19. K. Sekimoto, Prog. Theor. Phys. Suppl. 130, 17 (1998) 91
20. T. Hatano, S. Sasa, Phys. Rev. Lett. 86, 3463 (2001) 91
21. S. Sasa, H. Tasaki, J. Stat. Phys. 125, 125 (2006) 91
22. T. Komatsu, N. Nakagawa, Phys. Rev. Lett. 100, 030601 (2008) 91
23. T. Komatsu, N. Nakagawa, S. Sasa, H. Tasaki, Phys. Rev. Lett. 100, 230602 (2008) 91
24. M. Horodecki, J. Oppenheim, R. Horodecki, Phys. Rev. Lett. 89, 240403 (2002) 91
25. Y.G. Sinai, Sov. Math. Dokl. 4, 1818 (1963) 91
26. D. Ruelle, Thermodynamic Formalism (Addison-Wesley, Reading, 1978) 91
27. T.L. Hill, Thermodynamics of Small Systems, Parts I and II (Dover Publications Inc., New
York (Original: W.A. Benjamin, Inc. New York, 1963, 1964), 1994) 91
28. L.F. Cugliandolo, J. Kurchan, L. Peliti, Phys. Rev. E 55, 3898 (1997) 92
29. J.R. Taylor, An Introduction to Error Analysis, (Sect. 8.5) 2nd edn. (University Science Books,
California, 1997) 91
Chapter 3
Fluctuations in Chemical Reactions
We will survey several concepts and notions related to chemical reactions at various
scales and from various viewpoints. In the later chapters we will use them in the
energetics of the fluctuating world.
The notion of a molecule belongs to the far from equilibrium states. Chemical
reactions are then the transitions among different steady states of molecules, which
often accompany the conversion of degrees of freedom between the translational
degrees of freedom and the intramolecular ones.
In the survey of macroscopic reaction theory we limit ourselves to the case where
the reactions are described in terms of the concentrations of the chemical com-
ponents and the rate constants. Chemical equilibrium is related to the equilibrium
thermodynamics. Inversely, macroscopic open systems can be described as reaction
systems. Even within a macroscopic description of chemical reactions, the charac-
teristic scale is a useful notion. We mention as examples the buffer solutions and the
MichaelisMenten kinetics.
At more microscopic scales chemical reactions are described in terms of the
number of molecules (i.e., integers) of each chemical species. Stochastic processes
of discrete systems are the general framework of chemical reactions on this scale.
The master equation is the most often used description of discrete stochastic pro-
cesses. The detailed balance condition is the equilibrium condition expressed by
the probabilities of discrete states and the transition rates among these states. The
discrete version of Langevin equation is an alternative and equivalent method to the
master equation. The continuous Langevin equation (Chap. 1) can be represented as
a limit of discrete processes. We take the same examples as those we have taken in
the macroscopic description: A + B AB, the open system, and the Michaelis
Menten kinetics. It will be useful to find the similarities and differences between the
two different scales of description.
Sekimoto, K.: Fluctuations in Chemical Reactions. Lect. Notes Phys. 799, 93131 (2010)
DOI 10.1007/978-3-642-05411-2 3 c Springer-Verlag Berlin Heidelberg 2010
94 3 Fluctuations in Chemical Reactions
anion than monovalent cations do. This is a qualitative way to understand why the
overcharging around a charged colloid is more effective by multivalent counterions
than the monovalent ones [1].
The stability of molecular states is due to the high free-energy barrier (relative
to kB T ) associated with the destabilizing reactions. Since the reaction rate is mainly
governed by the exponential Boltzmann factor, the reactivity is efficiently changed
if the free-energy barrier is modified. This exponential dependence enables the near
discrete switching on/off of reactions in inorganic and organic matter. The energetic
aspect of controlling the barrier height will be discussed later (Chap. 7).
On the molecular level, the description of chemical reactions as well as open
system has to incorporate the transfer of degrees of freedom. Concomitantly the
energetics of chemical reaction and open system on the fluctuation level must take
into account the change in degrees of freedom (see Sect. 4.2.3).
d[AB]
= k[A][B], (3.1)
dt
where the parameter k is the rate constant of this reaction. The macroscopic equation
for the rate of production, like (3.1), is called a rate equation. k[A][B] indicates
the frequency of the collisions, or encounter, between an A molecule and a B
molecule. The dimension and magnitude of k depend on the choice of the unit of
concentration, [A], etc. In physical chemistry one usually uses the molar concen-
tration, or molarity, [no. of mols of solute]/[no. of liters of solution].2 In this book,
however, we will adopt the following unit, unless stated otherwise explicitly:
2 The use of mol in the definition is reasonable in the sense that macroscopic observers cannot
count the number of solute molecules.
3.2 Macroscopic Description 97
(no. of X molecules)
[X] . (3.2)
(volume in cm3 of solution)
We choose this unit with a view to maintain continuity with the more microscopic
level discussions in later sections.
If we also take into account the reverse reaction, i.e., A + B AB, we introduce
another rate constant, k , which has different dimensionality from that of k:
d[AB]
= k[A][B] k [AB]. (3.3)
dt
The second term on the right-hand side, k [AB], implies that an isolated AB
molecule has an average inverse lifetime, k . The law (3.3) is valid in the limit of
(i) dilute solution and (ii) near equilibrium. We will discuss a more general case of
nondilute solution in Sect. 3.3.3.
Application of the zeroth law of thermodynamics to the whole system, the closed
reaction system plus the thermal environment, implies that this isolated system
will reach the thermal equilibrium after infinitely long time. The steady state
of (3.3), i.e.,
[AB] k
= (equilibrium). (3.6)
[A][B] k
This type of equilibrium condition, where the powers of concentration appear in the
denominator and numerator, is called the law of mass action. The ratio of the kinetic
parameters on the right-hand side is called the equilibrium constant, which depends
only on the thermodynamic parameters, as we will see below.
If the reaction is not closed, for example, under the steady injection of A and B
and the removal of AB, the stationary condition, (3.5), does not give equilibrium.
A + B = AB . (3.7)
On the other hand, the chemical potential of the solute molecules in a dilute
solution is
d[AB]
= k[A][B] k [AB] kout [AB] + kin ,
dt
d[A] d[B]
= k[A][B] + k [AB] = . (3.10)
dt dt
Here the exchange with the particle reservoir of [AB] is characterized by the two
rate constants, kout and kin . The stationary condition, (3.5), yields the following
two independent relations as the law of mass action:6
5 Note that NA = [A]V etc., where V is the volume of the system. Under the isothermal and isobaric
condition, the second law requires (ET S+P V ) = G W = 0 for any spontaneous changes.
d NA and d NB are dependent on d NAB according to (3.4).
6 Ifthe particle reservoir exchanges (only or also) A molecules, the equation for d[B]/dt still
assures the law of mass action. If the system exchanges all the species of molecules, the law of
mass action will no more hold.
3.2 Macroscopic Description 99
[A][B] k kin
= , [ AB] = (equilibrium). (3.11)
[ AB] k kout
At the equilibrium, the total Gibbs free-energy, G(T, p, NA , NB , NAB ) + G res (T, p,
NAB,res ), should be minimized under the constraints of particle conservation, d NA +
d NAB + d NAB,res = 0 and d NB + d NAB + d NAB,res = 0. This yields7
A + B = AB = AB,res , (3.12)
Remarks
The chemical reaction theory emphasizes the transformation of the mass, but
the energetic aspects (change of free energy, endothermic, or exothermic, etc.) are
treated separately using thermodynamics. If we use stochastic energetics, both mass
transformation and energetics are discussed on the basis of a single event of the
reaction.
The chemical potential is a quantity on the level of the description where the
particles are anonymous. The chemical potential (e.g., (3.8)) does not represent the
free energy carried by an individual molecule. A is the energetic interpretation
of the relative probability of the arrival of anonymous molecule A, [ A]eA /kB T ,
0
A /kB T
using the form of the Boltzmann factor, e .
It often occurs that two or more reactions related to the same molecular species
have widely different equilibrium constants. That the equilibrium constants depend
exponentially on the chemical parameters, like 0A in (3.9) and (3.13), explain this.
Because of this aspect, we expect that a chemical reaction system can exhibit
qualitatively different behaviors in different regimes. The following two simple
examples demonstrate how the phenomena of different scales are treated.
7 We may use the method of Lagrange multiplier: The requirement of d[G(x, y, z) + G res (z)] = 0
under the constraints, d(x + z + z) = 0 and d(y + z + z) = 0, is equivalent to the constraint-
free requirement, d[G(x, y, z) + G res (z) x (x + z + z) y (y + z + z)] = 0, with x and
y being unknown, called the Lagrange multipliers. The parameter set, (x, y, z, z), stands for
(NA , NB , NAB , NAB,res ). Then A = x , B = y , AB = x + y , and res = x + y .
100 3 Fluctuations in Chemical Reactions
h (b0 + h y) = K w , (3.14)
hy
= Ka, (3.15)
a0 y
Kw
b0 a0 = C(h, a0 , K a ) h, (3.16)
h
8 Inmolar unit, what would be the denominator, [H2 O] = 55.5, is roughly of order 1 and conven-
tionally suppressed.
3.2 Macroscopic Description 101
(A,pKa) = (1,5)
pH
14
12
10
ba
1.0 0.5 0.5 1.0
The new regime, called the buffer regime, is h/K a 10pH+pK a 1 in (3.16).
If we assume h a0 and Khw a0 in (3.16), h/K a is written as Kha b0ba 0
0
. It
b0 a0
implies that, as far as b0 1, the pH of the solution is kept at around pH
pK a log10 K a . The above assumptions are consistent since ah0 Ka0a = Y 1 and
Kw
ha0
KKa wa0 = YX 1. The buffering regime is, therefore, realized due to the double
inequalities, X Y 1.
A simple 1:1 reaction between a Substrate (S) molecule and a Product (P)
molecule catalyzed by an Enzyme (E) is E + S E + P. When we are inter-
ested in the dependence on the enzyme concentration, [E], we do not simplify this
scheme as S P. One of the representative schemas of enzymic reaction is the
MichaelisMenten kinetics which has been introduced around 1913. The schema of
MichaelisMenten kinetics is
E + S ES E + P. (3.17)
In this schema we assume the following circumstance, see Fig. 3.2 (A):
1. The direct reaction, S P, is slow enough to be ignored.
2. There is a transition state called EnzymeSubstrate complex (ES).
3. The total concentration of the enzyme, [E]tot = [E] + [ES], is finite.
4. At most one substrate particle, S, can interact at any time with an enzyme particle
(protein).
5. [P] is much smaller than the equilibrium value, so that the backward reaction,
ES E + P, can be neglected.
These hypotheses might look very particular among many other possibilities.
Nevertheless we will argue later that the above schema (3.17) is rather general from
the viewpoint of the timescales of reaction.
MichaelisMenten Equation
Let us denote by kcat the rate constant of the forward reaction, ES E + P. See
Fig. 3.3 (left). We seek the rate of production,
d[P]
= kcat [ES], (3.18)
dt
when the preceding step of complex formation has reached the steady state: 9
d[ES]
= k+ [E][S] (k + kcat )[ES] = 0. (3.19)
dt
This equation leads to the relation10
[E][S] k + kcat
= . (3.20)
[ES] k+
From this equation, we eliminate [E] using [E] = [E]tot [ES] and then we substitute
the resulting [ES] as function of [S] into (3.18). We then obtain the desired result,
which is called the MichaelisMenten equation (See Fig. 3.2 (B)),
Vmax [S]
v kcat [ES] = , (3.21)
K M + [S]
where we have introduced the maximal production rate, Vmax , for [S] = +, and
the so-called MichaelisMenten constant, K M :
k + kcat
Vmax kcat [E]tot , KM . (3.22)
k+
10 Although (3.20) takes of the form of the law of mass action, the system is not in equilibrium.
The exception is when kcat k holds. The approximation of neglecting kcat in (3.19) is called
the rapid-equilibrium assumption.
11 The latter representation is called the HanesWoolf plot.
104 3 Fluctuations in Chemical Reactions
and the product P for this key reaction is a substrate of the following (downstream)
reaction(s). Thus, if the product of the upstream reaction is supplied faster than it
is converted into P, then the substrate S stays among E + S and ES states, and the
quasi-equilibrium (rapid equilibrium), E+S ES, is realized. Also, if the substrate
of the downstream reaction is consumed faster than it is converted from S, then the
product P has little probability to be bound to E to form ES, and the unidirectionally
reaction, ES E+P, is realized. These two features are what constitute the schema
of MichaelisMenten kinetics. Therefore, the MichaelisMenten schema describes
a general feature of the key reactions. The formula (3.21) shows that the key reaction
is controlled by the total number of the enzyme, [E]tot .12
If a reaction has the substrate concentration, [S], much smaller than the K M , then
this reaction cannot be controlled by [E]tot , and the reaction is not the key reaction. If
a lot of enzyme [E]tot is injected in a key reaction, then there would be a shortage of
the substrate and the process is no more controlled by this reaction, i.e., the reaction
ceases to be the key reaction. In brief, the MichaelisMenten kinetics works when
and where the reaction in question is among the key reactions in the network. In
biochemical reaction network the change of the activity of enzymes13 may change
dynamically the locations of the key reactions.
The notion of the key reaction described above is not limited to the bulk enzy-
matic reaction in solution. The production rate of the form of (3.21) is found in
other conditions: for the reaction with a surface catalyst the rate of production obeys
the form of (3.21). If the concentration [S] is higher than K M , then most catalyst
molecules are occupied, and it limits the production rate. For crystal growth from
vapor or from solution competition between bulk diffusion and surface kinetics leads
to the growth velocity in the form of (3.21).
There is another interpretation of the MichaelisMenten equation, which also
explains why MichaelisMenten-type behavior is found in a variety of situations.
See Fig. 3.3 (right). If we look back to the derivation of (3.21), we find that
the mathematical origin of its saturating feature is found simply in the bistable
transition between the states, E+S and ES. See Fig. 3.2. Wherever this (quasi)
equilibrium to-and-fro exists, any observable linearly related to the occupied frac-
tion of the ES state should show MichaelisMenten type saturation.14 For exam-
ple, if S and E are, respectively, the ATP-hydrolyzing motor protein and its fil-
ament, then the ATP consumption rate should obey the curve of (3.21) since
ATP hydrolysis is catalyzed by the motorfilament interaction. In this case
the products, ADP and inorganic phosphate (Pi), are not fragments of a motor
protein.
Transition Rate
We denote by Ii = { j}i the set of indices of the states {S j } that the state Si of the
system can make direct transitions.16 The so-called transition rate, wi j , from the
discrete state Si to the discrete state S j ( j Ii ) is defined as follows:
Suppose that a system is in the state Si at a time t.
The (conditional) probability that the system makes the transition to a different
state S j during an infinitesimal time lapse, dt(> 0), is wi j dt.
As a result, the probability that the system remains in the same state Si at t + dt is
(1 jIi wi j dt).
We recall the first passage time, introduced in Sect. 1.3.3.3: The transition from
the state i to the other state can correspond to the exit from a basin of a potential
energy U (x). This was the idea of Kramers to calculate the transition rate [3].17
1
The average first passage time (FPT) is related to ( jIi wi j ) . To assess the
individual transition rate, wi j , we will need to solve the first passage problem
under constraints. In the context of FPT, the Markov approximation amounts to the
neglect of the initial position dependence of the first passage time.
16 By definition, i Ii .
17 A review [5] surveys many papers after [3] up to 1990.
3.3 Stochastic Description 107
Ji j = Pi wi j P j w ji = J ji . (3.23)
Master Equation
Let us denote by Pi (t) the probability to find the system at time t in the state
Si . Given the concept of the probability flux (3.23), the redistributed probabilities
{Pi (t + dt)} should satisfy
Pi (t + dt) = Pi (t) Ji j dt.
j
Then we have the following evolution equation for {Pi (t)} called the master equation:
d Pi (t)
= Ji j , (3.24)
dt j
Steady State
The steady state (or the stationary state) of the master equation is defined such that
d P j (t)/dt = 0 for every state, S j . The steady state does not imply the flux-free state:
Ji j = 0. A simplest model may be the three-state system with the transition rates
being w12 = w23 = w31 w > 0 and w21 = w32 = w13 w > 0
18 The expression (3.24) has the same structure as the equation of continuity or the mass conserva-
tion, where the sum on the right-hand side is the divergence of the flux.
108 3 Fluctuations in Chemical Reactions
n
Pi
D( P Q) Pi ln ( 0). (3.25)
i=1
Qi
The continuous version of this quantity has been introduced in Sect. 1.2.3.2, where
the nonnegativity of this quantity has been shown.19 When the time-discretized
probability P converges to the steady-state probability Q, the D( P Q) mono-
tonically decreases to 0. Below is a brief derivation. The technical details are given
in Appendix A.3.1.
1. If the time and the states are discretized, the evolution of the probability P by the
master equation can be formally written as a Markov chain, i.e., the discrete-time
discrete state Markov process,20
n
P {Pi } K P { Ki j Pj } (i = 1, . . . , n), (3.26)
j=1
21 That K P approaches K Q does not imply the uniformization of P and Q by the Markov evolu-
tion, K.
110 3 Fluctuations in Chemical Reactions
Poisson Noise
A particular realization of Poisson noise ( ) is as follows:
(t) = (t t ), (3.29)
where (z) is the Dirac delta function and {t } with t < t+1 represents the time of
spike events, which take place randomly. The spike events are a Markov processes
and, therefore, characterized only by the mean rate of spiking per unit of time, w.
Within a very small time interval, (t, t + t], the normalized probability for having
n spikes is given by the Poisson distribution:
t+t
ewt
P (s)ds = n = (wt)n . (3.30)
t n!
t+t
For n t (s)ds we can verify n = wt. We generalize this definition to all
the time slices, and we allow for the dependence of the mean spiking rate w(a) on
the external parameter, a. When a is varied as a function of time, (t) = w(a(t)).22
If 1(t) and 2 (t) are two independent Poisson noises, we can identify 1(t)2 (t) = 0
within any integration over time t. The reason is that, for a given time interval,
(t, t + t], the probability that both processes give rise to at least one spike is
O(t 2 ). In the limit of t 0 such events are negligible (i.e., measure 0). Here-
after we often omit for stochastic processes for simplicity of notation.
To apply the Poisson process for transition between discrete states, we assign
i.i.d. Poisson process to each distinct transition:
t+t
n i, j (t, t + t) = i, j (s)ds, (3.31)
t
and i, j (s) are independent Poisson noises with i, j (t) = wi j (a(t)).23
We denote, following [16], the states that the system can take by the base vectors
|i, etc., i is a discrete index. These states can depend on the systems parameter a.
We introduce also the dual base vectors, i|, etc., so that i| j = i j for the same
parameter a.
By |i(t) we denote the state of the system at time t. i(t) is among the index
mentioned above. We define that, unless the systems state undergoes transition,
t
|i(t) = 0 even if the systems parameter a depends on time.
22 We
$ i n % can(everify the following formula for the characteristic function:
= e 1)wt . The characteristic functional of the Poisson noise ( ) is
i
e
t
ei 0 (t ) (t )dt = exp 0 (ei(t ) 1)w(a(t ))dt , where (t) is an arbitrary smooth real function.
t
23 The characteristic function ei 0t i, j (t )i, j (t )dt is equal to
exp 0 (eii, j (t ) 1)wi j (a(t ))dt .
t
3.3 Stochastic Description 111
d
|i(t) = (| j |i(t)) i(t), j (t), (3.32)
dt j
where dtd |i(t) (|i(t + dt) |i(t))/dt, and the multiplicative Poisson noise,
i(t), j (t), obeys (3.31). The symbol means the Ito-type multiplication: The vec-
tor |i(t) on the right-hand side is nonanticipating with respect to the variation of
i(t), j (t).
In order to see how (3.32) works, let us assume that the first future spike in the
Poisson noise among {i(t), j (t)} is at t = t with j = j . Then the integration of
(3.32) from the present time t up to t + 0 yields
where the last factor 1 comes from the time integration of (t t ). We then update
the systems state to |i(t ) = | j .
Another confirmation is that (3.32) reproduces the master equations (3.24) and
(3.23). We denote the probability for the state |i at the time t by P j (t) j,i(t)
with j,i(t) = j|i(t). Then we have
d Pj
= ( j, j,i(t) ) i(t), (t)
dt
= ( j, j,i(t) )k,i(t) k, (t)
k
= [Pk wk j (a(t)) P j w jk (a(t))]. (3.33)
k
Equation (3.32) is analogous to the Langevin equation and the characteristic func-
tional of its noise (t). Gillespie has proposed an approximate approach with finite
t and then coarse grained it to have a Langevin equation [17].
pii (t): the probability that the system stays continuously in the state between
the time t and t + t.
pi j (t) ( j Ii ): the probability that the system has ceased to stay in the state
Si to make the first transition to the state S j between the time t and t + t.
The solution for pii (t) and pi j (t) ( j = i) with any t( 0) are
pii (t) = exp{ wi j t},
j Ii
wi j
pi j (t) = [1 exp{ wi j t}]. (3.36)
j Ii wi j j I i
These probabilities satisfy the sum rule, pii (t) + jIi pi j (t) = 1.
We can verify that those (t , j ) defined by (1)(2) reproduces (3.36). Using
the rules (1)(2), the probability corresponding to pi j (t) writes
Prob T j ( y j ) t, and T j ( y j ) T j ( y j ) ( j Ii ) pi j (t), (3.37)
where we defined the functions T j (y) by ewi j T j (y) = y. Using the uniform distri-
bution of { y j } jIi , we have
where (z) = 0 for z < 0 and (z) = 1 for z 0. The integrations are simple to
do,25 and we find for pi j (t) the identical expression pi j (t) of (3.36).
or, equivalently,
We call such steady state the state of detailed balance. Whether a system has the state
of detailed balance depends on the transition rates because there are more equations
{Ji j = 0} than the number of components of {P j }.
wi j
25 We use the formula j Ii ( j = j) wi j +1= 1
wi j j Ii wi j .
114 3 Fluctuations in Chemical Reactions
eq
The above parameterization allows us to represent both {P j } and {wi j } in the
context of the thermal activation-assisted transition in a free-energy landscape.
eq eF j /kB T i, j Fi
Pj = F /k T , wi j = exp . (3.40)
i e kB T
i B
F
Fig. 3.4 Interpretation of
(3.40). The squares on the F F
corners represent the
free-energy levels, Fi , while
the plateaus between the
nearest states represent the F
barriers, i, j
29 Once we fix the s value, the difference F has a physical meaning of the activation
i, j i
barrier and contains no arbitrary additive constants.
116 3 Fluctuations in Chemical Reactions
Nonequilibrium Processes
Once the transition rates are fixed using the detailed balance condition of equilib-
rium, we can proceed to study the master equation under nonequilibrium conditions.
This implies several different things:
Transient nonequilibrium states: Keeping the transition rates of the form (3.40)
unchanged, we study the relaxation of the probabilities P j (t) starting from
nonequilibrium initial ones. Also we solve the discrete Langevin equation
with these transition rates and look for stochastic processes on the state space.
Nonequilibrium settings I: We may vary F j s, i, j s, or kB T as function of time
in the transition rates of (3.40). In this case, the instantaneous equilibrium
distribution and the landscape picture are still valid according to (3.40). The
P j (t) evolves toward a temporary equilibrium state, although there can be a
lag of relaxation.
Nonequilibrium settings II: We modify each activation barrier, i act
j i, j
Fi , disregarding the DB conditions, based on physical arguments.
In fact each activation barrier often has a physical justification of its own,
independent of the compatibility with the global equilibrium states. It is like
that the macroscopic rate constant k for the reaction A+BAB can be used
either in (3.3) or in (3.1), which is far from equilibrium. Therefore, we can
combine these transition rates to build up a reaction network having nonequi-
librium steady states. A simple example is given in Sect. 3.3.3.3
Since the number of combinations of [i j], i.e., n(n 1), is more than
the degrees of freedom left by the DB condition, n(n + 1)/2 1, the modi-
fied transition rates can no more be represented by a single-valued landscape
like Fig. 3.3. The system now allows steady-state circulations of probability
flux.30
Relation between the transient nonequilibrium states and the nonequilibrium
settings II: When we model the chemical coupling schematized by Fig. 2.7,
we can model either the whole closed system, i.e., the chemical engine plus
the four particle reservoirs, or the chemical engine as an open system.
The former viewpoint is not very practical but formally simple: we assume
the transition rates satisfying detailed balance. The global landscape is fixed
like in Fig. 2.8. The nonequilibrium process is regarded as a transient process
toward the equilibrium of the whole closed system.
The latter viewpoint is more practical but we must use the nonequilibrium
settings of type II. The reservoirs states are no more taken into account, and
only the chemical potentials enter as a parameter for the chemical engine.
In the former formal point of view, the systems evolution is relaxing and
downhill on the average. In the latter point of view, the chemical engine
makes a stochastic cyclic transitions in its (reduced) state space.
5 4 3 2 1 j=0 +1 +2 +3 +4 +5
time in the left region ( j < 0), where the diffusion takes more time? The key to
avoid this trap is to be aware of the opposing effects to this argument. The state S j=0
situated between the high barrier W to the left and the low barrier w to the right. If
the system is in this state, the (conditional) probability of the transition S0 S1 is
smaller than that of the transition S0 S+1 . Therefore, the chance that the system
enters into the j < 0 states is relatively small, though the residence time in j < 0
states are relatively large. In equilibrium, these two effects exactly cancel.
1 dU (xt )
d xt = dt + 2Dd Bt , (3.41)
d xt
where we denote by xt the value of x at time t and D is the diffusion coefficient. The
probability density for xt is (x xt ). Then the conditional probability density for
xt+dt = xt + d xt , given that x = xt at t, is (x xt )(x xt+dt )/(x xt ). For
x = x , this conditional density gives the flow of probability to x within the time
dt. Therefore, the transition rate wxx should be related to this density through
31 The derivation uses the technique called the KramersMoyal expansion. See, for example,
[3, 25, 26, 7, 8].
118 3 Fluctuations in Chemical Reactions
1 dU (x)
wxx = (x x) + D (x x). (3.43)
dx
where f (x) and g(x ) are arbitrary functions of good properties.33 A straightforward
integrations by parts of (3.44) with (3.43) leads to D = kB T / .
stochastic parameters, and see how the number of molecular species are distributed
at equilibrium.
NA NB
w NAB NAB +1 = k , (3.45)
V
where k is a constant independent of the number of molecules or volume. The
approximation leading to (3.45) is that every A molecule and B molecule is dis-
tributed randomly in the volume V and that the chance to find a pair of A and
B molecule within an atomic distance is NAVNB up to the relative error of
O(NA /V, NB /V ). The factor k should include the activation factor, or the prob-
ability that a collision between an A and a B molecules leads to the formation of an
AB molecule.
Rate Constant
We will relate the transition rate w NAB NAB +1 to the rate constant of the macroscopic
description of the reaction, described in Sect. 3.2.1.1. If the state transition, NAB
(NAB + 1), occurs the probability w NAB NAB +1 dt for an infinitesimal time dt, NAB
should increase approximately by w NAB NAB +1 dt, that is d NAB /dt w NAB NAB +1
= k NA NB /V , where we have used (3.45). Dividing each part of this equation by
V , we have
120 3 Fluctuations in Chemical Reactions
The rightmost the equation is of the same form as the formula for the macroscopic
reaction. Therefore, we identify the coefficient k in (3.45) with the rate constant of
the reaction, A + B AB.
Next we take into account the backward state transition, (NAB + 1) NAB , also.
This corresponds to the reaction AB A + B. We assume that the transition rate,
w NAB +1NAB , is proportional to the number of AB molecules before the transition,
therefore,
d[AB]
k[A][B] k [AB]. (3.48)
dt
We, therefore, identify the coefficient k in (3.47) with the rate constant of the reac-
tion, AB A + B.
JNAB NAB +1 P(NAB , t)w NAB NAB +1 P(NAB + 1, t)w NAB +1NAB
= JNAB +1NAB . (3.49)
d P(NAB , t)
= JNAB 1NAB JNAB NAB +1 . (3.50)
dt
min{NAtot ,NBtot }
P(n, t) = const. (3.51)
n=0
Equation (3.52) includes macroscopic reaction equation for [ AB], i.e., for NAB =
min{NAtot ,NBtot }
n=0 n P(n, t). See, for example, [79] for the methods to derive the macro-
scopic equation from (3.52).
under fixed values of NAtot and NBtot . This condition gives the probabilities of the
equilibrium state:
122 3 Fluctuations in Chemical Reactions
where N is the normalization constant, and the three parameters, ( NA , NB , NAB ) are
solutions to the following three equations:
( NA v0 /V )( NB v0 /V ) k
= , NA + NAB = NAtot , NB + NAB = NBtot , (3.54)
( NAB v0 /V ) k
where v0 is a constant with the dimension of volume. The formula (3.53) looks
apparently the product of three independent Poisson distributions. In fact, these
distributions are not independent for the closed system, because of the constraints
NA = NAtot NAB and NB = NBtot NAB . Equation (3.54) gives a function of the
single variable, NAB .
Nevertheless the form of the product of Poisson distributions is the consequence
of only the DB condition, whether or not the system is closed. When the number
of NA and NB is adjustable by their particle environments, (3.53) describes the true
product of Poisson distributions.
Chemical Potential
In order to see how the chemical potential is interpreted in this simple reaction
scheme, A + B AB, we will compare the result of P eq (NAB ) with the prediction
from statistical mechanics.
We define 0A , 0B , and 0AB as the free energies of individual molecules, A, B,
and AB, respectively, which reflect the effect of the kinetic energies and the internal
degrees of freedom of each molecule in the form of the free energy:
0M = kB T ln z M (M = A, B, AB), (3.55)
where z M are the partition functions of a molecular species M.35 These free energies
can depend on the temperature. We then assign to each state of the system, NAB
( min{NAtot , NBtot }), the following potential energy:
where NA NAtot NAB and NB NBtot NAB .36 We then write down the canonical
(configurational) partition function of the form of eU (NAB )/kB T . We have
min{NAtot ,NBtot }
Z (T, V, NAtot , NBtot ) ANA BNB AB
NAB
= , (3.56)
NAtot ! NBtot ! NAB =0
NA ! NB ! NAB !
0M
where M (V /v1 )e kB T , and v1 is a constant. The factorials come from the combi-
natorial number, NAtot C NA NBtot C NB NAB !, for forming NAB of AB molecules out of
NAtot of A molecules and NBtot of B molecules. According to the statistical mechanics
of Gibbs, we can identify each term in the sum on the right-hand side of (3.56) with
the relative probability. This probability has exactly the form of (3.53), except for a
factor of normalization. Therefore, we find
M = NM .
We have thus recovered the result of macroscopic thermodynamics, (3.9), and (3.55)
gives the microscopic meaning for 0A , 0B , and 0AB .
35 Strictly speaking, z M is made nondimensional by a factor related to the atomic specific volumes.
36 The relation between the potential energy and the landscape representation (Sect. 3.3.1.5) will
be discussed in the next Sections.
37 We put v0 = v1 .
124 3 Fluctuations in Chemical Reactions
This is a simple model of physical adsorption of molecule A, where the system (S)
is a 2D substrate with a volume V .
Following the protocol of Sect.3.3.2, we build up the master equation of this
model.
1. The state of the system is distinguished by the number of particles in the open
system, NA N .
2. The transition rate w NN+1 represents the average rate at which a particle enters
the open system, while w N+1N reflects the rate at which any one particle in the
open system leaves for the environment. We will not assume the properties of
dilute solutions, but leave the transition rates very general [33].
3. The probability flux is JN N +1 = PN w N N +1 PN 1 w N 1N .
4. The master equation for PN is then
d PN
= JN N +1 JN N 1 . (3.59)
dt
5. We assume that the system is in equilibrium. We then impose the detailed balance
condition (3.40). It relates w N N +1 with w N +1N :
FN +1 FN
w N +1N = w N N +1 exp . (3.60)
kB T
S S
R RI RII
(A) (B)
Fig. 3.6 (A) A system consisting of an open subsystem, S, and a reservoir, R. (B) A system con-
sisting of an open subsystem, S, and two reservoirs, RI and RII
3.3 Stochastic Description 125
dN
= w N N +1 w N +1N . (3.61)
dt
The relation (3.60) is not very convenient because FN concerns the whole system.
By a physical argument we will rewrite (FN +1 FN ) in terms of chemical potential.
We assume that the coupling between the open system and the reservoir is short-
sys
ranged so that the free energy is additive: FN = FN + FNrestot N + (indep. of N ),
where N tot is the total number of particles in the whole system (SR), and the last
term represents the interface between the open system and the environment. The
chemical potentials of the open system and the reservoir are defined, respectively,
sys
by FN / N and res FNresres / N res . Then we have
FN +1 FN = res . (3.62)
dN res
= w N N +1 1 exp .
dt kB T
where
Unlike the law of mass action, (3.64) allows the rate (d[A]/dt) to depend on [A]
in nonlinear manner. In the van der Waals model of fluids, depends on [A] in
a sigmoidal way. As a result, there can be several equilibrium concentrations [A]
satisfying ([A]) = res for certain value(s) of res . This leads to the phase changes
between the cooperative adsorption phase and the dilute adsorption phase.39 In
addition to the multi-equilibria, formula (3.64) describes the relaxation and hys-
teresis of the concentration [A].
N ,N +1 (+ FN + FN +1 ), + + = 1, (3.66)
is also of O(V 0 ) for arbitrary choice of + (= 1 ). Then (3.61) with (3.40) can
be rewritten using this and the chemical potentials of (3.62):
d N + ( res ) + ( res )
= exp exp . (3.67)
dt V kB T kB T
and denote the state of the whole system by the number of particles in the system,
N N A , and that in the reservoir RI , N NAResI . (Note that the number of
particles in reservoir RII , N AResII depends on N and N ).
The transitions we consider are
(N , N ) (N 1, N 1): migration of an A molecule between S and RI .
(N , N ) (N 1, N ): migration of an A molecule between S and RII .
One of the easiest ways of modeling the transition rates for these transitions is
to consider the special case where one of the two reservoirs is practically inacces-
sible due to an extremely high energy barrier between the system and the reservoir.
We can then use the previous results (3.63) and (3.65) for the transition rates for
the migration with the reservoir which is not blocked. A physical hypothesis that
we take here is that the transition rates between the system and a reservoir are
unchanged whether or not the migration with the other reservoir is blocked.41 If
this is a good approximation, we can repeat the same argument by exchanging the
blocked reservoir and unblocked one, we obtain the following model:
We recall the note on the Relation between the transient nonequilibrium states
and the nonequilibrium settings II (Sect. 3.3.1.5). If we describe the process in the
whole (N , N ) space, we may construct a landscape that matches the above tran-
sition rates. (In fact the state space is separated into slices according to the total
number of particles.) But it is not a practical description. We rather use a (reduced)
representation where we look at only the open system and regard the particle envi-
ronments as stationary reservoirs. Then each line of (3.69) represents the rate of
entrance from RI , of departure to RI , entrance from RII , and of departure to RII ,
respectively. Here the chemical potential depends on N .
On the macroscopic level, the equation for [A] corresponding to (3.64) is
d[ A] ResI ResII
= kin
ResI
1e kB T
+ kin
ResII
1e kB T
, (3.70)
dt
tP
n
tP = (tr(k) + tv(k) ). (3.71)
k=1
Setup of Problem
We suppose that one can only observe the events of the product release, that is the
sequence of tP . Also we assume that one can do the observation with different values
of the substrate concentration, [S]. We introduce a Markov model which contains
the following statistical parameters:
and
tP 1
= [S] + . (3.73)
tP 2 2tP
2 Tr
For the derivation, see Appendix A.3.2. The first result (3.72) is the rate of pro-
duction per enzyme. This could have also been obtained from the macroscopic pro-
duction rate, v, of (3.21) if we knew the total molar concentration of the enzyme,
130 3 Fluctuations in Chemical Reactions
[E]tot . By contrast, the second result, (3.73), contains intrinsically the information of
the stochastic data. From the data of tP and tP 2 for various concentrations of the
substrate, [S], the formulas (3.72) and (3.73) allow us to obtain the three constants,
, and Tr 1 , and q. Especially, n = (1 q)1 is obtained.
3.4 Discussion
We have analyzed chemical reactions using two levels of descriptions, macroscopic
deterministic and continuous description and mesoscopic stochastic and discrete
description. The latter description includes the former as a result of coarse grain-
ing. But if the total number of a molecular species is not large, the coarse-grained
description is not valid.
Even the discrete stochastic description may not be valid if the transition rate
depends strongly on the parameters which are not represented by the chemical for-
mula, such as the orientations or configurations of participating molecules and the
internal states of solvent molecules. For example, a rapid water exchange between
the hydration shell of a molecule and the surrounding fluid water is beyond the
description of the previous sections.
Depending on what spatiotemporal scale is decisively important for the reaction
rate, we should choose different methods, such as the Langevin equation, molecular
dynamics simulation, or density functional description. Except for the full quantum
descriptions of whole atoms and electrons, we always ignore some details as rapidly
(often said adiabatically) following degrees of freedom, but the justification of
such hypothesis separation of fast and slow degrees of freedom is often done a
posteriori by the comparison of model results with experimental observations.
References
1. R. Messina, C. Holm, K. Kremer, Phys. Rev. E 64, 021405 (2001) 96
2. P.W. Atkins, Physical Chemistry, 4th edn. (Oxford University Press, Oxford, 1990) 103
3. H.A. Kramers, Physica 7, 284 (1940) 103, 106, 117
4. P. Faccioli, M. Sega, F. Pederiva, H. Orland, Phys. Rev. Lett. 97, 108101 (2006) 103
5. P. Hanggi, P. Talkner, M. Borkovec, Rev. Mod. Phys. 62, 251 (1990) 106
6. C.W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry for Natural Sciences,
3rd edn. (Springer, 2004) 107
7. N.G. van Kampen, Stochastic Processes in Physics and Chemistry, Revised edn. (Elsevier
Science, 2001) 107, 117, 121
8. R. Kubo, K. Matsuo, K. Kitahara, J. Stat. Phys. 9, 51 (1973) 107, 117, 120, 121
9. K. Kitahara, Adv. Chem. Phys. 29, 85 (1973) 107, 121
10. L.D. Landau, E.M. Lifshitz, Statistical Physics (Part 1): Course of Theoretical Physics, Vol 5,
3rd edn. (Butterworth-Heinemann, 1980), sects. 7 108
11. T.M. Cover, J.A. Thomas, Elements of Information Theory (John Wiley & Sons, Inc.,
New York, 1991) 108, 109
12. M. Paniconi, Y. Oono, Phys. Rev. E 55, 176 (1997) 109
13. Y. Oono, M. Paniconi, Prog. Theor. Phys. Suppl. 130, 29 (1998) 109
References 131
In this chapter we introduce the concept of heat in the physics of the Langevin
equation, as well as its discrete version.
In the fluctuating world, some important aspects of the system cannot be described
in terms of the ensemble average of physical observables. Einstein was among the
first to study the connection between fluctuations and energy in the fluctuating world
[1]. Then Kramers described this connection in the context of stochastic processes.
In his seminal paper [2], he represented a chemical reaction as a process in which the
state of molecule(s) undergoes fluctuations and eventually changes its state qualita-
tively. That is, the molecules state (reaction coordinate) which was originally found
near the bottom of a potential (U (x)) can overcome a potential barrier, see Fig. 4.1.
Kramers analyzed this situation by constructing a FokkerPlanck equation, called
the Kramers equation, and derived the reaction rate (or transition rate).
Now, our interest is how much energy is exchanged between the system and
the thermal environment along an individual realization of such process. Does the
energy exchanged vary from one realization to other, or not? Knowing the con-
servation law of energy, we may argue that the energy absorbed from the ther-
mal environment equals to the height of the potential energy barrier. In order
to show explicitly the transferred energy and answer to the above question, the
FokkerPlanck (Kramers) equation is not appropriate, because this equation deals
with an ensemble of stochastic processes, but not an individual one. Therefore, our
study of energetics should be based on the Langevin equations1 :
dp U (x, a) p dx p
= + (t), = (underdamped) (4.1)
dt x m dt m
or
U (x, a) dx
0= + (t) (overdamped), (4.2)
x dt
if the inertial term is negligible.
1 In the following, we shall omit the symbolof X , etc., denoting a random variable or a stochastic
process, unless some confusion is possible.
Sekimoto, K.: Concept of Heat on Mesoscopic Scales. Lect. Notes Phys. 799, 135174 (2010)
DOI 10.1007/978-3-642-05411-2 4
c Springer-Verlag Berlin Heidelberg 2010
136 4 Concept of Heat on Mesoscopic Scales
Fig. 4.1 Kramers picture of chemical reaction: A new state is reached once a potential barrier is
overcome (dashed arrow) with the help of the thermal fluctuations
4.1 Framework
System: Of the whole world, a part which is properly cut out is called the system.
Its state is represented by x (and p) in (4.1) or (4.2). These equations describe the
evolution of the state of the system. We consider that the potential energy and the
kinetic energy belong to the system.2
Thermal environment: The background system to which the system is connected is
called the thermal environment. This environment is characterized in itself by a
single parameter, the temperature T . The conservation laws for the total energy,
mass or volume, and momentum have no effect on the state of the thermal envi-
ronment. The environment returns instantaneously to its equilibrium state and
keeps no memories of the systems action in the past. The interaction between the
system and the thermal environment is characterized by the friction coefficient
as well as temperature. The strength of the thermal random force, (t), is specified
by these two parameters.3
External system: It is an agent which is capable of controlling macroscopically
the system through the parameter a of the potential energy U (x, a). The term
external indicates that the evolution of the parameter a is not determined by
Eqs. (4.1) or (4.2). In the case of coupled Langevin equations, e.g., of the system
described by the state variables (x1 , x2 ), the x2 could take the position of a, that is,
the external system with respect to the subsystem, x1 . For the moment, however,
we reserve the concept of the external system for the variables whose dynamics
are not determined by stochastic equations.4
as motion. The fluctuating world is between these two limits. Let us try to imagine
the mechanical processes undergone by a Brownian particle.
The law of action and reaction: Depending on the velocity of a Brownian parti-
cle d x/dt, this particle should receive the unbalanced number of collisions with
environment, e.g., solvent molecules, more from forward than from backward. The
total transfer of momentum per unit time is the restoring force d x/dt and the
random thermal force (t) with zero mean. We suppose that the law of action and
reaction holds always between the Brownian particle and its environment: when the
environment exerts a force d x/dt + (t) on the particle, the particle exerts the
reaction force, ( d x/dt + (t)), on the environment.
Concept of heat: Let us denote by d x(t) the evolution of x over a time interval
dt, according to the Langevin equation. It is determined for each realization. The
work done by the particle on the environment upon this change d x(t) is equal
to the product of this variation and the above mentioned reaction force, that is,
( d x/dt + (t)) d x(t). We adopt the Stratonovich-type product in the above.
(Justification is given below.) This work can be either positive or negative. When
positive, it represents energy lost by the system. From the standpoint of the environ-
ment, energy ( d x/dt + (t)) d x(t) is lost as work to the Brownian particle. In
any case, ( d x/dt + (t))d x(t) is the energy transferred from the environment to
the system. We then define this energy transfer as heat [3]. Although the microscopic
motions in the thermal environment are not explicitly represented in the Langevin
equation, the law of action and reaction allows us to identify how much energy has
been transferred.
Sign convention and the formula of heat: As in macroscopic thermodynamics
(Chap. 2), we assign the positive sign for the energy received by the system. For
instance, when a positive amount of work is done on the system by the thermal
environment, we say that the system receives a positive amount of heat d Q > 0.
The equation defining d Q then reads
dx
d Q + (t) d x(t). (4.3)
dt
When inertia is taken into account, we replace d x/dt above by p/m. We shall use
the same sign convention for the work done by the external system on the system
(see below). In (4.3) and hereafter, we add d dash (d ) to the heat Q or the work
W when it concerns the process during an infinitesimal time, dt. We distinguish
this from d and reserve the latter for the differential.5 In general d Q is not
differentials.
5 That is, the differentials of a monovalent function, like d x(t), or the total differentials of multi-
valued function, like dU (x, a).
4.1 Framework 139
dp dp p p2
d x(t) = dt = d ,
dt dt m 2m
U U
d x(t) = dU (x(t), a(t)) da.
x a
We recall that Stratonovich calculus allows us to use the usual rules of calculus about
the differentials. After substituting these two expressions in the last expression of
d Q, we obtain
p2 U
d + U (x, a) = d Q + da. (4.4)
2m a
p2
E + U (x, a) (4.5)
2m
U
dW da. (4.6)
a
Then (4.4) is expressed in the form shown in Fig. 4.2:
d E = d Q + d W. (4.7)
Equation (4.7) expresses the balance of energy concerning the system. It is anal-
ogous to the first law of thermodynamics; both are based on the principle of
energy conservation in nonrelativistic system.6 We would stress that (4.7) holds
for each realization of the stochastic process. A purely mechanical energy balance,
d E = d W, will hold in case that a can be changed suddenly, that is, very fast
with respect to the characteristic relaxation time of x, but slow enough so as not to
violate the time resolution of the Langevin model.
Energy balance without inertia: The definition of heat (4.3) and the Langevin equa-
tion (4.2) leads to the following result:
U
dU (x, a) = d Q + da. (4.8)
a
The calculation is very similar to and even simpler than the previous case. The
energy balance relation is
Compared to (4.7), the kinetic energy term is missing from the complete differen-
tial form. However, if the temperature T depends on the variable x, the evolution
equation (1.94) in Sect. 1.3.2.1 leads to the general form of energy balance (4.7).
The above is the basis of stochastic energetics, i.e., the energetics of a single
realization of stochastic process associated with the Langevin equation. The assign-
ments of E, d Q, and d W are all consistent with the law of mechanics and the usual
definition of work. And the law of energy balance is not postulated but is derived
from the Langevin equation.
The choice of Stratonivich-type calculus for d Q (4.3) was crucially important
for the derivation of the law of energy balance. In the overdamped case, the temporal
integration of (4.3) requires some care. The separate integral of d x/dt d x(t) or
of (t) d x(t) does not assure finite results.
6 Can we conceive the similar problem about a relativistic system? We should remember that the
thermal environment imposes a particular reference frame to define d x/dt = 0, i.e., the Langevin
equation does not support the Galilean invariance.
4.1 Framework 141
4.1.2.2 Remarks
General
1. Energetics about a single realization is more detailed than energetics based on
the FokkerPlanck equation. The latter describes an ensemble statistics at each
instant of time. For the purpose of studying the energy exchange associated with
a particular fluctuation event the latter method is not adequate. For example,
an exceptionally large heat absorption should be associated to a rare event of
climbing over an energy barrier. This characteristics will be masked if we used
the ensemble statistics because such rare fluctuations in the ensemble are not
synchronized in time (See also Sect. 4.2.2 below).
We know, however, that the FokkerPlanck equation contains the same informa-
tion as the Langevin equation. There is no contradiction between this statement
and the previous one. The above mentioned difference concerns their solutions,
i.e., the difference between the instantaneous probability distributions and the
single realization of stochastic process.7
2. In some literatures after [3] the heat for the overdamped system was defined
using the following identity:
U U
dU (x, a) = dx + da. (4.10)
x a
p U U
d E( p, x, a) = dp + dx + da
m x a
dp U U
= + dx + da. (4.11)
dt x a
U
The last expression gives the definition of heat such as d Q = dp dt
+ x
d x,
with the quantities in [ ] being the negative of inertial and potential forces. These
are mathematically correct as an identity. Einstein treated heat in the above men-
tioned way when he formulated statistical mechanics.8 Also this way will be
adopted for the energetics of the discrete Langevin equation (Sect. 4.1.2.6).
The definition of heat (4.3) is, however, more general for the system with inho-
mogeneous temperature T (x) [4]: The overdamped Langevin equation, inter-
preted in the Stratonovich sense, is (see (1.94))
7 Roughly speaking, the correspondence between ( P(x, t)/t)dt and {d E, d W , d Q} is far
from 1 to 1, while the mapping {d x(t), da(t)} {d E, d W, d Q} in the overdamped case is almost
bijection.
8 Y. Oono pointed this out to the author.
142 4 Concept of Heat on Mesoscopic Scales
dx kB T (x)
= U (x, a) + + (t), (4.12)
dt x 2
where the random thermal force, (t), is multiplicative. In this case, the definition
of heat, (4.3), leads to energy balance (4.7), where the energy E is redefined
as E kB T (x)/2 + U (x, a), not as (4.5). Also the definition based on the
law of action reaction, (4.3), provides a clearer view of heat when stochastic
energetics is generalized for the fluctuating hydrodynamics (Sect. A.4.7.1) and
for the suspension of hard spheres (Sect. A.4.7.2).
About Heat
1. Long before the introduction of the concept of heat [3], an attempt to relate the
Langevin equation to the thermal physics had been made [5]. They derived the
following expression about the energy, d E = d [ p 2 /2m + U (x, a)]:
2 p2 kB T p
dE = dt + d W + 2 kB T d Bt . (4.13)
m 2m 2 m
Here before d Bt denotes the product of Ito type. See Appendix A.4.1 for
the derivation. Equation (4.13) is mathematically equivalent to (4.7), but the
nonanticipating term ( d Bt ) was explicitly sorted out. As mentioned above,
Ito-type calculus is preadapted to the ensemble average, and the first term on
the right-hand side has an allusion that the Langevin dynamics generates only
the relaxation processes to equilibrium, like the equipartition. By contrast, (4.7)
emphasizes balance of energy for individual processes.
2. Can heat defined in the above be derived by the projection method, the method
which transformed the microscopic Hamiltonian dynamics to the Langevin equa-
tion (see Sect. 1.2.1.5)? There are simple cases where we can identify energies or
heat on two levels: In the model from which Zwanzig demonstrated the nonlinear
Langevin equation (Sect. 1.2.1.3), the potential energy U0 (X ) in the resulting
Langevin equation (1.41) was nothing but the proper energy of X in the starting
Hamiltonian dynamics, apart from an additive constant that reflects the elimi-
nated microscopic degrees of freedom. A similar situation will be realized when
a Brownian particle in a fluid moves under an optical trapping potential [6].
However, it is not the case in general. While the law of energy conservation
holds in both levels of description, what are meant by energy and by heat differ
generally from one level to the other.9 The full analysis of energy and heat at
different levels of description will be done in Chap. 6 [7]. Here we note only two
points:
(1) The potential energy Ueq ( A) in the Langevin equation (1.42) obtained by
projection method is in fact a (constrained) free energy (see (1.45)).
About Work
1. In Sect. 4.1.1.1 we prohibited the external system to control the interaction
between the system and the thermal environment. One reason is now clear from
the above formalism: our definition of work is based on the potential force.
Related but more fundamental reason is that the stochastic energetics is con-
scious about the scales of description: For example, let us consider the rota-
tional Brownian motion of a vane in a fluid medium (cf. Fig. 4.6). The friction
coefficient can be modified by changing the shape of the vanes.11 In order
to assess the work to change the friction coefficient for this motion, we need
structural and mechanical information about the interface between the system
and the thermal environment. Such a requirement conflicts with our assumption
about the thermal environment: the latter should be memory-free, nonstructured,
and uncorrelated. A solution for the above example would be to include some
region of the thermal environment near the interface as a part of the extended
system. See Sect. 7.1.1.5 for a concrete procedure of redefining the system.
2. The work d W has been defined such that it is 0 if the control parameter a is
constant. For example, suppose that a represents a constant external force g0 > 0
applied to a Brownian particle (position x(t)) in the direction of x > 0. Accord-
ing to our definition, the external system applying this force does not do work
while the particle drifts toward x(t) . We count (g0 d x) as a part of the
change of the internal energy of the system. Therefore, we regard that the particle
dissipates its potential energy g0 x. In contrast,
g if an external system increases
the force strength from g = 0 to g0 , then 0 0 (x(t)) dg(t) should be counted
as the work on the system, d W .12
4.1.2.3 Examples
Deformation of an Ideal Chain
Let us consider an ideal chain consisting of many microscale rigid rods joined
together by completely flexible joints (Fig. 4.3). The only interactions are steric
repulsion among rods. Therefore, the systems potential energy, U ({x i }), is constant
10 The impatient reader could give a quick glance at Sect. 6.2 (especially the Eq. (6.23)) of Chap. 6,
where F stands for U (x, a) here.
11 This example has been brought by T. S. Komatsu.
12 Jarzynski discussed this issue [8] in the context of Jarzynski nonequilibrium work relation
(Sect. 5.4.1).
144 4 Concept of Heat on Mesoscopic Scales
xN+1 = a
x0 = 0
Fig. 4.3 Ideal chain consisting of rigid rods (gray rectangles) and free joints (filled discs) immersed
in a thermal environment. One end, x N +1 a, can be externally controlled, while the other end,
x 0 , is fixed at x = 0
(i) How does a short-ranged bonding interaction (of some A range) cause a large-
scale conformational change of proteins (of 10-nm range)?
(ii) What supplies the energy for the proteins to deform prior to the release of the
bonding energy?
Ubond
Udeform
Fig. 4.4 Schematic presentation of the jump-and-catch transition (top) and its energetics (bot-
tom). A protein (shaded object) has an intramolecular binding pair (filled disc and filled crescent).
The height of rectangles shows the energies Ubond (upper law) and Udeform (lower law), respectively.
The process of jump (from left to right) requires to borrow energy from the thermal environment
conformation change, Udeform . In order to achieve the transition from the open
state (left) to the closed state (right), the thermal environment can transfer heat
d Q to the protein.13 The supplied energy is stocked as the potential energy for
deformation, dUdeform = d Q > 0. If bond formation is unsuccessful, the protein
conformation will return to the relaxed state, and the stocked energy is returned
to the thermal environment as heat. If bond formation occurs, the binding energy
of the bond dUbond > 0 is released to the thermal environment, while the
deformation energy dUbond is retained in the proteins. If the gain of the bond
energy |dUbond | overcomes the energy cost of deformation, dUdeform , we have
(dUdeform + dUbond ) < 0, and the transition is stabilized.
In the above, the thermal environment worked as a bank of (free) energy. It
allowed the protein to explore large deformations and find a short-ranged binding
pair which were originally far apart. While the average energy of fluctuations is
kB T 4pN nm per a degree of freedom, rare fluctuations may attain the heat
transfer of more than 10kB T . If the protein waits long enough time, e.g., ms, it
can attain large conformational fluctuations.14
Borrowing energy from the thermal environment is a characteristic feature of
the fluctuating world. This principle is fully used to extract work from the thermal
or chemical energy. Energetics of mesoscopic thermal and chemical engines are
discussed in Sects. 8.1.1 and 8.1.2. Stochastic energetics brings much more infor-
mations with respect to the ensemble theory of statistical mechanics.
(A) (B)
+
++ ++
x (t ) + ++
++ x(t)
+ + + + ++ ++ + + + + + +
Fig. 4.5 Schematic setups for measuring dU and d W . (a) Optically trapped bead [6] and (b) brass
wire-held pendulum [9]. The dashed lines are light for detecting x(t). The parallel arrows (a) and
rectangle (b) denote the optical+electrostatic and mechanical mechanisms, respectively, for the
restoring forces
15 For example, [6] had an error of only kB T /4 vs. about 10kB T of total energy variations.
16 The effect of inertia is too short lived ( 10 ns) to be captured.
4.1 Framework 147
dU (x) dx
0= + (t), (4.14)
dx dt
where (t) = 0 and (t) (t ) = 2 kB T (t t ). Our object is to solve this
equation by using discretization in time with a finite time step, h. We fix the total
time interval of integration. The total number of steps N is, therefore, N h1 .
The discretized solution with an arbitrary ( ) is said to converge to the true solu-
tion of (4.14) with the same ( ) if the difference between the discretized (rectilinear)
x(t) approaches indefinitely to the true x(t) in the limit of h 0. Let us take the
simplest convergent scheme of discretization:
dU (xt )
0= h (xt+h xt ) + wt,t+h , (4.15)
dx
where17 wt,t+h 2 kB T (Bt+h Bt ) with Bt being the Wiener process, and
xt xt + (1 )xt+h with being arbitrary for the moment except that it is nor-
mal to choose such that 0 1. In fact (4.15) is an integral of (4.14) from t to
t +h. There is, therefore, a value [0, 1] for which (4.15) is exact.18 We can show
1
that, with any choice of , the error in x(t) by this scheme is O(h 2 ). (See Appendix
A.4.2.) In other words, the difference between the approximate x(t) of (4.14) and
the true x(t) of (4.15) decreases as O(N 2 ) for N . The explicit and lowest
1
order Euler scheme uses = 0. Therefore, this scheme is convergent with the error
in x(t) being O(N 2 ).
1
1
Q (xt+h xt ) + wt,t+h (xt+h xt ). (4.16)
h
Similarly, the change in the potential energy should be expressed as
t+h
17 If we abuse the integral and derivative, wt,t+h = t (s)ds and (t) = d Bt /dt.
18 But such depends on x, ( ), and t.
148 4 Concept of Heat on Mesoscopic Scales
n
Q E i (a(t )) E i1 (a(t )) . (4.21)
=1
d
|t = (| j |t ) t , j (t).
dt j
4.1 Framework 149
Through a process, the energy of the system changes as E t (a(t)). Using the discrete
Langevin equation, the energetics defined above can be described simply.
We first express the energy E t (a(t)) using an energy operator, E(a), defined by
E(a) E i (a)|ii|. (4.22)
i
2 p 2 kB T p
= dt + 2 kB T d Bt (4.25)
m 2m 2 m
19 We remind that the vectors | j or j| are nominative, and their time derivatives are always 0,
even if the physical state j changes through the parameter a(t).
20 For the Langevin equation without inertia, the kinetic energies exchanged between the system
and the thermal environment are hidden, but should exist. That d Q is linear in the imbalance
150 4 Concept of Heat on Mesoscopic Scales
This simple formula gives an illusion of stability, but may mask what occurs in
the system. For example, suppose that the system starts from a metastable state of a
p2
potential U (x) as in Fig. 4.1. Until the system escapes from the metastable valley, 2m
remains kB2T , but the net release of the heat will occur from the escaping particle
beyond the potential barrier.
of the kinetic energy comes from the linearity of the Langevin equation in . The NavierStokes
equation of hydrodynamics is also the linear theory from this point of view.
21 A byproduct of (4.27) is the equilibrium relationship, when the parameter a is fixed: U 2 =
x
2 eq
kB T xU2 . This equality can also be obtained using the canonical equilibrium distribution
eq
eU/kB T .
4.2 Generalization 151
d Q E E
= Jx + Jp d X d P (with inertia effect),
dt X P
d Q U
= Jx d X (without inertia effect).
dt X
(4.30)
The expressions of the average heat (4.30) have long been found in the context of
FokkerPlanck equation [13].
The first equation of (4.30) includes, as special case, the result for the purely
mechanical system. In that case, E = H , (Jx , J p ) = ddtX P, ddtP P , and (X, P)
obey the Hamiltonian equation, ddtX , ddtP = ( H/ P, H/ X ). We then have
E
J + EP J p 0, that is, the heat transfer in the first equation of (4.30) vanishes
X x
identically.
4.2 Generalization
pi 2
E({xi , pi }, a) = + U ({xi }, a), (i = 1, . . . , n), (4.31)
i
2m i
where m i is the mass of the ith degrees of freedom, and a is external control param-
eter(s). The Langevin equation for the ith degree of freedom is
d xi pi
=
dt mi
dpi pi U
= i + 2i kB T i i (t), (4.32)
dt mi xi
where i are the friction constants associated with the coupling between the system
and the ith thermal environment of the temperature Ti .23 i (t) are white Gaussian
random noises with zero mean and i (t) j (t ) = i, j (t t ).
The heat from the ith thermal environment to the system, d Q i is defined as
pi
d Q i = + i (t) d xi , (4.33)
mi
while the work d W is defined as before,
U
dW = da. (4.34)
a
It is easy then to verify that the balance of energy:
d E = dW + d Qi , (4.35)
i
Following exactly the same procedure as we did in Sect. 4.1.3.1 for a single
thermal environment, we can derive the following formula of the average heat flux
from ith thermal environment to the system: (cf. (4.30))24 :
d Q i E E
= Ji,x + Ji, p d X d P (with inertia effect),
dt Xi Pi
d Q i U
= Ji,x d X (without inertia effect), (4.36)
dt Xi
where Ji,x mPii P and Ji, p U Xi
+
i m i P i kB T i Pi P are the ith flux com-
Pi
ponents of the FokkerPlanck equation associated with X i and Pi , in the case with
inertia, and also Ji,x 1i U Xi
P + k T
B i Xi
P is the ith probability flux associ-
ated with X i in the case without inertia. The derivation is given in Appendix A.4.4.
The interpretation of terms like XEi Ji,x + PEi Ji, p as an analog of d Q = (dp/dt+
U / x) d x(t) (see the paragraph above (4.4)) would be far fetched, because the
derivation of (4.36) uses integral by parts. The former, therefore, is not adapted to a
local interpretation.
23 If some internal degrees of freedom do not directly interact with the thermal environments,
we assign i = 0 for those degrees of freedom.
24 Here d X d X and d P d P .
i i i i
4.2 Generalization 153
About the average heat fluxes, a type of H-theorem has been derived [13]. It
concerns the Shannon entropy S of the system:
S P log Pd. (4.37)
d S 1 d Q i 1 i Pi P 2
= P + kB T i d Xd P
dt i
kB T i dt P i kB T i m i Pi
(underdamped),
d S 1 d Q i 1 i (Ji,x )2
= dX (overdamped).
dt i
kB T i dt P i kB T i
1 d Q i
0 (stationary state). (4.39)
i
kB T i dt
25 We use (P/t) log Pd = i Ji i logP, where i denotes the derivatives with respect
Pi U
to ith degree of freedom. We use also i mi Xi
X i Pi
Pd = 0, assuming that the
surface integral vanishes.
154 4 Concept of Heat on Mesoscopic Scales
Model
In Fig. 4.6 two thermal environments are at the temperatures T and T , respectively.
In each of these a vane is immersed. The coupling of these vanes to their thermal
environments is characterized by the friction coefficients and , respectively. The
two vanes are joined through a harmonic torsional spring with the elastic constant,
K . We will neglect the inertia effect for a moment. Let us denote by x and x the
rotation angles of the vanes in the environment of temperature T and T , respec-
tively. The Langevin equations for the vanes are
where (t) and (t) are two independent Gaussian white noisees with zero mean
and (t1 ) (t2 ) = 2 kB T (t1 t2 ) and (t1 ) (t2 ) = 2 kB T (t1 t2 ). We will
use the dot (x etc.) to mean the time derivative.
We define heat by d Q = ( x + ) d x and d Q = ( x + ) d x . The
balance of energy is in a time differential form,
d Q d Q d K 2 (x x )2 d K
+ = (x x ) . (4.41)
dt dt dt 2 2 dt
The first and second terms on the right-hand side of (4.41) shows d E/dt and
d W/dt, respectively.
We can also show the following relation:
d Q d Q
= K (x x ) ( (t) + (t)). (4.42)
dt dt
Moreover, we can derive the Langevin equation for the relative rotation angle,
x x , from the original coupled Langevin equation:
d 1 1 (t) (t)
= + K + . (4.43)
dt
These three equations give the heat currents ddtQ and ddtQ for each realization of
t t
( ( ), ( )). Since Q(t) 0 (d Q/dt)dt and Q (t) 0 (d Q /dt)dt are quadratic
functionals in the Gaussian stochastic processes, and , detailed statistics of Q(t)
and Q (t) are available through the generating function eQ(t) , etc. Using this fact,
for example, the theorem of heat fluctuation [14] has been tested [15]. This theorem
4.2 Generalization 155
is a development of the so-called fluctuation theorem (FT) [16, 17]. See also [18]
and a recent review to date [19]. We will not go into details of these statistics. Below
we will discuss the average and variance of heat currents.
Note: The original coupled Langevin equation is decoupled by using the relative
angle, , and the center of diffusion X ( x + x )/( + ). The latter obeys
a free rotative Brownian motion, ( + ) X = (t)+ (t).26 Therefore, the diffusion
of X does not contribute to the heat currents.
d Q d Q K (t) d $ %
+ = (x x )2 , (4.44)
dt dt 2 dt
d Q d Q
= K (t)(kB T kB T ). (4.45)
dt dt
In the steady state with K constant, the right-hand side of (4.44) vanishes. Then we
have the average heat currents:
d Q d Q K
= = (kB T kB T ). (4.46)
dt dt +
The above result describes the heat conduction mediated by a mechanical spring.
When T = T the average heat current vanishes. Also if or goes to infinity,
there is no conduction. The average heat current increases with K . It means that the
stiffer spring can transmit energy more efficiently from one thermal environment to
the other. According to (4.46) the heat current diverges in the limit K . This
apparently unphysical result is a sign of the abuse of the Langevin equation without
inertia: In the limit of high stiffness of the spring, the timescale ( + )/K
characterizing the variation of = (x x ) is beyond the time resolution of the
modeling by such Langevin equation. In order to avoid the spurious divergence, the
model has to take into account the effects of inertia, for example, due to the moment
of inertia of the vanes. In the limit of K = , appropriate Langevin equations are
as follows:
where I and I are the moments of inertia of the vanes, f (t) is the force of
action/reaction through the rigid spring, determined so that x(t) x (t) = 0 at
d Q d Q
= = (kB T kB T ). (4.49)
dt dt ( + )(I + I )
the cool environment T is indispensable. It is the hot environment that exerts the
random force on the vane, but the motion of the vane is not incoherent like Joules
spoon. To see this point clearer, let us analyze the behavior of x(t) of the system
(4.40) putting T = 0. Under this assumption we can solve the second equation of
(4.40) for x (t):
+
K K s
x (t) = e x(t s)ds. (4.50)
0
Substituting this result into the first equation of (4.40), we have the following
equation for x(t):
+
K K s
x(t) = K x(t) e x(t s)ds + (t). (4.51)
0
The integral on the right-hand side of (4.51), i.e., x (t) of (4.50), shows a feedback
control in the restoring force. As x follows adaptively to the motion of x in the
past, the restoring force, K (x(t) x (t)), is weakened as compared with a spring
with fixed end (x (t) =const.). As the result, the vane x(t) returns less energy to the
environment T than what was injected from this environment. A related model is
found in [20].
Based on (4.51) we can design a cooling device in the fluctuating world: we
somehow trap a Brownian particle (position x(t)) around the trapping center (the
position x (t)). If x(t) can be observed and if the feedback circuit can change x (t)
quickly enough to realize (4.50), then we can extract energy around the Brownian
particle as if the particle is in contact with another environment with T = 0.
Partition of Heat
Now we suppose that T = T for the two thermal environments while they have
no direct contact with each other. The object is to find how much heat the spring
absorbs from respective thermal environments when the spring constant K (t) is
slowly lowered. Analogous situation was mentioned in Sect. 2.2, see, Fig. 4.7.
There, the volume of the cylinder V (t) played a role of the inverse square root of
the spring constant, K (t)1/2 . If we pull up the piston quasistatically, what will be
the partitioning of the heat supplied from these environments?
Conventional thermodynamics did not answer to this question.29
The average heat currents are found from the relation (4.45) with T = T and
(4.44). The result is
d Q d Q K (t) d $ %
= =
(x x )2 . (4.52)
dt dt + 2 dt
This shows that whatever the change of K (t) is, the heat currents are partitioned
according to the inverse ratio of and . More heat is exchanged with the envi-
ronment of smaller value of . In particular, for the slow change of K (t), we may
have a concrete result: Using the equipartition law, K (t)(x x )2 /2 = kB T /2, we
integrate (4.52) to obtain the average cumulate heat from respective environments,
Q and Q :30
.
K init
Q = Q =
kB T ln (quasistatic), (4.53)
+ K fin
where K init and K fin are the initial and final values of K (t). If the spring is softened
indefinitely (K (t) +0), the external system extract arbitrarily large amount of
energy from the thermal environments, in the proportion of 1 vs. 1 .31
Diffusion of Heat
Even in the case of T = T and K = const., heat flows randomly between the two
thermal environments. The cumulated heat undergoes a Brownian motion around the
average steady change. For the case of K = const., this fluctuation is characterized
by the thermal diffusion constant D through the following relation:
,
t t 2 -
dQ d Q
dt dt = 2Dt. (4.54)
0 dt 0 dt
K kB T + kB T kB T + kB T
D= . (4.55)
+ + +
30 We integrate K (t) d kB T
2 dt K (t)
= kB T d
2 dt
ln 1
K (t)
.
31 The energy in the spring is kB T /2, independent of the stiffness.
32 The expression of D in the Japanese Version [21] should be corrected as that of (4.55).
4.2 Generalization 159
See Appendix A.4.5 for the derivation. This formula is invariant under the exchange
of T and T . At high temperature and/or for very slow change of K (t), the diffusion
of heat may mask the prediction about the partition of heat, (4.53).
We revisit the thermal ratchet motors (Sect. 1.3.4) and discuss several aspects of
energetics. Figure 4.8 recapitulates the Feynman pawl and ratchet and Buttiker and
Landauer ratchet.
where U (x, y) = U1 (x (y)) + U2 (y) f x, and (t) and (t) are mutu-
ally independent white Gaussian random noises with zero mean and (t) (t ) =
(t) (t ) = (t t ). For T = T this model can rotate on average the ratchet
wheel x in one direction. Here the load f is regarded as a part of the system, and its
T T
b U(x)
x
Fig. 4.8 (a) Feynman ratchet and pawl and (b) Buttiker and Landauer ratchet. (The same figures
as Figs. 1.9 and 1.10. See those legends.)
160 4 Concept of Heat on Mesoscopic Scales
f P=vf
1
0 0 0
v v v
Fig. 4.9 Schematic presentation of the load f , power P = v f , and the energy efficiency of
motors as function of its velocity v. The force at v = 0 is the stall force
potential energy ( f x) is counted in the systems potential energy. This system is,
therefore, autonomous with no external parameters.
After Feynman, there have been many variations of his model, and we will not
go into details of those model. We only summarize in Fig. 4.9 general qualitative
features of this model, in terms of the average working velocity v = x, load f ,
average power output P = v f , and the efficiency , i.e., the ratio of P to input
energy from high temperature bath (T) normalized by the Carnot maximum value,
|T |/T (see (2.24)).
We will discuss mainly two aspects of energetics: the stalled state where the
load is just strong enough that the motor moves neither forward nor backward, and
around the equilibrium state where the efficiency by linear nonequilibrium thermo-
dynamics becomes undetermined (of 0/0 type).
33 That is, the ratio of the available work to the supplied heat from the high-temperature (T )
thermal environment is max 1 T /T .
4.2 Generalization 161
In conclusion, there is a leakage of heat at the finite rate at the stalled state. The
efficiency of energy conversion then vanishes at the stalled state. Experience shows
that, if the parameters of the model are well tuned, efficiency can attain almost 1
(theoretical maximum) just off this stalled state while the efficiency at the stalled is
0. The efficiency has, therefore, a dip at this state. From practical viewpoint, if the
size of the system is increased, the effect of jiggling may become negligible, and
the dip in the efficiency may be too narrow to be observed. Discrete modeling may
also suppress or lessen this jiggling.
Jp T L pq f /T
= , (4.57)
Jq L pq T 2 T /T 2
34 Because of detailed balance at equilibrium, forward cycles and backward cycles occur
equally likely. Due to this cancellation this phenomena cannot be captured by macroscopic
phenomenology.
162 4 Concept of Heat on Mesoscopic Scales
step
L pq = qstep . (4.58)
step
This argument is confirmed numerically and is also justified from the calculation of
GreenKubo formula [28].35
the analysis showed the efficiency of energy conversion up to 136 [4, 31]. The
maximum is attained in the stall state.
If the inertia is taken into account, the model is (1.111), i.e.,
dp U (x) p dx p
= f + (t), = , (4.60)
dt x m dt m
35 Note that the potential energy of the system U (x, y) does not contribute to the quantum of the
transferred heat, qstep , except for the loads potential energy, f step . The remaining part of U (x, y)
returns to the original value, on the average, after a quantum cycle is completed.
36 i.e., the Carnot efficiency up to max = 1 T /T .
4.2 Generalization 163
where (t) is the white Gaussian noise with (t) (t ) = 2 kB T (x)(t t ). The
analysis of the latter model concluded that the efficiency should drop to 0 at the
stalled state [3234], in contradiction to the model without inertia.
This discrepancy implies that the limit of m 0 is not equivalent to the model
with m = 0. If we put m = 0 in the Langevin equation, the equation changes the
order of differential. Such a change may drastically affect the behavior of x(t), and
the energetics of Buttiker and Landauer ratchet is sensitive to this change.
Scale of Description
To better understand the cause of discrepancy and to know the true efficiency, we
first remember that we should respect the time resolution of each Langevin equation
(see, Sect. 1.3.2). In fact, both models abuse the Langevin equations beyond their
validity range. When a particle moves across the discontinuity of the temperature,
the environment for the particle changes within a infinitesimal time, shorter than
any finite time resolution that is associated with a Langevin model.
Next we note the role of the velocity relaxation time. This time is finite for the
model with inertia, p m/ , while it is assumed to be infinitesimal for the model
without inertia. We have a physical argument (see below) that, if we keep track of
this timescale p , the sudden jump of temperature is buffered and the model with
inertia gives a meaningful result. We should not take the overdamped limit first.
Physical argument
When a particle switches its thermal environment, a thermalization of the kinetic
energy of the particle occurs. On the average, the particles kinetic energy will
relax toward its (new) equipartition value. Through this process, an irreversible heat
exchange takes place with a new thermal environment [35, 32]. This is the cause of
heat leakage.
To discuss more quantitatively, suppose that a particle moves from the temper-
ature T to the temperature T . When the particle cross the border, its motion is
almost ballistic during the time p m/ . During this period, the particle keeps
its original velocity vth kB T /m. Then thermalization occurs, and the particle
exchanges energy, which is about the difference of the equipartition kinetic energy
|kB T kB T |/2. This energy does not depend on m.
In summary, as far as the particles kinetic energy is concerned, the temperature
border is smoothed. The effective temperature gradient is |T T |/(vth p )
m 1/2 .[36, 32]. The heat leakage through the thermalization of the kinetic energy
occurs with this temperature gradient. At the stalled state this heat leakage continues
at a finite rate. The efficiency is, therefore, 0 at the stalled state.
The conclusion is parallel to the Feynman pawl and ratchet wheel. But the mech-
anism of the leakage in the two models is very different. The maximum of the
efficiency is realized near but off the stalled state, and less than the Carnot limit.
164 4 Concept of Heat on Mesoscopic Scales
Kinetic Argument
Is there a remedy for the model without inertia to recover the above mentioned
leakage [33, 34]? Below is the argument against this possibility: for a Wiener pro-
cess, or a overdamped free Brownian motion (0 = d x/dt + ), there is no
characteristic timescale. The implication is that, once it visits a point, it revisits the
same point indefinitely many times within a short period. Therefore, the to-and-fro
of the particle at the temperature boundary can occur indefinitely many times, before
the bias field drives the particle off the boundary. It is, therefore, impossible to incor-
porate the thermalization of kinetic energy in the overdamped model. If the space is
discretized [37], this problem disappears apparently. But justification is needed for
the discretization.
Proof by Numerics
The above physical argument supporting the Langevin model with inertia was finally
verified by a careful molecular dynamic simulation [38]. Their analysis of energy
transfer confirmed the divergence m 1/2 of the heat leakage ( Q kin ) due to the
kinetic energy carried by the particle. See, Fig. 4.10. Their simulation used 2D hard-
core gas as thermal environments of two different temperatures. The gas is dense
enough that the Knudsen number (mean-free path/system size) is < 1. The Q kin
m 1/2 behavior holds well even for p = m/ as small as the microscopic time,
i.e., the inverse collision frequency of gas particles against the Brownian
particle.37
6 104
Underdamped Langevin
MD simulation
1/2
const (m/m0)
4 10 4
Q kin
Q kin . It diverges as
(m/m 0 )1/2 , where m and m 0 2 104
are, respectively, the mass of
the Brownian particle and the .
mass of the gas particle
constituting the model 0
0 10 20 30 40 50 60
thermal environment (Figure m/m0
by the courtesy of R. Kawai)
37 This result also justifies the Langevin equation with inertia down to the microscopic timescale.
The validity of this equation at the intermediate scale, 10102 p , will be scrutinized in
Sect. 6.3.1.
4.2 Generalization 165
In conclusion, the study of Buttiker and Landau ratchet gives a lesson that an
apparent reasonable result (Carnots limit) can be wrong, and apparently singular
result can be true.
dx j E tot
0 = + j (t) ( j = 1, . . . , N ), (4.61)
dt x j
E ({x },a )
x(t )
Fig. 4.11 Top: Open system and its particle environment c , which is outside of . Thick
curve is a trajectory of the particle which passes through the system. Bottom: The energy of the
system, E ({x},a) (thick curve and thick line segments). In the absence of particles, the energy of
open system is zero, E({ }, a) = 0
166 4 Concept of Heat on Mesoscopic Scales
with { j (t)} being mutually independent white Gaussian random forces with zero
mean and correlations j (t)k (t ) = 2 kB T j,k (t t ).38
However, we explicitly describe only those particles in the system . Once a
particle exits from the system, we lose track of it, i.e., we lose its degree of freedom.
In summary, the open system is a region, and the particles belong to this system
while they are in this region.
Therefore, the number of particles in the system, n(t), is also a random variable.39
If there are more than one species of particles, we will distinguish them and denote
x1() , etc.
38 The generalization to the case with inertia is straightforward, but we will not discuss here.
39 For those who know second quantization, this state space may be reminiscent of the Fock space.
In the Fock space, however, the identity of individual particles is lost because of the symmetriza-
tion or antisymmetrization of the product states of single particle. In the present state space, the
particles identity is preserved while particle remains in the system.
4.2 Generalization 167
parameter a, we require that U1 (x, a) depends on a only for x .40 The reason
is that we have excluded the direct interactions between the external system and the
systems environment. The one-particle energy U1 has an arbitrariness of an additive
constant.
If only one particle passes through the system , the systems energy will change
as shown in Fig. 4.11 (bottom). If there are N particles in the entire system, the
energy of the system, E, is represented as (see (4.62))
N
E= U1 (x j , a) (x j ), (4.63)
j=1
where we have introduced the (single-particle) characteristic function (x) for the
zone :
)
1 (x ),
(x) = 0 (x ). (4.64)
x x
Fig. 4.12 The situation where a particle in the environment (x ) is within the interaction distance
(illustrated by the dashed circle) of another particle in the system (x) across the system boundary
. We will then count the interaction energy as a part of energy of the system, E (see the text)
40 That is, U1 /a = 0 if x .
41 is considered to be a closed set.
168 4 Concept of Heat on Mesoscopic Scales
close to each other across the boundary of , and that interactions between them
are not negligible. There is no a priori reason to include or ignore this interaction
energy as a part of energy of the system, E. The same is true for the interaction
energy involving more than two particles. For macroscopic thermodynamics, the
premise of the extensive property of thermodynamic quantities justified to exclude
such ambiguity.42 In the study of mesoscopic systems, however, we cannot avoid
this boundary effect.
We will then take up the following definition for the energy of the system. If
a cluster of particles are interacting with each other, and if at least one of those
particles is in , we count this interaction energy in the energy of the system, E.
Leaving the details of analysis in Appendix A.4.6, the result is
N
N
N
E= U1 (xi , a)(1) (xi ) + U2 (x j , xk )(2) (x j , xk )
i=1 j=1 k= j+1
N
N
N
+ U3 (x j , xk , xl )(3) (x j , xk , xl ) + . . . , (4.65)
j=1 k= j+1 l=k+1
where U2 (x j , xk ) and U3 (x j , xk , xl ) are properly defined two and three body inter-
actions, etc., and (2) (x j , xk ), (3) (x j , xk , xl ), . . . , takes the value 1 if at least one of
their argument takes the value in and 0 otherwise. Because of the sharp boundary
of the system , the energy of the system E can vary discontinuously.
The above definition of the systems energy is not the unique choice. Nor this
choice is proven to be the best one. Apparently, this definition needs modifica-
tions when there are two open systems 1 and 2 which share a part of their
boundaries.
For later use, we write down also the total energy E tot of the entire system
c :
N
N
N
N
N
N
E tot = U1 (xi , a) + U2 (x j , xk ) + U3 (x j , xk , xl ) + . . . .
i=1 j=1 k= j+1 j=1 k= j+1 l=k+1
(4.66)
Remark : Necessity of steric repulsion for a single-particle binding site. On the
level of description by the Langevin equations, the hard-core repulsive interaction
between particles should be explicitly accounted for as a part of the term U2 , even if
the single-particle binding potential, U1 (x, a), affords only room for single particle.
Otherwise, more than one particles, e.g., {x1 , x2 }, can enter the same binding site
at the same time. As is the case with usual hard-core repulsion, the last interaction
has no direct contribution to the energy of the system: it only restricts the available
phase space of x1 , x2 , etc.
42 In all cases, we exclude the long-range interactions: we exclude unscreened electrostatic inter-
action and gravitational interaction.
4.2 Generalization 169
(in) dx j
dQ (x j ) + j (t) d x j . (4.67)
j
dt
Recall that E tot depend on a only through U1 (x, a) and the latter depends on a only
when x .
Because the particles can migrate during the time dt, and because the energy E
should be updated upon the migration of particles, we do not expect the relation
d E = d Q (in) + d W (wrong). In order to compare d E and d Q (in) + d W , we use,
in addition to the Langevin equation (4.61), the following identity, valid for any
functions f (x, a) and (x):
f (x, a) f (x, a)
d[ f (x, a) (x)] da + d x (x) + f (x, a)d(x). (4.70)
a x
We then find the law of energy balance for the open system [39]:
d E = d W + d Q (in) + d Q (mig) + d Q . (4.71)
Below, we will briefly describe the additional terms, d Q (mig) + d Q .43 We used
the notation Q for these terms because these energies are not directly controlled
by the external system.
d Q (mig) : Heat due to the migration of particles. This term accounts for the energy
caused by the migration of particles. In the identity (4.70), this energy comes from
the last term on the right-hand side. This energy is as follows:
N
N
N
d Q (mig) U1 (x j , a)d(1) (x j ) + U2 (x j , xk )d(2) (x j , xk ) + . . . . (4.72)
j=1 j=1 k= j+1
x
x x x
x
Fig. 4.13 The first (A1) and second (A2) terms on the right-hand side of (4.72) for d Q (mig) , and
the term (4.73) for d Q (B). These terms appear when a particle changes its position from the
gray disc to its adjacent black disc in each figure
Figure 4.13 (A1) and (A2) illustrates the first and second terms, respectively. In
( p)
(4.72), d (x j1 , . . . , x jn ) takes nonzero value when, during the infinitesimal time
( p)
dt, a particular p-particle cluster comes to participate in the system (d = 1) or
( p)
ceases to belong to the system (d = 1).
d Q : Heat due to the interaction with the particles just outside the boundary. As
illustrated in Fig. 4.13 (B), the systems energy changes also by the displacement of
those particles which are outside but interacts with a particle (or particles) inside
.For example, the terms in d Q attributed to the two-particle interaction are
N
N
N N (k= j)
(2) U2
U2 U2
(2) dx j + d xk (1) (x j ) dx j, (4.73)
j=1 k= j+1
x j xk j=1 k=1
x j
4.3 Discussion
44 For example, we do not say unconditionally such as an electron carries the Fermi energy, F ,
(=chemical potential of electrons) when it moves across a junction.
4.3 Discussion 171
framework is applied, i.e., the assignment of system and its environment, definition
of heat, and the relation of energy balance.
Hydrodynamic fluctuations have become more and more accessible through the
Brownian particle in a real fluid [40] or through a molecular dynamic simulation
of fluid jet [41]. The fluid dynamic equation with spontaneous thermal noise is
described by a Langevin equation for field. This subject is called fluctuating hydro-
dynamics and is developed by Landau and Lifshitz [42]. The energetic aspects
associated with these fluctuation phenomena can be formulated along the principle
described in Sect. 4.1. In Appendix A.4.7.1 we derive the formula of energy balance
for the fluctuation hydrodynamics, with several simplifying restrictions.
The stochastic motion of suspended hard spheres has been described by the
Langevin equation. In this Langevin equation, there is cross-coupling between the
force on a sphere and the velocities of the other spheres, due to hydrodynamic
interactions. In Appendix A.4.7.2we derive the formula of energy balance for the
suspension of hard spheres, based on the recent model equation by [43, 44].
Stochastic motion expressed by curvilinear coordinates, or the stochastic motion
on a curved manifolds (surface, curve, etc.) are described by Langevin equations
with multiplicative noises. That is, the amplitudes of thermal random force depend
on the variable of the equation. In Appendix A.4.7.3 we derive the formula of energy
balance for the Langevin equation on the manifold.
The Langevin equation has the microscopic basis of the projection technique
(Sect. 1.2). But in order that its derivation and the Markov approximation justify the
Langevin equation, the eliminated degrees of freedom, i.e., the thermal environment,
should behave not far from the equilibrium fluctuations.
A question is how far we can extend the use of the Langevin equation beyond
the equilibrium fluctuations. Can we study the transient or steady nonequilibrium
phenomena, especially their energetics?
There are many examples where stochastic energetics are applied to the nonequi-
librium phenomena, and their studies brought, in most cases, physically sound and
interesting insights about the nonequilibrium phenomena. Below is an incomplete
list of those studies:
Theoretical analysis of the energetics of the ratchet models. (There are many
papers and we refer only two reviews, [45, 46].)
Theoretical demonstration of the fluctuation theorem about the heat [47, 48]
Definition of pathwise entropy [48].
Numerical analysis of the Feynman pawl and ratchet [26, 49].
Experimental assessment of the energy balance.
172 4 Concept of Heat on Mesoscopic Scales
Still, we could mention the case where Langevin modeling is invalid under
nonequilibrium setup. Suppose that we measure the random force on a Brownian
particle while we apply a constant force onto it, e.g., in the positive x direction. In
this steady nonequilibrium state, the spatial symmetry of x is apparently broken.
We, therefore, expect a broken symmetry in the statistics of the random force (i.e.,
the skewness in the force distribution). However, the thermal random force of the
Langevin equation, (t), is always symmetric, by definition.
Therefore, we should be conscious about what type of nonequilibrium settings
we can study using the Langevin equation and its stochastic energetics. There is
no unique criterion for this point, partly because it depends on our exigence, partly
because our knowledge of nonequilibrium phenomena is incomplete. Still the reflec-
tion on the above successful examples and also on the cases of abuse suggests the
following (evident) thing:
This method works if the non-equilibrium is in the system, but it does not work
if the non-equilibrium is at the interface between the system and the environ-
ment.
In fact, the absence of skewness in the random force mentioned above is apparently
the latter case.
45 For the Langevin equation without inertia, both the velocity response and the velocity correla-
tion diverge in the short-time limit, |t t | 0. But the difference remains finite.
46 A drawback to this beautiful formula is that v (x) is not a local observable as function of x
s
and t.
References 173
4.3.3 Comments
Can we associate entropy with each trajectory?
Entropy and its production along a stochastic trajectory has been proposed to derive
the integral fluctuation theorem [48]. This somehow contradictory notion, entropy
vs. individual sample, is in fact well defined in the context of the path ensemble
average. In this book we limit to disuss those observables which are determined or
measured for an isolated realization.
References
1. Einstein, Investigation on the theory of the Brownian Movement, (original edition, 1926) ed.
(Dover Pub. Inc., New York, 1956), chap.V-1, pp. 6875 135
2. H.A. Kramers, Physica 7, 284 (1940) 135
3. K. Sekimoto, J. Phys. Soc. Jpn. 66, 1234 (1997) 138, 141, 142, 160
4. M. Matsuo, S. Sasa, Physica A 276, 188 (2000) 141, 162
5. H. Hasegawa, T. Nakagomi, J. Stat. Phys. 21, 191 (1979). 142
6. V. Blickle, T. Speck, L. Helden, U. Seifert, C. Bechinger, Phys. Rev. Lett. 96, 070603 (2006) 142, 146, 172
7. K. Sekimoto, Phys. Rev. E 76, 060103(R) (2007) 142
8. C. Jarzynski, C. R. Physique 8, 495 (2007) 143
9. S. Joubaud, N.B. Garnier, S. Ciliberto, J. Stat. Mech. 2007/09/, P09018 (2007) 146, 172
10. K. Schwab, E.A. Henriksen, J.M. Worlock, M.L. Roukes, Nature 404, 974 (2000) 146
11. D.R. Schmidt, R.J. Schoelkopf, A.N. Cleland, Phys. Rev. Lett. 93, 045901 (2004) 146
12. K. Kawasaki, in Phase Transitions and Critical Phenomena, Vol.2, ed. by C. Domb and M.S.
Green (Academic, New York, 1972) 149
13. H. Spohn, J.L. Lebowitz, Adv. Chem. Phy. 38, 109 (1978) 151, 153
14. C. Jarzynski, D.K. Wojcik, Phys. Rev. Lett. 92, 230602 (2004) 154
15. F. van Wijland, Phys. Rev. E 74, 063101 (2006) 154
16. D.J. Evans, E.G.D. Cohen, G.P. Morriss, Phys. Rev. Lett. 71, 2401 (1003) 155
17. G. Gallavotti, E.G.D. Cohen, Phys. Rev. Lett. 74, 2694 (1995) 155
18. J. Kurchan, J. Phys. A 31, 3719 (1998) 155
19. J. Kurchan, J. Stat. Mech. P07005 (2007) 155
20. H. Qian, J. Math. Chem. 27, 219 (2000) 157
21. K. Sekimoto, Stochastic Energetics (Iwanami Book Ltd., 2004, in Japanese) 158
22. R.P. Feynman, R.B. Leighton, M. Sands, The Feynman Lectures on Physics vol.1 (Addison
Wesley, Reading, Massachusetts, 1963), 46.146.9 160
174 4 Concept of Heat on Mesoscopic Scales
Where is the boundary between the system and the external system?
We do not describe the dynamics of the external system. It implies the hypothesis
that the external system is not influenced by the systems state and its dynamics.
In reality, however, it is very difficult to impose a fixed position of a rigid wall:
first of all, there does not exist microscopically rigid container, and the container is
also subject to thermal fluctuations. In order to take into account the fluctuation of
the wall surface, one may regard the materials composing the wall surface as a part
of a new enlarged system. Then the question is where we separate the system and
the external system and how we define a (see Sects. 5.2.3.3 below). To the authors
knowledge, no systematic argument is developed about the condition of the external
control parameter.
We deal with the cases where x represents a very few, typically a single degree of
freedom. We assume that stable control parameter(s) a can be found. Such simplifi-
cations allow us to concentrate on the principal things. We must, however, remember
all the above mentioned ambiguities when we consider a modeling of experimental
setups.
1 Microscopically, the interaction with the rigid wall is also due to the force field by the wall onto
the systems variable.
5.2 Work Under Infinitely Slow Variation of Parameters 177
tribution at a given temperature T and at the fixed value of a. For this system, those
equilibrium states specified by other values of a have not particular meaning.2 It is
the external system and its controllability against the system that define the process
from one equilibrium to the other one. It is the force U/a that tells which equi-
librium state is preferred by the system among the candidates parameterized by a.
Remarks
(1) We have excluded direct interactions between the external system and the ther-
mal environment.
Therefore, in U/a the potential energy U can always be replaced by the
total energy U tot of the system plus environments. When we apply the frame-
work to an experimental setup, the above point should be verified. For example,
if we apply an electric field to a protein motor, the influence of the field on the
surrounding water as thermal environment is not a priori counted in the original
formalism of stochastic energetics.
(2) Keeping the value of a constant is in general not equivalent to keeping constant
the force on the system, U (x, a)/ x.
The only exception is if U (x, a) can be written in the form of U0 (x)(a) x.
Otherwise, keeping U (x, a)/ x constant needs the adaptation of a(t). Such a
feedback control from x(t) to a(t) introduces a correlation between the protocol
of a(t) and the particular realization.3
(3) Force, U (x, a)/a, depends on the resolution of a.
For example, if we added small details u(a) = sin( a ) to U (x, a), the
resulting force would be changed by cos( a ).4
2 The system does not know what boundary conditions are variable and what others are not.
3 cf. the nanomachine to absorb the heat (Sects. 4.2.1.2).
4 cf. the coarse graining about x, Sects. 1.3.2.2.
178 5 Work on the Mesoscopic Systems
ai af
T T
Fig. 5.1 (Left) Brownian particle (thick dot) in thermal environment (temperature T ) is trapped in a
harmonic potential (thick curves). The external system changes the profile of this potential. (Right)
Trapping of Brownian particle under laser tweezer. The focusing controls the profile of trapping
potential
U (x, a) dx
+ + (t) = 0, (5.1)
x dt
where (t) is the Gaussian white random noise with zero mean and (t) (t ) =
2 kB T (t t ).
We will calculate the work done to change a from ai to af taking very long oper-
ation time, op . A similar model has been considered in Sects. 4.2.1.2. But here
we do without the ensemble averages over realizations. The work W to change the
parameter a(t) from ai to af is
a(op )=af
U (x(t), a) 1 a(op )=af
W = da(t) = x(t)2 da(t). (5.2)
a(0)=ai a 2 a(0)=ai
In the above, the value of W is evaluated with a particular realization of x(t) obeying
the Langevin equation (5.1). Therefore, W is a random variable. We will see that,
in the limit of op , the random variable W converges to a single value which
depends only on ai and af , not the protocol of a(t).
We formulate the limit of slow process as in Sects. 1.3.3.2. That is, first we define
a protocol a(s) that takes unit time, i.e., a(0) = ai and a(1) = af . And then we
expand this protocol to the time op by a(t) a( top ).5 Using this representation
of the protocol, the last integral of (5.2) is then written as
1
1 d a(s)
W = x(op s)2 ds
2 0 ds
s=1
1 d a(s) 1 a(s) x(op s)2
= ds d(op s). (5.3)
s=0 a(s) ds op ds 2
timescale of (5.1), i.e., /min(a). If op ds is much larger than this timescale, i.e.,
N op ds/( /min(a))
1, then x(op s) in the integrand experiences large num-
ber ( N ) of uncorrelated, or i.i.d., fluctuations. Then the underlined part of the
integral in (5.3) is the longtime average of ax(t)2 /2 with a being virtually fixed.6
By the footnote below (1.105), this temporal average converges to its canonical
average, kB T /2. Therefore, we have
s=1 /
kB T d a(s) af
W = kB T ln (op ). (5.4)
2 s=0 a(s) ai
For finite N , the typical error is estimated to be N 1/2 by the central limit theo-
rem (see Sects. 1.1.2.3).7 The result (5.4) is instructive for two reasons:
1. The result is definite. Although x(t) varies temporally and differs from one real-
ization to another, the work W takes asymptotically the same value.
2. The result corresponds to the statistical mechanical result. In the Gibbs statisti-
cal mechanics, the Helmholtz free energy F(a, ) ( 1/kB T ) of the present
system is
1
F(a, ) = kB T ln + (terms independent of a). (5.5)
a
The work W obtained above is equal to the difference of this free energy,
F(af , ) F(ai , ). This correspondence is what we expect for any quasistatic
thermodynamic process.
Gibbs statistical mechanics considers the ensemble of realizations. Any results
from that framework is, therefore, the statistical average over the ensemble. The
approach of the FokkerPlanck equation also yields results about the ensemble of
realizations. The above analysis, however, dealt with only a single realization. For
N the convergence is of probability 1 due to the property of self-averaging
for slow process. In this sense, the thermodynamic structure appeared due to the
(strong) law of large numbers (of realizations).8
s+ ds
6 (op ds)1 A(s, op s) d(op s) is approximated by (op ds)1 opops op A(s, T ) dT, and the last
expression is then approximated by the long time average about T . This approach is called the
method of multiple scale. See, for example, Chap. 6 of [1].
7 This argument owes to C. Jarzynski in the context of his analysis of ergodic adiabatic invariant
[2, 3].
8 Y. Oono, private communication.
180 5 Work on the Mesoscopic Systems
a(op )=af
U (x, a)
W = da(t). (5.6)
a(0)=ai a (x,a)=(x(t),a(t))
We will show that, in the limit of slow variation of the operation (op ), the
work W converges to the difference of the Helmholtz free energy, F, defined by
F F(af , ) F(ai , ), e F(a,) eU (X,a) d X. (5.7)
converges to the following integral over the parameter a in the limit of slow variation
of a(t):
af
I (, a)eq da (op ), (5.9)
ai
where (, a)eq is defined in (1.109), that is, (, a)eq (X, a)P eq
(X, a; T )d X, and P eq (X, a; T ) is the canonical probability distribution at kB T =
1 with a given value of a
eU (X,a)
P eq (X, a; T ) . (5.10)
eU (X ,a) d X
We show next that the limit in (5.9) is unique, independent of the protocol a(s)
between ai and af . Applying the above general formula to (X, a) = U (x, a)/a,
we have U (x, a)/aeq in the integrand. This can be rewritten by using the so-
called Ehrenfest formula:9
U (x, a) U (x, a) eq F(a, )
P (X, a; T ) d X = . (5.11)
a eq a a
W F (op ). (5.12)
9
We can verify (5.11) by differentiating the normalization condition of the canonical distribution,
(F(a,)U
e(F(a,)U (X,a)) d X = 1 a
e (X,a))
d X = 0.
5.2 Work Under Infinitely Slow Variation of Parameters 181
In the limit of slow variation of a(t), we can replace the integrand in the rightmost
of (5.13) by the canonical average, dE(a)/daeq , thus
af
dE(a)
W da (op ). (5.14)
ai da eq
where Ftot is the Helmholtz free energy of the entire system, defined by 12
Ftot /kB T 1
e = eEtot /kB T d Ntotx, (5.17)
Ntot !
where Ntot is the number of particles in the entire system, and the integral runs over
the entire system for each particle.
10 See Sects. 3.3.1.3 and 4.1.2.6. We wrote W instead of W for the consistency of notations
between (5.12) above and (5.15) below.
11 We use e(F(a,)E j (a)) = 1 e(F(a,)E j (a)) = 0.
j j a
12 For the facility of calculation we put the factor (Ntot !)1 and render Ftot extensive. See below.
182 5 Work on the Mesoscopic Systems
c c c
We need to relate Ftot to the thermodynamical quantity of open system. For this
purpose we take the volume of the environment, c , to infinity in Ftot , while
keeping only the part which is relevant to the open system. See in Fig. 5.2. The
result is
Ftot
f c (T, ) = lim .
c
c
Then we define J by subtracting from Ftot the (asymptotic) free energy of the envi-
ronment, c f c (T, ):
J (a, t, ) lim Ftot c f c (T, ) . (5.19)
c
W J (op ). (5.20)
eJ/kB T = e(F n)/kB T
(n)
, (5.21)
n=0
5.2 Work Under Infinitely Slow Variation of Parameters 183
where F (n) is the Helmholtz free energy of the open system when it contains n
particles;
1
eF /kB T
eEn /kB T d n x.
(n)
(5.22)
n! ()
The relations (5.16) and (5.18) are the key steps through which the chemical
potential (in J ) enters the energetics based on the Langevin equation.
(i) In the study of the thermodynamic processes in the classical regime, increas-
ing all the materials by, for example, twice is not a thermodynamic process. It
contradicts the conservation of massenergy. Therefore, the absolute value of
the thermodynamic functions are not observable, and its extensivity is merely
a convenient choice. However, the extensivity is imposed on the differences of
the thermodynamic observables.
(ii) In quantum mechanics, the individuality of identical particles is lost upon col-
lision due to the particlewave duality. But also in classical mechanics, the
individuality is lost when we describe the chemical reactions or the processes
of open systems in terms of the number of each molecular species.
Thus the factor of n! in (5.22) appeared without evoking quantum mechanics.
13 The duality asserts that the permutations among identical particles do not make new quantum
states.
184 5 Work on the Mesoscopic Systems
We define quasistatic process by the limiting process that realizes the convergence
of (5.12) or (5.15). In such processes, the work W does not depend on the protocol
defined by a(s) (0 < s < 1), whatever is the number of components of a.14
A quasistatic process is reversible or retractable in the sense that the process
that retraces the same pathway in the parameter space costs no work. The work F
to go is exactly compensated by the work (F) to return.
If the control process makes a closed loop and returns to the initial point, i.e.,
those a(s) satisfying a(0) = a(1), the quasistatic work is 0.15 Whether or not such
a process leaves any change after closure of the trajectory is a subtle question.
We will discuss it later (5.2.3.4). That the Langevin equation derived by using the
Markov approximation can realize the reversible process emphasizes the importance
of being conscious about the scale of description.
In macroscopic thermodynamics, the quasistatic process is characterized such
that at each instant of time the system realizes the equilibrium state under a given
constraints. However, the equilibrium state is defined as the state which is realized
in the system after infinitely long time under a given constraints. These two state-
ments are incompatible unless we define the limiting procedures unambiguously.
But macroscopic thermodynamics does not describe the temporal changes.
On the level of Langevin equation, one could consider the closeness to the
quasistatic process by comparing the probability density P(X, a(t), t) obtained
through the FokkerPlanck equation with the canonical equilibrium distribution,
P eq (X, a; T ), by using a suitable measure such as the KullbackLeibler distance,
D(P||P eq ) (see (1.81)). However, the general theory in the previous section gives
natural and operational criterion of the quasistatic process: We note that the differ-
ence W F in the continuous process is
a(op )=af
U (x(t), a) U (x, a)
W F = da(t). (5.23)
a(0)=ai a a=a(t) a eq
Thus we measure the approach to the quasistatic process by the effect of the replace-
ment of force by its instantaneous equilibrium expectation value in the integral
(5.23). That is
U (x(t), a) U (x, a)
(5.24)
a a=a(t) a eq a=a(t)
or
dE(a) dE(a)
,
da t (5.25)
a=a(t) da eq a=a(t)
in the integral by a(t). This is a much more specific criterion than the comparison
of the probability densities. The quantity U (x(t),a)
a U (x,a)
a will
a=a(t) eq a=a(t)
appear again in the context of asymptotic estimation of the error, W F for
the nonquasistatic process (Sects. 5.3.1).
Z (ai )
W = F = kB T log (quasistatic).
Z (af )
As Z (a) is decreasing function of |a| (i.e., the chain is less flexible for large |a|), the
external system does a positive work W to stretch the chain. According to the law of
energy balance, d E = d W + d Q, this work is immediately released to the thermal
environment (d Q = d W ), because the energy of the ideal chain is constant.
dx p dp Upis UT
= , = , (5.26)
dt m dt x x
where m is the mass of the molecule, and Upis (x, xpis ) stands for the interaction
energy between the molecule and the piston tip. xpis is the position of the piston tip.
We will define UT (x, xT ) below.
Systemthermal environment interface: We introduce a thermal wall (left vertical
wall of the chamber). This is mechanically coupled to the thermal environment. We
assume the overdamped Langevin equation for the position xT of the thermal wall:
dxT UT
0 = + T (t) , (5.27)
dt xT
where T (t) is the white Gaussian random noise with zero mean and T (t)T (t ) =
2 kB T (t t ). UT = UT (x, xT ) represents the interaction energy between the
Bath
a
XT X Xpis
Fig. 5.3 Schematic setup of piston and cylinder system for a single particle. The particle (thick dot:
position x) is confined within a volume (central rectangle) enclosed by (1) a thermal wall (left ver-
tical wall: position xT ) which is linked to the thermal environment (shaded rectangle bath) and
is supported by a fixed point (spring to the left of the thermal wall), (2) cylinder walls (upper and
lower horizontal walls), and (3) piston tip (right vertical wall: position xpis ), which is connected
to the controlled point (open circle: position a) through a coupling potential (spring between xpis
and a)
5.2 Work Under Infinitely Slow Variation of Parameters 187
molecule and the thermal wall as well as the supporting potential energy of the
thermal wall.
Systemexternal system interface: We assume that the piston is just a microscopic
tip. This tip is connected through a spring to the macroscopic apparatus (the right-
most open circle at a). The position xpis and momentum ppis of the piston tip obey
the Newton equation:
where m pis is the mass of the piston tip. Uel = Uel (xpis , a) is the internal energy of
the piston and depends only on xpis a.The energies, UT (x, xT ) and Upis (x, xpis ),
are expected to behave like sharp repulsive walls, blowing up as x xT or xpis x
decrease to 0, respectively.
Balance of energy and quasistatic work: The balance of energy is found to be
( 2
+
p2 ppis
d + + UT + Upis + Uel = d Q + d W, (5.29)
2m 2m pis
where
d xT Uel
d Q + T (t) d x T , dW da. (5.30)
dt a
According to the general theory of Sects. 5.2.1.2, the quasistatic work for the dis-
placement of the macroscopic apparatus a is the change of the Helmholtz free
energy, F(a, ), where
F(a,)
e = C() e[UT (x,xT )+Upis (x,xpis )+Uel (xpis ,a)] d x d x T d x pis . (5.31)
The pressure P thus defined is the time-averaged force that the macroscopic appara-
tus receives at a. Using the law W = F for the quasistatic process, we can identify
this pressure P with the thermodynamic pressure:
F(a, )
P= . (5.33)
a
188 5 Work on the Mesoscopic Systems
xint
Volume: It is not evident how a can be related to the volume of the chamber.
Unless the repulsive walls in UT (x, xT ) and Upis (x, xpis ) are infinitely steep (i.e.,
the rigid wall), the volume cannot be unambiguously defined. The ambiguity due
to the finite gradient of the potential energies, UT / x and Upis / x, is important
when the systems spatial extent, xpis xT , is small. a xpis depends on a as well
as temperature. Therefore, we cannot identify da with dxpis . That is, we cannot
replace dW = Pda by dW = Pdxpis .
In summary, controlling displacement and controlling force for small systems are
not equivalent. In statistical mechanics, the saddle-point method or DarwinFowler
method assured that these two are equivalent. (See a related discussion in [5].)
Figure 5.3 was a toy model. But a somehow similar setup is used to measure the
interaction between a single head of myosin (protein motor) and an actin filament,
see Fig. 5.4. The motorfilament interaction occurs at xint while the AFM apparatus
controls the position a. The motor head is bound to the latter by a needle at xndl . If
the position of xndl is optically measured, the result reflects both the motorfilament
interaction and the thermal fluctuations of the motor and of the needle. An optical
technique has been developed to suppress the thermal fluctuation of measuring
devices (down to 5K!) [6]. The energetics of the feed-back-controlled system is
discussed in [7].
Fig. 5.5 Profile of potential U (x, a) vs. x between the left-particle environment (L) to the right
one (R). Thick arrow indicates the progress of the potential profile. (Figure adapted from Fig. 2 of
[8])
(i) Even in the quasistatic process, we cannot always approximate P(x, t) of the
FokkerPlanck equation by the equilibrium
Peq (x, a; T ). If we did it,
density,
the probability flux J [P] 1 Ux
P + k P
B x is always 0, and we would
T
17
not have transport.
(ii) In the quasistatic limit, work has a potential function, i.e., the free energy
F(a, ). However, the probability flux J is not necessary the gradient of a
potential function.
Remarks:
1. It is essential that the potential profile undergoes a cyclic change. If the profile
change is a simple go-and-back along the same pathway, there is no net transport.
The leftright asymmetry of the potential profile is a necessary but not sufficient
condition.
2. Experimental demonstrations of such work-free transport must face the fluctu-
ating part of the transport: in taking the quasistatic limit, the fluctuation in the
number of transported particles per cycle will diverge as tcyc with the time
spent for a cycle, tcyc . The prefactor of tcyc can be decreased by raising the
energy scales of the potential U . However, it then makes the condition for the
quasistatic process more and more stringent, or it requires larger cycle time, tcyc .
More discussion will be given in Chap. 7.
3. The work-free quasistatic transport can be realized also in conventional thermo-
dynamics by, for example, using the (macroscopic) Carnot cycle. In the formula
of the reversible efficiency, rev = (Th Tl )/Th (see Sect. 2.3.3), the case Th = Tl
17 This is general remark when we use linear nonequilibrium thermodynamics. The flux of energy
or mass is caused by their small spatial gradients across the local regions in which the equilibrium
is assumed.
190 5 Work on the Mesoscopic Systems
assures the transport of heat between the two thermal environments of identical
temperature without work rev = 0.
4. The work-free transport discussed above does not contradict the second law of
thermodynamics: the heat or particles transported between the environments of
the same temperature or chemical potential cannot be the source of later work.
Wirr W F. (5.34)
See Fig. 5.6. For continuous process described by Langevin equation, it is written
as (see (5.12))
af
U (x(t), a(t)) U (x, a(t))
Wirr = da(t). (5.35)
ai a a eq
For finite op , the irreversible work Wirr is a random variable, whose value varies
from one realization to the other. In this section and next section we deal with the
average of Wirr over the ensemble of paths, Wirr .
a
0 ai af
Fig. 5.6 Cumulated work (solid curves) along the change of parameter, ai af ai . The
difference from the quasistatic work (dashed curves) gives the irreversible work, Wirr
5.3 Work Under Very Slow Variation of Parameters 191
1 (eq)
P (x) g(x, x ; a) = (x x ). (5.40)
x x
The first term on the right-hand side is positive and a functional of the rescaled
protocol, a(s). We can, therefore, define the lower bound of this integral as S(ai , af ):
P (x, a). Since diffusion is a purely relaxing phenomena, the spectra of the above operator
are all negative except of a single 0, corresponding to the constant eigenfunction. As the inverse
operator of this diffusion operator, the Green function g, is a symmetric function with respect to x
and x . Moreover, the spectra of g are the inverse of the diffusion operator and, hence, all real and
negative. From the last fact, (a) is positive definite.
192 5 Work on the Mesoscopic Systems
1
d a d a
S(ai , af ) min (a) ds , (5.42)
a 0 ds ds
where the minimum is sought for with all the continuous rescaled protocols a(s)
under the conditions, a(0) = ai and a(1) = af . Since S(ai , af ) does not depend on
the protocol between these end points, this is what we should have in (5.36).
From the standpoint of the external system (Ext), the quasistatic work W is appar-
ently stored in the system as the increment of the potential energy, F (Sect. 2.2).
There is a parallelism: the external system Ext does not see the degrees of freedom
of the system on the one hand, and the system (Sys) does not see those degrees of
freedom in the thermal environment on the other hand.23
21 cf. The standard deviation of Wirr decreases with op as O((op )1/2 ). See Sects. 5.2.1.1.
22 If (a) has anisotropy, the orientation of the route as well as its location should be optimized.
23 Below is an example of how the world looks differently from different viewpoints: study of
the fluctuations of cell motility is an activity measurement for biologists but passive measure-
5.3 Work Under Very Slow Variation of Parameters 193
The study of the irreversible work Wirr now provides a further parallelism: (a)
for the Ext corresponds to (friction coefficient) for Sys. In fact, if we combine
the results of Sects. 5.2 and 5.3, the average work W done through a slow change
of the parameter a is
op
da da
W = F + (a) dt
dt dt
af 0
F da
= da(t) + (a) , (5.43)
ai a dt
with an error of O((op )1 ). The second line of (5.43) allows the interpretation
that Ext applies the force, aF + (a) da
dt
, onto Sys. See Fig. 5.7. By the law
of actionreaction, the external system receives the potential force, aF , and the
friction force, (a) da
dt
. Therefore, (a) is the friction constant for the parameter
a(t).
To take into account the deviation of W from W , we introduce a noise term
(t) such that
af
F da
W = da(t) + (a) (t) , (5.44)
ai a dt
af
where the noise term behaves as24 ai (t)da(t) O((op )1/2 ). We can rewrite
(5.44) as
af
da
F = (a) + (t) da(t) + W, (5.45)
ai dt
The last expression is similar to the law of energy balance for the Langevin equation,
d E = d Q + d W . Schematically, the parallelism is
U(x,a)
system
bath
a
x
Fig. 5.7 (Left) A system and its control by an external parameter a. (Right) The system viewed
from the external system as a black box
ment for rheologists. Study of the response of cell against external perturbations is a passive
measurement for biologists and active measurement for rheologists.
24 cf. Footnote 21.
194 5 Work on the Mesoscopic Systems
Sys Ext
x a
U (x) F(a, )+?
(a)
random thermal force: system noise:
time resolution op
In the above schema, we have not considered the dynamical evolution of the param-
eter a(t). Therefore, ? in the above is not specified. On the last row, the time
resolution of the Langevin equation should be larger than the baths relaxation time
on the one hand, and op should be larger than the systems relaxation time on the
other hand. If the time resolution op is too large, there is no noise . A question
is whether there is a smooth limit from stochastic energetics to macroscopic ther-
modynamics when the size of the system goes to infinity. The parallelism across
different scales can also be found between the system and its subsystem. This issue
is addressed in the next chapter.
Note: Throughout this section we have assumed that the temperature is constant. The
process including the time-dependent temperature, T (t), has been studied 0 in Matsuo
(1999, unpublished paper). The author showed the Clausius inequality, d Q/T
0, using stochastic energetics and statistical entropy.
First it has been demonstrated for a thermally isolated system to which work is
added mechanically. Later the relation turned out to be valid more generally. We
show a brief demonstration using a form of the FeynmanKac formula.25 Those
who are not interested in the mathematical details may skip to Sects. 5.4.1.2.
25 The description of this section is based on [11] and the series of lectures by C. Jarzynski at
Institut Henri Poincare (Oct. 2007).
5.4 Work Under the Change of Parameter at Arbitrary Rates 195
where
P[X |x0 , t0 ] is the probability for the path X over the time interval [t0 , t], and
(x0 ,t0 ) . . D X denotes the path integral with the initial condition, (x 0 , t0 ). We define
.
the generating operator L(at ) of K (x, t|x0 , t0 ) through
K (x, t|x0 , t0 )
= L(at )K (x, t|x0 , t0 ). (5.48)
t
where P[X |x0 , t0 ] is always the path probability for the path X governed by L(at ).
wt can depend on x and on a. G(x, t|x0 , t0 ) satisfies G(x, t0 |x0 , t0 ) = (x x0 ).
From (5.49) G(x, t + dt|x , t) can be written as
To go to the second line (5.48) has been used. This formula will be used later.
FeynmanKac Formula
A simple version of Feynman and Kac tells26 that G(x, t|x0 , t0 ) obeys
G
= (L(at ) + wt )G. (5.51)
t
That is, G(x, t|x0 , t0 ) is generated by L(at ) + t .
26 In physicists language, the general FeynmanKac formula gives the (Feynmans) path integral
representation of the solution of an SDE of Ito type. In the path integral, the action in the expo-
nential is the sum of the kinetic part x 2 and the potential part U (x, a). To apply to the SDE
of Stratonovich type, the action should be corrected by 2 U/ x 2 .
196 5 Work on the Mesoscopic Systems
Substituting (5.50) into (5.52) and keeping up to the order of O(dt), we have
G(x, t + dt|x0 , t0 ) (1 + wt dt + dtL(at ))G(x, t|x0 , t0 ). This means (5.51).28 The
G(x, t|x0 , t0 ) defined above is, therefore, the Greens function of (5.51).
Evolution of eHt
We apply the FeynmanKac formula to the case where a Hamiltonian Ht
H(x, at ) satisfies L(at )eHt = 0 for each t. We then define wt by
H(x, at ) dat
wt . (5.53)
at dt
From wt , the work done to the system by an external system during the interval
[t0 , t] is given by
t
Wt,t0 = wt dt . (5.54)
t0
27 The ChapmanKolmogorov equation means the following. The totality of the paths from (x0 , t0 )
to (x, t) is given as the sum of those paths that pass through a gate at x g at a fixed time tg , then
summed over all x g . For the Markov process the probability weight for the paths from (x0 , t0 ) to
(x, t) via (x g , tg ) can be factorized into those weights of each segments.
28 If (5.51) is a FokkerPlanck equation, then (5.49) gives its formal explicit solution using
P[X |x0 , t0 ]. If (5.51) is a Schrodinger equation, (5.49) again gives its formal explicit solution
in the same manner [12].
5.4 Work Under the Change of Parameter at Arbitrary Rates 197
e Ft0 eHt = P[X |x0 , t0 ]eWt,t0 D X e(Ft0 Ht0 ) d x 0 . (5.56)
(x0 ,t0 )
Here the average is taken over the paths starting from canonical equilibrium. This
relation works very well for small systems [13, 14]. With increasing number of
degrees of freedom, this method requires a lot of data for a good statistics. The
reason is that very rare events for W is dominantly important in the average because
of its exponential dependence, eW [15].
The precision of canonical initial condition in the above is important: a
counterexample has been demonstrated for the microcanonical initial condition
[16]. We come back to the implication of this example later (see, the end of
Sect. 7.1.3).
The inequality (5.60) includes the case of quasistatic process, where Wirr eq = 0.
Recall that in Sects. 5.2.1.2 we obtained a stronger statement, Wirr = 0, for an
individual quasistatic process.
Remark. There is a different definition of the average irreversible work, which is not
directly related to the work measurement [17]. Using the probability density P(x, t),
we define the statistical entropy, S P ln P d x, and then the quasi-free energy,
F U T S, where U = U Pd x. Then the following inequality is proven:
d W d F J [P]2
= d x 0, (5.61)
dt dt P
where J [P] is the probability current of the FokkerPlanck equation. The formula
(5.61) has essentially the same content as the H-theorem (4.38) in Chap. 4.
The so-called (a version of) fluctuation theorem (FT) or Crooks relation for the
irreversible work, Wirr , is [18, 19].
PR (Wirr )
= eWirr , (5.63)
PR (Wirr )
where PR (Wirr ) is the probability density for Wirr . Integration of eWirr PR (Wirr ) gives
the average, eWirr , while the integration of PR (Wirr ) gives unity. Therefore, the
Jarzynski nonequilibrium work relation for the discrete process is derived:
$ %
eWirr eq
= 1. (5.64)
Finally the nonnegativity of the average irreversible work, Wirr eq 0, is derived.
(some
n can be redundant) is always found above nthis graph. Especially this center of mass,
n n
( n1 i=1 z i , n1 i=1 f (z i )), is vertically above ( n1 i=1 z i , f ( n1 i=1 z i )).
5.5 Discussion 199
5.5 Discussion
dx
+ (t) K [x a(t)] = 0. (5.65)
dt
The general solution for x(t) of (5.65) is x(t) = a(t) + eK t/ (ai a(t)) +
t K s/
(K / ) 0 e (a(t s) a(t))ds. We compare the two protocols: (i) a smooth
linear protocol a(t) = ai + Va t with Va constant and (ii) stepwise protocol with the
a(t ) a(t )
ai ai
t t
Fig. 5.8 (Left) Smooth linear protocol a(t) = ai + Va t and (Right) stepwise protocol a(t) =
ai + a[Va t/a]
same average rate, a(t) = ai + a[Va t/a], where [z] denotes the integer part of
z( 0). See Fig. 5.8. Using the above solution, the average work per unit time is
(i) Va 2 and (ii) W = Va 2 (K t/2) coth( K t/(2 )) with t = a/Va . The second
W converges smoothly to the first in the limit of the fine steps t 0.
References
1. A.H. Nayfeh, Perturbation Methods (Wiley-VCH, New York, 2004) 179
2. C. Jarzynski, Phys. Rev. A 46, 7498 (1992) 179
3. C. Jarzynski, Phys. Rev. Lett. 71, 839 (1993) 179
4. T. Shibata, K. Sekimoto, J. Phys. Soc. Jpn. 69, 2455 (2000) 182
5. T. Hondou, Europhys. Lett. 80, 50001 (2007) 188
6. T. Aoki, M. Hiroshima, K. Kitamura, M. Tokunaga, T. Yanagida, Ultramicroscopy 70, 45
(1997) 188
7. H. Qian, J. Math. Chem. 27, 219 (2000) 188
8. J.M.R. Parrondo, Phys. Rev. E 57, 7297 (1998) 188, 189
9. K. Sekimoto, S.I. Sasa, J. Phys. Soc. Jpn. 66, 3326 (1997) 190
10. C. Jarzynski, Phys. Rev. Lett. 78, 2690 (1997) 194, 197
11. G. Hummer, A. Szabo, PNAS 98, 3658 (2001) 194
12. R.P. Feynman, A. Hibbs, Quantum Mechanics and Path Integrals (McGraw-Hill Inc. U.S.,
1965) 196
13. F. Ritort, C. Bustamante, I. Tinoco Jr., PNAS 99, 13544 (2002) 197
14. D. Collin et al., Nature 437, 231 (2005) 197
15. G.E. Crooks, C. Jarzynski, Phys. Rev. E 75, 021116 (2007) 197
16. K. Sato, J. Phys. Soc. Jpn. 71, 1065 (2002) 197
References 201
17. H. Spohn, J.L. Lebowitz, Adv. Chem. Phy. 38, 109 (1978) 198
18. J.L. Lebowitz, H. Spohn, J. Stat. Phys. 95, 333 (1999) 198, 199
19. G.E. Crooks, Phys. Rev. E 61, 2361 (2000) 198
20. D.J. Evans, E.G.D. Cohen, G.P. Morriss, Phys. Rev. Lett. 71, 2401 (1003) 199
21. D.J. Evans, D.J. Searles, Phys. Rev. E 50, 1645 (1994) 199
22. G. Gallavotti, E.G.D. Cohen, Phys. Rev. Lett. 74, 2694 (1995) 199
23. J. Kurchan, J. Phys. A 31, 3719 (1998) 199
24. G.E. Crooks, Phys. Rev. E 60, 2721 (1999) 199
25. C. Maes, J. Stat. Phys. 95, 367 (1999) 199
26. G. Bochkov, Y. Kuzovlev, Sov. Phys. JETP 45, 125 (1977) 199
27. G. Bochkov, Y. Kuzovlev, Sov. Phys. JETP 49, 543 (1979) 199
28. G. Bochkov, Y. Kuzovlev, Physica A 106, 443 (1981) 199
29. G. Bochkov, Y. Kuzovlev, Physica A 106, 480 (1981) 199
30. C. Jarzynski, C. R. Physique 8, 495 (2007) 199
31. K. Kawasaki, J.D. Gunton, Phys. Rev. A 8, 2048 (1973) 199
32. Y.B. Band, O. Kafri, P. Salamon, J. Appl. Phys. 53, 8 (1982) 199
33. T. Schmiedl, U. Seifert, Phys. Rev. Lett. 98, 108301 (2007) 199
Chapter 6
Heat Viewed at Different Scales
This short chapter deals with the heat from different levels of description. We shall
call the heat defined by the stochastic energetics mesoscopic heat and the heat
defined by conventional statistical thermodynamics calorimetric heat. The latter
should measured by conventional calorimetric methods. The main question of this
chapter is to understand the connection between these two types. Simple thought
experiments will help clarify the problem. Work against a viscous medium can be
defined with the aid of the notion of transformation of the heat from one scale to
another scale.
Sekimoto, K.: Heat Viewed at Different Scales. Lect. Notes Phys. 799, 203220 (2010)
DOI 10.1007/978-3-642-05411-2 6
c Springer-Verlag Berlin Heidelberg 2010
204 6 Heat Viewed at Different Scales
was coarse grained to eliminate the periodic potential, Um (x), which has a small
period, . The resulting Langevin equation retaining only the slow modulation force,
a0 V (x), could be written as (1.102), that is,
Since generally we have the inequalities, (0) = , the heat Q eff thus defined is
different from the heat Q that we define for the original Langevin equation, d Q =
1
{ x + [2 kB T ] 2 (t)} d xm (t).
xN+1= a
x0 = 0
Fig. 6.1 A chain of beads (thick dots except at the ends) and springs (shaded rectangles). Each
spring can be a purely mechanical spring, or it can be a purely entropic ideal chain. [The same
figure as Fig. 4.3.]
except at the ends x 0 and x N +1 . Each bead is supposed to be a free joint for the
two neighboring springs. We fix one of the end at the origin, x 0 0, while we
control the other end, x N +1 a. The overdamped Langevin equations for the ith
bead (1 i N ) write as
where the random thermal forces, i (t), are i.i.d. Gaussian white noises with zero
mean and i (t) i (t ) = 2 kB T 1(t t ).1
The mesoscopic heat from the heat bath is
N
Q= [ x i + i (t)] d x i (t). (6.5)
i=1
(U/ a)d a.
According to the result of Chap. 5, the quasistatic work W to pull the chain from
a = ai to af is F F(af ) F(ai ), where the Helmholtz free energy F(a) is
defined by
K N
3N /2 K
i=0 (x i+1 x i )
2
N
2 a2
e F(a,)
= e 2 d xi = e 2(N + 1) . (6.6)
i=1
K
1The symbol 1 is the 3D unit tensor. In (6.4) we adopted the model of Rouse type [2], where the
hydrodynamic effects ([3, 4], see also Sect. A.4.7.2), are ignored for simplicity.
2 In order to avoid confusion with Ito-type product, we suppress the dot for the 3D scalar
product.
206 6 Heat Viewed at Different Scales
Therefore,
af
K
W = F = a2 (quasistatic process). (6.7)
2(N + 1) ai
This result is valid independent of the origin of K . As for calorimetric heat, however,
we need to consider the origin of the spring before comparing d Q defined above
with the calorimetric measurement. It is enough to consider the quasistatic process
to show the essential point:
Case 1: If the springs are made of thin stainless steel wire, the spring constant
K can be regarded to be independent of the ambient temperature. Then the
change of the average energy U can be identical to the prediction from
the average energy Um of equilibrium statistical mechanics:
( F) 3N kB T K
Um = = + a2 . (6.8)
2 2(N + 1)
By comparing the Um from this result with (6.7), we find that Um =
U = F, which then implies that Q is equal to the calorimetric heat,
Q m , and Q m = Q = 0 for the quasistatic process. This result could be
confirmed by local calorimetric measurements or by numerical calculation.
Case 2: If each spring consists of an ideal polymer chain immersed in a fluid
environment, then the calorimetric results will be different. Each spring is
purely entropic with spring constant, K = knbB T2 , where n is the number
of monomers in the chain, and b is the so-called Kuhn length. The second
equality of (6.8) is no more valid since K depends on the temperature. From
(6.6) with K = const., the product F is found to be independent of tem-
perature. Then (6.8) yields U = 0 independent of a. If we denote by Q m
the calorimetric heat expected from the calorimetric energy conservation, we
have Q m = W = F.
The relations between the mesoscopic heat and the calorimetric heat, therefore,
depend on whether or not the spring constant K is of entropic origin. The argument,
including the nonquasistatic case, will be developed in the next section.
Example 4. Stochastic thermodynamics of chemical reaction networks
For chemical reaction networks an alternative interpretation of work and heat has
recently been proposed [5]. While it shares the basic idea of describing the ener-
getics of individual stochastic process, their interpretation is different from our
6.1 * Introduction What Is Heat? 207
framework in Sect. 4.1.2.6.3 For illustration purpose, we take a very simple situation
as example.
A system is a vesicle which occupies the volume , and is subset of the entire
volume (0) . The vesicle is surrounded by the reservoir of particles of species A.
The reservoir, occupying the volume (0) \ , is big enough that it is characterized
by a chemical potential, (0)
A [6]. The particle A can pass through the vesicle mem-
brane. Using the notations of chemical reaction, A X means that a particle A
enters the vesicle from the reservoir, while X A means the reverse process. The
state of the system is specified by the number of X particles, n X (t). We denote by
n A (t) the number of A particles in the reservoir. Therefore, n Xdt(t) + n Adt(t) = 0. The
energy eX is assigned to each particle inside the vesicle. The change of energy, E,
during t = 0 and t = t is therefore
t n X (t)
dn X (t)
E = eX dt = eX dn X (t). (6.9)
0 dt n X (0)
The heat Q chem is then defined so that the energy balance, E = Wchem + Q chem ,
be satisfied4 :
n X (t) n A (t)
Q chem = eX dn X (t) + (0)
A dn A (t). (6.11)
n X (0) n A (0)
Problems
Through the previous examples, we have seen different definitions of heat, depend-
ing on the scale of description, the assignment of energy, or identification of the
state variables of the system. One might say that there are different semantic
3 The authors of [5] use the word stochastic thermodynamics since they also introduce entropy
for each point on each trajectory.
4 [5] adopted the opposite-sign convention for the heat from the present text.
208 6 Heat Viewed at Different Scales
The energetics of the stochastic process has shown the balance of energy on the
mesoscopic level. The heat and the energy defined are, however, generally different
from their macroscopic counterparts. In this section we show that this discrepancy
can be removed by adding to these quantities the reversible heat associated with the
mesoscopic free energy.
6.2.1.1 Background
Calorimetric heat is defined by its measurement. Usual measurement monitors the
temperature of a probe material (thermometer). Calorimetric heat is then deduced
from this temperature data, using various physical principles, such as heat diffusion,
radiation, etc. It is the microscopic degrees of freedom of the probe material that
equilibrate with the sample. Therefore, calorimetric heat is calculated from micro-
scopic theories or simulations based on statistical thermodynamics.
Mesoscopic heat depends on the description level of the stochastic phenomena.
Throughout this section, we use the notations with the tilde ( F, f , Q, etc.) to mean
those concepts of stochastic energetics.
If the mesoscopic energy comes from interactions with an external field
(e.g., laser tweezers [7]) or with a nonentropic restoring force (e.g., a brass wire
holding a pendulum [8]), then mesoscopic heat can be identified with the calorimet-
ric (i.e., microscopic) heat. Contrastingly, they are different when mesoscopic heat
contains the entropic contribution due to microscopic degrees of freedom which
have been projected out to achieve the mesoscopic description. We have seen some
examples in Sect. 6.1.2. Below we present another thought experiment analogous to
the jump-and-catch mechanism in (Sect. 4.1.2.3).
F (x, a, )
0 x a
0 x (t) a x
(m)
(a) (b) U (x a)
Fig. 6.2 (a) A bead (position: x, a big gray disk) is leashed by an ideal polymer chain (thin curve)
at the origin 0 (thick dot). A laser tweezer to trap the bead is focused around a distant point a. (b)
The mesoscopic potential energy F(x, a, ) in (6.18) [solid curve] and the microscopic potential
energy that accounts for the calorimetric measurement, (6.25) [dashed curve]
polymer is an ideal chain. We further assume that the arrangement is such that the
bead undergoes bistable transitions, either being trapped around x = a when the
polymer chain is stretched to the distance |a|, or wandering around x = 0 when
the chain is relaxed and fluctuating. We then assume that the stochastic behavior
of the bead is well described by a Langevin equation for the bead position, x(t).
The main question is how much heat is released to or absorbed from the thermal
environment when the bead switches from one of the bistable states to the other.
The bistable states can be represented by a double-well mesoscopic potential
energy F(x, a, ) for the bead. See Fig. 6.2(b). However, it is only the optical trap
that realizes a potential hole of microscopic energy; the ideal chain exerts purely
entropic restoring forces. (Recall that the kinetic energy of ideal chain is inde-
pendent of the conformation of the chain.) Microscopic theory should predict that
calorimetric heat depends only on the microscopic potential energy of the optical
trap. The framework of stochastic energetics predicts, however, that mesoscopic
heat is absorbed from the environment whenever the bead climbs up the potential
barrier of F(x, a, ) and is released to the environment during the downhill motion.
5 We could start from an entirely isolated whole system, except that the argument is more compli-
cated.
210 6 Heat Viewed at Different Scales
where the suffices x, y of Trx,y indicate the degrees of freedom over which the
trace should be taken. We also introduce F(x, a, ), sometimes called Landau free
energy, by eliminating only the degree(s) of freedom, y:
Then, from (6.12), (6.13), and (6.14), we can verify the relations:
6 [9] studied two-component (fast and slow) Brownian system, using the FokkerPlanck equation.
6.2 * Calorimetric Heat vs. the Heat of Stochastic Energetics 211
f (x, a, ) = Tr y [e(FH ) f ],
f (a, ) = Trx [e(F F) f ] = Trx,y [e(FH ) f ]. (6.15)
The second quantity with the objectivity is the energy (not the mesoscopic energy).
In addition to H (x, y, a), we define
E(x, a, ) = Tr y [e(FH ) H ],
E(a, ) = Trx [e(F F) E] = Trx,y [e(FH ) H ]. (6.16)
F
E F = T . (6.17)
T
dx F(x, a, )
0= + (t) , (6.18)
dt x
212 6 Heat Viewed at Different Scales
where is the friction constant for x, and (t) is the white Gaussian random force
with zero mean and the correlation, (t) (t ) = 2 kB T (t t ). From the stand-
point of x, the fast fluctuations of the heat and energy due to the change in y are
averaged over the time resolution of (6.18). As in the static case summarized above,
it is the mesoscopic energy, F(x, a, ), that gives the bias for the variable x. The
mesoscopic energy balance along a particular realization of the stochastic process is
d F = d W + d Q, (6.19)
where we use d (not d ) to mean the total differential at constant temperature, i.e.,
d dx + da , (6.20)
x a
F
d W da, (6.21)
a
dx F
d Q + (t) d x = d x. (6.22)
dt x
The core logic is the following: If the Langevin description (6.18) is a good
model of a phenomenon, then the eliminated degree(s) of freedom y are supposed to
follow x and a rapidly enough. From standpoint of y, the process of x is always qua-
sistatic. It means that the heat released can be captured by the change in the pertinent
entropy, F/ T [10]. (A related argument is also found in [11].) In order to con-
vert the mesoscopic heat d Q into the calorimetric heat, d Q m , it is, therefore, suffi-
cient to add to both d Q and d F the differential of the term found in (6.17), that is8
F
d Q d Q m d Q T d ,
T
F
d F d E d F T d . (6.23)
T
Now the mesoscopic energy balance equation (6.19) is converted to the new equa-
tion that includes only calorimetric heat and the quantities with objectivity:
d E = d W + d Q m . (6.24)
The last expression holds for a particular realization of the Langevin equation (6.18).
Equation (6.24) could be directly verified experimentally or calculated using the
original Hamiltonian H . By the definition of the total differential (6.20), the term
T d( F/ T ) in (6.23) has no cumulative effects for cyclic processes.
dU (m) = d W + d Q m
= [U (m) (x a)d a] + [U (m) (x a)d x], (6.25)
Further Considerations
The change of F(x, a, ) through the change of x is a quasistatic work for the fast
degrees of freedom, y. The ideal chain should, therefore, release heat even though
the displacement of the bead x(t) is spontaneous. This statement looks somewhat
paradoxical. But it does not contradict with the above analysis; it is the thermal
environment that does the work to displace the bead, gathering the energy nearby.9
The released heat T d S (p) is, therefore, compensated.10 If one can measure the
heat at a very short distance, the local transfer of calorimetric heat around the chain
and the bead should be observed (Fig. 6.3(a)). By contrast, if there is no bead at
x(t), there is no such local transfer of heat (Fig. 6.3(b)). This should be checked
experimentally.
9 To move a mesoscopic object, there should be the fluctuations, most probably, of the length scale
of the object. See Sect. 6.3.1 below.
10 Another heuristic argument could be to assume thermophoresis of the bead due to local warm-
ing up of solvent around the chain.
214 6 Heat Viewed at Different Scales
0 x 0 x
(a) (b)
Fig. 6.3 Ideal chains with a bead and without bead (b). One end of the chains is fixed (0: filled
square). The slow displacement of the bead (gray disc) in (a) accompanies the heat transfer
between the chain and the neighborhood of the bead (thick bidirectional arrow), while the dis-
placement of the free end (x) in (b) does not accompany the heat transfer
discrete stochastic process. We first adapt the notations of Eqs. (4.18), (4.19), (4.20),
and (4.21) as follows to conform with the discrete case:
F = W + Q, (6.26)
with
These relations correspond to (6.19) in the continuum case. In order to have the
counterpart of (6.24) for the discrete process, we can again use the correspondence
relations (6.23). As the result, the energy balance relation,
E = W + Q m , (6.30)
holds with
F
E F T , (6.31)
T
F
Q m Q T , (6.32)
T
the moment, mesoscopic potential energy and heat have only begun to be assessed
experimentally [7, 8, 12, 13]. The direct measurement of the fluctuating observ-
able heat, d Q m , will be a future experimental challenge. The possibility to measure
directly d Q is an open theoretical problem. We might need to extend our notion of
the measurement so that the approach taken by the above references, i.e., the deduc-
tion of heat from the trajectory data is the mesoscopic measurement of the heat.
The random force on a Brownian particle in a fluid reflects the temporarily coher-
ent momentum transfer from the fluid molecules. Such spontaneous movements of
fluid can be approximately modeled by the Langevin equation for the fluid, or, fluc-
tuating hydrodynamics [14] (see also Appendix A.4.7.1). The evolution equation
of fluctuating hydrodynamics is the NavierStokes equation complemented by the
random dipolar source of momentum. The thermal motion of a Brownian particle is
the passive reaction to the fluctuations in the fluid [15].11 Although the fluctuating
dipolar force (i.e., stress) in the fluid is assumed to be a white Gaussian process, the
velocity of the Brownian particle has a long-term memory.
What causes memory in the velocity of Brownian particle? 12 Two mechanics
coexist. The first factor is the inertia of the Brownian particle (and the fluid which is
directly entrained by this particle). This effect is effectively included in the Langevin
equation as the inertia term. The decay time13 is characterized by p m/ , where
m is the mass of the particle and = 6 R is the Stokes friction constant with R
and being the radius of the particle and the viscosity of the fluid, respectively. If
11 The result of [14] has been also found before [24] cf. [25].
12 See also Sect. 1.1.3.1 Remark about the inertia effect.
13 By solving m v = v + (t), the evolution of the particle velocity v(t) contains the convolution
of the random force (t) with the memory kernel, (t)et/ p .
216 6 Heat Viewed at Different Scales
the mass density of the particle is p , the time p is 2R 2 p /(9). The second factor is
the inertial motion of fluid around the Brownian particle. The momentum carried by
a locally coherent motion of fluid can decay only by diffusion, since the momentum
conservation law prohibits its individual decay. The diffusion coefficient of the fluid
momentum is the kinetic viscosity, /f , where f is the mass density of the
fluid. The decay of the average velocity in this region is then characterized by the
time, f = R 2 / = R 2 f /.14 If the mass density of the Brownian particle p is
comparable to that of the fluid, f , we have p f .
Moreover, the fluids coherent momentum decays slowly (algebraic). This slow
decay reflects the fact that the spatial range of this coherent motion changes in time:
suppose that the fluid had an initial momentum P0 around the Brownian particle.
After the time t, this momentum spreads by diffusion over the range (t)1/2 . The
mean local velocity v of the fluid is inversely proportional to the mass of the fluid
within this range, i.e., v P0 /(f (t)d/2 ), where d is the spatial dimensionality.
This algebraic decay t d/2 is called the (hydrodynamic) longtime tail [16, 17].15
This memory effect, which is neglected by the Langevin equation, is important over
the timescales of f 102 f .16 Recent experiments to trace the Brownian particle
[18, 19] used a bead, for example, with R 0.5 m. In that case f = 0.25 s.
They have verified theoretical predictions [20] with the time resolution of s.
However, if we used a protein of R 10 nm, f would be 104 s. This is too
small to resolve experimentally at present.17 An evidence of the longtime tail is seen
in the mean square displacement (MSD) [x(t)x(0)]2 of the particle position x(t).
See Fig. 6.4 for the schematic behavior of the MSD. Due to the longtime tail effect,
the ideal diffusion behavior [x(t) x(0)]2 = 2Dt 18 has an algebraic correction
[2123, 15]:
0 t
Fig. 6.4 Schematic representation of MSD, [x(t)x(0)]2 , normalized by its diffusion form, 2Dt,
is plotted against time t. Due to the (hydrodynamic) longtime tail, the approach of the MSD toward
the latter value is algebraic, much slower than exponential
14 If we replaced the fluid with a gas in the Knudsen limit, there is no such memory [26, 27].
15 Rahman [16] found an anormaly in the fluctuation spectra in his molecular dynamic simulation.
Then Alder and Wainwright [17] showed the power low decay of velocity correlation function.
16 Even at 102 f coherent momentum is spread over only 10R.
17If a protein motor of several nanometer size wishes to make use of the fluctuations that last for
s, one strategy would be to link itself with a bigger object of m size. cf. Sect. 1.3.1.3.
18 We project on the x-axis.
6.3 Change in the Scale of Heat 217
/
[x(t) x(0)]2 f
=12 + (t > f ), (6.34)
2Dt t
A question has been raised about the efficiency of the molecular motor protein.
When the motor protein hydrolyses an ATP molecule and swings a long fila-
ment, should we count the dissipated heat due to the filament motion as a part
of work done by the motor? If yes, how can one estimate the performance of
this work? This problem is relevant for scale-dependent description of fluctuations
and heat.
F1ATPase is a protein complex that synthesizes ATP from ADP and Pi (inorganic
phosphate) using the rotation of its central axial rod ( -subunit) as the source of
work [28]. F1ATPase can also work inversely, that is, as a rotary motor consuming
the free energy of ATP hydrolysis [29]. To demonstrate this motor function, fluores-
cent actin filament was attached to the -subunit [30] (Fig. 6.5). In the presence
of ATP (of concentration < M), the filament was observed to rotate stepwise
against the viscous friction of the solvent. Each ATP hydrolysis amounts to the 2/3
rotation of the filament. An efficiency of (free) energy conversion, , was defined
as the ratio of the dissipation, 2 st , to the corresponding free-energy cost of
ATP hydrolysis, G hyd 20kB T . Here is a friction coefficient of the filament,
st is the stepping time of the 2/3-rotation of -subunit, and = 2/(3st ) is
the mean angular velocity. The result [30] showed that the efficiency is fairly
close to unity, suggesting indirectly that the motor may well work inversely as the
synthesizer.
Although apparently sound and useful, the above definition of efficiency requires
justification on a physical basis. In general, the square of an average quantity, like
Fig. 6.5 Schematic setup of the F1ATPase (central cluster of oval modules) as rotary motor, which
swings a long actin filament (long bar). The rotation of the actin filament is driven by the rotat-
ing axe ( -subunit: short vertical bar) of the F1ATPase. The spatial scale of the actin rotation
(dashed loop) is of several microns and is intermediate between the mesoscopic scale of the rotary
motor and the macroscopic scale
218 6 Heat Viewed at Different Scales
where the task is required to finish within a given time. The above definition explains
why the square of the average velocity 2 appears and can be useful for designing
an optimum engine under a given task.
Still, a question remains: If we regard as fluctuating variable of mesoscopic
scale, the least dissipating process, = const., is by no means realizable in the pres-
ence of thermal noise. Should we use such an idealized limit to define the efficiency?
Is there an alternative, hopefully scale conscious, interpretation for the 2 ?
There is an insightful analysis on this issue [33]. The argument was done in the
context of a thermal ratchet model (Sect. 1.3.4). This model is described by the
Langevin equation with inertia,
p p
p = (U0 (x, t) L x) + (t), x = , (6.36)
x m m
p p
dp = (U0 (x, t) L x)dt dt + dwt , dx = dt, (6.37)
x m m
U0 (x, t) 1 p2 p
dW = dt, d Q = dt + dwt , (6.38)
t S 2m m
19 After its sawtooth shape, this type of potential energy is named ratchet potential.
6.3 Change in the Scale of Heat 219
where S 2 m
. When the potential U0 (x, t) is externally changed in time, the model
can transport the particle x(t) against the load L. The crucial step is that the heat
d Q is split into the large-scale part, d Q L and the small-scale part, d Q S [33],
so that d Q = d Q L + d Q S , with
d Q L = v 2 dt,
m kB T dt p
d QS = [v(t)]
2
+ dwt , (6.39)
2 2 S m
20 We might recall that the eolian generator of electricity can operate under the winds even if the
weeklong average velocity of wind is 0.
220 6 Heat Viewed at Different Scales
References
1. G. Goupier, M. Saint Jean, C. Guthmann, Phys. Rev. E 73, 031112 (2006) 204
2. I.P. Rouse, J. Chem. Phys. 21, 1272 (1953) 205
3. B.H. Zimm, J. Chem. Phys. 24, 269 (1956) 205
4. V. Lisy, J. Tothova, A.V. Zatovsky, J. Chem. Phys. 121, 10699 205
5. T. Schmiedl, U. Seifert, J. Chem. Phys. 126, 044101 (2007) (2004) 206, 207
6. N.G. van Kampen, Stochastic Processes in Physics and Chemistry, Revised ed. (Elsevier Sci-
ence, 2001) 207
7. V. Blickle, T. Speck, L. Helden, U. Seifert, C. Bechinger, Phys. Rev. Lett. 96, 070603 (2006) 208, 215
8. S. Joubaud, N.B. Garnier, S. Ciliberto, Journal of Statistical Mechanics 2007/09/, P09018
(2007) 208, 215
9. A.E. Allahverdyan, T.M. Nieuwenhuizen, Phys. Rev. E 62, 845 (2000) 210
10. K. Sekimoto, Phys. Rev. E 76, 060103(R) (2007) 212
11. R.A. Blythe, Phys. Rev. Lett. 100, 010601 (2008) 212
12. S. Toyabe, T.N.H.-R. Jiang, Y. Murayama, M. Sano, Phys. Rev. E 75, 011122 (2007) 215
13. S. Toyabe, M. Sano, Phys. Rev. E 77, 041403 (2008) 215
14. L.D. Landau, E.M. Lifshitz, Fluid Mechanics (Course of Theoretical Physics, Volume 6), 2nd
edn. (Reed Educational and Professional Publishing Ltd, Oxford, 2000) 215
15. E.J. Hinch, J. Fluid Mech. 72, 499 (1975) 215, 216, 217
16. A. Rahman, Phys. Rev. 136, A405 (1964) 216
17. B.J. Alder, T.E. Wainwright, Phys. Rev. Lett. 18, 988 (1967) 216
18. B. Lukic et al., Phys. Rev. Lett. 95, 160601 (2005) 216
19. B. Lukic et al., Phys. Rev. E 76, 011112 (2007) 216
20. H.J.H. Clercx, P.P.J.M. Schram, Phys. Rev. A 46, 1942 (1992) 216
21. V. Vladimirsky, Y. Terletzky, Zh. Eksp. Teor. Fiz. 15, 259 (1945) 216
22. A.V. Zatovsky, Izv. Vuzov. Fizika 10, 13 (1969) 216
23. E.H. Hauge, A. Martin-Lof, J. Stat. Phys. 7, 259 (1973) 216
24. V. Vladimirsky, Y. Terletzky, Zhur. Eksp Teor. Fiz. 15, 259 (1945) 215
25. V. Lisy, J. Tothova, cond-mat/0410222 215
26. C. Van den Broeck, R. Kawai, P. Meurs, Phys. Rev. Lett. 93, 090601 (2004) 216
27. C. Van den Broeck, R. Kawai, Phys. Rev. Lett. 96, 210601 (2006) 216
28. H. Itoh et al., Nature 427, 465 (2004) 217
29. H. Noji, R. Yasuda, M. Yoshida, K. Kinosita Jr., Nature 386, 299 (1997) 217
30. R. Yasuda, H. Noji, K. Kinosita, Jr. M. Yoshida, Cell 93, 1117 (1998) 217
31. I. Derenyi, M. Bier, R.D. Astumian, Phys. Rev. Lett. 83, 903 (1999) 218
32. H. Wang, G. Oster, Europhys. Lett. 57, 134 (2002) 218
33. D. Suzuki, T. Munakata, Phys. Rev. E 68, 021906 (2003) 218, 219
Chapter 7
Control and Energetics
For a system external agents act as a main source of free energy, on one hand, and
also as agents that control the processes, on the other hand. For example, when
one drives a car, the car (system) needs two kinds of external agents; fuels (oil and
oxygen) and a driver. From the viewpoint of energetics, the work for controlling the
processes of a macroscopic system is usually negligible as compared with the main
work, such as the combustion of fuel to keep a car going. It is, therefore, reasonable
that standard textbooks of thermodynamics describe the Carnot heat engine without
mentioning the work of attaching or detaching the engine with the heat baths.
In fluctuating mesoscopic systems, however, the situation is different. The work
to control a fluctuating system can be an important part of the total work exchanged
between the system and its external agents. Ignorance of this type of work would
easily lead to paradoxes. If one were to invent a perpetual machine, and if one would
check the consistency of this machine with the second law of thermodynamics, it
would be better to make a mesoscopic model, because the energetics of control
appears naturally in the mesoscopic description.
There have been many studies on the work related to control:
Paradox of Maxwells demon [1]: the demon which makes use of thermal fluc-
tuations to realize a perpetual machine of the second kind (see Sect. 4.2.1.2).
Thermodynamics of computation [2, 3]: theories revealing the minimal irreversi-
ble work to operate a binary digit memory (Sect. 7.1.2 below).
Feynman ratchet and pawl [4]: a model of autonomous heat engine (see Sect.
1.3.4.2).
Motor proteins [5]: autonomous chemical engines of molecular scale, such as
linear or rotatory molecular motors or ion pumps. (See Chap. 8).
Signal transducing proteins [6]: G-proteins, etc., which share a universal molec-
ular architecture with the motor proteins [7].
One of the main questions about the energetics of control is Can any type of
operations to a system be done quasistatically? (Sect. 7.1), because we know that
quasistatic work is recoverable. Among the processes of control, there are certain
important cases where the process can never be done quasistatically by construc-
tion. We will call such processes essentially nonquasistatic processes. There are two
Sekimoto, K.: Control and Energetics. Lect. Notes Phys. 799, 223253 (2010)
DOI 10.1007/978-3-642-05411-2 7
c Springer-Verlag Berlin Heidelberg 2010
224 7 Control and Energetics
distinct mechanisms that disable a quasistatic process, and these two mechanisms
come into play often together:
In 1960s Landauer [2], Bennett [3], and others have elucidated the minimal irre-
versible work required for a cycle of operations on a single-bit memory. While the
operation of a single-bit memory includes essentially nonquasistatic processes of
(Case 1), this minimal irreducible work and the above mentioned reducible irre-
versible work should be distinguished (Sect. 7.1.2).
The control of free-energy transducers will be discussed separately in Chap. 8.
Another important question of control concerns detection under fluctuations
(Sect. 7.2). How can a mesoscopic system detect external signal particles with max-
imum certainty and minimum cost? While a gate correlates actively the objects
of control with the subject system, the sensor or detector correlates passively the
system with its surroundings. The two principles for avoiding the thermal noises
from the detection use, respectively, (i) the steric repulsion or (ii) the compensation
of interaction energies.
denote by sys this time. The ratio between these two timescales is sometimes called
the Deborah number, De sys /op .
(i) The relaxation time of the system, sys (a), depends on the system parameter, a.
(ii) During a characteristic time of operation, op , an external system changes the
value of a.
(iii) The change in a is such that the relaxation time sys (a) changes from the full
relaxation regime, sys (a) op , to the regime of freezing, sys (a)
op , or
the inverse. (Fig. 7.1 right)
sys(a)
sys(a)
log(time)
sys(a)
log(time)
ac a op
Fig. 7.1 (Left) sys (a) vs. a. The figure shows the case where sys (a) diverges for a ac . (Right)
Change of sys (a) across the operation time op through the change of parameter a. The time axes
should be regarded as logarithmic scale
U (x, a) dx
0= + (t) (x tot ). (7.1)
x dt
226 7 Control and Energetics
U(0, a)
x
L/2 L/2
tot
U (0, a)
x
/2 /2
Fig. 7.2 A Brownian particle (thick black dot) in a zone tot {x| L/2 x L/2} along the
x-axis. The curves show the profiles of the potential energy U (x, a) with ai a < a (top) and
U (x, a ) with a < a af (bottom) vs. x. The maximum value of the potential is U (0, a) kB T
and U (0, a )
kB T , respectively. The profile of the potential U (x, a) changes only in the zone,
{x| /2 x /2}
If the height of the potential barrier, U (0, a), is small, i.e., U (0, a)/kB T 1, the
time sys is approximately the diffusion time, 0 L 2 /(2D) (D = kB T / ), for a
Brownian particle to visit almost entire zone tot . By contrast, if U (0, a)/kB T
1
the relaxation time of the system, sys (a), is essentially the inverse rate of ther-
mally activated transition across the potential barrier. According to (3.40) the rate is
roughly given by
1 1 U (0,a)
e kB T , (7.2)
sys (a) 0
where we have ignored the corrections in the prefactor due, for example, to the
details of the potential profile.
Now suppose that U (0, ai ) kB T for the initial value ai and U (0, af )
kB T for
the final value af . Then sys (a) increases from sys (ai )/0 1 to sys (af )/0
1
during the operation time op . For example, if 0 = 109 s and op = 1 h, the
crossover point sys (a) op is attained for U (0, a) 30kB T . In the regime of
sys (a)
op the particle is practically localized within either side of the poten-
tial barrier. Through the crossover of the timescales, the accessible phase space of
the system shrinks almost discontinuously. Moreover, we cannot precisely predict
which subspace the particle will be confined to. The probabilities that the parti-
cle is confined to the left (x < 0) and that to the right (x > 0) of the barrier
are 1/2:1/2.
protocol of raising the potential barrier.2 We will denote by a (ai < a < af ) the
value of the parameter a at the crossover of timescales, i.e., sys (a ) op .3 For ai
a(t) < a the almost quasistatic operation is realizable while for a < a(t) af it
is impossible in any case because sys (a(t))
op .
The optimal strategy to minimize the irreversible work is that we spend most
of the operation time in raising the potential barrier up to U (0, a ) so that the
associated work is almost recoverable when the barrier will be lowered. During
this stage, the probability density P(x, t) for the particle position x is close to the
canonical distribution over the entire zone tot , P eq (x, a; T ) = e[F(a,T )U (x,a)]/kB T ,
where eF(a,T )/kB T tot eU (x,a)/kB T d x. The associated (almost quasistatic) work
W[ai ,a ] is (see (4.6))
The error in (7.3), i.e., the irreversible work associated with this slow process, is
approximately proportional to the inverse of the time spent, i.e., op 1 , as dis-
cussed in Sect. 5.3.1.
After this almost quasistatic work, we will raise the barrier height from U (0, a )
to U (0, af ) in a short time. The average work W [a ,af ] associated with this step
during a < a(t) af is estimated to be
Leaving the derivation to Appendix A.7.1, the result (7.4) can be qualitatively
explained: Since the height of the potential barrier U (0, a ) is already very large
with respect to kB T , the particle almost surely escapes from the region where the
barrier is rising. Therefore, in most cases virtually no work is needed to raise the
barrier. However, in rare cases, with the probability of (/L) eU (0,a )/kB T , a par-
ticle happens to remain within the zone and is lifted up. The last process costs
the work of the order of U (0, af )U (0, a ). Therefore, we have (7.4). The last work
is not recoverable because in the reverse operation from af to a , the probability to
find the particle within at a = af is extremely less than P eq (x, a ; T ).
Equation (7.4) tells that W [a ,af ] /kB T can be made very small due to the expo-
nential factor, eU (0,a )/kB T . If we take 0 = 1013 s and U (0, af )/kB T 103 , then
op = 1 ms is sufficient to satisfy W [a ,af ] /kB T < 107 . That is, we need the
irreversible work of less than 105 % of kB T and the time of 1 ms to establish the
barrier which a particle will not pass before 0 eU (0,a )/kB T > 10400 days. Moreover,
2 We discuss here about the irreversible work being directly related to a single action of raising the
barrier: This work should be distinguished from the irreversible work associated with the cyclic
operation to erase a single-bit memory. See Sect. 7.1.2.
3 Practically, it is better to include a safety factor to define sys (a )/op 10 102 rather than
sys (a )/op 1. We, however, ignore this point for the simplicity of the argument.
228 7 Control and Energetics
this 107 kB T of work is not typical work: for the most realizations, W is practically
0 ( 107 ), while for one realization out of e10 10430 the work is W 103 kB T .
3
dx p dp U (x, a) (x, y, )
= , = , (7.5)
dt m dt x x
y
( x,y, )
m
x
U(x,a)
dy (x, y, )
0 = + (t) . (7.6)
dt y
In (7.6) and in Fig 7.3 we assumed that y is a rotational degree of freedom (i.e.,
angle) so that its Brownian motion is confined within a periodic domain of period
2. We assume (x, y, ) such that (see Fig 7.4) for = 0 there is no energetic
coupling between x and y, i.e., (x, y, 0) = 0, and for = 1 the variation of y is
highly correlated with that of x. For 0 < < 1, change of (x, y, ) is supposed
to be monotonous with .
When the parameter is decreased down to the decoupling limit = 0, the
crossover of timescales is unavoidable with whatever large op .8 We will use the
clutch mechanism to model the Carnot heat engine (Sect. 8.1.1) and the total work
to operate parameters will be calculated in Sect. 8.1.1.2.
>>kBT
~ kBT
<<kBT
xy
0 2 4 6
Fig. 7.4 Potential function used for the clutch mechanism, (x, y, ) vs. x y, is shown for 0
(bottom curve), 0 < < 1 (middle curve), and = 1 (top curve)
7.1.1.6 Remarks
Generality of the Crossover of Timescales and Spatial Scales
Figure 7.2 represents a nave oversimplified picture of the glass transition, where
lowering temperature or changing other fields splits the equilibrium states into more
than one (meta) stable states.
The crossover of timescales has long been known in quantum chemistry. The
BornOppenheimer approximation assumes that the electrons state follows quickly
enough the movement of atoms or ions [8]. But when the two atoms are grad-
ually separated, the rate of the electronic transition between the atoms becomes
smaller and smaller with the increase of the interatomic distance. Therefore, the
timescale of the electronic transition diverges during this process. Eventually the
BornOppenheimer approximation becomes invalid.
The crossover of spatial scales has been studied in quantum mechanics. The
WentzelKramersBrillouin (WKB) approximation [9] assumes that the de Broglie
wavelength of the particle, 2/k, is much shorter than the length scale over which
this wavelength varies. The de Broglie wavelength depends on the total energy
E and the potential energy U (x) through k = 2m(E U (x)), where is the
Planck constant/2 and m is the mass of the particle. The turning point x is
defined by E = U (x ). At this point the de Broglie wavelength diverges. The WKB
approximation becomes, therefore, inevitably invalid as x approaches the turning
point.
Auto-Adjusting Timescales
In the above examples the timescales are controlled by external operations. When
the systems relaxation time sys depends on the systems state, sys can be adjusted
by itself in the vicinity of the timescale of operation op . A very simple model of
aging and plastic flow show that sys approaches to op from below and from above,
respectively. The models are described in Appendix A.7.2.
The basis of this subject was founded by Landauer [2] and Bennett [3] (see a survey
[1]). They have shown that the work of overwriting a memory is no less than the
work which Maxwells demon can extract. Therefore, no perpetual machine of
the second kind can be constructed on the basis of memory operations. Stochastic
energetics provides an explicit formulation of the minimal irreversible work along a
single realization of memory operation.
D
A
A D
B C
Asymmetry
B C
Fig. 7.5 A cycle of operation of overwriting, which consists of the erasure, E : A B, and the
subsequent writing, W : B C D A
232 7 Control and Energetics
dx U (x, a, b)
= + (t), (7.7)
dt x
where is the friction coefficient, (t) is the white Gaussian process of zero mean
with (t) (t ) = 2 kB T (t t ). The operation requires two parameters, a and b
[2]. The parameter a controls the asymmetry of the potential to bias the memory to
take a desired state, while the parameter b controls the barrier height of the potential.
U (x, a, b) = ax + b(x 2 1)2 is an example.
We split the x-axis into the left half 0 = (, 0) and the right half 1 =
(0, ). Then the states = 0 [ = 1] of the memory are assigned to x 0
[x 1 ], respectively.
By the dashed horizontal line and the shaded zone in Fig. 7.5 AD, we denote the
zone of (free) energy which the particle practically never attains within op if it starts
from the minimum of the potential. For example, in Fig 7.5A, the particle remains
within one of the valleys, that is, the state of the memory is stable. Likewise, in
Fig 7.5D the memory is in = 1 state. By contrast, in Fig 7.5B, C, the systems
relaxation time sys is less than op . The memory can flip between 0 and 1.
The counterclockwise cycle in Fig. 7.5, ABCDA, is the operation of
overwriting. This operation is decomposed into the process of erasing memory, E,
and the process of writing memory, W:
Erasing, E : A B,
Writing, W : B C D A. (7.8)
Suppose that the system has retained a memory state, , in A. Through the process
E, this memory is lost. Then by the process W, the memory is forcedly reset to the
state = 1. In order to reset to the state = 0, the profiles of C and D should be
replaced by their mirror images C* and D*, respectively.
B ( ll) a
(01) (10)
(hh) : sys /op
1 and sys /op
1
(01) (10)
(h) : sys /op
1 and sys /op 1
(01) (10)
(h) : sys /op 1 and sys /op
1
(01) (10)
() : sys /op 1 and sys /op 1. (7.9)
Now the two control parameters (a, b) are essentially reflected in ln[sys(01)
/sys
(10)
]
and ln[sys sys /op
(01) (10) 2
], respectively. Therefore, the operations of overwriting and
the timescales can be related as shown in Fig. 7.6. In this schema, the important
part of the cycle is the vertical process, AB. While the external system does the
same operation along a = 0 line, the memory state at A depends on the previous
operation, either from D or from D*. Therefore, it is often said that this erasure
process E is the origin of irreversibility.
system loses control of the state after the process, A B, as discussed in the
above. If upon the operation of BA the memory comes back to = 0, then the
subsequent operations, ADCB, will cost a work much more than kB T . This
is an example of the consequence of the discontinuous change of accessible phase
space (Sect. 7.1.1.6).
Furthermore, we will see below that the total work for the overwriting, AB
CDA, is nonzero, bounded below by kB T ln 2 per cycle.
The work for the writing process, WBCDA , can therefore be given as the quasistatic
work: Let us introduce the configurational free energies, F(a, b; 0 1 ) and
F(a, b; ), corresponding to U (x, a, b) and U eff, (x, a, b), respectively, by
F(a,b;0 1 )/kB T
e eU (x,a,b)/kB T d x, (7.11)
0 1
F(a,b; )/kB T
eU
eff,
e (x,a,b)/kB T
dx (7.12)
0 1
= eU (x,a,b)/kB T d x.
because (7.11) and (7.12) give eF(a,b;0 1 )/kB T = 2 eF(a,b; )/kB T . This is the
main result of [2, 3]: Every operation of overwriting costs at least kB T ln 2 of
irreversible work. By the stochastic energetics the result holds for an individual
process with the probability of one in the limit of slow operation.
The heat associated with this cycle has been analyzed using the stochastic ener-
getics [10].11 In the language of information, the memory specified through the
11 [10] used the Bennetts definition of copying of memory. Since this definition does not com-
plete a cycle, their results differ apparently from those in the text. However, these formulations are
mathematically equivalent.
236 7 Control and Energetics
preceding writing process is lost irreversibly during the erasure process. The statisti-
cal entropy associated with the variable has been increased by ln 2. The ensemble
averaged irreversible work has been given in [11]. More recently [12] incorporated
mechanical approach of the type of Jarzynski nonequilibrium work relation, and
confirmed both analytically and numerically kB T ln 2 as the lower bound of the aver-
age irreversible work. [12] also analyzed the case where U (0, a)/kB T is moderately
large.
D A
Fig. 7.7 The data bit (left column), and the energy involving movable bit and the interaction
energy between the two bits, U (x, a, b, ) (right)
13 We assume that the profile of U (x, a, b; ) and that of U (x, a, b; 1 ) are of the mirror images
along the x-axis.
238 7 Control and Energetics
We realize that the complete cycle is essentially the same as the cycle of overwrit-
ing a bit memory shown in Fig. 7.5. The minimum cost of making copy of the data
bit is, therefore, kB T ln 2. The only difference is that the asymmetry of the potential
energy in the copying process is not determined by the control parameter a but is
determined by the memory ( ) of the data bit. To conclude, we need the work to
be no less than kB T ln 2 per movable bit. Therefore, we cannot make a perpetual
machine.
where H (x, p; a) is the systems energy as a function of position x and the momen-
tum p. We ignore the work of detachment and attachment with the thermal environ-
ments, since they were shown to be reducible as small as needed.
If the system is macroscopic, the choice of a2 amounts to equalizing the systems
temperature to T2 before the contact with the second heat bath. Under this condition,
the process is reversible in the limit of quasistatic adiabatic operation, and the work
W is 0.14
For the mesoscopic system, several aspects are different from the macroscopic
case:
1. The work W fluctuates from one realization to the other, however, slowly the
operations are done.
14 Unlike the Carnot cycle, the parameter a is not changed under isothermal condition.
7.1 Limitations of Quasistatic Operations 239
(a)
(i) (ii) (iii) (iv)
T1
a1 a1 a2 a2
T2
(b)
(i) (ii) (iii) (iv)
1
a1 a1 a2 a2
Fig. 7.8 A protocol including the detachment and attachment with heat baths (a) or with a particle
reservoirs (b). The numbers, (i), etc., refer to those in the text. The systems parameter a is changed
in the processes (ii)(iii), where the system is isolated from these environments. (The figure is
modified from Fig. 1 of [13])
15 Roughly speaking, the systems energy at the end of return operation, (iii)(ii), is on average
higher than the initial thermalized state. Therefore, during the operation (ii)(i) the heat (Q) =
W is dissipated to the thermal environment.
240 7 Control and Energetics
16 For quantum system, this condition leads to the relations among the energy levels, E (a )
E 0 (a ) = (a, a )[E (a) E 0 (a)], where (a, a ) is a given function. For the open system the list
of exceptional cases is not exhaustive.
17 A similar argument should be made for the open system.
7.2 Detection and Control Under Fluctuations 241
18 For the canonical ensemble, the similar operation immediately contradicts the second law of
thermodynamics. But it is possible for microcanonical ensemble.
19 Some claims that the motor proteins like F1ATPase are working at about 100% efficiency.
242 7 Control and Energetics
7.2.1.1 Half-Sensors
Let us distinguish the two notions about the detection, which we call IN/OUT and
ON/OFF. A signal particle is said to be IN if it is present at the detection site
of the sensor, and otherwise it is OUT. The sensor is assumed to take two internal
states ON and OFF. The ON [OFF] states of the sensor are designed to be positively
correlated to the signal particles state, IN [OUT], respectively. IN and OUT are the
objective fact about the presence or absence of the particle, while ON and OFF are
the perception by the sensor about the particle.
Figure 7.9(a) shows those correlations in the ordinary sensor. The two shaded
zones are prohibited to occur and these constraints allow one-to-one correspondence
between the sensors states and the arrival of signal particles.
In the mesoscopic scale, the unbiased and error-free detection is still possible
if we sacrifice one of these two prohibition zones. We define two types of half-
sensors by the tables Fig. 7.9b, c. In response to the particles state IN/OUT, these
sensors are multivalued functions in the range ON/OFF. The half-absence sensor
can have the state OFF only when the particle is OUT. Thus if the sensor thinks
that the particle is OFF, the signal particle is surely OUT without bias or errors.
But this absence sensor can miss the particles absence OUT by thinking it be
IN. The half-presence sensor works completely complementary manner to the
half-absence sensor.
The half-sensors can be designed to work reversibly (see below).20 The pos-
sibility of the usual sensor (Fig. 7.9a) on the mesoscopic scale is not known. A
usual sensor can be made as the composite of Fig. 7.9b, c. This corresponds to the
product of two Boolean variables. Since the product of two Boolean variables is
an irreversible operation, the function of such composite sensor might need some
free-energy resource (see [17]).
The utilities of the two types of half-sensors are different. For example, if the
signal particle is toxic for a system, the system can use the OFF sign of the
half-absence sensor. This OFF signal is analogous to the regulation of gene tran-
scription by a repressor protein [18], since the transcription is surely prohibited
internal state
internal state
ON
ON
ON
dd
internal state
dd
i
i
rb
rb
fo
fo
en
n
OFF
OFF
OFF
de
dd
id
i
rb
fo
fo
Fig. 7.9 The functions of the usual sensor (a) and the two kinds of half-sensors (b) and (c). The
black squares indicate the forbidden situations
20 Therefore, the half-sensors are different from to know, which is irreversible and essentially
nonquasistatic process.
7.2 Detection and Control Under Fluctuations 243
by the repressor protein bound to the operator site on the DNA. By contrast, if the
particle is the food for a system, it can use the ON sign of the half-presence sensor.
This is analogous to the activation of the repressor protein, since the activation of
the repressor surely requires the signal particle. Also the uptake of ATP molecule
by a molecular motor will be analogous to the half-presence sensor.
OUT
a
ON 1
0
x
OFF
en
dd
IN U (x,a) = const.
bi
r
fo
ON
1
1 0 +1
a x
Fig. 7.10 Left: Schematic representation of the half-absence sensor. The thick bar (a: the leftmost
position) and the thick dot (x) repel with each other by steric repulsive interaction. Right: On the
(x, a)-plane the shadowed region is accessible without bias
21 The total energy of the ligandsensor system, U (x, a), is written as U (x, a) = U for x a +1
0
and U (x, a) = otherwise. The sensors states are assigned as OFF for a 1+ and as ON for
1 a < 1+, respectively, with a small > 0.
244 7 Control and Energetics
OUT a
OFF 0 x
IN
OFF
ON
1 1
a 0 x U(x , a) = const.
Fig. 7.12 Left: Schematic representation of the half-presence sensor. The filled square (a: sensor
tip) and the thick dot (x) attract each other for |x a| 1, while the displacement in the a > 1
region provokes a restoring force. Right: On the (x, a)-plane the shadowed region is accessible
without bias
22 The cost may be either energetic such as of mechanical deformations or entropic such as of
folding. The detail of the process may contain the aforementioned jump-and-catch process.
23 The total energy of the ligandsensor system, U (x, a), is the sum of short-range strong attractive
interaction energy, M(a x) (M
kB T ), between the ligand particle and the sensor tip and
also strong restoring potential M(a). Here (z) ( < z 0) is (z) = 0 for z 1 and
monotonically increasing from (1) = 0 to (0) = 1.
7.2 Detection and Control Under Fluctuations 245
Remarks
Discretization of state and the half-presence sensor: In looking at Fig. 7.12, one
might wonder if the sensor is actually bijective, like Fig. 7.9a. However,
the discretization of the sensors state and the particles position requires
the gray zone of (IN, OFF), because the judgement of OFFON requires
the presence of particle (IN) so that an energy more than kB T is exchanged
between the attracting potential and the restoring potential.
Timescale of sensing: The diffusion of small signal particles can be too rapid for
the sensor to follow adiabatically.25 This is also the reason to distinguish the
particles position and the sensors state.
Relation between the half-absence sensor and half-presence sensor: The half-pre-
sence sensor is, in some sense, on the basis of the half-absence sensor. In
fact, for the latter sensor, the object that interacts sterically with the signal
particle (the thick bars in Fig. 7.10 Left) must be sensed elsewhere.
Signal sensing must have a consequence: An isolated sensor serves nothing. The
state of the sensor must be coupled to its downstream mechanisms, e.g., by
accelerating an enzyme reaction, modifying the access of another molecule
to the system, etc. Through this coupling, the sensing process can become
biased and irreversible.26
Technical Remarks
Reaction coordinate of the presence sensor: In Fig. 7.12 right, the detection of the
particle occurs through an unbiased corridor corresponding to the unbiased
valley of the total potential energy. Therefore, a single reaction coordinate
x x a (0 x < ) (7.19)
24 i.e., M(1 + ) kB T . The consistency of these assignments with the above definition can
be verified by scrutinizing the graphs of M{(a) (a x)} vs. a for various values of x.
25 Biological sensor may be designed to avoid this.
26 Sect. 7.2.1.4 is an example.
246 7 Control and Energetics
Fig. 7.13 Left: the (x, a) pathway between the right environment (x > 0) and the left environment
(x < 0) through the detection site x = 0. Right: extended reaction coordinate x, where the a
coordinate is duplicated to 1 x 1
x
x= max(|x| 1, 0), a = min(|x|, 1). (7.21)
|x|
See Fig. 7.13 Right. For |x| 1, the pair of points {+x, x} should be iden-
tified as a single physical state, (x(x), a(x)). This technical trick is convenient
for discussing the coupled transport in Chap. 8.
balanced in an unbiased sensor. The main body of the sensor is pushed to the left by
the particle, while it is pushed to the right by the restoring force of the sensor tip.
In terms of the flux of momentum in the x direction, Fig. 7.14 represents a per-
manent circulation of the momentum [20].27 A similar mechanical analysis can be
done for the induced fit in Fig. 7.11.
27 The +x-oriented momentum flows toward +x along the attractive force and toward x along
the repulsive (restoring) force.
28 A similar phenomenon has been observed for the binding of ATP-activated kinesin molecules
to a microtubule, called the cooperative binding, [24].
248 7 Control and Energetics
1 0 +1
a x
Exchange of binding
1 0 +1
a x
1 1
~
x1 x~1
0 1 0
1
Fig. 7.16 Left: Allosteric transition by the coupling, (a1 a2 )2 . Right: Exchange of binding by
the coupling, (a1 + a2 + 1)2 . (x1 , x2 ) are the reaction coordinates (7.19) for the two half-presence
sensors. The shaded region is the region of a constant energy
7.2 Detection and Control Under Fluctuations 249
stable until the second particle arrives at the half-sensor in the unoccupied (OFF)
state. Once the two half-sensors bind their signal particles, the state variables (a1 , a2 )
can diffuse along the line of a1 + a2 = 1. When (a1 , a2 ) arrive either at a1 = 1
or a2 = 1, it is possible that one of the particles leaves the detection site. See
Fig. 7.15 Bottom.29 As a result, the switch between (OFF1 ,ON2 ) and (OFF1 ,ON2 )
can be realized without passing the (OFF1 ,OFF2 ) state. Figure 7.16 Right represents
the process of exchange of binding on the plane of reaction coordinates, (x1 , x2 ). If
there is a bias between (a1 , a2 ) = (1, 0) and (0, 1), the arrival of the particle with
higher affinity can expel the previously bound particle by the allosteric effect.
Margins of Operation
A small positive number has been introduced to represent the fact that the poten-
tial barriers should be located outside but close to the detection sites. Although the
irreversible work of operating the gate can be made negligible, this , as well as
29 This is the first passage time (FPT) problem. See Sects. 1.3.3.3.
250 7 Control and Energetics
the small introduced in the model of sensors (Sect. 7.2.1), constitute the source of
small leakage when we construct free-energy transducers.
When we discretize the model in the form of the reaction networks, neglecting
this marginal effect can sometimes lead to unphysical results. Especially caution
must be maintained when one claims a strictly tight coupling or a 100% efficiency
for a model of autonomous free-energy transducer.
7.3 Discussion
30 The average is taken over all realizations starting from a pertinent equilibrium ensemble.
7.3 Discussion 251
Complementarity relation: The average irreversible work Wirr. and the time spent
for the process, t, satisfy the relation parallel to (5.36) [26]:
(i) Unlike absolute temperature, chemical potential has a no absolute zero and
admits indefinitely negative values.
(ii) Unlike a heat engine, the transferred particles of a particle engine do not neces-
sarily carry energy.
H L
Fig. 7.18 A Carnot cycle
to extract work. The particle
(thick dot) enters from a
dense particle environment of
the chemical potential H (a) (d)
(a). The piston is pulled
while the system is closed
(b). The particle exits to the
dilute particle environment of
the chemical potential H
(c). The piston is pushed
while the system is closed
(d). If the dilute environment
is vacuum (L = ), the (b) (c)
infinite work can be extracted
in principle a
252 7 Control and Energetics
References
1. H.S. Leff, A.F. Rex, Maxwells Demon: Information, Entropy, Computing (A Hilger (Europe)
and Princeton U.P. (USA), 1990) 223, 231
2. R. Landauer, IBM J. Res. Dev. 5, 183 (1961) Reprint: 44, 261269 (2000) 223, 224, 231, 232, 235
3. C.H. Bennett, Int. J. Theor. Phys. 21, 905 (1982) 223, 224, 231, 235, 236
4. R.P. Feynman, R.B. Leighton, M. Sands, The Feynman Lectures on Physics vol.1 (Addison
Wesley, Reading, Massachusetts, 1963), 46.146.9 223
5. E. Eisenberg, T.L. Hill, Science 227, 999 (1985) 223, 247
6. C. Branden, J. Tooze, Introduction to Protein Structure, 2nd edn. (Garland Science, New York,
1999), 13 223
7. F.J. Kull, R.D. Vale, R.J. Fletterick, J. Muscle Res. Cell Motil. 97, 877 (1998) 223
8. M. Born, R. Oppenheimer, Annalen der Physik 84, 457 (1927) 230
9. A. Messiah, Quantum Mechanics (Dover, Mineola, 1999; reprint) 230
10. S. Ishioka, N. Fuchigami, Chaos 11, 734 (2001) 235
11. K. Shizume, Phys. Rev. E 52, 3495 (1995) 236
12. R. Kawai, J.M.R. Parrondo, C. Van den Broeck, Phys. Rev. Lett. 98, 080602 (2007) 236
13. K. Sato, K. Sekimoto, T. Hondou, F. Takagi, Phys. Rev. E 66, 06119 (2002) 238, 239
14. H. Tasaki, cond-mat/0008420 v2 (2000) 239
15. I. Procaccia, D. Levine, J. Chem. Phys. 65, 3357 (1976) 240
16. K. Sato, J. Phys. Soc. Jpn. 71, 1065 (2002) 241
17. R. Landauer, Science 272, 1914 (1996) 242
18. B. Lewin, Genes IX (Jones and Bartlett, U.S., 2007) 242
19. D.E. Koshland, Sci. Amer. 229, 52 (1973) 244
20. K. Sekimoto, in Chemomechanical Instabilities in Responsive Materials (NATO Science for
Peace and Security Series -A: Chemistry and Biology), ed. by P. Borckmans, et al., (Springer,
New York, 2009) 247
31For the ideal gas particle, the quasistatic isothermal work (W ) is (W ) kB T ln(Vf /Vi ), with
p kB T /V the pressure and Vi [ Vf ] are the initial and final values of volume, respectively.
32 It may take a long time Vf .
References 253
21. B. Alberts et al., Essential Cell Biology, 3rd edn. (Garland Pub. Inc., New York & London,
2009) 247
22. J. Monod, J. Wyman, J.P. Changeux, J. Mol. Biol. 12, 88 (1965) 247
23. D.E. Koshland, G. Nemethy, D. Filmer, Biochemistry 5, 365 (1966) 247
24. E. Muto, H. Sakai, K. Kaseda, J. Chem. Biol. 168, 691 (2005) 247
25. C. Jarzynski, Phys. Rev. Lett. 78, 2690 (1997) 250
26. T. Shibata and S. Sasa, J. Phys. Soc Jpn. 67, 2666 (1998) 251
Chapter 8
Free-Energy Transducers
Sekimoto, K.: Free-Energy Transducers. Lect. Notes Phys. 799, 255279 (2010)
DOI 10.1007/978-3-642-05411-2 8 c Springer-Verlag Berlin Heidelberg 2010
256 8 Free-Energy Transducers
where x and p are, respectively, the position and momentum of the oscillator. In this
thought experiment we will not talk about how to isolate the main system. We will
verify below that the usage of the harmonic potential avoids irreversible work upon
relaxation from non-Gibbs ensembles of fluctuations (Sect. 7.1.3).
We use the two clutches which have been introduced in Sect. 7.1.1.5
(cf. Fig.7.4). The clutch H (xyH , H ) controls the contact with the high-temperature
heat bath TH , and the other clutch L (x yL , L ) controls the contact with the
low-temperature heat bath TL . We denote by yH and yL the auxiliary variables
TH, H TL, L
yL
yH
Fig. 8.1 Schematic
composition of a mesoscopic
m x
Carnot engine. (Adapted H (xyH, H) L (xyL, L)
k
from Fig. 1 of [1])
2 About this fact, people sometimes say that the second law of thermodynamics is locally broken.
If is so, it applies also to the free Brownian motion, whose local unidirectional motion takes place
against the viscous force.
8.1 Externally Controlled Free-Energy Transducers 257
where the thermal random forces H (t) and L (t) are white Gaussian processes with
zero mean and correlation (, = H or L):
(t) (t ) = 2 T (t t ).
Protocol of Control
A cycle of the operation for the parameters is depicted in Fig. 8.2. As the macro-
scopic Carnot cycle, two isothermal processes and two adiabatic processes (along
the k-axis) alternate. However, at every instance, there intervenes an explicit control
B0
A0 AL
BH
C0
DL
D0
Fig. 8.2 Protocol of control CH L
for the mesoscopic Carnot 0 1
cycle (arrowed loop).
Horizontal paths (k =const.)
correspond to the attachment
and detachment with thermal H
environments. (Adapted from 1
Fig. 2 of [1])
3 Unlike Sect. 7.1.1.5, we explicitly assumed that the clutches depend on the differences of the
variables, x yH or x yL .
258 8 Free-Energy Transducers
W ( AL A0 ) + W (D0 DL ) = 0,
W (B0 BH ) + W (CH C0 ) = 0. (8.5)
Adiabatic Processes
In the limit of slow control of the external parameters, the only source of stochastic-
ity is the detachment from the thermal environment. Upon detachment from the heat
bath of temperature T , the energy of the main system is chosen randomly according
to the canonical equilibrium distribution of this temperature. The probability density
of the energy, Peq,e (E; T ), is obtained as
Peq,e (E; T ) = const. dx dp eH (x, p;k)/kB T (E H (x, p; k))
1 E/kB T
= e . (8.6)
kB T
8.1 Externally Controlled Free-Energy Transducers 259
After the detachment, the oscillators energy E changes with the quasistatic
change of parameter k. Classical mechanics says [5] that the so-called the action
integral, I (E, k) remains constant (adiabatic invariant) during this process. Here
I (E, k) is defined by
/
E m
I (E, k) dqdp = . (8.7)
H (x, p;k)E 2 k
when the spring constant k is changed quasistatically, k k . Then the work done
to the oscillator during an adiabatic process k k is
Wad E E = k /k 1 E. (8.9)
The average of the adiabatic works, Wad (A0 B0 ) and Wad (C0 D0 ), is, there-
fore, given by (8.9) and (8.6):
. .
kB kD
Wad (A0 B0 ) = TL 1 , Wad (C0 D0 ) = TH 1 . (8.10)
kA kC
As for the energy distribution (8.6), the quasistatic adiabatic process (8.8) pre-
serves the initial form of Gibbs equilibrium distribution with only a shift in temper-
ature4 :
/
k
Peq,e (E; T ) Peq,e (E ; T ). (8.11)
k
This result shows that the irreversible work of attachment to the new thermal envi-
ronment (Sect. 7.1.3) can be completely avoided by a proper choice of the ratio
k /k. Such a choice corresponds, naturally, to the adjustment of the temperature
parameters:
. .
kB kD
TL = TH (A0 B0 ), TH = TL (C0 D0 ). (8.12)
kA kC
4The energy distribution afterthe adiabatic process, P (E ), should satisfy, Peq,e (E; T )dE =
P (E )d E . Substituting E = k/k E into this equation, we obtain (8.11).
260 8 Free-Energy Transducers
By these choices, there is no net energy transfer upon reattachment with heat baths.
Also, (8.12) leads to the cancellation of the average of the total adiabatic works
(see (8.10):
Isothermal Process
The quasistatic isothermal work to change k is given as the change in the Helmholtz
free energy, F (T , k, ) (Sect. 5.2.1.1). From the form of (8.4), we have
TH kC TL kA
WH (BH CH ) = log , WL (DL AL ) = log , (8.14)
2 kB 2 kD
TH TL kC
WH (BH CH ) + WL (DL AL ) = log (< 0). (8.15)
2 kB
The result is negative since the engine does work to the external system.
TH kB
Q H (BH CH ) = WH (BH CH ) = log (> 0). (8.16)
2 kC
in the limit of slow operation of mesoscopic Carnot cycle. In conclusion, the meso-
scopic Carnot cycle can attain Carnots maximal efficiency. The works of the control
are nonzero but they cancel after an entire cycle.
8.1 Externally Controlled Free-Energy Transducers 261
The first line on the right-hand side is the net work extracted from the cycle, and the
second line is the average extractable work in the ideal case. Therefore, = 0,
but in general ()2 0. The question is whether the fluctuations in can break
the second law of thermodynamics. To find nthe answer, we introduce the cumulative
gain over n consecutive cycles, Rn i=1 i , where i is the gain of the ith
cycle. The theory of unbiased random walk5 tells that for some n the value of Rn
will become positive with probability 1. Can we stop at that moment ( Rn > 0)
and repeat the same operation again and again? If we could have Rn > 0 for an
extensive number of times with respect to the total time of experiment, then the
mesoscopic Carnot cycle constitutes a perpetual machine of the second kind, which
is impossible. Again the theory of unbiased random walk (discrete step size = 1
with probability 1/2 for each sign) tells that the probability of having Rn 1 for the
3/2
first time at n = (2m 1) is f 2m 2m[2(2m2)!
m1 (m1)!]2 m . This f 2m has a long tail
such that the mean number of (2m 1) is infinite, m=1 (2m 1) f 2m = . That
is, there is no chance to accumulate the gains ( Rn > 0) for extensive times.
Here we consider the isothermal open Carnot cycle6 that works between two
solute particle reservoirs. In Sect. 7.3.1.2 we have already presented a model (see
Fig. 7.18). There the principle was directly analogous to the mesoscopic Carnot
engine (Fig. 8.1). In both cases heat is absorbed while it does the work, being
mediated by the kinetic energy of gas particle. This mechanism using the kinetic
energy is, however, not likely to be a prototype of molecular machine, notably for
biological motors. Another principle using the binding energy is presented below.
In this mechanism the heat is supplied after the work is done.
The objects of this section are:
1. to understand that the machinery uses the rareness of the particles to extract
heat from the thermal environment,
2. to see that the machinery can work with the maximal efficiency allowed by the
second law of thermodynamics.
3. to verify that there is no need to stock the energy in the system.
H L
H L
(A) (D)
Fig. 8.3 Setup and typical
process of an open Carnot a
cycle. A particle (thick dot)
migrates from the dense
reservoir (left box) to the
dilute reservoir (right box).
The systems binding strength
(affinity) to the particle is
controlled by a. The
particles access is controlled
by H and L . (A) and (C) (B) (C)
are open isothermal processes
and (B) and (D) are closed
isothermal processes
a
H L
A
D
B
C
H L
1 0 1
Fig. 8.4 The protocol of the control parameters (arrowed loop). AD correspond to those in
Fig. 8.3. Horizontal arrows (a =const.) correspond to the attachment and detachment with particle
environments
Protocol of Control
Three parameters (a, H , L ) are changed through the protocol of Fig. 8.4, which
is essentially the same as Fig. 8.2 of (closed) Carnot cycle. We assume that all the
changes are done slowly enough.
The main work is extracted during the stage of binding ((A) and (B) in Fig. 8.4),
while heat is supplied later for the unbinding ((C) and (D) in Fig. 8.4). The dense
reservoir pushes the particles into the main system despite the low binding energy
(A), while the dilute reservoir favors the particles diffusing out despite the large
binding energy (C).
where F (n) is the free energy of the system (5.22) when it contains n particles.
This definition should reflect the position of the sensor tip in Fig. 8.3.9 For the
later usage, we summarize several equilibrium properties of the system. When the
system is open to the -reservoir ( = H or L), the equilibrium probability to find a
particle in the system, P(1) , is10
e( a)/kB T
P(1) (a) = = 1 P(0) (a). (8.20)
1 + e( a)/kB T
e J = eF /kB T
+ e(F )/kB T
(0) (1)
(8.21)
or11
a
J = F (0) (a) + kB T log 1 + e kB T
= F (0) (a) kB T log 1 P(1) (a) . (8.22)
W (B C) = PH(0) (aB )(F (0) (aC ) F (0) (aB )) + PH(1) (aB )(F (1) (aC ) F (1) (aB ))
= F (0) (aC ) F (0) (aB ) PH(1) (aB )(aC aB ),
W (D A) = F (0) (aA ) F (0) (aD ) PL(1) (aD )(aA aD ). (8.23)
where we have used (8.22). These works are definite in the quasistatic limit, despite
the fact that they are represented in terms of P(1) (a).
aA aD = (aC aB ) = H L . (8.25)
This result shows that this open Carnot cycle extracts the maximal work defined by
(2.25), where the coefficient of (H L ) is the number of transported particles per
cycle.14 As the process is reversible, the inverse operation can work as a particle
pump at maximal efficiency.
Conclusion
The open Carnot cycle could extract energy from the thermal environment of a
unique temperature. This energy was used to unbind the particle. Since the work has
already been extracted upon binding the particle, the thermal energy is effectively
converted into work. With whatever large binding energy, the thermal energy can
release a bound particle if only we wait for sufficiently long time. The rareness of
particle in the dilute particle reservoir assures that the particle once unbound has
little probability to return. One may say that open Carnot cycle synthesizes the first
and second laws of thermodynamics.
Thanks to this mechanism, the main system need not stock high energies. In
principle the constraint (8.25) does not exclude the possibility of a > 0 (energy
stockage) [7, 8]. But molecular architectures to stock high energy seem less feasible
than that to bind at low energies. The 3D X-ray structural data of protein motors
show no magic structures to temporarily stock the energy.15 Stochastic energetics
will help decrypt the structurefunction relationship of the protein motors.
Introduction
Autonomous systems are the systems which control themselves. In biology auto-
nomous systems are found from proteins to societies. As a module the autonomous
system works more or less independently of the others for a range of spacetime.
The collaboration among autonomous modules can create higher order functions
and also protect themselves on the one hand, but it can impose constraints on the
constituting modules on the other hand. The organisms consisting of modular struc-
ture are robust against growth and accidental injury.16
In the rest of this chapter we focus on mesoscopic autonomous systems that
convert a form of free energy into another form in the presence of thermal fluc-
tuations. The Feynman ratchet wheel and pawl [2] and the ButtikerLandauer
ratchet [3, 4] are classical examples (Sects. 1.3.4 and 4.2.2). Symmetry and the
Curie principle capture a physical basis of autonomous free-energy transducers
(Sect. 1.3.4.4). Protein motors, ion pumps, signal-processing proteins, etc., are bio-
logical autonomous free-energy transducers [9]. Many studies have been done to
understand the structurefunction relationship of specific biological free-energy
transducers, such as myosin and kinesin. But, at present, only tentative studies
have been made to explain the functions of these motors on the common structural
basis [10].
15 Inside the nucleotide pocket of a motor protein, the cleavage of the high-energy bond between
-Pi and -Pi of an ATP molecule occurs almost at equal free energy.
16 The object-oriented architecture of information systems uses these properties.
8.2 Autonomous Free-Energy Transducers 267
In this section we discuss free-energy transducer as a black box from the functional
point of view.
17 Rigorously speaking ATP-hydrolyzing protein motors are catalysts only when it does not work
against macroscopic external forces.
268 8 Free-Energy Transducers
pumping [14]). In the latter case the substrates {ATP, water} and the products {ADP,
Pi} are, respectively, identified as the h-side and on the -side of the fuel particles.18
Mechanotrasduction by motor proteins, such as myosin or kinesin, is also simi-
lar to the above schema: The displacement d x along the filament can be regarded
as the transport of the filament monomers (G-actin monomers for myosin, tubulin
dimers for kinesin, etc.) between an anterior monomer reservoir and a posterior
monomer reservoir, see Fig. 8.6. If a mechanical load L is applied between the
myosin molecule and the actin filament, the two reservoirs have different chemical
potentials corresponding to the work done by a single-step displacement under the
load.
The thermodynamic constraint on the function of this transducer is the decrease
of the total Gibbs free energy,
G tot G F + G L < 0,
with G F (Fh F )(n F )(< 0) and G L (Lh L )n L (> 0) (see Fig. 2.8).
Here Fh , etc., are the chemical potentials of the particle environments, and n L of
the L particles are transported upward (L Lh ) at the expense of the transport of
n F of the F particles downward (Fh F ).
18 The correspondence is not exact because of the four species of particles in place of two. How-
ever, the stoichiometry relation of the reaction allows us a simplified representation of Fig. 8.5.
8.2 Autonomous Free-Energy Transducers 269
where 01LF , etc., are the free-energy levels of the transducer in each state.
The master equation for the schema Fig. 8.7 yields the steady-state rate of the
counterclockwise reaction cycle, :
The chemical kinetics approach has recently been put forward to describe single-
molecule experiments, taking into account applied forces [18] and intermediate
19 Other sequence of migration is also possible. In an ion pump [13], for example, the in-and-out
of a proton occurs in between the in-and-out of a lactose molecule.
20 Section 3.3.1.5 Nonequilibrium settings II.
270 8 Free-Energy Transducers
steps [19, 20]. How well this phenomenology works depends on the choice of repre-
sentative chemical states and of allowed transitions among them, and also depends
on the identification of basic parameters from separate but consistent experiments.
It is beyond the scope of the chemical kinetics approach to study what structural
basis constitutes the representative chemical states and what mechanisms support
state transitions.21
it is possible to realize such cycle.23 Some people think that a class of biological
transmembrane transporters function with a single principal degree of freedom. This
hypothesis is called the alternative access model [21, 13].
However, the transducer with a single allosteric coupling may not be the simplest
case, because a single degree of freedom must play double roles of detector and
gate.24
Contrastingly, recent structural analyses of a single head of myosin suggests two
independent internal degrees of freedom [22, 23]. Conceptually the transducer with
two allosteric degrees of freedom is simpler than that with a single allosteric degree
of freedom. The actions of sensor and detector at an interface of F particles can
be directly communicated to those at the interface of L particles and vice versa if
there are two allosteric degrees of freedom. We, therefore, focus on the free-energy
transducer with two allosteric couplings.
gate sensor
(L,l) (L,h)
(A) (B)
Fig. 8.9 A coupling rule between the sensor and the gate. The symbols for the gate are the same
as Fig. 8.3. (A) The detection of F particle (thick dot) allows the access of L particles from their
dense reservoir (L,h), and nondetection of F particle (open dashed circle) allows the access of L
particles from their dilute reservoir (L,). (B) The completely symmetric rule applies for the socond
allosteric coupling
8.2 Autonomous Free-Energy Transducers 273
(F,h)
0F
1F
(F,l)
0F
(L,l) (L,h)
0L 1L 0L
Fig. 8.10 Schematic representation of the function of bidirectional control by the rules of Fig. 8.9.
The horizontal axis is the extended reaction coordinate for L particles, where the left [right] of the
detection site correspond to the dilute [dense] environment, respectively. The vertical axis is the
extended reaction coordinate for L particles, where the zone above [below] the detection site cor-
respond to the dense [dilute] environment, respectively. Thick vertical [horizontal] borders of the
gray region represent the blockade by the gate by the coupling rule Fig. 8.9(A) [(B)], respectively
a particular realization such as a dashed curve in Fig. 8.10, the occupancies of the
particles complete a clockwise cycle along the schema of Fig. 8.7, that is,
Throughout this cycle the potential energy is constant, and this free-energy trans-
ducer is, therefore, purely entropic.
26 Section 7.2.2, Margins of operation. The was defined in Sect. 7.2.1.2, Interaction potential of
the half-presence sensor, and the was defined in Sect. 7.2.2, Gate made by the potential energy
barriers.
8.2 Autonomous Free-Energy Transducers 275
at the stalled condition, F,h F, = L,h L, , the efficiency of free-energy
conversion drops to 0.
~
xF x~F
x~L ~
xL
27 L particle cannot go through the mutant transducer from the right (L, h) reservoir to the left
(L, ) reservoir without the assistance of the to-and-fro transitions of F particle: 0F 1F 0F .
28 This dichotomy is invariant under the exchange of the roles of Fuel and Load: the diagram of
Fig. 8.10 is symmetric with respect to the two axes.
276 8 Free-Energy Transducers
(F,h) (F,l)
(L,l) (L,h)
Fig. 8.13 (A) Typical process of particles migration under the rules of Fig. 8.9. (B) The case
under the complementary rules
(L, h) exchange occupancy, like the exchange of binding (see Sect. 7.2.1.4) but with
delay. In (B) it is the reservoirs (L, ) and (F, ) that do a similar exchange. Biolog-
ical motors seem to prefer type (A) (see the next section).
Information-Theoretic Approach
The above model of free-energy transducer uses two agents of communication. If
the gate of an F particle is directly correlated to the sensor of this particle, for exam-
ple, the system realizes feedback loop for an L particle, that is, the detection of
this particle influences the accessibility of this particle. For feedback control, the
information-theoretic analysis has been done [30, 31]. The generalization to the
bidirectional control might be done.
U2(x)
U1(x)
x1 x2
Fig. 8.14 Ratchet model driven by position-dependent switching of potentials. When x(t) is
detected within the zones of thick lines (around x1 , x2 , etc.) it is more probable for the actual
potential function to be switched to the other potential function
8.2 Autonomous Free-Energy Transducers 277
The study of biomolecular machines started from the macroscopic level in physiol-
ogy and then in biochemistry. More recently, the atomic structure of those molecular
machines has been largely uncovered by X-ray crystallography on by NMR. Also,
single-molecule experiments have been developed since the late twentieth century.
Full molecular dynamic studies including quantum aspects are also developing.
The more the spatiotemporal resolution is improved, the more complex charac-
teristics of the molecular machines are found. Through these discoveries, simple
models made of spring and binding block, or of combustion engine and lever arm,
etc., are obliged to be refined, modified, or sometimes abandoned. A single protein
may accept different coarse-grained models according to different working condi-
tions and regulations.
But at the same time, structural analysis has revealed that some local atomic
structures are highly universal among a superfamily of biomolecules. For exam-
ple, the notion of the common ancestor [32] of myosin, kinesin, and G-proteins
(see [9]) stemmed from such observations.
A gap of our understanding is between this local structural universality and the
global functional diversity among the members of a superfamily. For example, the
nucleotide pocket of these molecular motors resembles each other very closely.
Nevertheless, the relative timing between the ATP hydrolysis cycle and the fila-
ment binding/unbinding cycle is known to be very different between a conventional
myosin head [33, 34] and kinesin [9].29
Another question is how these biomolecular machines can work under strong
thermal fluctuations. Recent single-molecule experiments have started to uncover
how individual molecular machines behave in nanometer and microsecond ranges.
On the scales of thermal fluctuations many notions used in macroscopic world need
to be reconsidered like the laws of motion, force, stability, chemical potential, heat,
etc. For example, structural biology papers use the notions of gate, sensor, switch,
etc., to consider the causal sequence of molecular events. However, the bidirectional
control described above distinguishes between two types of causalities: one is the
(quasi) instantaneous linkage of events through the allosteric coupling between a
sensor and a gate, and the other involves the event-waiting states whose subsequent
transition depends on the thermal fluctuations.
In summary, the study of intramolecular processes of biomolecular machines
should clarify the structurefunction relations of the machines based on the physi-
cal concepts adapted for the fluctuating world. One goal is to establish a common
29 The myosin head detaches from the actin filament upon binding with an ATP molecule, while
kinesin head with ATP binds to microtubule filament rather strongly. The isomerization confor-
mational changes associated with intramolecular hydrolysis reaction takes place when the myosin
head interacts with an actin filament only weakly, while the isomerization of a kinesin head occurs
when the head is ready to detach from the microtubule.
278 8 Free-Energy Transducers
References
1. K. Sekimoto, F. Takagi, T. Hondou, Phys. Rev. E 62, 7759 (2000) 255, 256, 257
2. R.P. Feynman, R.B. Leighton, M. Sands, The Feynman Lectures on Physics vol.1 (Addison
Wesley, Reading, Massachusetts, 1963), 46.146.9 255, 266
3. M. Buttiker, Zeitschrift der Physik B 68, 161 (1987) 255, 266
4. R. Landauer, J. Stat. Phys. 53, 233 (1988) 255, 266
5. L.D. Landau, E.M. Lifshitz, Mechanics (Course of Theoretical Physics, Volume 1), 3rd edn.
(Reed Educational and Professional Publishing Ltd, Oxford, 2002) 259
6. Y.G. Sinai, Probability Theory: An Introductory Course (Springer, New York, 1992) 261
7. R.D. Vale, F. Oosawa, Adv. Biomys. 26, 97 (1990) 266
8. M. Nishiyama, H. Higuchi, T. Yanagida, Nat. Cell Biol. 4, 790 (2002)
9. B. Alberts et al., Essential Cell Biology, 3rd edn. (Garland Pub. Inc., New York & London,
2009) 266
10. K. Sekimoto, C. R. Physique 8, 650 (2007) 266, 277
11. E. Eisenberg, T.L. Hill, Science 227, 999 (1985) 266, 271, 278
30 For example, the arrival of an L particle from the dilute reservoir is detected by the myosin at
the lever arm and its comoving units, while for the kinesin the target of detection is not the tubulin
but seems to be internalized.
References 279
12. B. Alberts et al., Essential Cell Biology: An Introduction to the Molecular Biology of the Cell
( 12.)(Garland Pub. Inc., New York, 1998) 267, 270
13. J. Abramson et al., Science 301, 610 (2003) 267
14. Y.C. Kim, M. Wilkstrom, G. Hummer, PNAS 104, 2169 (2007) 267, 269, 271
15. P.G. Bergmann, J.L. Lebowitz, PR 99, 578 (1955) 268
16. T.L. Hill, Free Energy Transduction and Biochemical Cycle Kinetics, Dummy ed. (Springer-
Verlag, 1989) 269
17. T. Shibata, K. Sekimoto, J. Phys. Soc. Jpn. 69, 2455 (2000) 269
269
18. G.I. Bell, Science 200, 618 (1987) 269
19. A.B. Kolomeisky, B. Widom, J. Stat. Phys. 93, 633 (1998) 270
20. A.B. Kolomeisky, M.E. Fisher, Biophys. J. 84, 1642 (2003) 270
21. C. Tranford, Annu. Rev. Biochem. 52, 379 (1983) 271
22. P.D. Coureux, H.L. Sweeney, A. Houdusse, EMBO J. 23, 4527 (2004) 271
23. M.A. Geeves, K.C. Holmes, Adv. Protein Chem. 71, 161 (2005, chap. V) 271
24. K. Sekimoto, Physica D 205, 242 (2005) 271, 275
25. M. Porto, M. Urabakh, J. Klafter, Phys. Rev. Lett. 85, 491 (2000) 276
26. H. Sakaguchi, J. Phys. Soc. Jpn. 67, 709 (1998) 276
27. I. Derenyi, R.D. Astumian, Phys. Rev. E 59, R6219 (1999) 276
28. F. Julicher, J. Prost, Phys. Rev. Lett. 75, 2618 (1995) 276
29. K. Sasaki, R. Kanada, S. Amari, J. Phys. Soc. Jpn. 76, 023003 (2007) 276
30. H. Touchette, S. Lloyd, Phys. Rev. Lett. 84, 1156 (2000) 276
31. H. Touchette, S. Lloyd, Physica A 331, 140 (2004) 276
32. F.J. Kull, R.D. Vale, R.J. Fletterick, J. Muscle Res. Cell Motil. 97, 877 (1998) 277
33. R.W. Lymn, E.W. Taylor, Biochemistry 10, 4617 (1971) 277
34. D.R. Trentham, J.F. Eccleston, C.R.Q. Bagshaw, Rev. Biophys. 9, 217 (1976) 277
35. K.C. Holmes, R.R. Schroder, H.L. Sweeney, and A. Houdusse, Phil. Trans. R. Soc. B 359,
1819 (2004) 278
Appendix A
M C k N M C nk
Pr ob[k = k] P(k, n|M, N ) = ,
N Cn
k
f (k; ) e , (0 k < : integer, > 0 : real),
k!
is called the Poisson distribution. The parameter is the average number of red
balls (k) found among a vast number (n) of chosen balls. The probability f (k; )
is maximum at around k = but has tails above and below this average number.
The Gaussian distribution is particular first because it contains only two parameters,
e.g., the average and the variance, and second it has a quadratic form in the exponen-
tial. The second property gives rise to a general relation between the fluctuation of a
random variable and the response of its average against the change in the parameters
of the Gaussian distribution [1]. The similar relation has been well-known for gen-
eral probability distribution related to thermal equilibrium, but the quadratic expo-
nential nature of the Gaussian distribution provides this relationship irrespective of
the origin of the Gaussian distribution, either equilibrium or nonequilibrium.
Suppose that a random variable x obeys the Gaussian distribution, with the prob-
ability
density, P(x; ) (x x) , where is a parameter and O(x) =
O(x)P(x; )d x. The average x and the variance a2 [x x ]2 are, there-
fore, functions of .
Because of the quadratic exponential character of Gaussian distribution, the
effect of the change of parameter, + , on the probability distribution must
2 2
take the form, P(x; + ) = P(x; ) euy+vy /euy+vy , where y x x
and the coefficients u and v depend generally on and .3 [1] showed that the
effect of small change of on the average, x , has a simple relationship with the
variance or the fluctuation, [x x ]2 . In fact
2
xeuy+vy
x+ =
euy+vy 2
y + u (1) y 2 + v (1) y 3 + O(()2 )
= x +
1 + u (1) y + v (1) y 2 + O(()2 )
= x + u (1) y 2 + O((a)2 ), (A.1)
3In we should write y x x rather than y, but we abuse the latter for simplicity of
notation.
Appendix A 283
xa+ x
lim = u (1) [x x ]2 . (A.2)
0
Therefore, we have
x x 2 =0
= . (A.3)
kB T
The comparison of (A.2) with (A.3) shows that the inverse of u (1) plays the role
of an effective temperature. This fictitious temperature can be measured as a true
temperature if we attach to the system a small monitoring system (thermometer)
isolated from the thermal environment [2].
4 cf. It is not always the case; for the Lorenzian distribution, p(a) = (/ )(a 2 + 2 )1 , the second
moment is divergent.
284 Appendix A
Since eia is the Fourier transform of (A N a ), the last limiting property implies
that PN (A N ) ( A N a ) is the distribution.5 This is called the (weak) law of
large numbers. A stronger law has also been studied, but we will not discuss it here.
In order to study convergence more precisely, we focus on the narrowing peak
of PN (A N ). In this purpose we introduce, instead of the empirical average, A N , the
following standardized deviation:
N
A N a ai N a
W = i=1
, (A.5)
N2
N
$ %
where 2 = (a a )2 is the variance
of ai . We introduce the characteristic func-
tion for W as Q N () = ei W . By the independence of ai , we have Q N () =
i(ai a )
N
e N2 . By developing the exponential function in terms of ,6 we have
2 N 2 2
2
Q N () 1 2N + O N 2 e 2 for N . As e 2 is the Fourier
transformation of the standard normal distribution, ew /2 / 2 , the above result
2
implies the following detailed but universal behavior of the sharp peak (de Moivre
Laplace theorem):
/ / 2
c2 w2
2 2 e
lim Prob c1 < A N a < c2 = dw. (A.6)
N N N c1 2
theorem, belong to the subject of the large deviation theory. The Cramers theorem
tells how the probability of the large deviations behaves: (i) the probability Pd (x) of
the large deviation, x = A N a , obeys the law, 8
From (1.8), (1.9), and (1.18) together with (1.22), we find that the right-hand side
of (A.9) is eDq t . On the other hand, by (1.23) the characteristic function for X
2
Dq 2 t
also yields e . Since the characteristic function contains (almost) the same
informations as the probability density, we find the equality between P(X, t) =
(X [x(t) x(0)]) and (1.23).
The method of projection operators is summarized without going into details, and
the description is rather formal.
The gross variables (i.e., slowly varying physical observables of our interest),
A(x, p), are functions of the phase space point, {x, p}.9 From classical Hamiltonian
mechanics, we know that the time evolution of A(x, p) is governed by the Liouville
equation, which is linear in A(x, p):10
9 Except for very simple cases like the one studied in the previous section Sect. 1.2.1.4, the observ-
able quantities of fluctuating system in general depend on many degrees of freedom. For example,
the density of gas particles found in a specified small but macroscopic volume could be defined in
terms of all the particles coordinates.
10When we study directly the evolution of A(x, p), the arguments {x, p} denote the phase space
point at the initial time t = 0, that is, {x(0), p(0)} = {x, p}. The function A(x, p) at time t
286 Appendix A
A = iL A. (A.10)
Here the linear operator, called the Liouville operator, L, is defined using the Pois-
son bracket11 { , } and the Hamiltonian H as L A { A, H }.
We will denote H the functional space, in which the projection operators act.
The essence of the method of projection operator is to represent the evolution
of the observable A, after having (linearly) projected it into a subspace, H ( H),
spanned by a set of slow variables at the initial time, which are represented by an
j } in H [3]. The action of the projection oper-
orthogonal set of the unitary vectors {e
ator, P, can be represented as P X = j e j (e j , X) under the properly defined scalar
product, (X, Y ), in the functional space H. The temperature of the environment
composed of the nonslow variables appears in the definition of the scalar product
or, more precisely, in its weighing function.12 The formal procedure to obtain the
projected evolution equation of A starts by decomposing the Liouville equation,
A = iL A, into the two equations for P A and (1 P) A:
P A = iPL[P A + (1 P) A],
t
(1 P) A = i(1 P)L[P A + (1 P) A].
t
connects this phase point to the value of A at that time. This viewpoint is parallel to the Lagrange
picture in hydrodynamics or the Heisenberg picture in quantum mechanics. (As a complementary
picture, if we defined A(x, p) as a fixed function of {x, p}, the temporal change of the value of A
would be represented
as A(x(t), p(t)).)
11 {A, B} A B A B , where i is the index of all the microscopic degrees of freedom.
i xi pi pi xi
12 The scalar product of A and B in H is defined by ( A, B) k T (kB T )
1
B 0
Tr eq eH A eH B d, where eq = Z 1 eH/kB T is the equilibrium distribution or the density
operator at temperature T .
Appendix A 287
14 It was Zwanzig who first proposed the equation with memory effect. [6].
15 In Appendix to Chap. 4, we describe the energetics of the Langevin equation on a manifold
(Sect. A.4.7.3).
288 Appendix A
d Wt d=1 O (xt ) dWt (in Itos sense) eliminates the cross-correlation, and
we may write d Wt = (xt )d Bt , where (x) is the th diagonal component of
the positive diagonal matrix (x) and Bt is the th component, in this local frame,
of d-dimensional Wiener process. Since the principal directions of the symmetric
matrix g(x) depend on x, the suffix at different points of x has no simple relations.
The sum of the second and third terms on the right-hand side of (1.59) yields
up to dt 16
d xt d xt
f (xt ) + f (xt ) d xt = f (xt + )d xt
2 2
f (xt+dt ) + f (xt )
= d xt
2
= f (xt ) d xt . (A.12)
As is seen from (A.13), formulas of the calculus of Stratonovich type take apparently
the form of real analysis (integration by parts, integration by substitution, etc.). This
is a merit of the calculus of Stratonovich type as compared with the Ito type.
These two types of calculus are, however, a matter of choice. The conversion
from one type to the other is always possible and is given by
b(xt , t)
f (xt ) d Bt = f (xt ) d Bt + f (xt )dt. (A.14)
2
To understand this, we can rewrite the left-hand side as
16 Since d xt dt, we keep only up to the (d xt )2 terms in accordance with (1.52).
Appendix A 289
and substitute into f (xt+dt ) f (xt ) (= d f (xt )) the formula (1.58). For f (xt ) =
b(xt , t), the formula (A.14) can be easily generalized to give b(xt , t) d Bt =
b(xt , t) d Bt + b(x2t ,t) b(x t ,t)
xt
dt . Then we arrive at (1.60).
d(X xt )
b(xt )2
= a(xt ) (X xt ) + (X xt ) dt b(x t ) (X xt ) d Bt
2
1
= [a(X )(X xt )] + [b(X ) (X xt )] dt b(xt ) (X xt ) d Bt .
2
2
Taking the path average of each term, the terms including d Bt vanish by the rule
of the Ito-type calculus. We then have
1
dP(X, t) = [a(X )P(X, t)] + [b(X ) P(X, t)] dt.
2
2
In the case with inertia, we apply the Itos lemma to (X xt )(P pt ). We find
d[(X xt )(P pt )]
p
= (X xt )(P pt )
t
dt
( m +
pt U
+(X xt ) (P pt ) + dt + 2 kB T d Bt
m x x=xt
1
+ (X xt ) (P pt )(2 kB T )dt. (A.16)
2
By taking the path average of each term, we arrive at (1.74) in the text. As noted
in the footnote below (1.61), we need not to distinguish between d Bt and d Bt in
this case even if and/or T depend on xt .
dn dn n
17 Note that dy n
[(y)(y x)] = dy n
[(x)(y x)] = (x) dy
d
n (y x).
290 Appendix A
dX
= V(X (t), (t)).
dt
This can be a Langevin equation, where X = (x, p) or x; a Hamiltonian equation,
where X represents all the positions and momenta and we ignore (t); or any other
Markovian evolution equation. We adjust the additive constants in X such that its
equilibrium average without external force vanishes.
The above equation (without external force) can be formally solved as an initial
value problem. We denote the solution as follows:
where () in the integrand denotes the average taken over the random thermal noise
() but at a fixed initial value of X.
In the linear response theory, the relaxation function of X(t) against a force on
this variable is defined through the evolution of X(t) given that a weak constant
external force h on X has been applied until t = t0 and is switched off at t = t0 .
Because of the exponential Boltzmann factor form of the canonical distribution, the
equilibrium density in the presence of the weak force takes the form
X 0 h X0 h
e kB T
P (X 0 ; T ) 1 +
eq
P eq (X 0 ; T )
kB T
() X0 h
X(t) = X ( (); X 0 ) 1+ P eq (X 0 ; T ) d X 0 h
kB T
1
= C X,X (t t0 ) h. (A.19)
kB T
On the right-hand side, the coefficient of h is called the (linear) relaxation function,
R X,X (t t0 ).
In conclusion, we have the following formula, which is essentially the fluctuation
dissipation relation:
where we took into account the causality R X,X (t) = 0 if t < 0. The generalization
to the quantum case requires the noncommutativity of X(t) and X 0 . The general-
ization to the fluctuations of currents needs some derivation and integration [7],
but we do not go into those details for simplicitys sake. From the above derivation,
it is clear why this relation does not depend on the particular evolution model. We
notice that neither the concrete form of P eq (X 0 ; T ) nor the detailed balance con-
ditionshave beenused except for the fact that the force perturbs P eq (X 0 ; T ) by a
factor 1 + XkB0T h . This suggests the generalizability of the above relationship to
the nonequilibrium steady states.
t
t s
U (x(s), a(s)) t
We substitute this expression for the p(t ) in the time integral of the second equation
in (1.31) over [t, t + t],
t+t
[x(t + t) x(t)] = ( /m) p(t )dt . (A.20)
t
Step (2): The estimation of the resulting double integral requires some care: we
use the following identity:
t+t t t t+t t+t t+t
dt ds f (s, t ) = ds dt f (s, t )+ ds dt f (s, t ). (A.21)
t 0 0 t t s
Also we use the development of the functions of x(s) in powers of [x(s) x(t)], like
d T (x)
T (x(s)) = T (x(t)) + [x(s) x(t)] + . (A.22)
d x x=x(t)
The first and the second lines on the right-hand side of (A.23) are, respectively,
of the order O((t)1/2 ) and O(t). [Bt+t Bt ]2 is known to obey the so-called
2 -distribution of one degree of freedom. Those Bt s in the first and the second line
of (A.23) represent the same realization (path).
Step (3): In the limit of t 0, the term {[Bt+t Bt ]2 t} vanishes because
of the law d Bt2 = dt (see (1.52)). We then arrive at (1.93) and (1.94). When we solve
numerically these equations, the formula (A.23) is useful.
This form is essentially identical to so-called Milstein scheme [8] to solve the
SDE of the form (1.93). It means that the coarse-grained form (A.23) is not a result
peculiar to the elimination of the inertia effect.
1 (n+1)t
nt t
In (X, a) (X x(t))dt d X a
t nt op op
(, a)eq [a((n + 1)t) a(nt)]. (A.24)
Adding up these integral of segments, we find the result (1.108) in the text. If a has
n(> 1) components, the product da is understood as the scalar product.
P
= LP, (A.25)
t
this absorbing boundary imposes the Dirichlet boundary condition on the probability
density, P(x, t) (x x(t)) :
P(x, t) = 0, x , (A.26)
If (A.25) is solved under this boundary condition and the concentrated initial
condition, P(x, 0) = (x x0 ), (x0 ), the total probability to find the particle
in , i.e., P(x, t)d x, decreases in time for t > 0.
= (x0 ). (A.31)
Appendix A 295
y
D
C
B
x
Fig. A.1 Integrals of dz = a(x, y)d x + b(x, y)dy along the path A B C and along the path
A DC
296 Appendix A
z(ADC) z( ABC)
= [z(DC) z( AB)] [z(BC) z( AD)]
1 1 1 1
[a( x, y) a( x, 0)]x [b(x, y) b(0, y)]y
2 2 2 2
a b
[ y]x [ x]y. (A.32)
y x
a b
Then the relation = assures the existence of a function z(x, y).
y x
E E = E + e N , S S = S + s N , (A.33)
where e and s are arbitrary constants. Evidently the transformations of (A.33) are
consistent with (2.4).
The less evident consequences of the above transformations (A.33) are the
changes in other complete thermodynamic functions derived through the Legendre
transformation 21 :
F E TS F = F + (e T s )N
G E T S + pV G = G + (e T s )N . (A.34)
The last formula defines how the chemical potential, , should transform: With the
relation G = N , we have
= + (e T s ). (A.35)
21 From here up to the end of this section we will take the example of a single-component gas
system: The generalization of the proofs is not difficult.
Appendix A 297
d E = T d S pd V + d N . (A.36)
We rewrite this relation in terms of E, S, and , using (A.33) and (A.35). The
result is d E = T d S pd V + d N , which is identical form to the original one.
S(E, V, N )
= . (A.37)
T N
S S
= [S( E e N , V, N ) + s N ] = (e ) + + s. (A.38)
N N E T
Remark: When we deal with the multicomponent systems with chemical reac-
tions, we must respect the constraints among the e s and s s. Where the quantum
effects appear, the arbitrariness of the additive constants is more restricted. When
several chemical substances participate in a chemical reaction, the third law of ther-
modynamics (Sect. 2.1.2) about the uniqueness of the ground state at T = 0 induces
the relations among the specific quantities of those substances. See, for example,
Chap. 9 of [11].
22 The demonstration below follows [12]. Y. Oono pointed this out to me.
298 Appendix A
n
where i 0, i=1 i = 1. By substituting f (t) = t ln t, i = bi / nj=1 b j , and
ti = ai /bi in (A.39) with ai > 0 and bi > 0, we have the log sum inequality:
n
n
aj
n
ak
a j ln a j ln k=1
n . (A.40)
j=1
bj j=1 l=1 bl
n
Using i=1 K i j = 1 on the left-hand side, we arrive at (3.27).
We then have the average and the variance: tv = Tv , and tv 2 tv 2 = Tv 2 .
(2) The transition rate Tv 1 is proportional to the concentration of the substrate
(around the active site of the enzyme). It is, therefore, Tv 1 = [S] as written
above.
(3) That the termination of the complex ES is a Markov process and is independent
of its fate, either E+S or E+P. The probability Pr (tr ) that the complex ES exists
for the period tr is
Pr (tr ) = etr /Tr .
We then have the average and the variance: tr = Tr , and tr 2 tr 2 = Tr 2 .
23 The strictly convex function is defined such that 12 ( f (x) + f (y)) f ( x+y
2
).
24To obtain this, one should solve d Pv (tv ) = Tv 1 dtv with the initial condition, Pv (0) = 1.
Similar argument applies to Pr (tr ) below.
Appendix A 299
(4) That the probability by which the complex ES dissociates into E+P, rather than
into E+S, is independent of the past history, so that the probability PP (n) that
the product P is formed in the nth formation of the complex ES is
PP (n) = (1 q)q n .
We then have the average and the variance: n = (1 q)1 and n 2 n2 =
(1 q)2 q.
With the assumption made above, the sum in (3.71), that is, nk=1 (tr(k) + tv(k) )
consists of independent random numbers, tr(k) , tv(k) , and n. Therefore, we have the
mean value of t P and that of t P 2 as
, -
n
n
t P = PP (n) (tr + tv ) = n[tv + tr ],
(k) (k)
(A.42)
k=1 k=1
, n 2 -
n
t P =
2
PP (n) (tr + tv )
(k) (k)
k=1 k=1
= n [tv +
2
tr ]2 + n[tv 2 tv 2 + tr 2 tr 2 ]. (A.43)
By substituting the averages and variances of tv , tr , and n, into (A.42) and (A.43),
we finally obtain the two concise equations (3.72) and (3.73) in the text.
p U
dp = dt dt + 2 kB T d Bt .
m x
We multiply each term of this equation by ( p/m). Here the product of Stratonovich
type with pt and that of Ito type are related via pt ( pt + dpt /2). In the dpt /2
we can reiterate the above SDE. Finally, using (d Bt )2 = dt, we arrive at (4.13), i.e.,
2 p2 kB T p
dE = dt + d W + 2 kB T d Bt .
m 2m 2 m
In order to recover the energy balance equation (4.7) from (4.13), we apply the
identity, d( p 2 /2m) = ( p/m) dp in (4.13). Then we rewrite dp using the SDE
(1.61). Finally we rewrite p d Bt using (A.14), where x should be reread as p, and
b( pt , t) should be reread as 2 kB T .
300 Appendix A
This is also the expansion in powers of (xt+h xt ) = O(h 0.5 ).25 Therefore, the error
in (4.14) is O(h 1.5 ). The approximate solution x(t) is basically the sum of (4.14),
and the cumulated error in x(t) is, therefore, O(h 1.5 ) N = O(h 0.5 ).
About the energetics, we calculate r = U Q defined in the main text. With
the discretization of the solution (4.14), the heat (4.16) is Q = dUd(xxt ) (xt+h xt ).
We expand U (xt ) and U (xt+h ) in U , as well as U (xt ), around xt (xt + xt+h )/2
up to second order. We then have
1 d 2 U (xt )
r = (xt+h xt )2 + o(h)
2 d x 2 = 12
1 d 2 U (xt ) 2kB T
h + o(h), (A.45)
2 dx2
where we have used (xt+h xt )2 h to go to the second line, and o(h) is such that
limh0 o(h)/ h = 0. Therefore, the cumulated error should become N O(h) = O(1)
unless we choose = 12 .
U U 1 2U
d Bt = d Bt + 2kB T / dt. (A.46)
x x 2 x2
25In a very short time, xt undergoes free Brownian motion, and therefore, (xt+h xt ) 1
w
t,t+h
=
O(h 0.5 ).
26 cf. (1.65).
Appendix A 301
U
Because in the rightmost term is nonanticipating with respect to d Bt , we obtain
x
the result (4.27) in the text.
Following exactly the same procedure as in Sect. 4.1.3.1 for a single thermal envi-
ronment, we can show the following expression for d Q i :
2i pi2 kB T i pi
d Qi = dt + 2 kB T i d Bi,t , (A.48)
mi 2m i 2 mi
where the Wiener processes {Bi,t } are mutually independent and satisfy (d Bi,t )2 =
dt.
If the effect of inertia is negligible, we will use, instead of (4.32), the following
Langevin equation:
d xi U
0 = i + 2i kB T i i (t), (A.49)
dt xi
Given these formula of d Q i , we rewrite the terms dt, the terms which are not
nonanticipating, using probability fluxes and probability density. We introduce the
following identities:
E E 2i Pi2 kB T i Pi
Ji,x + Ji, p = P i kB T i P , (A.51)
Xi Pi m i 2m i 2 Pi mi
2
U 1 U 2U kB T i U
Ji,x = kB T i 2 P P , (A.52)
Xi i xi xi i X i X i
302 Appendix A
Combining (A.48) with (A.51) and also (A.50) with (A.52), we arrive at (4.36) in
the text.
( ) = ( T ) , ( + ) = ( T ) + ,
which yield
d Q = (2 T )dt + 2 T d Bt , d Q = (2 T )dt 2 T d Bt .
In the double integral to calculate the left-hand side of (4.54), we can replace
d Bu 1 d Bu 2 by (u 1 u 2 )du 1 du 2 . We also use the formula about the average of the
product of four Gaussian variables:
U1 (x2 , a)+ U2 (x1 , x2 ) when these two particles are well isolated from all the
other N 2 particles.
3. In the same manner the p-particle interaction energy U p (x1 , . . . , x p ) ( p 3) for
a cluster of p particles is defined as the energy to be added to all the contribution
of U1 , . . . , U p1 , when this cluster of p particles is well isolated from the (N p)
other particles. By definition, U p (x1 , . . . , x p ) ( p 2) vanishes unless all these
positions are close to each other.
( p)
4. The characteristic function for the p-particle cluster, (x1 , . . . , x p ), ( p 1),
can be written in terms of (x) as follows:
( p) p
(x1 , . . . , x p ) 1 j=1 [1 (x j )]. (A.53)
( p)
5. Using U p (x1 , . . . , x p ) and (x1 , . . . , x p ), the energy of the open system, E, is
written as (4.65) in the text.
v
+ v v = f ext p1 , (A.55)
t
where p is the hydrostatic pressure and f ext is the distant force per mass exerted
by an external system. The symmetric tensor represents the deviatoric (traceless)
stress, with ( p1 + ) being the stress.27 This is [13] ( and are the Cartesian
indexes)
( +
v v 2 v
= 2 + + s , (A.56)
x x 3
x
is the deviatric thermal random stress, obeying white Gaussian statistics with the
following moments:
We denote by V (t) a connected volume which moves with the mass of the fluid.
The equation of continuity (A.54) is, simply,28
d
d V = 0. (A.59)
dt V (t)
(A.55) is29
6
d
v d V = f ext d V ( p1 ) n d S. (A.60)
dt V (t) V (t) V (t)
The surface integral on the right-hand side cancels with its neighbors because the
outward unit normal vector n of the neighboring volume is oriented opposite to that
of V (t).
We introduce further simplifying assumptions that the heat capacity and the ther-
mal conductivity of the environment are large enough that the temperature T is
kept constant. We define u by d u = p d 1 . This represents the change of the
Helmholtz free energy in the isothermal process. By taking the scalar product of
each term of (A.55) with the velocity v in the Stratonovich sense and also using
(A.54), the balance of the total energy of the fluid is 30
2
2
v v
+ u + + u v + ( p1 )v = : (v)+ f extv. (A.61)
t 2 2
0
28 The derivation uses the kinematical identity, dtd V (t) Ad V = V (t) A d V + V (t) A v n d S, and
0 t
the integral theorem of Gauss, V B n d S = V B d V. V (t), is the surface of the volume V (t),
and n is the unit outward normal vector on this surface.
29The derivation uses the kinematic identity mentioned before with (A.54), and also rewriting,
v (v) = (vv) + v v.
30 S.I. Sasa pointed out an early mistake of the author about the treatment of the terms including
. The symbol hereafter is not for Ito calculus, but forthe
scalar product. The symbol : of
A : B for symmetric any rank-2 tensors A and B means A B . The above derivation
uses (v) : = v : : v.
Appendix A 305
Hereafter, all the v in : (v) can be replaced by its symmetric deviatric part31
due to the deviatric and symmetric property of . The integral form of (A.61) is
d v2
+ u d V = : (v)d V + f ext vd V
dt V (t) 2 V (t) V (t)
6
( p1 ) : (v n)d A. (A.62)
V (t)
As in (A.60) the surface integral in the second line on the right-hand side cancels
with its neighbors.
Stochastic Energetics
The object is to rewrite (A.62) so that the structure of the energy balance, d E =
d Q + d W , is visible.
For the fluctuating fluid system, the system is fluid particle, or macroscopically
small connected element of mass that contains a large number of fluid molecules.
The local state of this system is characterized by the density field and the velocity
field v. The (local) thermal environment of the system consists, therefore, of the
degrees of freedom in the fluid less the fields and v.
When there is no mechanical support within the fluid, the momentum conserva-
tion, or the Galilee invariance, prohibits an isolated spontaneous force as a vector.
The random thermal force, therefore, takes the form of the force dipole (a sym-
metric tensor). This is the entity of s . The s arises as a result of coincidental
spatiotemporal local coherence in the molecular motions.
of (A.56) is the force dipole on the system from the environment. The two
terms on the right-hand side of (A.56), therefore, play the roles of ( ddtx ) and (t)
of the Brownian motion.
Let us regard V (t) as the volume occupied by a system, i.e., a fluid particle.
Equation (A.62) then describes the energy balance for the system. As mentioned
above, the surface integral in the second line on the right-hand side cancels with
its neighbors. The remaining terms look similar to the form of the energy balance,
d E = d Q + d W.
We should, however, remember the second remark about the work (Sect. 4.1.2.2):
If f ext is constant, for example, the work by the external system is 0, while f ext v
in (A.62) does not vanish. We will then rewrite (A.62) in the case that the external
force has a potential, i.e.,
31 1 [v
2
+ (v)t 23 ( v)1].
32 The derivation uses the aforementioned kinetic identity and (A.54).
306 Appendix A
d ext da
f ext v d V = ext d V + d V.
V (t) dt V (t) V (t) a dt
We finally have
v2 u
d + u d V = dt : (v)d V + da dV
V (t) 2 V (t) V (t) a
6
dt ( p1 ) : (v n)d A
V (t)
u
= dt ( ) vd V + da dV
V (t) V (t) a
6
dt pv nd A. (A.64)
V (t)
d 2 xi U ({x j }, a)
m = + Mi (t), (A.65)
dt 2 xi
d xi
Mi (t) = 0 G(xi x j ) M j + Ri (t), (A.66)
dt j(=i)
where 0 = 6 a with being the viscosity of the solvent and G represents the
hydrodynamic interaction (of Oseen and dipole types):
3 a x x 1 a 3 x x
G(x) = 1+ + 13 . (A.67)
4 x x x 2 x x x
Ri is the Gaussian Markov random force on the ith particle, with zero mean and
20 kB T (t t )1 (i = j)
Ri (t)R j (t ) = . (A.68)
20 kB T (t t )G(xi x j ) (i = j)
m d x
2
E= + U ({x j }, a),
i
2 dt
d Q = Mi d x i ,
i
U ({x j }, a)
dW = da. (A.69)
a
Here all the kinetic energy of the solvent is ignored.
34 A concise and clear introduction to the metric geometry is found in a book of Dirac [20]. Below,
we will also use g as the determinant of g and g as the inverse matrix of g .
308 Appendix A
where time is discretized and represented by the indices m and n, etc. As for the
multiplicative noise n , the limit of t 0 should be interpreted as of the Ito type.
That is, the random force n takes place between tn and tn+1 while xn is the value
at tn . (See, Examples 1 and 2 of Appendix Sect. A.1.6).
The above SDE is equivalent to the following FokkerPlanck equation for the
scalar density = g P with P being the probability density:
1 U
= gg + k B T . (A.72)
t x x g x g
For a being fixed, the canonical equilibrium distribution P eU/kB T is the station-
ary solution of the above FokkerPlanck equation.
We can obtain the energy balance, U = W + Q, by operating (t)1 g
(xm+1 xm ) on each term of (A.70)36 :
x xm kB T ( gg ) m
Q = m+1 + +
g (xm+1 xm )
t g(xm ) x t
1 U (xm+1 , am+1 ) U (xm , am )
= + (xm+1 xm ). (A.73)
2 x x
By direct expansion around (x, a) ( 21 (xm + xm+1 ), 12 (am + am+1 )), we can ver-
ify that error in the energy balance, U = W + Q, is O(t 3/2 ). Recall that
the error must be at most O(t 3/2 ) to assure the correct energetics (Sect. 4.1.2.5).
35Hereafter we
use Einstein summation convention for the Greek indices appearing pairwise;
A B implies A B .
36 On the curved space (Riemannian manifold), the first-order scheme for Brownian motion is
given in [21, 22], but we will need the order of t 1.5 accuracy. [cf. A.B. Cruzeiro, C. Alves,
Monte-Carlo simulation of stochastic differential systems a geometrical approach (preprint).]
Appendix A 309
Therefore, the above result assures the convergence with t 0 to the correct
energy balance over a finite time interval.
For illustration purpose, let us introduce a toy model that imitates (very roughly) the
growth of actin gel. See Fig. A.2:
1. The system consists of Hookean springs (wiggly lines) on the substrates (base
line), which are vertically joined by beads (open circles).
2. The external force a(t) is applied on the uppermost end of the system (filled
circle).
3. The insertion and removal of the springs occur only at the substrate (x = 0).
4. The positions of the beads are denoted by {x1 , . . . , xn } from top to bottom, where
n is also a stochastic variable. When there are n springs, xn+1 0.
5. The natural length i and the elastic stiffness ki of a Hookean spring are deter-
mined when the spring is inserted. They depend on the force on the lowermost
spring before the insertion (state-to-function information).
While {xi } specify the state of the gel, {i } and {ki } are the parameters of the
gels function as an elastic body. Upon the creation of (n + 1)th spring, the state
variable xn+1 , and other positions, are determined after the functional parameters,
(kn+1 , n+1 ), are specified (function-to-state information).
Through the loop of (state-to-function information) and (function-to-state infor-
mation), the growth of an open system carries the memory of its past state.37
x1
k1
x2
x1 xn 1
k1 kn 1
x1 x2 xn
k1 k2 kn
Fig. A.2 A simplified model of gel growth from the substrate. From the left to the right, there are
n = 1, 2, and n springs incorporated into the gel. When the (n + 1)th spring is inserted, not only
the variable xn+1 for a new joint appears, but also a new energy function, (kn+1 /2)(xn+1 n+1 )2 ,
with new parameters kn+1 and n+1 appear
37 The usage of the words state and function may not be proper since the state influences the
function also. Lambda calculus [23] describes object/subject duality.
310 Appendix A
Since there is no thermodynamic operation to change the parameters sup and Nsup ,
we can arbitrarily choose the additive constant in its free energy,
Nsup
Fsup /kB T Nsup ! eFtot /kB T E c /k T
e = e Nsup Ntot B d Nsup Ntot x, (A.75)
N =0
(N sup N tot )! ( sup \ tot )
tot
where Ftot is the Helmholtz free energy of the entire system, defined by
1
eFtot /kB T = eEtot /kB T d Ntotx, (A.76)
Ntot ! tot
We
approximate the integral (A.75) assuming ideal gas
E Nsup Ntot /kB T Nsup Ntot
(sup \tot ) e d x = (sup \tot A) Nsup Ntot , where A does not
depend on sup \tot or the number of particles. Using the approximation, Nsup !/
(Nsup Ntot )! Nsup Ntot , we have
Nsup
Ntot
Nsup
eFsup /kB T eFtot /kB T (sup \tot A) Nsup .
Ntot =0
sup \tot A
Identifying Ntot
sup \tot A
= e/kB T , we obtain
Nsup
e Fsup /kB T
e [Ftot Ntot ]/kB T
e(indep. Ntot ) .
Ntot =0
Appendix A 311
Although the last exponent is not extensive with respect to Nsup or to sup , there
is no causality because our operations will not change these parameters. We thus
justified (Ntot !)1 in the definition (A.76).
Ntot
1
eFtot /kB T = eF /kB T eE Ntot n /kB T d Ntot n x,
(n) c
n=0
(Ntot n)! ( c )
where F (n) has been defined in (5.22) in the main text. Repeating the same calcula-
tion as above, we obtain
Ntot
Ftot /kB T Ntot /kB T
e(F n)/kB T
(n)
e e .
n=0
By transferring the global factor on the right-hand side to the left-hand side, we
finally obtain
Ntot
Ftot Ntot
e(F n)/kB T .
(n)
exp (A.77)
kB T n=0
From this expression, (5.19) with the definition (5.21) is derived by taking the limit
of Ntot tot . The equilibrium probability Pn to find n particles in the
open system is found from (5.21), i.e.,
Pn = e(J F +n)/kB T
(n)
. (A.78)
The analysis of the transport is in some aspect analogous to that of the Berry phase
[24] in quantum physics. Since J [P (eq) ] = 0 as mentioned in the text, we need
to take into account the first-order correction in the time derivative, a(t), to this
probability:
P(x, t) = P (eq) (x, a(t)) + (x, a(t)) a (t) + O(|a|2 ), (A.79)
312 Appendix A
The average number of particles, N , that cross a fixed spatial point (e.g., at x0 ) to the
right per cycle of operation is given by the time integral of J [P]x=x0 . From (A.80)
this time integral can be written as an integral over a along the closed trajectory on
the parameter space [25]:38
6
N = J [ (x0 , a)] d a (quasistatic limit). (A.81)
{J [ (x0 , a)]} is not of the gradient form with respect to the parameter.39 Its line
integral along a closed loop generally, therefore, does not vanish. The result for
N is, therefore, generally nonzero. Although the integral of (A.81) looks purely
static, with no dependence on a, this effect is kinetic: J [P] includes the kinetic
coefficient .
P (eq)
= LP1 . (A.82)
t
The left-hand side of (A.82) is O((op )1 ). From the explicit form of P (eq) , this term
satisfies
P1 then depends on the time only through a(t) and its time derivative.
U (eq)
The average work W up to O((op )1 ) is given as W = a
(P + P1 )d x da.
The average irreversible work Wirr is then
t=op
U
Wirr = P1 d x da(t) + O((op )2 ). (A.84)
t=0 a
where L(a) operates on the x variable to the right. g depends on and a through
this equation. We require that g vanishes at the infinite boundaries.
Using the explicit form of the Gibbs distribution, and also the explicit form of
the FokkerPlanck operator, L(a), we can rewrite (A.85) as (5.40) in the text.
+O((op )2 ). (A.88)
Substituting the expression (A.83) for P (eq) /t, we have, after some rearrange-
ment, the symmetric form (5.37) in the text.
We will consider a particular realization of the process, which starts from the state
i 0 at time = 0 and jumps at time t from the state i 1 to the state i (1 n)
until = t, where t < t+1 .42 We denote by Prob[i( ), a( )] the path probability
for the realization of such history i( ), with the condition that the process starts
with the canonical equilibrium state corresponding to the given initial parameter,
eq
Pi0 (a(0)). The path probability is constructed as follows:
eq
(i 0 )
Prob[i( ), a( )] = Pi0 (a(0))M[0,t 1]
wi0 i1 (a(t1 ))M(i[t11,t) 2 ] wi1 i2 (a(t2 ))
42 Hereafter in this section, the in i( ) and a( ) implies being a generic time and i( ) means a
particular realization or a history.
Appendix A 315
M(i) (i)
[t ,t ] = M[tt ,tt ] . (A.91)
The path probability for the realization of the time-reversed process i( ) under the
time reversed protocol a( ), in condition that the process starts with the canonical
eq
equilibrium state corresponding to the given (reversed) initial parameter, Pin (a(t)),
is then given by
eq (i )
Prob[i( ), a( )] = Pin (a(t))M(i[tnn ,t]
)
win in1 (a(tn ))M[tn1
n1
,tn ]
Then we can use the detailed balance condition (5.62) to eliminate all the transition
rates:
eq eq eq eq
Prob[i( ), a( )] Pi (a(t1 ))Pi1 (a(t2 )) Pin1 (a(tn ))Pin (a(t))
= 0eq eq eq eq . (A.94)
Prob[i( ), a( )] Pi0 (a(0))Pi1 (a(t1 ))Pi2 (a(t2 )) Pin (a(tn ))
A.5.4.4 A lemma
We introduce R[i( ), a( )] through the definition,43
Prob[i( ), a( )]
eR[i( ),a( )] . (A.95)
Prob[i( ), a( )]
This quantity is a functional of the history, i( ), under a given protocol of the exter-
nal parameter, a( ), if we know how the right-hand side depends on i( ) and a( ).
At least we know the property,
which holds immediately from the definition (A.95). The probability distribution
of R[i( ), a( )] can be deduced through the average using the path probability,
Prob[i( ), a( )]:
Since the potential barrier is raised beyond U (0, a ), we can expect that P(x, t)
inside the barrier region is smaller than P eq (x, a ; T ).44 Then we can have an
upper bound of W [a ,af ] :
af
U (x, a)
W [a ,af ] P eq (x, a ; T ) d x da
a a
x
[U (0, af ) U (0, a )] P eq (x, a ; T )d x, (A.101)
x
There are very interesting cases where sys adjusts itself to approach and remains in
the proximity of op . Such cases are observed when a system contains the internal
feedback mechanism from the fluctuating variables (like x in (7.7)) to the control
parameters (like a or b, ibid). The case where sys approaches op from below,
will be referred as aging,45 while the case where sys approaches op from above,
A.7.2.1 Aging
Figure A.3 shows a typical situation of what we call aging. For each double well,
a particle (a black dot) is attached to a spring under tension. The tension favors
the thermally assisted transition from the left well to the right well. If the distance
44 We admit that it is not rigorous: The exception will occur if the tail of the barrier is too much
extended, and the particle excluded from the barrier top becomes stagnant in that tail region.
45 Aging can be defined in different ways depending on the context. Here we adopt the version of
P. G. de Gennes (Lecture at the College de France, 2002).
318 Appendix A
(a)
(b)
(c)
Fig. A.3 A schematic model of aging: Both the masses (black dots) and the double-well potential
are mobile. The time proceeds from (a) to (c). The relaxation time sys increases with the (obser-
vation) time t
between the two walls is fixed, the transition in one of the double wells leads to
the diminution of the tension on the springs. Thus the system remains in (b) for a
time longer than in (a) until the further transition to, for example, the state (c). In
this manner, as op proceeds with the time of observation, the transition is observed
when sys op is attained. We, therefore, have sys = t. As consequence, the ten-
sion of the spring, to which sys is exponentially related, decreases logarithmically
in time.
(a)
(b)
(c)
(d)
Fig. A.4 A schematic model of plastic flow: The time proceeds from (a) to (d)
References
1. K. Sato, Y. Ito, T. Yomo, K. Kaneko, PNAS 100, 14086 (2003) 282
2. L.F. Cugliandolo, J. Kurchan, L. Peliti, Phys. Rev. E 55, 3898 (1997) 283
3. H. Mori, Prog. Theor. Phys. 33, 423 (1965) 286
4. K. Kawasaki, J. Phys. A 6, 1289 (1973) 286, 287
5. S. Nordholm, R. Zwanzig, J. Stat. Phys. 13, 347 (1975) 286
6. R. Zwanzig, Phys. Rev. 124, 983 (1961) 287
7. R. Kubo, M. Toda, N. Hashitsume, Statistical Physics II: Nonequilibrium Statistical Mechan-
ics, 2nd edn. (Springer-Verlag, Berlin, 1991) 291
8. G. Milstein, Numerical Integration of Stochastic Differential Equations (Kluwer Academic,
Dordrecht, 1995) 292
9. H. Qian, J. Math. Chem. 27, 219 (2000) 293
10. T. Shibata, K. Sekimoto, J. Phys. Soc. Jpn. 69, 2455 (2000) 296
11. I. Prigogine, R. Defay (translated by D.H. Everett), Chemical Thermodynamics (Longmans,
Green & Co., London, 1954) 297
12. T.M. Cover, J.A. Thomas, Elements of Information Theory (John Wiley & Sons, Inc., New
York, 1991) 297
13. L.D. Landau, E.M. Lifshitz, Fluid Mechanics (Course of Theoretical Physics, Volume 6), 2nd
edn. (Reed Educational and Professional Publishing Ltd, Oxford, 2000) 303
14. E.J. Hinch, J. Fluid Mech. 72, 499 (1975) 306
320 Appendix A
15. P.N. Pusey, R.J.A. Tough, J. Phys. A: Math. Gen. 15, 1291 (1982) 306
16. M. Tokuyama, I. Oppenheim, Phys. Rev. E 50, R16 (1994) 306
17. M. Tokuyama, I. Oppenheim, Physica A 216, 85 (1995) 306
18. R. Mochizuki, Prog. Theor. Phys. 88, L1233 (1992) 307
19. G.G. Batrouni, H. Kawai, P. Rossi, J. Math. Phys. 27, 1646 (1986) 307, 308
20. P.A.M. Dirac, General Theory of Relativity (John Wiley & Sons, Inc., New York, 1975) 307
21. A. Cruzeiro, P. Malliavin, A. Thalmaier, C. R. Acad. Sci. Paris, Ser. I 338, 481 (2004) 308
22. A. Cruzeiro, P. Malliavin, Stochastic Process. Appl. 116, 1088 (2006) 308
23. H.P. Barendregt, The Lambda Calculus (Studies in Logic and the foundation of mathematics,
vol. 103), Revised edn. (Elsevier, Amsterdam, 1984) 309
24. M.V. Berry, Proc. Roy. Soc. London, A 392, 45 (1984) 311
25. J.M.R. Parrondo, Phys. Rev. E 57, 7297 (1998) 312
26. U. Seifert, J. Phys. A: Math. Gen. 37, L517 (2004) 314
27. D. Gillespie, J. Comput. Phys. 22, 403 (1976) 314
28. J.L. Lebowitz, H. Spohn, J. Stat. Phys. 95, 333 (1999) 316
29. G.E. Crooks, Phys. Rev. E 61, 2361 (2000) 316
30. C. Maes, On the Origin and the Use of fluctuation Relations for the Entropy: Poincare Sem-
inar 2003, ed. by J. Dalibard, B. Duplantier, V. Rivasseau (Birkhauser Verlag, Basel, 2004),
pp. 145191 315, 316
31. G.E. Crooks, J. Stat. Phys. 90, 1481 (1998) 315
32. T. Ooshida, K. Sekimoto, Phys. Rev. Lett. 95, 108301 (2005) 318
33. Y. Miyamoto, K. Fukao, H. Yamao, K. Sekimoto, Phys. Rev. Lett. 88, 225504 (2002) 318
34. K. Sekimoto, in Chemomechanical Instabilities in Responsive Materials (NATO Science for
Peace and Security Series A: Chemistry and Biology), ed. by P. Borckmans, et al., (Springer,
New York, 2009) 318
Index
A E
Adiabatic invariant, 259 Effective temperature, 283
Allosteric effect, 247 Einstein relation, 19, 22, 118
Allosteric transition, 247 Empirical average, 12, 14
Alternative access model, 271 Enthalpy-entropy (H -S) compensation, 90
Autonomous system, 266 Entropy, 71, 82, 173
Environment, 68
B Enzyme, 101
Buttiker-Landauer ratchet, 58, 162, 255 Equilibrium state, 70
Bidirectional control, 271 Essentially non-quasistatic processes, 223
Binomial distribution, 281 Exchange of binding, 247, 276
Bit memory, 231 Extensive variables, 68
Black box, 79, 192, 267 External system, 69, 78, 137, 140, 234, 270
Born-Oppenheimer approximation, 230
Brownian motion, 3 F
Butler-Volmer equation, 126 Feynman ratchet, 58, 156, 159, 255
First law, 70
C First passage time (FPT), 56, 57, 106, 249
Calorimetric heat, 203 Fluctuation theorem (FT), 198, 314
Central limit theorem, 13, 179 Fluctuation-dissipation (FD) relation, 4, 48,
Chemical coupling, 84 118, 172, 290
Chemical formula, 95 FokkerPlanck equation, 44
Chemical potential, 77, 85, 99 Fokker-Planck equation, 4
Chemo-mechanical coupling, 82 Forth law, 71
Common ancestor, 277, 278 Free Brownian motion, 17
Complete thermodynamic function, 72 Fundamental relation, 72
Cooperative binding, 247
Crossover of timescales, 225
Curie principle, 61 G
Gaussian white noise, 3
Generalized efficiency, 218
D
Detailed balance (DB), 47, 105, 113, 121, 127,
269 H
Diffusion, 3, 21 H-theorem, 46, 153, 198
Diffusion coefficient, 22, 53 Half-absence sensor, 242
Diffusion equation, 23 Half-presence sensor, 242, 262, 271
Discrete Langevin equation, 105, 109, 213 Heat, 70, 137, 146, 153, 169, 202
Double layer, 167 Heuns method, 43
321
322 Index
I P
Independent, 11, 16 Poisson distribution, 281
Independently and identically distributed Poisson noise, 109
(i.i.d.), 12 Potential theory, 295
Induced fit, 244 Probability density, 7
Intensive variables, 71 Probability flux, 45, 107
Ito type, 37 Projection operators, 26
Itos lemma, 4, 39 Protein friction, 23
J Q
Jarzynskis non-equilibrium work relation, Quasi-static, 57, 70, 177, 184
194, 195
R
K Random variable, 6
Kramers equation, 44 Ratchet models, 58, 61, 276
Kramers-Moyal expansion, 117 Rate constant, 96, 120
Kullback-Leibler distance, 46, 108, 239
S
L Second law, 71, 175, 261
Lagrange multiplier, 99 Self-averaging, 70, 179
Landauer and Bennett, 231
Steins lemma, 109
Landauers blow torch, 52
Steric interaction, 243
Langevin equation, 3, 26, 109
Stochastic calculus, 4
Large deviation property, 13
Stochastic differential equation (SDE), 4, 38
Law of large numbers, 12, 20, 179
Stochastic process, 4, 10
Legendre transformation, 72
Stratonovich type, 37, 138
Linear non-equilibrium thermodynamics, 81
Structure-function relationship, 277
Log-normal, 41
System, 67, 137, 165, 266
Long-time tail, 19, 216
Lyapnov, 46
T
M Thermal environment, 14
Margins of operation, 249, 274 Thermal random force, 3, 13
Markov approximation, 3, 31, 106 Thermodynamic function, 70
Markovian process, 31, 146 Thermodynamic variable, 70
Markovian property, 31 Thermophoresis, 51
Master equation, 107 Third law, 71
Maxwell relation, 74 To know, 236
Maxwells demon, 61, 69, 156 Transition rate, 106
Mesoscopic heat, 203
Michaelis-Menten, 102 V
Milstein scheme, 292 Vant Hoff correction term, 212
Momentum conservation, 216 Vant Hoff equation, 80
Viscous friction coefficient, 16, 126
N
Non-anticipating, 38, 111 W
Numerical error, 42 Wentzel-Kramers-Brillouin (WKB)
approximation, 230
O Wiener process, 35
Objectivity, 210 Work, 69, 140, 143, 179
Onsager coefficient, 161
Open system, 76, 82, 124, 126, 165, 181, 239, Z
251, 261 Zero-th law, 70