The Thermodynamics

of Small Driven Systems

Von der Fakultät Mathematik und Physik der Universität Stuttgart

zur Erlangung der Würde eines Doktors der
Naturwissenschaften (Dr. rer. nat.) genehmigte Abhandlung

vorgelegt von

Thomas Speck
aus Magdeburg

Hauptberichter: Prof. Dr. Udo Seifert

Mitberichter: Prof. Dr. Siegfried Dietrich

Tag der Einreichung: 31. Oktober 2007

Tag der mündlichen Prüfung: 14. Dezember 2007

II. Institut für Theoretische Physik der Universität Stuttgart

2007
Abstract
In the present thesis, a thermodynamic theory is developed for driven soft matter sys-
tems like polymer solutions and colloidal suspensions. Within this theory, notions like
the work applied to the system externally and the heat dissipated into a heat reservoir
will be defined consistently. Going beyond mean values, all quantities are defined along
a single trajectory. This thermodynamic approach is completed through a nonequilibrium
generalized entropy.
Recent theoretical progress has led to the formulation of a class of nonequilibrium fluc-
tuation relations holding in great generality beyond the well-understood linear response
regime. We derive these relations and show how they are connected to the thermody-
namic quantities introduced before. Since these quantities are defined on the level of
single trajectories, stochasticity of the system implies probability distributions for them.
The nonequilibrium fluctuation relations like the Jarzynski relation for the work then ba-
sically restrict these probability distributions. In order to obtain the still nonuniversal
distribution functions, we show how these functions can be obtained through deriving
equations of motion for the joint probability including the stochastic state of the system.
As a third issue, we show how Onsager’s principle leading to the fluctuation-dissipation
theorem may be generalized in nonequilibrium steady states. Explicitly, we derive the
nonequilibrium fluctuation-dissipation theorem for the stochastic velocity and show how
the equilibrium form of the fluctuation-dissipation theorem may be restored.

3
Contents

Abstract 3

Kurzfassung 9

Notations and Symbols 13

1 Introduction 15
1.1 The theoretical perspective: A short history . . . . . . . . . . . . . . . . . . . . 15
1.2 The experimental perspective: What are small systems? . . . . . . . . . . . . 18
1.3 Organization of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.4 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2 Dynamics 23
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 State space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.1 Continuous state space . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.2 Discrete state space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Stochastic processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.1 Markov processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.2 Non-Markovian processes . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4 Overdamped continuous dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.2 Smoluchowski equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4.3 Local mean velocity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.4 Langevin equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3 Thermodynamics of Small Systems 31

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Work and heat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.1 Quasi-static processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.2 An experimental study: The case of a single colloidal particle . . . . 33
3.3 Work and heat in the presence of nonconservative forces and external flows 35
3.3.1 An introductory example . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.2 General expressions for work and heat . . . . . . . . . . . . . . . . . . 37
3.4 Entropy production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

5
Contents

3.4.1 Equation of motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.4.2 Dissipation function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.5 Steady state thermodynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.6 An illustration: a dumbbell in simple shear flow . . . . . . . . . . . . . . . . . 42
3.7 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4 Probability Distributions of Work, Heat, and Entropy 45

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Nonequilibrium fluctuation relations: Path integral formalism . . . . . . . . . 45
4.2.1 Trajectory involutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.2.2 Connection to heat dissipation . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.3 Reversed dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.4 The relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3 Evolution of path functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.3.1 Generating function and fluctuation relations . . . . . . . . . . . . . . 53
4.3.2 Transition functional and work distribution . . . . . . . . . . . . . . . 54
4.3.3 Experimental work distribution . . . . . . . . . . . . . . . . . . . . . . . 55
4.3.4 Housekeeping heat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4 Extension to non-Markovian processes . . . . . . . . . . . . . . . . . . . . . . . 58
4.4.1 Substitute operator for a moving trap . . . . . . . . . . . . . . . . . . . 60
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5 Discrete Systems 63
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2 The system: A single defect center in diamond . . . . . . . . . . . . . . . . . . 63
5.3 The transition functional . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.3.1 Probability distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.4 Entropy production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6 Linear Response Theory 69

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2 Fluctuation-dissipation theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2.1 Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2.2 General Markov processes . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.3 The nonequilibrium velocity fluctuation-dissipation theorem . . . . . . . . . 72
6.3.1 Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.3.2 Restoring the equilibrium form and a violation function . . . . . . . . 74
6.3.3 Generalized Einstein relation . . . . . . . . . . . . . . . . . . . . . . . . 75
6.4 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

A Appendix 79
A.1 Projected processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

6
 Contents

A.2 Derivation of the Fokker-Planck equation . . . . . . . . . . . . . . . . . . . . . 80

A.3 Time evolution operator: The harmonic oscillator . . . . . . . . . . . . . . . . 82
A.4 Multiplicative noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
A.5 Derivation of the Smoluchowski equation . . . . . . . . . . . . . . . . . . . . . 84
A.5.1 Reduced time evolution operator . . . . . . . . . . . . . . . . . . . . . . 84
A.6 Evolution of time-local functionals . . . . . . . . . . . . . . . . . . . . . . . . . 86

Bibliography 89

Danksagung 97

7
Kurzfassung
In den letzten Jahren hat die Erforschung von Nichtgleichgewichtsphänomenen große
Fortschritte mit der Formulierung so genannter Fluktuationstheoreme oder auch Nicht-
gleichgewichts-Fluktuations-Relationen gemacht. Zum ersten Mal wurde ein Fluktua-
tionstheorem für die Entropieproduktion in einer simulierten gescherten Flüssigkeit, also
einem Nichtgleichgewichtssystem, gefunden. Zwei Besonderheiten traten bei diesem Com-
puterexperiment hervor. Zum einen gab es Simulationsläufe, bei denen sich die Teilchen
der Flüssigkeit kurzzeitig spontan ordneten, die Entropie also verringert wurde. Zum
anderen wurde gezeigt, dass eine negative Entropieproduktion nicht willkürlich auftritt,
sondern dass vielmehr die Wahrscheinlichkeitsverteilung p(σ) der Entropieproduktions-
rate σ die Beziehung p(−σ)/p(σ) ∼ e−σt im Limes langer Zeiten t erfüllt. Diese Rela-
tion fand viel Beachtung, da sie eine der wenigen, allgemein gültigen Aussagen über das
Nichtgleichgewicht jenseits des Linearen Response darstellt. Eine unmittelbare Folge der
Beziehung ist, dass der Mittelwert der Entropieproduktionsrate in Übereinstimmung mit
dem zweiten Hauptsatz der Thermodynamik immer nichtnegativ ist. Zentrales Thema
dieser Dissertation ist eine Reihe ähnlicher Relationen, welche aus einem allgemeinen
Zugang abgeleitet werden.
Während diese neuen Nichtgleichgewichtsrelationen generell gelten, so haben sie doch
vor allem Bedeutung für Systeme, in denen Fluktuationen eine essentielle Rolle spielen.
Dies sind vor allem kleine Systeme, wobei klein nicht notwendigerweise bedeutet, dass
das System nur aus wenigen Teilchen besteht oder von molekularer Größe ist. Vielmehr
ist entscheidend, dass die Energie, die mit der Umgebung als Arbeit oder Wärme aus-
getauscht wird, so klein ist, dass ihre Fluktuationen beobachtbar sind. So kann durch-
aus ein makroskopisches Objekt wie ein Torsionspendel als experimentelle Realisierung
eines kleinen Systems verstanden werden. In erster Linie zielt die zu entwickelnde The-
orie jedoch auf Systeme, die der weichen Materie zugeordnet werden können. Das sind
zum einen Moleküle mit biologischen Funktionen wie z.B. DNA- und RNA-Moleküle, Pro-
teine wie das Muskelprotein Titin oder auch Motorproteine, die für aktive Zellprozesse
zuständig sind wie Kinesin und Myosin; zum anderen sind das Polymere und Kolloide.
Besonders kolloidale Systeme nehmen einen wichtigen Platz in der experimentellen Über-
prüfung der neuen Nichtgleichgewichtsrelationen ein und ein Großteil der existierenden
Experimente arbeitet mit kolloidalen Partikeln. Systeme der weichen Materie, wie die
eben genannten, sind in der Regel in eine viskose, inkompressible Flüssigkeit suspendiert.
Diese Flüssigkeit hat eine wohldefinierte, konstante Temperatur und agiert in unserem
Zugang als das Wärmebad, deren Freiheitsgrade wir weder kontrollieren noch deren Dy-
namik explizit beschreiben können. Die Flüssigkeit muss nicht notwendigerweise ruhen,

9
Kurzfassung

sondern kann durch äußere Einwirkungen stationär oder auch zeitabhängig fliessen. Von
einem hydrodynamischen Standpunkt aus zeichnen sich Systeme der weichen Materie
durch eine niedrige Reynolds-Zahl aus, d.h. Trägheitseffekte sind vernachlässigbar und
die Dynamik ist überdämpft, es gibt also keine Bewegung ohne Kraft. Die Gesamtheit
aller Partikelpositionen oder allgemein der Freiheitsgrade wird als Konfiguration bezeich-
net.
In der vorliegenden Arbeit wird die überdämpfte Dynamik von Systemen der weichen
Materie als stochastisch angenommen und auf zwei Ebenen modelliert. Die zeitliche
Entwicklung der Wahrscheinlichkeitsverteilung wird durch die Smoluchowski-Gleichung
beschrieben. Als entscheidende Größe tritt in dieser die lokale mittlere Geschwindig-
keit der Teilchen auf, welche auf zwei Weisen physikalisch interpretiert werden kann.
Auf der einen Seite ist sie der Mittelwert der stochastischen tatsächlichen Geschwindig-
keit, wobei aber nur über die Untermenge an Trajektorien gemittelt wird, die durch eine
gegebene Konfiguration verlaufen. Andererseits muss ein Unterschied zwischen lokaler
mittlerer Geschwindigkeit und der Geschwindigkeit des äußeren Flusses durch eine Kraft
bedingt sein, welche sich aus drei Teilen zusammensetzt: einer konservativen Kraft auf-
grund eines Potentials, möglichen nichtkonservativen Kräften und der thermodynami-
schen Kraft, welche das Erreichen der Gibbs-Boltzmann-Verteilung als Gleichgewichtsver-
teilung sicherstellt. Eine alternative Beschreibung der Dynamik auf der Ebene einzelner
Trajektorien ist die Langevin-Gleichung. Hierbei wird die Interaktion mit dem Wärme-
bad durch einen stochastischen Kraftterm direkt modelliert. Für eine korrekte Beschrei-
bung der Dynamik in Gegenwart hydrodynamischer Wechselwirkungen muss diese sto-
chastische Kraft jedoch durch eine zusätzliche Drift vervollständigt werden.
Ausgehend von zwei Axiomen stellen wir in Kapitel 3 eine Thermodynamik kleiner
Systeme auf. Das erste Axiom definiert Arbeit als kontrollierbare Energieänderung des
Systems, welche von außen verursacht wird. Das zweite Axiom besagt, dass entlang
jedweder Trajektorie Energieerhaltung und damit der erste Hauptsatz der Thermody-
namik gilt. Die (negative) Differenz zwischen Änderung der potentiellen Energie und
der Arbeit wird dann als dissipierte Wärme, also nicht kontrollierte Energieänderung, in-
terpretiert. Im Gegensatz zur klassischen Thermodynamik ist das System nicht durch
einige wenige Zustandsvariablen charakterisiert, sondern durch seine Konfiguration. Eine
vollständige Beschreibung makroskopischer Systeme ist damit weder möglich noch sinn-
voll. Die Prozessgrößen Wärme und Arbeit werden für stochastische Systeme, wie wir
sie im Blick haben, zu Zufallsgrößen mit einer Verteilung. Für ein System, das aus dem
Gleichgewicht durch zeitabhängige Änderung eines Kontrollparameters getrieben wird,
demonstrieren wir dieses Konzept experimentell. Dafür messen wir die Trajektorie eines
einzelnen kolloidalen Teilchens, welches sich vor einer Glasplatte bewegt. Das Teilchen
wird von einer stark fokussierten Laserfalle auf eine eindimensionale Bewegung senkrecht
zur Glasplatte unter Einwirkung der Gravitation eingeschränkt. Eine zweite, defokussierte
optische Falle drückt dann das Teilchen zeitabhängig gegen die gleich geladene Glasplatte.
Im nächsten Schritt erweitern wir die Theorie auf nichtkonservative Kräfte sowie äußere
Flüsse. Während die Arbeit, die von den nichtkonservativen Kräften verrichtet wird, ein-

10
fach Kraft mal Weg ist, ist die Arbeit des äußeren Flusses schwieriger zu bestimmen. Aus-
gehend von dem Grundsatz, dass die Arbeit invariant unter einem zeitabhängigen Wech-
sel des Bezugssystems sein muss, finden wir als Generalisierung, dass nun die konvektive
Ableitung an die Stelle der partiellen Zeitableitung tritt. Vervollständigt wird unsere ther-
modynamische Beschreibung durch die Einführung einer generalisierten Systementropie
definiert auf der Ebene der Einzeltrakjektorien. Somit ist auch die Entropie eine Zufalls-
größe mit einer Verteilung. Die totale Entropieproduktion spaltet sich auf in einen Term,
der die Änderung der Systementropie beschreibt und einen Anteil, der in das umgebene
Medium, also die Flüssigkeit, dissipiert wird. Da das Medium eine wohldefinierte Tem-
peratur aufweist, kann der Zusammenhang zwischen dissipierter Wärme und Entropieän-
derung über die Clausius Formel hergestellt werden.
Als drittes Element diskutieren wir unseren Formalismus für stationäre Nichtgleichge-
wichtszustände. Einer phänomenologischen Theorie folgend wird in einem stationären
Nichtgleichgewichtszustand die Wärme dissipiert als Reaktion auf die Arbeit, die auf-
gewendet werden muss, um diesen Zustand aufrecht zu erhalten. Dieser Anteil der
Wärme, die Housekeeping-Wärme, muss dann in Übergängen zwischen stationären Zustän-
den unterschieden werden von der Überschusswärme. Auch für diese Größen leiten wir
explizite, trajektorienabhängige Ausdrücke her.
Das Kapitel 4 ist zweigeteilt. Im ersten Teil werden die bekannten, historisch be-
dingt verschieden hergeleiteten Nichtgleichgewichtsrelationen aus einem vereinigenden
Zugang abgeleitet. Grob zerfallen diese für eine Prozessgröße R (z.B. Arbeit oder Wärme)
in zwei Klassen, die integralen Relationen in der Form 〈e−R 〉 = 1, welche die mögli-
chen Verteilungen beschränken ähnlich der Normierungsbedingung, sowie ferner die de-
taillierten Relationen, p(−R)/p(R) = e−R , welche die Wahrscheinlichkeitsverteilung von
Ereignissen mit negativem R aus der von Ereignissen mit positivem R bestimmen. Es
wird gezeigt, dass die bekannten Relationen (und im Prinzip unendlich viele weitere)
alle aus dem Verhalten des antisymmetrischen Anteiles der Pfadwahrscheinlichkeit einer
Trajektorie unter Zeitumkehr folgen. Die Vielfalt der Relationen hat zwei Gründe: zum
einen haben wir die Freiheit, die Wahrscheinlichkeitsdichte des finalen Zustandes frei zu
wählen und zum anderen kann die Operation der Zeitumkehr auf den Prozess, die Dy-
namik des Systems, oder auf beide angewendet werden. Explizit gezeigt und diskutiert
werden allerdings nur die Relationen, die Aussagen über eine der im vorhergehenden
Kapitel eingeführten thermodynamischen Größen machen.
Die detaillierten Fluktuationsrelationen bestimmen zwar die Verteilung von Ereignis-
sen mit negativem R, allerdings ist die Verteilungsfunktion p(R ¾ 0) immer noch eine
nicht-universelle Funktion, über die wir mit den bisherigen Mitteln keine Aussagen tref-
fen können. Im zweiten Teil des Kapitels 4 wenden wir uns daher Bewegungsgleichungen
für die Verteilungsfunktion zu. Der gewählte Zugang kann wie folgt motiviert werden: Die
Dynamik des Systems wird beschrieben durch die Smoluchowski-Gleichung. Die zeitliche
Änderung einer Prozessgröße hängt nur ab von der momentanen Konfiguration des Sys-
tems. Wir betrachten daher die Verbundwahrscheinlichkeit, dass ein Wert von R in der
Zeit t akkumuliert wurde und sich das System zum Zeitpunkt t in einem bestimmten

11
Kurzfassung

Zustand befindet. In anderen Worten ist die Verbundwahrscheinlichkeit die Wahrschein-

lichkeit eines Wertes R für die Menge aller Trajektorien, die in einem bestimmten Zustand
enden. Für diese Verbundwahrscheinlichkeit lässt sich eine Bewegungsgleichung ableiten
und die Verteilung für R folgt dann durch Integration über alle möglichen Endzustände.
Die Bewegungsgleichung für die Verteilung erlaubt darüber hinaus, die verschiedenen
Fluktuationsrelationen zu beweisen. Explizit geben wir die Bewegungsgleichungen für
Arbeit und Housekeeping-Wärme an.
Als System, das von unserer Agenda abweicht aber trotzdem sehr interessante Eigen-
schaften aufweist, diskutieren in Kapitel 5 eine Fehlstelle in einem Diamant. Die Fehlstelle
kann für unsere Zwecke als ein diskretes Zwei-Niveau-System dargestellt werden. Das
Besondere an diesem System ist, dass die Entropieproduktion im Medium abstrakt defi-
niert ist, da ein Wärmebad mit einer definitierten Temperatur fehlt. Trotzdem können die
Konzepte, die wir bislang für Systeme in Kontakt mit einem Wärmebad entwickelt haben,
auch auf diskrete Systeme angewendet werden.
Die Lineare Response-Theorie ist ein zentraler Bestandteil der statistischen Mechanik
und wird im letzten Kapitel für Systeme der weichen Materie untersucht. Im Gleich-
gewicht verknüpft das Fluktuations-Dissipations-Theorem die Antwortfunktion einer Ob-
servablen auf eine kleine Störung mit der Ableitung der Korrelationsfunktion dieser Ob-
servablen mit ihrer durch die innere Energie konjugierten Größe. Das wohl bekannteste
Beispiel für ein Fluktuations-Dissipations-Theorem erhält man für die Geschwindigkeit
eines freien Teilchens, dessen integrierte Version zur Einstein-Relation führt. Diese ver-
knüpft die Mobilität des freien Teilchens mit der Diffusionskonstanten durch die Temper-
atur der Flüssigkeit.
Während sich die Antwortfunktion auch auf das Nichtgleichgewicht ausdehnen lässt, so
ist hingegen die physikalische Interpretation eines generalisierten Fluktuations-Dissipa-
tions-Theorems weitaus schwieriger. Der Grund dafür ist, dass nun die Korrelationsfunk-
tion nicht mehr mit der gleichgewichts-konjugierten Observablen, sondern mit dem ab-
strakten, generalisierten Potential gebildet wird. Wir zeigen, dass es dennoch möglich
ist, das generalisierte Fluktuations-Dissipations-Theorem für die Geschwindigkeit durch
physikalisch messbare Größen auszudrücken, und finden als neue konjugierte Größe die
relative Geschwindigkeit, also die Differenz zwischen aktueller stochastischer und lokaler
mittlerer Geschwindigkeit. Als Anwendung diskutieren wir die Verallgemeinerung der
Einstein-Relation auf einen stationären Nichtgleichgewichtszustand numerisch und ex-
perimentell.
Zusammenfassend haben wir in dieser Arbeit eine Theorie entwickelt, die als zentralen
Punkt Aussagen über die Wahrscheinlichkeitsverteilungen thermodynamischer Größen
wie Arbeit und Wärme in kleinen, getriebenen Systemen macht. Als essentielle Bausteine
dieser Theorie haben wir konsistente Definitionen von Prozessgrößen wie Arbeit und
Wärme sowie Entropie gegeben und ihre Wahrscheinlichkeitsverteilungen durch die An-
gabe von Bewegungsgleichungen bestimmt.

12
Notations and Symbols
α ................... discretization parameter
β ................... any scalar function; friction kernel
γ .................... control parameter
Γ .................... protocol
" .................... small parameter (small time step, etc.)
"˙ . . . . . . . . . . . . . . . . . . . . strain rate
µ, µ0 . . . . . . . . . . . . . . . . mobility tensor, bare mobility
ψ, ψs , ψeq , ψ" . . . . . . state (probability distribution), steady state, equilibrium state, per-
turbed state
Φ ................... nonequilibrium generalized potential
A ................... dynamical free energy
A, B, B . . . . . . . . . . . . . . general observable; conjugate observable, in nonequilibrium
D, D0 . . . . . . . . . . . . . . . diffusion tensor, bare diffusion coefficient
d, dα . . . . . . . . . . . . . . . . drift vector, spurious drift
f, F . . . . . . . . . . . . . . . . . . nonconservative, total force
F ................... equilibrium free energy
g .................... generating function
I .................... violation function
kB . . . . . . . . . . . . . . . . . . . Boltzmann’s constant
L̂, L̂s , L̂eq , λ . . . . . . . . . time evolution operator, in a steady state, in equilibrium; their
eigenvalues
m ................... conditional moment; mass
m ................... mean displacement
N ................... number of particles
p .................... momentum
p, ρ . . . . . . . . . . . . . . . . . probability, joint probability distribution
P ................... transition probability
P ................... path probability
P̂ . . . . . . . . . . . . . . . . . . . projecting operator
Q, Q hk , Q ex . . . . . . . . . . total, housekeeping, excess heat
r .................... position
R, r . . . . . . . . . . . . . . . . . skew-symmetric part of the action, its explicit value
R, Req . . . . . . . . . . . . . . . response function, in equilibrium
∆s, ∆sm , ∆stot . . . . . . . change of system, medium, total entropy
t .................... time

13
Notations and Symbols

τ .................... time scale (e.g., relaxation time )

T , kB T . . . . . . . . . . . . . . temperature, thermal energy
U ................... inner energy
Û . . . . . . . . . . . . . . . . . . . propagator
v, vs , δv . . . . . . . . . . . . . local mean velocity, in a steady state; relative velocity
W ................... work
W, Ws . . . . . . . . . . . . . . dissipation function, in a steady state
x .................... microstate (collection of particle positions)
X ................... trajectory
Y ................... transition functional
Z ................... partition function

14
1 Introduction
Thermodynamics is a branch of physics which emerged in the early 19th century in con-
junction with the development of steam engines [1]. Classical thermodynamics is a phe-
nomenological theory describing processes in macroscopic systems composed of many
particles on the order of the Avogadro constant 1023 . Somewhat counter-intuitively, this
large number allows to describe macroscopic systems at equilibrium by a set of relatively
few state variables like pressure, energy, temperature, etc. Later, the microscopic founda-
tions of thermodynamics have been laid by the work of Boltzmann, giving thermodynam-
ics a probabilistic interpretation based on the motion of molecules and atoms and leading
eventually to the formulation of equilibrium statistical mechanics [2].
In contrast to equilibrium, our understanding of nonequilibrium is still far from com-
plete despite constant theoretical progress. Of course, in equilibrium there are still open
questions left to be answered, but from a fundamental point of view the Boltzmann-Gibbs
distribution gives a complete description of any system in equilibrium with a large heat
reservoir. In the sense of this universality, a comparable theory for nonequilibrium systems
is lacking and may even, in full generality, never arise.
For a certain class of small driven systems as reviewed in [3, 4, 5, 6], however, unex-
pected and significant progress occurred during the last decade with the formulation of
exact relations like the fluctuation theorem [7, 8, 9, 10], the Jarzynski relation [11], and
various extensions [12, 13, 14, 15, 16], all holding beyond the well understood linear
response regime. Characteristically, these systems consist of a few degrees of freedom em-
bedded in a heat bath of well-defined temperature. External mechanical forces or flows
drive these few degrees of freedom out of their equilibrium while still in contact with an
equilibrium bath of well-defined temperature. The common theme of all these nonequilib-
rium relations is that they apply to whole probability distribution instead of just to mean
values. This shows that fluctuations play a prominent role and we only have the chance
to verify and use these relations in what we will call small systems.

1.1 The theoretical perspective: A short history

The study of thermal fluctuations dates back to the work of Einstein on Brownian mo-
tion [17] in 1905, where the first fluctuation-dissipation relation for a sedimenting sus-
pension under the influence of gravitation was given. Einstein’s work revealed that (Gaus-
sian) fluctuations determine essential properties of equilibrium systems like transport
coefficients, e.g., the mobility of a Brownian particle. Transport coefficients have been

15
1 Introduction

Figure 1.1: Scanning electron micrograph image of titania colloidal beads. Colloids are
mesoscopic particles which are suspended in a liquid with typical particle sizes between
10 and 1000 nm. They have developed into an ideal system to experimentally test new
nonequilibrium relations. (Picture from [Link]

calculated first by Nyquist for linear electrical circuits [18] and then extended by On-
sager [19, 20] to thermal systems through the formulation of the fluctuation-dissipation
theorem. Since then, the focus has gradually shifted towards nonequilibrium systems.
However, nonequilibrium concerns a much wider range of phenomena compared to equi-
librium and its state is far from complete lacking a universal description comparable to
the Boltzmann-Gibbs distribution.

In 1993, Evans, Cohen, and Morriss gave the study of nonequilibrium systems a new
turn [7]. They investigated a simple two-dimensional sheared fluid numerically [21]. In
these simulations, the Newtonian equations of motion for all particles of the fluid are
integrated. Driving the fluid would lead to an increasing kinetic energy and additional
constructs, called thermostats, have to be used. These thermostats constrain, e.g., the ki-
netic energy of all particles and keep the so defined temperature fixed through introducing
additional degrees of freedom. The phase space volume then is not conserved anymore
and its contraction rate can be related to the entropy produced in one or more thermostats
mimicking thermal baths. Two things were observed. First, sometimes a negative entropy
production was found, i.e., entropy in the bath was annihilated. Second, these negative

16
 1.1 The theoretical perspective: A short history

events do not occur randomly but they obey the symmetry

p(−σ)
∼ e−σt , (1.1)
p(+σ)

where σ is the phase space contraction rate and p(σ) its probability distribution. The
symbol ∼ indicates that the ratio approaches the right hand side in the long time limit.
Two years later, Gallavotti and Cohen placed the fluctuation theorem on solid mathe-
matical ground [8]. They proved it under certain conditions, particularly the chaotic hy-
pothesis. Loosely speaking, the chaotic hypothesis assumes that a nonequilibrium steady
state is a hyperbolic attractor having both positive and negative Lyapunov exponents.
Regarding this hypothesis as a technical ingredient, is was conjectured [22] that the fluc-
tuation theorem holds more generally. Kurchan gave the first hints how the fluctuation
theorem arises for stochastic dynamics [9] and how it is connected to the entropy produc-
tion in these systems. In a seminal paper, Lebowitz and Spohn gave the proof for general
Markov processes and discussed several systems governed by stochastic dynamics [10].
Stochastic systems are easier to handle in the sense that technical difficulties arising in
deterministic dynamics like ergodic issues and the nonexistence of a stationary measure
on strange attractors are bypassed.
Another branch of development started with the formulation of the Jarzynski rela-
tion [11] showing that fluctuation relations can be found also for other quantities than
the entropy production. The Jarzynski relation

〈e−W /T 〉 = e−∆F/T (1.2)

turns the second law

〈W 〉 ¾ ∆F (1.3)

for a process connecting two equilibrium states with a free energy difference ∆F into
an equality. Whereas the equality in (1.3) only holds for quasistatical processes, the
equation (1.2) is valid for any process and implies (1.3) through the Jensen inequal-
ity 〈exp(x)〉 ¾ exp(〈x〉). The brackets 〈· · ·〉 denote the average over the correspond-
ing probability distribution. A relation similar to (1.2) has been found already in the
70s [23, 24], albeit invoking a different definition of the work [25]. Crooks has then
derived a symmetry related to the fluctuation theorem (1.1) based on microscopic re-
versibility [12, 13]. The Jarzynski and Crooks relations have received the largest atten-
tion among the nonequilibrium fluctuation relation due to their practical importance. In
particular, they can be used to recover the free energy profile of some reaction coordi-
nate [26, 27, 28, 29, 30, 31, 32]. To assess their practical applicability in experiments and
numerical studies, convergence and statistical properties have been studied [33, 34].
The Hatano-Sasa relation [35] has conceptually advanced the study of nonequilibrium
steady states as it provides a bridge to steady state thermodynamics. Steady state ther-

17
1 Introduction

modynamics [36] is a phenomenological theory which promotes the splitting of the dissi-
pated heat Q = Q hk + Q ex into the housekeeping heat Q hk needed to maintain a nonequi-
librium steady state and the excess heat Q ex . From the Hatano-Sasa relation

〈e−Q ex /T −∆Φ 〉 = 1, (1.4)

a generalized second law

T ∆S ¾ −〈Q ex 〉

can be derived through the Jensen inequality. Here, ∆S = 〈∆Φ〉 is the change of (Shan-
non) entropy due to a transition between nonequilibrium steady states.
The short survey given above is far from complete and represents a highly personal
point of view as to which theoretical papers have contributed most to this exciting field.
The list of literature by now is large and continuously growing. Unfortunately, an all-
encompassing review has not yet been written and so for further information and refer-
ences, the interested reader is referred to the reviews [3, 5, 6, 37].

1.2 The experimental perspective: What are small systems?

Up to now we said that we aim to study small systems. This is not a very precise term and
the need arises for a more accurate definition. F. Ritort answers this question in [38] with
the definition:
Small systems are those in which the energy exchanged with the environment is a few times
kB T and energy fluctuations are observable. A few can be 10 or 1000 depending on the
system. A small system must not necessarily be of molecular size or contain a few number of
molecules.
Examples of small systems in this sense include single colloidal particles but also colloidal
suspensions (figure 1.1), motor proteins (figure 1.2), and large molecules like DNA, Titin,
or proteins (figure 1.3) to name just the most important ones. However, also apparently

Figure 1.2: A paradigm for small driven sys-

tems are biological systems like motor proteins.
The intracellular aqueous solution provides an en-
vironment with constant, well-defined tempera-
ture. Sketched is a possible experiment exerting
a force on a kinesin molecule which walks along
a microtubule. (Picture from [Link]
[Link])

18
 1.2 The experimental perspective: What are small systems?

Figure 1.3: Mechanical ma-

nipulation of small systems
has become feasible through
the development of optical
traps and the atomic force
microscope. An example
are single molecule pulling
experiments: (A) stretching
DNA, (B) unzipping DNA,
(C) mechanical unfolding of
RNA, and (D) mechanical
unfolding of proteins. (Fig-
ure taken from [38].)

macroscopic systems like a torsion pendulum may be regarded as small system and can
be used to verify the fluctuation theorem [39].
Direct manipulation of single molecules and colloids in experiments became feasible
with the advent of two devices, the atomic force microscope [40] and the optical tweez-
ers [41, 42]. Especially colloidal particles driven by time-dependent laser traps [43, 44,
45, 46] have become the ideal systems for quantitatively studying nonequilibrium fluc-
tuation relations experimentally for essentially two reasons. First, individual trajectories
can be recorded allowing thus to compute the probability distribution of observables in
contrast to ensemble quantities typically obtained in scattering experiments. Second, even
though the particles are driven into a genuine nonequilibrium state, the surrounding fluid
still faithfully behaves like an equilibrium bath. Consequently, a large number of exper-
imental tests have used single colloidal particles. The first test was reported by Wang et
al [43]. It claimed to demonstrate the fluctuation theorem for the entropy production
but closer inspection shows that it really demonstrated the Jarzynski relation and the
Crooks symmetry for the work. Later tests include [44, 46]. The Hatano-Sasa relation
was demonstrated successfully also using a colloidal particle [45].
The probably most exciting perspective is offered through the understanding and mod-
eling of biological systems. These systems are generically out of equilibrium. Still, for
most processes in cell biology taking place on the level of a single (or few) molecules, the
intracellular aqueous solution provides an environment with constant temperature. The
genuine source of nonequilibrium are not temperature gradients but rather mechanical
or chemical stimuli provided by external forces or imbalanced chemical reactions. Hence,
single molecule experiments and experimental studies of biomolecules like proteins or
DNA can be regarded as a paradigm for small systems. Indeed, the first real world test of

19
1 Introduction

Figure 1.4: Experimental test of the Jarzynski relation (1.2). Left: A molecule of RNA
is attached to two beads and subjected to reversible and irreversible cycles of folding
and unfolding. A piezoelectric actuator controls the position of the bottom bead, which,
when moved, stretches the RNA. An optical trap formed by two opposing lasers captures
the top bead. The difference in positions of the bottom and top beads gives the end-
to-end length of the molecule. The blowup shows how the RNA molecule (green) is
coupled with the two beads via molecular handles (blue). The handles end in chemical
groups (red) that can be stuck to complementary groups (yellow) on the bead. (Figure
and caption taken from [4], the described experiment is [47].) Right: Typical unfolding
(orange) and refolding (blue) force extension curves for the RNA hairpin from a similar
experiment [30]. The blue area under the curve represents the work returned as the
molecule switches from the unfolded to the folded state.

the Jarzynski relation has been the recovering of the free energy profile of RNA molecules
through recording the work in stretching experiments [47], see also figure 1.4. Other ex-
perimental and numerical studies on specific systems have followed. The Crooks relation
has been tested as well [30]. Directed transport in noisy environments such as cells is also
studied in the context of ratchets [48, 49, 50, 51] and Brownian motors [52, 53].
Since their first formulation, nonequilibrium fluctuation relations have been anticipated
to play an important role in our understanding of biological processes, in the design of
devices on the nanoscale, and in the development of a thermodynamical theory for finite,
small systems. The motivation of this thesis is to sketch such a theory for systems governed
by stochastic dynamics and embedded into an environment of constant temperature.

1.3 Organization of the thesis

The thesis is organized as follows. The three building blocks of the thermodynamics of
small driven systems are: a consistent formulation of work, heat, and entropy introduced

20
 1.4 Publications

in chapter 3, the discussion of the nonequilibrium relations in chapter 4 and the linear
response theory of chapter 6. But first, in order to lay the foundations, we define the
stochastic dynamics in chapter 2 of the systems we aim to study. We introduce the time
evolution operator in a very general way and then specialize to overdamped Markovian
dynamics. Our basic model throughout the thesis will be the Smoluchowski equation.
Only in chapter 5 we will turn off the main road and study the discrete dynamics of a
two-level system whose dynamics is governed by the master equation.
The appendix is organized separately. It contains the more complex calculations and is
ordered such that single calculations build upon former results derived in the course of the
appendix. The appendix does not depend on definitions nor results derived in the main
text. The main text uses and references equations and relations obtained in the appendix,
however, these references do not occur in the same order as in the appendix.
Throughout most parts of the thesis, we will work with dimensionless quantities. For
example, we assume that all lengths have been made dimensionless through choosing
an appropriate length scale. In particular, probability distributions are real numbers. The
Boltzmann constant kB is set to unity, making the entropy dimensionless. The temperature
T is measured in units of the thermal energy kB T . The only exception are values for
experimental parameters and measured quantities, which we give with the proper units.

1.4 Publications
Part of the work used to compile this thesis has been published previously:

• “The Jarzynski relation, fluctuation theorems, and stochastic thermodynamics for

non-Markovian processes”
T. Speck and U. Seifert, J. Stat. Mech. L09002 (2007)

• “Einstein relation generalized to nonequilibrium”

V. Blickle, T. Speck, C. Lutz, U. Seifert, and C. Bechinger, Phys. Rev. Lett. 98, 210601
(2007)

• “Measurement of Stochastic Entropy Production”

C. Tietz, S. Schuler, T. Speck, U. Seifert, and J. Wrachtrup, Phys. Rev. Lett. 97,
050602 (2006)

• “Restoring a fluctuation-dissipation theorem in a nonequilibrium steady state”

T. Speck and U. Seifert, Europhys. Lett. 74, 391 (2006)

• “Thermodynamics of a Colloidal Particle in a Time-Dependent Nonharmonic Poten-

tial”
V. Blickle, T. Speck, L. Helden, U. Seifert, and C. Bechinger, Phys. Rev. Lett. 96,
070603 (2006)

21
1 Introduction

• “Integral fluctuation theorem for the housekeeping heat”

T. Speck and U. Seifert, J. Phys. A: Math. Gen. 38, L581 (2005)

• “Experimental Test of the Fluctuation Theorem for a Driven Two-Level System with
Time-Dependent Rates”
S. Schuler, T. Speck, C. Tietz, J. Wrachtrup, and U. Seifert, Phys. Rev. Lett. 94,
180602 (2005)

22
2 Dynamics

2.1 Introduction
The principal difference between a system in equilibrium with a large heat reservoir and
a system driven into nonequilibrium is dynamics. In equilibrium, the probability to find
the system in a specific configuration, called the microstate, is solely determined by the
energy associated with this microstate. This is achieved through the Gibbs-Boltzmann
distribution, the basic relation of equilibrium statistical mechanics. Quite in contrast,
nonequilibrium systems must be characterized through their dynamics. Consequentially,
in our approach the basic equation

∂ t ψ(t) = L̂(t)ψ(t) (2.1)

describes the time evolution of a, by now abstract, state ψ(t) which is governed by the
operator L̂. Throughout the thesis, we differentiate between the state of the system con-
taining global information and the microstate the system actually occupies. The evolution
operator may depend on time but it is independent of ψ itself and hence (2.1) is a linear
equation. The status of this equation can be compared to the Schrödinger equation in
quantum mechanics.

2.2 State space

2.2.1 Continuous state space

For continuous systems, we aim to describe the dynamics of d degrees of freedom which
we collect in the microstate x ≡ {x 1 , . . . , x d }. The set of all microstates forms the state
space. The state function ψ(x, t) is nonnegative everywhere and normalized,
Z
ψ(x, t) ¾ 0, dx ψ(x, t) = 1. (2.2)

We distinguish between:
Deterministic dynamics
In the case of deterministic dynamics of an isolated system, the time evolution operator
becomes the Liouville operator which can be stated as L̂· = {H, ·}, where H is the system’s

23
2 Dynamics

Hamiltonian and {·, ·} denote the Poisson brackets [54]. Hence, L̂ is a differential operator.
The state space is the phase space of the system and the state function ψ(x, t) corresponds
to the phase space density.
For driven systems some constraints, e.g., in order to keep the kinetic energy fixed,
have to be imposed. This is achieved through so called thermostats [21]. Although these
systems could be included in proofs of nonequilibrium relations based on (2.1), we will
not investigate thermostated systems in detail.

Stochastic dynamics
If a small system is coupled to a heat bath then one possibility to tackle the evolution of
the small system is to regard its dynamics as stochastic. The state ψ(x, t) then emerges
as the probability density of a certain microstate and we simply call ψ(x, t) the distribu-
tion function. This kind of dynamics is often called Brownian dynamics or motion. The
picture behind this name is that of particles immersed in a fluid undergoing a “zig-zag”
motion. However, and although we will use single colloidal particles both as motivation
and to demonstrate our theory, the methods we will develop are much more powerful
and not restricted to particles. The operator L̂ is again a differential operator but takes
into account the interactions with the heat bath explicitly. For general stochastic pro-
cesses, the evolution operator is called the Fokker-Planck operator [55]. However, for
the overdamped dynamics of colloidal particles and polymers it is more common to call
it the Smoluchowski operator [56, 57]. The set of all microstates in this case forms the
configuration space and a single microstate is also called a configuration.

2.2.2 Discrete state space

Stochastic dynamics can also be formulated on discrete state spaces where the system
changes its microstate with a rate w(x → x 0 ). This is in a way more abstract since no heat
bath is required as the source of randomness. The state ψ(x, t) is now the probability of
the microstate labeled x and it is subject to the constraints
X
ψ(x, t) ¾ 0, ψ(x, t) = 1.
x

The dynamics are governed by the master equation

X X
∂ t ψ(x, t) = L x x 0 ψ(x 0 , t), L x x 0 = w(x 0 → x) − δ x x 0 w(x → x 00 ), (2.3)
x0 x 00

where the entries L x x 0 of the stochastic matrix are determined by the hopping rates with
w(x → x) = 0. The first term quantifies the “gain” and the second term the “loss”.

24
 2.3 Stochastic processes

2.3 Stochastic processes

When the system evolves in time, the microstate changes randomly. The chronological
sequence of microstates

X ≡ {x(t) : t 0 ¶ t ¶ t 1 } (2.4)

is then called a trajectory. On the level of the evolution of the state ψ, stochastic processes
can be classified as either Markovian or non-Markovian.

2.3.1 Markov processes

For any stochastic process there exists a propagator Û(t, t 0 ) which propagates an initial
state ψ(t 0 )

ψ(t) = Û(t|t 0 )ψ(t 0 ) (2.5)

towards a later time t ¾ t 0 with Û(t|t) = 1. The propagator determines the transition
probability through

P(x, t|x 0 , t 0 ) = Û(t|t 0 )δ(x − x 0 ).

The stochastic process is called a Markov process if its propagators form a semi-group with
property

Û(t|t 0 ) = Û(t|t 00 )Û(t 00 |t 0 ). (t ¾ t 00 ¾ t 0 ) (2.6)

In the present context, this equation is known as the Chapman-Kolmogorov equation. The
time evolution operator is then the infinitesimal generator of the semi-group,

∂
L̂(t) = Û(t 0 |t) . (2.7)
∂ t0 t 0 =t

The most important property of Markov processes is that they have no memory. Using the
semi-group property (2.6), we find as equation of motion for the propagators

∂ t Û(t|t 0 ) = ∂ t Û(t|t 00 )Û(t 00 |t 0 ) t 00 =t

= L̂(t)Û(t|t 0 ),

which does not depend on times before t 0 . Hence, all informations about the process is
already contained in the generator L̂. In particular, we can construct the joint two-point
distribution

ψ(2) (x, t; x 0 , t 0 ) = P(x, t|x 0 , t 0 )ψ(x 0 , t 0 )

and following the same route, higher n-point distributions can be constructed from the
transition probability alone.

25
2 Dynamics

The basic properties of the operator L̂ can already be discussed in a general way even
though we do not know it explicitly. To this end we realize that the states {ψ} form a real
vector space with 1-norm.1 Its dual space is the vector space of linear functionals

ψ̄[ψ] ≡ 〈ψ̄, ψ〉 7→ R

mapping the state ψ to a real number. We introduce the adjoint operator through

〈ψ̄, L̂ψ〉 = 〈 L̂ † ψ̄, ψ〉.

A bi-orthonormal set {ϕ̄k , ϕk } must fulfill both the orthonormal and the completeness
condition,
∞
X
〈ϕ̄k , ϕl 〉 = δkl , ϕ̄k (x)ϕk (x 0 ) = δ(x − x 0 ),
k=0

respectively. For discrete systems, the δ-function in the last equation should be replaced
by the Kronecker symbol δ x x 0 .
We can map a system with continuous configuration space onto a discrete state space
through expanding a state into the series
n
X
ψ(t) = ck (t)ϕk
k=0

with n coefficients ck (t) = 〈ϕ̄k , ψ(t)〉. An operator is then represented by a n × n matrix,

L kl (t) = 〈ϕ̄k , L̂(t)ϕl 〉.

Although the continuous system corresponds in principle to an infinite-dimensional state

space, we assume that it is well approximated by a large but finite number n. This allows
us to make the following statements. For a constant operator L̂, we define its eigenstates
through

L̂ϕk = −λk ϕk

with sorted eigenvalues λ0 < |λ1 | ¶ |λ2 | ¶ . . . The lowest eigenvalue λ0 = 0 is unique
and corresponds to the stationary state. This property is a consequence of the Perron-
Frobenius theorem [58]. Higher eigenvalues can be complex numbers but with positive
real part such that the corresponding eigenstates relax on time scales (ℜλk )−1 and there-
fore ψ(t) approaches the stationary state ϕ0 in the long-time limit. The smallest nonzero
eigenvalue λ1 defines the relaxation time τ ≡ 1/|λ1 |. If we exclude the stationary state
1
We choose the 1-norm because it is compatible with the normalization condition (2.2).

26
 2.4 Overdamped continuous dynamics

then we can assume that the inverse operator L̂ −1 exists. Further, the inverse is bounded
with operator norm

k L̂ −1 k ≡ sup k L̂ −1 ψk.
kψk=1

Since we have chosen the 1-norm, we can calculate the norm easily in the diagonal matrix
−1
representation L kl = −(1/λk )δkl , leading to
X
k L̂ −1 k = max −1
|L kl | = τ.
l
k=1

In appendix A.3, the eigenvalue spectrum of a simple operator L̂ is calculated explicitly.

In section A.1, the existence and boundness of L̂ −1 are the crucial ingredients for a time
scale separation.

2.3.2 Non-Markovian processes

Markov processes constitute a special, albeit the most studied, class of stochastic pro-
cesses. In general processes, the transition probability depends on the history before t 0
and the propagators do not form a semi-group. There is no general theory for these
non-Markovian processes. However, in section 4.4 we will show that in some cases the
evolution equation (2.1) can be generalized even to non-Markovian processes.

2.4 Overdamped continuous dynamics

The main focus in this thesis lies on soft matter systems composed of N particles at po-
sitions rk . We will neglect inertial effects which are only important at short time scales,
typically less than 10−7 s in soft matter systems. The microstate x = (r1 , . . . , rN ) there-
fore is the collection of all particle positions. Soft matter systems like colloidal particles
and polymers are usually immersed in an incompressible fluid acting as a heat reservoir
with well-defined temperature T . In addition, the fluid may flow according to an imposed
velocity profile u(r) caused by external stirring.

2.4.1 Notation
We first fix the notation used throughout the thesis and, in particular, in the appendix. To
ease the notational overhead, we will use the following operations on vectors a = (ak )
and matrices A = (Akl ): The dot product contracts vectors
X X
a·b≡ ak bk , a · A ≡ ak Akl = A T a,
k k

27
2 Dynamics

and the double-dot product contracts matrices

X
A:B≡ Akl Bl k = Tr(AB).
kl

For two matrices, the double-dot product is a shortcut for the trace of the matrix product.
The superscript T denotes the transposed matrix. Indices as in rk indicate the particle
number. We follow the convention to sum over same indices.

2.4.2 Smoluchowski equation

The time evolution of the probability density ψ(x, t) is governed by the Smoluchowski
equation [56, 57]
∂ t ψ + ∇k · (vk ψ) = 0. (2.8)
Equation (2.8) is written in the form of a continuity equation in order to preserve probabil-
ity which is demanded by the normalization constraint (2.2). The Smoluchowski equation
balances the local change of the probability with the divergence of the current vk ψ. Any
deviation of the velocity
vk = u(rk ) + µkl Fl = u(rk ) + µkl (−∇l U + fl − T ∇l ln ψ) (2.9)
from the imposed flow profile u(r) has to be caused by a force Fl exerted on the l th particle.
This force may lead to a distortion of the flow which couples back to the other particles.
These hydrodynamic interactions are taken into account through the symmetric mobility
tensors µkl (x). The mobility tensors are positive semidefinite,
rk · µkl rl ¾ 0, (2.10)
for any vectors rk . Moreover, we assume that the inverse tensors µ−1
km
defined through
µ−1 µ = 1δkl
km ml

exist.
The force Fk stems from three sources. First, the gradient of the potential energy U(x).
This potential energy can be split into an inner energy due to interactions and the contri-
bution of external potentials. Second, the system can be driven through nonconservative
forces fk which cannot be written as gradient of a potential. And third, we have to take
into account a “thermodynamic” force arising from the stochastic interactions between
system and the surrounding fluid. In equilibrium, i.e., in the absence of external flows
and nonconservative forces, detailed balance must hold which amounts to vk = 0. The
thermodynamic force then ensures that
∇k (U + T ln ψ) = 0
leads to the correct equilibrium distribution ψeq ∼ e−U/T , which is of course the Gibbs-
Boltzmann distribution.

28
 2.4 Overdamped continuous dynamics

x Figure 2.1: The local mean velocity v as deter-

v
mined in (2.12). It is the stochastic particle velocity
ẋ averaged over the subset of trajectories passing
through a certain microstate, here x in one dimen-
t sion.

2.4.3 Local mean velocity

The velocity vk defined in (2.9) has an immediate physical meaning as a local average.
In equilibrium, the joint probability Ψ0 (x, x p ) = ϕMB (x p )ψeq (x) of particle positions and
momenta x p ≡ (p1 , . . . , pN ) is the product of the Maxwell-Boltzmann distribution (A.19)
and the equilibrium distribution ψeq . If we drive the systems, position and momentum
become correlated. In appendix A.5, we calculate the deviation Ψ1 from the equilibrium
case in the overdamped limit. With the operator expression (A.24), it can be written as

Ψ1 ' − L̂0−1 L̂1 Ψ0 = T −1 pk · vk Ψ0 .

The joint probability therefore is

Ψ(x, x p , t) = [1 + T −1 pk · vk ]Ψ0 (x, x p , t). (2.11)

We define the mean local velocity as the conditional mean

R
dx p (pk /m)Ψ(x, x p , t)
〈ṙk (t)|x〉 = = vk (x, t), (2.12)
ψ(x, t)
which reduces to vk after inserting (2.11) and carrying out the Gaussian integration over
the momenta. In words, the local mean velocity is the average of the stochastic velocity
ṙk over the subset of trajectories passing through a given microstate x at time t.

2.4.4 Langevin equation

A description complementary to the Smoluchowski equation on the level of a single tra-
jectory is the Langevin equation [59]

ṙk = u(rk ) + µkl (−∇l U + fl ) + dαk + ζk . (2.13)

The first two terms are the external flow and the direct forces exerted on the kth parti-
cle, respectively. The noise term ζk models the thermal environment with short-ranged
correlations

〈ζk (t)ζlT (t 0 )〉 = 2Dkl δ(t − t 0 ) (2.14)

29
2 Dynamics

and zero mean.

The term dαk takes into account an additional drift caused by the hydrodynamic interac-
tions. We have to include it in order for the system to reach the correct equilibrium state
in the absence of driving forces and flows [60]. In appendix A.2, the Fokker-Planck equa-
tion is derived. Together with the moments from A.4, we can rewrite the Fokker-Planck
equation in form of the continuity equation

∂ t ψ + ∇k · (dk + αGl m : ∇m Gkl − ∇l · Dkl − Dkl ∇l ln ψ)ψ = 0.

 

Hence, for the parenthesis to coincide with the velocity (2.9), we make the two identifi-
cations

Dkl = T µkl , (2.15)

dαk = ∇l · Dkl − αGl m : ∇m Gkl . (Dkl = Gkm Gml ) (2.16)

The last equation quantifies the additional drift. It vanishes for spatially uniform mobility
tensors. As detailed in A.4, it depends on our interpretation of multiplicative noise through
the parameter α. Equation (2.15) is the famous Einstein relation. It relates the short-time
diffusion tensors
1
Dkl (r) ≡ lim 〈[rk (t + ") − rk ][rl (t + ") − rl ] T 〉
"→0 2"
with the mobility tensors through the temperature T of the heat bath.
The thus introduced Smoluchowski and Langevin equations define the dynamics of the
systems we aim to study. Other evolution equations are thinkable for interacting many-
body systems, e.g., on the level of the density [61, 62]. However, the investigation of
these equations is beyond the scope of this thesis and remains for future study. In the next
chapter, we will introduce a thermodynamical description independent of the underlying
dynamics leading to expressions for work and heat. In chapter 4, the Smoluchowski
equation will appear in evolution equations for the probability distribution of work and
heat. The Langevin equation will become crucial in the linear response theory developed
in chapter 6.

30
3 Thermodynamics of Small Systems

3.1 Introduction
The idea that even stochastic dynamics like the Brownian motion of a particle must con-
serve energy when coupled to a heat bath could have been formulated back in the days
of Einstein. Somewhat surprisingly, as late as in 1997 Sekimoto was the first to realize
this fact and to formulate the first law of thermodynamics along a single stochastic tra-
jectory [63, 64] using a one-dimensional Langevin equation. In this chapter, we will give
a coherent reformulation of thermodynamic notions like work and heat along a single
trajectory for general driven systems. We will then discuss the concept of a trajectory-
dependent entropy arising in this context. Furthermore, the role of nonequilibrium steady
states is elucidated and we attempt the construction of a steady state thermodynamics.

3.2 Work and heat

Every good theory starts with just a few axioms from which anything else can be deduced.
In our case, we will require the following:

(A1) Work is the energy change of the system caused externally.

(A2) The first law of thermodynamics holds, i.e., energy is conserved along single trajec-
tories as well as mean values.

These axioms regard the energetics of the system and should hold universally indepen-
dent of the underlying dynamics. They contain notions which we have to explain in more
detail. We consider a small system governed by overdamped motion. Suppose the poten-
tial energy of the system U(x; γ) depends on a parameter γ which can be controlled with
high precision. The system is manipulated externally through varying γ according to some
prescribed protocol Γ ≡ {γ(t) : t 0 ¶ t ¶ t 1 }. The infinitesimal change of the energy is
∂U
dU = dγ + (∇k U) · drk
∂γ
employing the chain rule. We identify as work the energy change
∂U
d̄W ≡ dγ (3.1)
∂γ

31
3 Thermodynamics of Small Systems

caused by some outer agent. In particular, it is not the mechanical work spent by the
conservative force −∇U arising from the potential.
Following the first law of thermodynamics, the heat dissipated is then
d̄Q ≡ −(dU − d̄W ) = −(∇k U) · drk . (3.2)
The sign of heat and work is convention and it is reversed if we compare it to classical
thermodynamics. The heat is positive for energy dissipated into the reservoir and negative
for energy which is extracted from the system’s surroundings whereas the work is positive
if put into the system. Both heat and work are process quantities, i.e., they are not exact
differentials and depend on the whole path Γ in parameter space and the trajectory x(t)
through configuration space. Small changes are therefore represented by d̄ instead of d
which is reserved for total differentials.
In classical thermodynamics, the internal energy is a function of a few extensive quan-
tities like volume and the particle number. In the case of small systems, these few ob-
servables are replaced by the relevant degrees of freedom x. To see this connection, we
identify the parameter γ = V with the confining volume of a gas which we control, e.g.,
through moving a piston. With the pressure p = −∂ U/∂ V , equation (3.1) then yields the
well-known expression d̄W = −pdV for the work. In contrast to macroscopic systems, the
dependence on the relevant degrees of freedom means that if the dynamics of the system
is stochastic then work and heat also will acquire a stochastic nature. It is then clear that
their actual values will depend on the temporal coarse-graining of the trajectory, i.e., our
capability to observe it experimentally.
From the definitions (3.1) and (3.2) we can deduce the following cases immediately. If
the parameter γ is constant then no work is performed. If the system is isolated then the
accessible microstates in configuration space U(x; γ) = E(γ) have the same energy and
hence no heat is dissipated. If γ is constant but the system is coupled to a heat reservoir
then heat exchange takes place. It is zero on average if the system is in equilibrium with
the bath.

3.2.1 Quasi-static processes

We drive the system from an initial state A to a final state B. We change the parameter
Γ : γA 7→ γB quasi-statical, i.e., so slowly that the systems passes through a sequence of
consecutive equilibrium states with Gibbs-Boltzmann distribution
1
ψeq (x; γ) = e−U(x;γ)/T , (3.3)
Z(γ)
where the partition function Z(γ) depends on the actual value of the control parameter.
Following (3.1), the mean work spent in such a transition
Z γB Z γB
〈W 〉 = dγ 〈∂γ U〉 = dγ ∂γ [−T ln Z(γ)] = F(γB ) − F(γA) ≡ ∆F
γA γA

32
 3.2 Work and heat

equals the change of free energy ∆F independent of the path through parameter space,
where the equilibrium free energy is defined as F(γ) ≡ −T ln Z(γ). This shows that the
definition of the work (3.1) is compatible with the statements obtained from equilibrium
statistical mechanics [2].

3.2.2 An experimental study: The case of a single colloidal particle

Let us look at a specific system which we studied experimentally. In this experiment,
particle trajectories were determined using total internal reflection microscopy (TIRM),
where a single colloidal particle is illuminated under evanescent field conditions. This
field is created by total internal reflection of a laser beam at a glass–water interface. The
scattered intensity of a bead near the interface is proportional to exp(−ζx), with ζ−1 '
200 nm the decay length of the evanescent field and x the particle–wall distance [65].
Measuring the scattered intensity of a fluctuating Brownian particle as a function of time
thus yields its vertical position with a spatial resolution of about 5 nm.
To drive the colloidal particle between two equilibrium states, it was subjected to the
light pressure of an optical tweezers incident from below while a tightly focused second
laser beam was directed vertically from the top, confining the particle motion to an one-
dimensional trajectory in x-direction (cf. figure 3.1). The control parameter in this exper-
iment is the intensity of the first laser. It was modulated according to a time-dependent
symmetric protocol γ(t) = γ(t s − t) in the interval 0 ¶ t ¶ t s , where t s is the pulse dura-
tion. To ensure that the system is out of equilibrium, t s must be smaller than the particle
relaxation time.
The total potential in which the particle is moving is given by [65]

U(x; γ) = A0 exp(−κx) + B0 x + C0 γx (3.4)

with four parameters A0 , κ, B0 , and C0 . The first term describes the double-layer inter-
action between the negatively charged colloidal particle and the likely charged wall with
A0 depending on the corresponding surface charges and κ−1 the Debye screening length,
which depends on the salt concentration in the fluid. The second term accounts for the

Figure 3.1: Experimental setup. A

polystyrene bead in aqueous solution is
moving above a glass surface. A tightly
focused laser beam confines the motion of
the bead to one dimension perpendicular
to the surface. A second, defocused optical
tweezers drives the particle by exerting a
force f (t) counterbalancing the gravitational
force g.

33
3 Thermodynamics of Small Systems

Figure 3.2: Measured tweezers intensity γ(t) and

particle trajectory x(t). During the first pulse the
particle is pressed towards the surface. During
the second pulse thermal fluctuations support the
particle and it is able to move away from the wall.
Hence the applied work is positive for the first
pulse and negative for the second. The pause
t p between two consecutive pulses of duration t s
must be longer than the particle relaxation time
to guarantee equilibration of the system.

weight of the particle and the additionally exerted light pressure from the upper tweez-
ers, which both depend linearly on the particle distance x [66]. The last term considers
the time-dependent optical forces induced by the lower tweezers. Through measuring the
equilibrium distribution ψeq (x) of the particle position, the potential up to an irrelevant
constant is determined as

U(x) = −T ln ψeq (x),

from which the parameters in (3.4) are obtained.

Integration of (3.1) along a single trajectory x(t) leads to the work functional
Z ts
∂ U(x(t); γ(t)) X
W= dt γ̇(t) = C0 ∆t γ̇(t n )x(t n ), (3.5)
0
∂γ n

where we have inserted the potential (3.4). The right hand side of equation (3.5) accounts
for the discrete sampling of the particle trajectory during our experiments with rate (∆t)−1
at times t n = n∆t. In figure 3.2, the sampled trajectory is shown with two pulses, one
pulse with a positive work and the other leading to negative work supported by thermal
fluctuations.
We want to test experimentally to which precision energy conservation of the particle
on its trajectory is maintained, thus demonstrating the interplay of applied work and

Figure 3.3: Experimental demonstration

of the first law of thermodynamics along
a single trajectory. The stochastic quanti-
ties W , −Q, and ∆U for about 100 periods
of the protocol γ(t) are shown. (b) Distri-
bution histogram of δ = W − Q − ∆U, the
experimentally observed “deviation” from
the first law.

34
 3.3 Work and heat in the presence of nonconservative forces and external flows

exchanged heat when the system is non-adiabatically driven. Therefore, in addition to

the work W exerted on the particle we need to determine its heat exchange Q with the
environment. Because the velocity autocorrelation of a Brownian particle decays on a
timescale of some 10 ns, the velocity
Z t n+1
1
¯ n) =
ẋ(t d t ẋ(t)
∆t t
n

determined from a trajectory measured with sample frequency 2 kHz is not identical to
the instant particle velocity ẋ. However, since ∂ U/∂ x varies on a timescale much larger
than ∆t, the heat along a single trajectory x(t) can be written as
Z ts
∂ U(x(t); γ(t)) X ∂ U(x(t ); γ(t )) t n+1
Z
n n
Q=− dt ẋ(t) = − dt ẋ(t)
0
∂x n
∂x tn
X ∂ U(x(t ); γ(t )) (3.6)
n n
= −∆t ¯ n ).
ẋ(t
n
∂ x

Introducing ∆U ≡ U(x(t s ); γ(t s )) − U(x(0); γ(0)), we finally obtain a stochastic version

W − Q − ∆U = 0

of the first law of thermodynamics. Figure 3.3(a) shows work W , heat Q, and change
of inner energy ∆U for the trajectory of a single particle where the protocol γ(t) was
repeated about 100 times. For W and Q maximal energies of about 15 kB T are observed,
whereas ∆U is on the order of a few kB T . Obviously, Q and W are not independent quan-
tities. Usually trajectories resulting in a large work W are also accompanied by a large
value of Q. But only when taking all three energies into consideration, the distribution of
the deviation shown in figure 3.3(b) is centered around zero, having a half-width of about
0.7 kB T . Assuming that the three terms have the same contribution to the total error, the
error of these energies is about one quarter of kB T .

3.3 Work and heat in the presence of nonconservative forces

and external flows
The approach to a coherent definition of work and heat for a driven transition between
equilibrium states given at the start of the previous section only depends on the internal
energy U(x; γ) of the system. In order to drive the system into a nonequilibrium steady
state, we must consider additional forces. These may arise from two sources: nonconser-
vative forces fk (x; γ) applied directly to the kth particle or an external flow u(r; γ). While
for the nonconservative forces the applied work increment is simply fk ·drk , the work spent
by flows requires a more careful analysis.

35
3 Thermodynamics of Small Systems

3.3.1 An introductory example

To get an idea of how the work spent by flows should look like, we consider a very simple
example. A colloidal particle at position x is trapped by optical tweezers which are moving
with a constant velocity γ̇ = u0 . The trap strength is k and hence the potential energy is
U(x; γ) = (k/2)(x −γ)2 . The motion of the particle neglecting hydrodynamic interactions,
e.g., with a wall, is governed by the Langevin equation

ẋ(t) = −τ−1 [x(t) − u0 t] + ζ(t), (3.7)

where the noise ζ(t) has correlations (2.14), µ0 is the bare mobility, and τ ≡ (µ0 k)−1 is
the relaxation time. The work rate reads
∂U
Ẇ = γ̇ = k(x − u0 t)(−u0 ).
∂γ

Now suppose that we change to the comoving frame through y(t) = x(t) − u0 t. The po-
tential energy U( y) = (k/2) y 2 then becomes explicitly time-independent and following
the definition (3.1), the work rate vanishes, Ẇ = 0. However, physical intuition tells us
that the work should really be the same in both frames. The solution to this apparent con-
tradiction is that in the comoving frame the fluid is not resting but moving with uniform
velocity u = −u0 . Demanding that the work is the same in any frame, we find

∂U
Ẇ = k(x − u0 t)(−u0 ) = k y(−u0 ) = u.
∂y

Hence, the work rate can be expressed as force times the flow’s velocity. The same can be
done for the heat,
∂U ∂U
Q̇ = − ẋ = −k(x − γ)ẋ = −k y( ẏ + u0 ) = − ( ẏ − u),
∂x ∂y

which vanishes if the particle moves along with the velocity of the flow as then no heat
due to friction is dissipated.

Figure 3.4: A colloidal particle is

trapped by optical tweezers. In
the left panel, the trap is moving
with velocity γ̇ = u0 . In the right
panel, the trap is stationary but
the fluid is moving with velocity
u = −u0 . The mean trails behind
a distance u0 τ, where τ is the re-
laxation time.

36
 3.3 Work and heat in the presence of nonconservative forces and external flows

3.3.2 General expressions for work and heat

The previous example shows that by choosing a coordinate system moving with the ex-
ternal flow, we can eliminate this flow in tradeoff for a time-dependent inner energy and
vice versa. The work, however, must be invariant under such a change of frame. For
a refinement of the expressions for work (3.1) and heat (3.2) taking into account both
external flows and frame invariance, consider the convective derivative

DU ∂U
≡ γ̇ + u(rk ) · ∇k U.
Dt ∂γ

Suppose the particles are noninteracting and moving deterministically according to ṙk =
u(rk ), i.e., in figure 3.5 they are moving along a fluid trajectory. Then the total change of
energy along these trajectories becomes the work increment, dU = d̄W = (DU/Dt)dt, as
no heat is dissipated. Particle–particle and particle–fluid interactions lead to dissipation
through leaving the trajectory prescribed through the fluid. Adding the contribution of
the nonconservative forces, the work therefore is found to be

DU
d̄W ≡ dt + fk · [drk − u(rk )dt]. (3.8)
Dt

The new expression for the heat then follows from the first law as

d̄Q ≡ −(dU − d̄W ) = [−∇k U + fk ] · [drk − u(rk )dt]. (3.9)

It is straightforward to check both that these expressions reduce to either (3.1) or (3.9)
in the absence of external flow and nonconservative forces and that the work is invariant
with respect to a change of frame. The deeper physical reason is that we can distinguish a
resting fluid from a moving fluid through dissipation. The fluid velocity therefore appears
explictly in the expressions for work (3.8) and heat (3.9).

Figure 3.5: A particle at r in a fluid element. The

fluid element moves along the sketched trajectory.
If the particle stays within the fluid element and
moves with velocity u(r), no heat is dissipated
and the work is the change of inner energy. How-
ever, due to interactions, the particle moves with
ṙ leaving the fluid element thus leading to the dis-
sipation of heat.

37
3 Thermodynamics of Small Systems

3.4 Entropy production

One cornerstone of equilibrium statistical mechanics is entropy [2]. For a system at equi-
librium with a large heat reservoir, it can be written as
Z
Seq ≡ − dx ψeq (x) ln ψeq (x)

based on the equilibrium probability distribution ψeq (x). Seeking a generalization of

the concept of entropy to driven systems on the level of a single trajectory, it has been
proposed [16] to define

s(t) ≡ − ln ψ(x(t), t). (3.10)

The quantity s is, like our definitions of work and heat, a trajectory based quantity. A
single stochastic trajectory x(t) is parameterized by the time t. The definition (3.10) then
assign a scalar value s at any time t to the trajectory x(t). Given two times t 0 and t 1 , we
can ask for the change of system entropy along this trajectory,
Z t1
∆s ≡ s(t 1 ) − s(t 0 ) = dt ṡ(t),
t0

where ṡ ≡ ds/dt. Suppose we start at t 0 in equilibrium and drive the system into another
equilibrium state. If we choose t 1 large enough such that the system has been equilibrated
again, we clearly have ∆Seq = 〈∆s〉. We therefore have a trajectory dependent expression
the mean of which agrees with what we know from equilibrium statistical mechanics.
Driving the system causes energy to be dissipated in form of heat into the surroundings.
This is in our case the fluid but more formally, one calls it the medium. Dissipation takes
place while the system is driven but continues even after the driving has stopped as the
system relaxes towards equilibrium, where the dissipated heat Q leads to an increase of
medium entropy ∆sm . Both can be related by Clausius’ famous formula
Z
d̄Q
∆sm = (3.11)
T

if we assume that the surrounding fluid itself is and remains in equilibrium with temper-
ature T . At the first glance, this seems like a severe restriction of the applicability of our
formalism. However, any physical realization of fluid plus system is itself embedded into
an environment acting as an equilibrium super reservoir. The dissipated heat then even-
tually leads to a change of entropy in the super reservoir, where equation (3.11) holds.
The sum

∆stot = ∆s + ∆sm

thus is the total entropy change of the universe.

38
 3.4 Entropy production

3.4.1 Equation of motion

The rate of change of the system entropy (3.10) is

ṡ(t) = −∂ t ln ψ − [ṙk − u(rk )] · ∇k ln ψ − u(rk ) · ∇k ln ψ

Ds 1 1
= + Fk · [ṙk − u(rk )] − [−∇k U + fk ] · [ṙk − u(rk )] .
Dt T T| {z }
Q̇=T ṡm

Going from the first to the second line, we have inserted the force following from (2.9).
The first term is the convective derivative of the entropy (3.10). The last term is the
dissipated heat rate, see (3.9). The total entropy production rate ṡtot = ṡ + ṡm therefore
becomes
1
ṡtot (t) = [∂ t + u(rk ) · ∇k ][− ln ψ] + [ṙk − u(rk )] · Fk . (3.12)
T
This rate can be positive corresponding to a production of entropy but it can also be
negative corresponding to the annihilation of entropy. In equilibrium, the rate (3.12)
vanishes.
Calculating the mean total entropy production rate, we first look at the mean of the
convective derivative,
Z Z Z
dx [∂ t ψ + u(rk ) · ∇k ψ] = ∂ t dx ψ + dx ∇k · [u(rk )ψ] = 0.

The first term vanishes due to the conservation of probability. The second term vanishes
because we look at (i) incompressible fluids with ∇ · u = 0 and (ii) neglect boundary
terms. The mean total entropy production rate then becomes

T 〈ṡtot 〉 = 〈[ṙk − u(rk )] · Fk 〉 = 〈[vk − u(rk )] · Fk 〉 = 〈Fk · µkl Fl 〉 ¾ 0 (3.13)

following equation (2.9). It is always nonnegative because the mobility tensors are posi-
tive semidefinite (2.10). The calculation of the mean is done in two steps: first we average
over all trajectories passing through a given microstate x which results in replacing the
actual velocity by the local mean velocity. We then average with the probability distribu-
tion ψ(x, t) leading to (3.13). The fact that 〈ṡtot 〉 is nonnegative can be regarded as a
manifestation of the second law of thermodynamics. The equal sign holds in equilibrium
only.

3.4.2 Dissipation function

In a phenomenological approach to the dynamics of polymers and colloidal suspensions,
one starts from the total dissipation rate functional [57]

K[ψ, {vk }] ≡ 〈W〉 + Ȧ − Ẇ f , (3.14)

39
3 Thermodynamics of Small Systems

where we regard the distribution function and the local mean velocities as independent
variables. This functional captures the three sources of dissipation. First, the nonnegative
dissipation function

1
W(x, t) ≡ [vk − u(rk )] · µ−1
kl
[vl − u(rl )] ¾ 0 (3.15)
2

takes into account dissipation due to friction. Second, relaxation of the system is related
to a dynamical “free energy”
Z
A(t) ≡ dx ψ[U + T ln ψ] = 〈U〉 − T S.

And third, the mean work rate done by the nonconservative forces is
Z
Ẇ f (t) ≡ 〈ṙk · fk 〉 = 〈vk · fk 〉 = dx ψvk · fk .

The mean is again evaluated in two steps, first over trajectories passing a certain mi-
crostate and then over all microstates.
The time derivative of the dynamical free energy becomes
Z Z
Ȧ = dx (∂ t ψ)[U + T ln ψ] = − dx [U + T ln ψ]∇k · (vk ψ)
Z
= dx ψvk · ∇k [U + T ln ψ]

after inserting the continuity equation (2.8) and one integration by parts where we neglect
the boundary terms. Putting everything together, we obtain
Z
1
 
K= dx ψ [vk − u(rk )] · µ−1
kl
[vl − u(rl )] − vk · [−∇k U + fk − T ∇k ln ψ] .
2

Remembering that we regard the velocities as independent, minimization of the func-

tional (3.14) amounts to

∂K
Z
¦ © !
= dx ψ µ−1 [v l − u(r l )] − F k = 0,
∂ vk kl

leading for distributions ψ(x, t) > 0 to the dynamics (2.9). Hence, the dynamics which
minimizes (3.14) implies that the total entropy production rate (3.13) equals two times
the dissipation function, T 〈ṡtot 〉 = 2〈W〉.

40
 3.5 Steady state thermodynamics

3.5 Steady state thermodynamics

In the definition of the entropy (3.10), the actual probability distribution ψ(x, t) enters.
In contrast, work (3.8) and heat (3.9) are defined on the basis of energy, nonconservative
force, and external flow only. Among the accessible states, the steady states play a promi-
nent role. We thus attempt a description in terms of transitions between nonequilibrium
steady states characterized by the external control parameter γ in analogy to transitions
between equilibrium states. The term steady state thermodynamics has been coined for a
phenomenological theory [36, 67] promoting the splitting of the heat into the housekeep-
ing and the excess heat, a concept which we will also use.
In contrast to equilibrium which is determined by the energy alone, a nonequilibrium
steady state is described by two quantities. First, the stationary distribution which we
write as

ψs (x; γ) = e−Φ(x;γ) (3.16)

introducing the pseudo-“potential” Φ(x; γ) in analogy to the Gibbs-Boltzmann distribu-

tion (3.3). Second, the stationary probability distribution ψs (x; γ) leads to a local mean
velocity (2.9) reading

vs k (x; γ) = u(rk ; γ) + µkl [−∇l U(x; γ) + fl (x; γ) + T ∇l Φ(x; γ)].

With these two quantities, we split the total heat rate (3.9)

dΦ
Q̇ = [vs k − u(rk )] · µ−1
kl
[ṙl − u(rl )] + T [γ̇∂γ + u(rk ) · ∇k ]Φ − T (3.17)
dt
into three terms. The structure we thus obtain is similar to the splitting of the en-
tropy (3.12). It involves, however, the stationary distribution ψs (x; γ) belonging to the
current value γ of the control parameter instead of the actual distribution ψ(x, t).
From (3.17) we define, first, the housekeeping heat with rate

Q̇ hk ≡ [vs k − u(rk )] · µ−1

kl
[ṙl − u(rl )] + T u(rk ) · ∇k Φ. (3.18)

It quantifies the heat which is dissipated in response to the work put into the system just
to keep it in a nonequilibrium steady state, hence the name “housekeeping”. Further, we
define the transition functional
t1
∂Φ
Z
Y [x(t); γ(t)] ≡ dt γ̇(t) (x(t); γ(t)) (3.19)
t0
∂γ

along a specific stochastic trajectory x(t). The transition functional vanishes for γ̇ = 0, i.e.,
if there is no transition. Finally, the last term in (3.17) gives rise to a pure boundary term

41
3 Thermodynamics of Small Systems

∆Φ ≡ Φ(x(t 1 )) − Φ(x(t 0 )) rooted in the different stochastic initial and final microstates
of the trajectory. The excess heat finally is defined as

Q ex ≡ Q − Q hk = T Y − T ∆Φ. (3.20)

It is the heat dissipated in a transition minus the boundary term which is always present.
If the stationary state corresponding to all values γ of the control parameter reached
during a transition is equilibrium, we have vs k = 0. The housekeeping heat vanishes
and (3.20) becomes the first law of thermodynamics with

T Y = W − ∆F, T ∆Φ = ∆U − ∆F

after inserting Φ(x; γ) = U(x; γ) − F(γ).

3.6 An illustration: a dumbbell in simple shear flow

Let us finally illustrate the concepts and notions we have introduced in this chapter in the
case of a simple solvable model: two particles connected by a linear spring and driven by
simple shear flow in two dimensions.
The two particles are at positions x = (r1 , r2 ). The external flow is
 
0 1
u(r) = "˙ r
0 0

with strain rate "˙. The inner energy U(r) = (k/2)r2 is a quadratic function of the particle
displacement r ≡ r1 − r2 = (x, y) T . We neglect hydrodynamic interactions and obtain as
local mean velocities

v1,2 = u(r1,2 ) − µ0 ∇1,2 [U + T ln ψ]

with bare mobility µ0 . We assume an infinite system and concentrate on the relative
coordinate r. The relative local mean velocity thus is

v(r) ≡ v1 − v2 = u(r) − 2µ0 ∇[U + T ln ψ] (3.21)

Figure 3.6: Two particles with displacement r

coupled by a linear force −kr in shear flow. The
velocity profile of the simple shear flow is linear
in y-direction with strain rate "˙.

42
 3.6 An illustration: a dumbbell in simple shear flow

with ∇1 U = ∇U and ∇2 U = −∇U. The stationary distribution is obtained from the

differential equation

0 = ∇ · (vs ψs ) = ∇ · [u(r) − 2µ0 kr − 2µ0 T ∇]ψs .

The equations of motion are linear and hence the solution is a Gaussian

exp[− 12 r · S−1 r]
ψs (r) = p (3.22)
(2π)2 det S

with covariance matrix [68]

T 1 + 2γ2 γ
 
S(γ) = . (3.23)
k γ 1

We have defined γ ≡ "˙τ/2, where the relaxation time is τ ≡ (2µ0 k)−1 . The system is
controlled by two parameters: the strength of the potential k and the strain rate "˙. We use
as the external control parameter the reduced strain rate γ. Equilibrium corresponds to
γ = 0 and we can force the system into different nonequilibrium steady states by varying
the strain rate "˙.
For the work rate, we find from (3.8)
2
X
Ẇ = u(rk ) · ∇k U = u(r) · ∇U = k"˙x y
k=1

with mean rate

〈Ẇ 〉 = k"˙〈x y〉 = 2(T /τ)γ2 ,

where the moment 〈x y〉 can be read off the off-diagonal of the covariance matrix (3.23)
directly. The local mean velocity in a steady state with respect to the flow becomes after
inserting (3.22) into (3.21)
γ
 
−γ −1
v (r; γ) − u(r; γ) = τ V(γ)r,
s −1
V(γ) ≡ .
1 + γ2 −1 γ
With
1 1
vs 1 − u(r1 ) = + [vs − u(r)], vs 2 − u(r2 ) = − [vs − u(r)],
2 2
the housekeeping heat rate (3.18) can be rewritten
1
Q̇ hk = [vs − u(r)] · [ṙ − u(r)] + T u(r) · ∇Φ
2µ0
k d
= kVr · [ṙ − u(r)] + T u(r) · S−1 r = [Vr]2 − ku(r) · Vr + T u(r) · S−1 r.
2 dt

43
3 Thermodynamics of Small Systems

as a function of the displacement r only. The center of mass undergoes free diffusion
and hence does not contribute to the housekeeping heat. When averaging with the corre-
sponding stationary probability distribution, the first and the last term become zero. Only
the second term is extensive in time and leads to the mean rate
” — γ 2T
〈Q̇ hk 〉 = k"˙ γ〈x y〉 + 〈 y 2 〉 = γ2 = 〈Ẇ 〉.
1+γ 2
τ

Hence, the work put into the system to maintain the nonequilibrium steady state is dissi-
pated as the housekeeping heat as expected.

3.7 Outlook
The equations of motion for work and heat in driven systems in the theory we have devel-
oped are determined through two quantities, the stationary probability and the stationary
local mean velocity. Up to now, experimental test have been mostly restricted to a single
degree of freedom. However, their natural extension to many-body systems is confronted
with the problem that the stationary probability beyond linear systems as in the previous
section is not known analytically. Moreover, this probability distribution is accessible nei-
ther experimentally nor through numerical simulations. Instead, one has to settle with
reduced information as is contained, e.g., in the pair correlations distribution. A ma-
jor challenge for the future development therefore is to find expressions and to develop
approximations for work and heat incorporating these reduced quantities for complex
systems.

44
4 Probability Distributions of Work, Heat,
and Entropy

4.1 Introduction
In this chapter, we first introduce a family of exact nonequilibrium relations all holding
far from equilibrium. All these relations restrict the probability distribution p(R) of a
stochastic trajectory functional R defined as the asymmetric part of the path action under a
given map. In the second part of this chapter, we then derive various equations governing
the evolution of the probability distribution p(R).

4.2 Nonequilibrium fluctuation relations: Path integral

formalism
In the first part of this chapter, we exploit another approach to nonequilibrium systems
complementary to the dynamics defined in section 2.4. The path integral formalism [69,
70] offers another perspective which is especially appropriate for the trajectory based
quantities like work and heat we have in mind. To this end, we assign a probability
density P(X ) > 0 to every trajectory X . The path-independent functional measure DX is
determined through the normalization condition
Z
DX P(X ) = 1 (4.1)

when summing over all possible trajectories. It will be convenient to introduce a path
action functional S(X ) through

P(X ) ≡ exp[−S(X )].

The path integral introduced in (4.1) indicates continuous dynamics. The developed for-
malism, however, will also hold for discrete dynamics where the path integral sums over
all trajectories of discrete jumps. The presentation in this section is particularly inspired
by the work of Chernyak et al [71] and Maes [72, 73].

45
4 Probability Distributions of Work, Heat, and Entropy

4.2.1 Trajectory involutions

The various nonequilibrium fluctuation relations we are going to discuss can all be de-
rived from a general “master approach”. The only ingredient is an involution I(X ) acting
on trajectories X with I 2 = 1. This means that mapping a trajectory twice will restore its
original form. An involution necessarily has unity Jacobian and therefore preserves the
functional measure of trajectories. The most obvious involution is the identity I(X ) = X .
Other possibilities include any permutation involutions of the microstates. More gener-
ally, for a discretized trajectory, I can be represented by a matrix. This matrix can be
diagonalized with entries ±1 on the diagonal while the determinant is equal to one.
The most studied involution is time reversal since it leads to expressions which allow
for a direct physical interpretation and in the following, we will restrict our discussion to
time reversal. A time reversed trajectory is defined as

X ? ≡ {x(t 1 + t 0 − t) : t 0 ¶ t ¶ t 1 }.

In equilibrium, the detailed balance condition

P(x 0 , t 0 |x, t; γ)ψeq (x; γ) = P(x, t|x 0 , t 0 ; γ)ψeq (x 0 ; γ) (4.2)

holds for all values γ of the control parameter. It is then easy to show that the prob-
ability of both original and mirrored trajectory are the same, P(X ) = P(X ? ). One can
therefore expect that the skew-symmetric part of the action is connected with, in general,
dissipation, and particularly with the entropy production.

Stationary processes
We will first turn to stationary processes, i.e., constant γ. The crucial ingredient is (two
times) the skew-symmetric part of the action under time reversal,

R(X ) ≡ S(X ? ) − S(X ) = −R(X ? ). (4.3)

The fluctuation relations essentially exploit this skew-symmetry in the calculation of the
mean of an arbitrary function G(R),
Z
〈G(R)〉 = DX G(R(X ))P(X )
Z
= DX G(R(X ))eR(X ) P(X ? ).

In the second line, we replace the original trajectory X by X ? and use that the functional
measure is preserved, leading to

〈G(R)〉 = 〈G(−R)e−R 〉. (4.4)

46
 4.2 Nonequilibrium fluctuation relations: Path integral formalism

This is our first master formula from which nonequilibrium relations can be derived by
specifying I(X ) and the function G(R). The two most important classes of nonequilibrium
relations are: the integral fluctuation relations

〈e−R 〉 = 1 (4.5)

for G(R) = 1 and the detailed fluctuation relations

p(−r)
= e−r (4.6)
p(+r)
for G(R) = δ(r − R) because of p(r = R) = 〈δ(r − R)〉. The latter relation determines
universally the probability of events leading to a negative R through the probability for
positive R. The function p(r ¾ 0), however, is a nonuniversal distribution depending on
the details of the system and the process. Further relations can be generated easily, e.g.,
for the moments G(R) = Rn we obtain

〈Rn 〉 = (−1)n 〈Rn e−R 〉.

Nonstationary processes
If we change the control parameter during the process then the path action S(X ; Γ) de-
pends on both the trajectory X and the protocol Γ. We combine both paths into the vector
X ≡ (X , Γ) with time reversal X? = (X ? , Γ? ). The original protocol is called the forward
processes and its time-reversal is called the backward processes. We then repeat the steps
leading to equation (4.4) replacing X by X. However, the average on the left hand side
is now carried out for the reverse process. We make this clear by modifying the master
formula (4.4) to

〈G(R)〉 = 〈G(−R)e−R 〉? . (4.7)

With G(R) = e−R , the integral fluctuation relation remains (4.5) whereas the detailed
fluctuation relation (4.6) becomes
p? (−r)
= e−r (4.8)
p(+r)
because we have to measure the probability distribution of R with respect to either the
forward or the reverse process. A special case constitute symmetric protocols Γ? = Γ for
which a distinction between forward and reverse process is not necessary anymore and
therefore (4.7) reduces to (4.4).
End-point distribution
Beside choosing an involution, we have an additional freedom which can be used to tailor
the functional R. To this end, we split the trajectory probability

P(X) = P(X|x 0 )ψ0 (x 0 )

47
4 Probability Distributions of Work, Heat, and Entropy

into a conditional probability for trajectories starting in microstate x 0 and the initial prob-
ability distribution ψ0 (x). We do the same for the transformed trajectory, leading to
P(X? ) = P(X? |x f )ψ1 (x f )
with another function ψ1 (x) specifying the probability of the final microstate x f , which is
the initial microstate of the reversed process. The skew-symmetric functional R becomes
the sum
ψ0 (x 0 )
R = R̄ + ln with R̄ ≡ S(X? |x f ) − S(X|x 0 ). (4.9)
ψ1 (x f )
This apparently leaves us with two functions ψ0 and ψ1 which can be chosen freely. How-
ever, in order to give the left hand side of the master formula (4.7) the meaning of an
average over an existing process, the function ψ0 is fixed as the initial state. A normal-
ized function ψ1 together with G(R) = e−R therefore leads to the integral fluctuation
relation (4.5) but with the freedom to choose a normalized else arbitrary function as the
distribution of final states.

4.2.2 Connection to heat dissipation

In the previous section, we have introduced a scheme exploiting a functional R from
which various exact relations follow. To complete this scheme, we have to establish a link
to observable physical quantities, and this link is provided by the dissipated heat. For an
explicit calculation, we assume the noise to be Gaussian with action functional
Z t1
1
S(ζ) = dt ζk (t) · D−1 ζ (t).
kl l
4 t
0

For simplicity, we only regard constant diffusion tensors Dkl . We replace the noise through
the Langevin equation (2.13) with
ζk = [ṙk − u(rk )] − µkl [−∇l U + fl ], (4.10)
leading to the conditional action S(X|x 0 ) which has to be augmented by the initial state
x 0 . In addition, a change of variables gives rise to a Jacobian.
We have already split the noise (4.10) into an asymmetric and a symmetric part under
time reversal. The velocities in the first square brackets change their sign for the reversed
process whereas1 the second term remains invariant as does the Jacobian. The skew-
symmetric functional (4.9) therefore reads
Z t1
1 Q
R̄ = S(X? |x f ) − S(X|x 0 ) = dt [ṙk − u(rk )] · [−∇k U + fk ] = (4.11)
T t T
0

1
This convention corresponds to the time-reversal of the external flow, i.e., we invert the velocities of all
fluid particles. Yet another form of the fluctuation relations would arise for invariant velocities u(r).

48
 4.2 Nonequilibrium fluctuation relations: Path integral formalism

with the heat (3.9). This is the sought link, which will turn in the discussion below the
yet abstract relations into nonequilibrium fluctuation relations for measurable quantities.

4.2.3 Reversed dynamics

Up to now we have considered the behavior of the system under the time reversal of
trajectory and protocol but obeying the same dynamics like the original process. We
go now one step further and also reverse the dynamics of the process which leads to a
modified path probability

P? (X) = exp[−S? (X)].

To this end, we introduce the adjoint transition probability P ? for a backward transition
through a generalization of the detailed balance condition (4.2),

P ? (x 0 , t 0 |x, t; γ)ψs (x; γ) = P(x, t|x 0 , t 0 ; γ)ψs (x 0 ; γ). (4.12)

We discretize a trajectory X into n + 1 microstates x k ≡ x(t k ) at times t 0 < t 1 < · · · < t n

with final microstate x n = x f . The same is done for the protocol. For a Markovian process
starting in a stationary state corresponding to γ0 , the conditional trajectory probability
can be expressed as the joint probability

P(X|x 0 ) = P(x n |x n−1 ; γn−1 ) · · · P(x 1 |x 0 ; γ0 )

of all these microstates {x k }. We now explore two obvious choices for the skew-symmetric
functional R:

Reversed dynamics
The first possibility is to only reverse the dynamics through considering

R(X) = S? (X) − S(X).

This amounts to

P? (X|x 0 ) = P ? (x n |x n−1 ; γn−1 ) · · · P ? (x 1 |x 0 ; γ0 )

 
n−1
Y ψ s (x ; γ ) (4.13)
= P(X? |x f ) 
k+1 k
.
k=0
ψ s (x k ; γ k )

Reversed dynamics and process

Another possibility is setting

R(X) = S? (X? ) − S(X),

49
4 Probability Distributions of Work, Heat, and Entropy

amounting to the time reversal of both the dynamics and the process. The corresponding
trajectory probability then reads

P? (X? |x f ) = P ? (x 0 |x f ; γ0 ) · · · P ? (x n−1 |x n ; γn−1 ).

Using (4.12) to replace the backward transition probabilities, we get
 
n−1
Y ψs (x k ; γk )
P? (X? |x f ) = P(X|x 0 )  . (4.14)
k=0
ψ s (x k+1 ; γ k )

4.2.4 The relations

We will now present and discuss the various existing fluctuation relations and show how
they arise from the general scheme we just developed.
Fluctuation theorem
Through Clausius’ formula (3.11), we have R̄ = ∆sm . Historically, the term fluctuation
theorem is connected to systems driven into a nonequilibrium steady state. We then define
the entropy production rate σ ≡ ∆sm /t along a trajectory of length t. If the temporal
boundary term in (4.9) is bounded or grows sublinear, then
1
lim R=σ
t→∞ t
and the fluctuation theorem (1.1)
p(−σ)
∼ e−σt
p(+σ)
follows from (4.6). The fluctuation theorem is also discussed in deterministic systems [8,
3], where, however, certain notions like isoenergetic and isokinetic ensembles as well as
technical issues like the chaotic hypothesis have to be distinguished carefully. For stochas-
tic systems, the fluctuation theorem can be regarded as a special case of the following
relation.
Total entropy production
The definition of the entropy (3.10) turns the fluctuation theorem into a relation for the
total entropy production holding for any trajectory length t. Using the freedom to specify
any distribution for the final microstates, we choose the actual distribution, ψ1 (x) =
ψ(x, t). We then find

R = ∆sm + ∆s = ∆stot

and consequently
p? (−∆stot )
= e−∆stot , 〈e−∆stot 〉 = 1
p(+∆stot )

50
 4.2 Nonequilibrium fluctuation relations: Path integral formalism

first derived in [16].

Jarzynski relation
The most prominent nonequilibrium fluctuation relation is arguably the Jarzynski rela-
tion derived in 1997 [11, 74]. Suppose we drive the system from equilibrium state A
to equilibrium state B. Furthermore, the stationary states belonging to the values of
the control parameter γ reached during the transition γA → γB are equilibrium states,
ψs (x; γ) = ψeq (x; γ). This has to be distinguished from the actual distribution ψ(x, t)
which is the solution of the Smoluchowski equation (2.8). We choose ψ1 (x) = ψeq (x; γB ),
the equilibrium state belonging to the final value of the control parameter. Then

ψeq (x 0 ; γA)
R = R̄ + ln = [Q + ∆U − ∆F]/T
ψeq (x f ; γB )

together with the first law of thermodynamics W = Q + ∆U leads from (4.5) to

〈e−W /T 〉 = e−∆F/T ,

the famous Jarzynski relation.

Crooks relation
The Crooks relation [12, 13] is the detailed version of the Jarzynski relation,

p? (−W )
= e−(W −∆F)/T .
p(+W )

Here, the distinction between forward and reversed process is crucial.

Hatano-Sasa relation
The first relation involving reversed dynamics is the Hatano-Sasa relation [35]. It can be
derived from (4.14) including the boundary term for a transition starting in a steady state
and with ψ1 (x) = ψs (x; γn ),

1 ψs (x 0 ; γ0 ) ψs (x 1 ; γ1 ) ψs (x n−1 ; γn−1 )
 
R = − ln ··· ψs (x n ; γn ) .
ψs (x 0 ; γ0 ) ψs (x 1 ; γ0 ) ψs (x 2 ; γ1 ) ψs (x n ; γn−1 )

If we use further the expression (3.16) for the stationary probability, we can rewrite the
square brackets as
( ) ( Z t )
n
X 1
∂Φ
exp − [Φ(x k ; γk ) − Φ(x k ; γk−1 )] → exp − dt γ̇ ,
k=1 t
∂γ
0

where the last expression follows in the limit n → ∞ for fixed trajectory length. Hence,
R = Y and the integral relation 〈exp[−Y ]〉 = 1 is known as the Hatano-Sasa relation. Of

51
4 Probability Distributions of Work, Heat, and Entropy

course, any other relation stemming from a function G(R) like the detailed relation (4.6)
for R = Y is possible, too.
Housekeeping heat
For the housekeeping heat, we start with (4.13) but set ψ1 = ψ0 , i.e., we discard the
boundary term. This leads to
P(X|x 0 ) ψs (x 0 ; γ0 ) ψs (x 1 ; γ1 ) ψs (x n−1 ; γn−1 )
 
R = R̄ = ln + ln ··· ,
P(X? |x f ) ψs (x 1 ; γ0 ) ψs (x 2 ; γ1 ) ψs (x n ; γn−1 )

where the first term is the heat, see (4.11). Evaluating the square brackets the same way
we did for the Hatano-Sasa relation, we obtain the result
Q Q hk
R= − Y + ∆Φ = .
T T
Hence, also the housekeeping heat fulfills a fluctuation relation.
Steady state thermodynamics
As the final result, we repeat the approach to the Jarzynski relation but generalized to
transitions between nonequilibrium steady states. We set ψ1 (x) = ψs (x; γB ) and obtain
from (4.9)
ψeq (x 0 ; γA)
R = R̄ + ln = Q/T + ∆Φ
ψeq (x f ; γB )

for the original dynamics. Hence, both the housekeeping and the excess heat (modulo the
boundary term) as well as their sum

Q hk /T, Q ex /T + ∆Φ, Q/T + ∆Φ

independently fulfill a nonequilibrium fluctuation relation.

This list can be further extended due to the arbitrariness of the end-point distribution,
e.g., the relation of Bochkov and Kuzovlev [23, 24] can be recovered by setting ψ1 to the
initial distribution [25, 37]. Nevertheless, we will stop here and turn to the second part
of this chapter.

4.3 Evolution of path functionals

The path functionals R for heat and work discussed so far have a common property, they
can be written as integral
Z t1
R(X ; Γ) = dt ṙ(x(t), ẋ(t); γ(t)) (4.15)
t0

52
 4.3 Evolution of path functionals

with a rate function ṙ(x, ẋ; γ) depending on the positions, possibly the velocities, and on
time through the control parameter γ. Integration is carried out along a single trajectory
X in the time interval t 0 ¶ t ¶ t 1 . We call this property time-local. It allows us to find
an also time-local evolution equation for the joint probability distribution ρ(x, r, t) of the
system to be in microstate x at time t ¶ t 1 and to have accumulated an amount r = R
since t 0 . This evolution equation is found by treating r as another degree of freedom. The
explicit calculation is presented in the appendix A.6. For the general ansatz

ṙ ≡ [ṙk − u(rk )] · ak + β

with yet arbitrary vectors ak (x; γ) and scalar function β(x; γ), we obtain as equation of
motion ∂ t ρ = L̂ρ with operator (A.28)

L̂ = L̂SM − ak · [v̂k − u(rk )] − ∇k · Dkl al + β ∂ r + ak · Dkl al ∂ r2 ,


(4.16)

where L̂SM is the Smoluchowski operator (A.26) and v̂k is the velocity operator (A.21).
Evolution operators similar to, or special cases of, the operator (4.16) have been derived
previously [75, 76, 77, 78].

4.3.1 Generating function and fluctuation relations

Our main tool for proving fluctuation relations based on the joint probability ρ(x, r, t)
will be the generating functions. First, we note that the probability distribution of R is
obtained through integration over all final microstates,
Z
p(r, t) = dx ρ(x, r, t).

The generating functions are defined as

Z +∞ Z
gz (x, t) ≡ dr e −z r
ρ(x, r, t), ḡz (t) ≡ dx gz (x, t) (4.17)
−∞

with g0 (x, t) = ψ(x, t). We distinguish between the generating function gz (x, t) still de-
pending on the microstate and its integral ḡz (t). The nonequilibrium fluctuation relations
will be expressed as conditions on ḡz but only for gz we find the explicit equation of
motion

∂ t gz = L̂z gz (4.18)

with operator

L̂z = L̂SM − ak · [v̂k − u(rk )] − ∇k · Dkl al + β z + ak · Dkl al z 2 ,


(4.19)

53
4 Probability Distributions of Work, Heat, and Entropy

which is obtained through inserting (4.16) into (4.17) and integration by parts with re-
spect to r. The boundary terms vanish since the probability distribution p(r) has natural
boundary conditions, i.e., p(r → ±∞) → 0.
The Taylor expansion of the generating function gz around z = 0 yields for the condi-
tional moments
Z
mn (x, t) ≡ dr r n ρ(x, r, t)

the relation

∂n
gz (x, t) = (−1)n mn (x, t). (4.20)
∂z n
z=0

An equation of motion for the conditional moments is then obtained through differentiat-
ing (4.18) with respect to z.
The nonequilibrium fluctuation relations can now be traced back to certain conditions
on the generating function. For any integral relation, we have
Z Z
〈e−R 〉 = dr e−r p(r, t) = dx g1 (x, t) = ḡ1 (t) = 1,

i.e., any at z = 1 normalized generating function fulfills an integral fluctuation relation.

Note that in contrast to the derivation based on the path integral, here we need no state-
ment about the time reversal of process or dynamics. Rather, the existence of integral
fluctuation relations restricts the set of possible operators (4.19) and therefore the set of
path functionals R. For the sake of brevity, we will restrict the discussion of detailed fluc-
tuation relations to steady states. Inserting (4.6) into the definition (4.17), we find the
symmetry
Z
ḡz (t) = dr e−z r+r p(−r, t) = ḡ1−z (t) (4.21)

after a change of variables r 7→ −r, leading to the detailed fluctuation relation. Due to
normalization, ḡ0 (t) = ḡ1 (t) = 1 for all times t.

4.3.2 Transition functional and work distribution

For the evolution equation of the transition functional Y , we set β = γ̇∂γ Φ and ak = 0,
leading to

∂Φ
L̂z = L̂SM − γ̇ z.
∂γ

54
 4.3 Evolution of path functionals

The normalized solution of the evolution equation at z = 1 involving this operator is

g1 (x, t) = ψs (x; γ(t)).
The transition functional reduces to the dissipated work W − ∆F if the stationary state
at any γ reached during the transition is equilibrium. Setting β = γ̇∂γ U, the condition on
the generating function in order to yield the Jarzynski relation becomes slightly different,
Z
ḡ1 (t) = dx g1 (x, t) = ZB /ZA,

where ZA,B is the partition function of the initial and final state, respectively. The solution
of (4.18) fulfilling this condition and also the initial equilibrium condition is

g1 (x, t) = (ZA)−1 exp[−U(x; γ(t))/T ]

with g1 (x, t 0 ) = ψeq (x; γA).

4.3.3 Experimental work distribution

In section 3.2.2, we introduced an experiment demonstrating the conservation of energy
along a single trajectory. In particular, we determined the work from experimentally mea-
sured trajectories. We now want to compare the distribution obtained from this measured
trajectories with the theoretical prediction obtained from the evolution equation for the
work distribution. To this end, it is advantageous to consider the equation of motion for
the conditional moments mn obtained through (4.20) from the evolution equation for the
generating function (4.18), reading
∂U
∂ t mn = L̂SM mn + nγ̇ mn−1 . (4.22)
∂γ
These equations form a hierarchy of inhomogeneous differential equations which couple
the evolution of the nth conditional moment to that of the (n − 1)th moment. The lowest
moment m0 (x, t) = ψ(x, t) is the solution of the Smoluchowski equation with initial
condition m0 (x, t 0 ) = ψeq (x; γA). All higher moments have initial condition mn (x, t 0 ) = 0
for n ¾ 1. The formal solution of (4.22) can be written as a Dyson series.
The equations (4.22) are to be solved numerically for the first three conditional mo-
ments. The Smoluchowski operator reads

L̂SM = ∂ x D⊥ (x) [∂ x U(x; γ)] + ∂ x



for the distance x between particle and glass surface with potential energy (3.4). The
fact that the particle is moving perpendicular but close to a surface drastically influences
the diffusion coefficient due to hydrodynamic interactions of the colloidal bead with the
surface. We take this into account through the distance-dependent diffusion coefficient

D⊥ (x) ≈ D0 [1 + a0 /(x − a0 )]−1 ,

55
4 Probability Distributions of Work, Heat, and Entropy

Figure 4.1: Experimental non-Gaussian

work distribution. The data was taken
from about 16000 trajectories, where the
average work done on the particle was
about 2.4 kB T . The solid line shows the
Pearson type III distribution (4.23) cor-
responding to the theoretically obtained
moments, see main text. Inset: Loga-
rithm of the ratio of the probability to find
trajectories with work −W to those with
work +W . The solid line shows the ex-
pected slope of −1.

where a0 is the radius of the particle [79].

For a comparison with the experimental histogram, we seek a probability distribution
function which roughly resembles the histogram, i.e., it is non-Gaussian with a skew-
symmetry towards larger work values. The most simple normalized non-Gaussian is the
Pearson type III distribution [80]
1 r − α p−1 r −α
   
p(r) = exp − (4.23)
βΓ(p) β β
with three parameters α, β > 0, p > 0 determining its form. In this expression, Γ(p) is the
gamma function. This function cannot be the exact distribution since it has a sharp lower
bound and negative events are cut off for r < α. However, it is a good approximation to
the measured histogram, see figure 4.1. The three parameters are easily obtained from
the three numerically determined moments through the relations
κ1 = α + β p, κn = pβ n Γ(n),
for n > 1 where κn is the nth cumulant. In particular, with
Z Z
κ1 (t) = dx m1 (x, t), κ2 (t) = dx m2 (x, t) − κ21 (t),
Z
κ3 (t) = dx m3 (x, t) − 3κ1 (t)κ2 (t) − κ31 (t)

the parameters become

2κ22 κ3 4κ32
α = κ1 − , β= , p= .
κ3 2κ2 κ23
For the comparison, we take a full cycle of the symmetric protocol with t = t s . In fig-
ure 4.1, also the Crooks relation is tested experimentally as shown in the inset. The
deviation is due to the poor statistics of large negative work values W ® −4 kB T .

56
 4.3 Evolution of path functionals

4.3.4 Housekeeping heat

The derivation of the evolution equation for the housekeeping heat is straightforward.
From the expression (3.18), we identify ak = D−1
kl
[vs l − u(rl )] and β = u(rk ) · ∇k Φ. We
can relate

T ak · Dkl al = [vs k − u(rk )]µ−1

kl
[vs l − u(rl )] ≡ 2Ws

to the dissipation function Ws (x; γ) in the steady state corresponding to γ. In the next
step, we expand the velocity operator (A.21) as follows,

v̂k = vs k − Dkl [(∇l Φ) + ∇l ].

The term proportional to z in (4.19) then reads

2Ws /T − [vs k − u(rk )] · [(∇k Φ) + ∇k ] − ∇k · [vs k − u(rk )] + u(rk ) · (∇k Φ).

Pulling the derivative in the second term in front and using ∇ · u = 0, we can simplify this
expression to

2Ws /T − 2∇k · [vs k − u(rk )] + [(∇k · vs k ) − vs k · (∇k Φ)] + 2u(rk ) · (∇k Φ).

The sum in the third term cancels, leading us to the evolution operator

L̂z = L̂SM + 2 ∇k · [vs k − u(rk )] − u(rk ) · (∇k Φ) − (Ws /T ) z + 2(Ws /T )z 2


(4.24)

for the housekeeping heat R = Q hk /T .

For the integral fluctuation relation at z = 1, we find
Z Z
∂ t ḡ1 (t) = dx L̂1 g1 = −2 dx u(rk ) · (∇k Φ)g1 = 0. (4.25)

For z = 1, the last two terms in (4.24) involving Ws cancel. The integration over the
Smoluchowski operator vanishes due to the conservation of probability. The second term
leads to a boundary term which also vanishes. Finally, the remaining term becomes a
boundary term for g1 (x, t) = ψs (x; γ(t)), the same choice as for the transition functional.
Moreover, for this choice the stricter result L̂1 g1 = 0 also holds. Hence, for the initial
condition ḡ1 (0) = 1 it follows ḡ1 (t) = 1 for all times t and the integral relation for the
housekeeping heat holds as expected. Note that without external flow, relation (4.25) is
valid for any choice of g1 (x, t).
In a nonequilibrium steady state, there are two limits where the generating function can
be approximated by the product form gz (x, t) ≈ ḡz (t)ψs (x). First, this ansatz becomes
exact for vanishing potential energy and an approximation if the nonconservative forces
or the flow forces are much larger than the conservative forces arising from the potential.

57
4 Probability Distributions of Work, Heat, and Entropy

Second, this form also holds near equilibrium for small nonconservative forces or flows.
The equation of motion reads
–Z ™
∂ t ḡz (t) = dx L̂z ψs (x) ḡz (t) = −〈ṡtot 〉z(1 − z)ḡz (t)

with mean entropy production rate 〈ṡtot 〉 = 〈ṡm 〉 = 2〈Ws 〉/T in a nonequilibrium steady
state. The solution for the generating function

ḡz (t) = exp −〈ṡtot 〉z(1 − z)t

 

is a Gaussian and therefore the probability distribution for both the housekeeping heat and
the total entropy production in these limits are Gaussians, too. Obviously, this solution
obeys the symmetry (4.21) and therefore the fluctuation theorem (4.6).

4.4 Extension to non-Markovian processes

The time-local property of the functional (4.15) can be used to extend the proofs in sec-
tion 4.3.1 of the nonequilibrium fluctuation relations through the generating function to
non-Markovian dynamics. The only condition on the non-Markovian process is that it still
has a unique stationary state.2
The crucial ingredient for an extension of the proofs is that the time evolution equa-
tion (2.1)

∂ t ψ(x, t) = L̂(t; t 0 )ψ(x, t) (4.26)

for the distribution ψ(x, t) still holds as in the case of a Markov process. It is some-
what surprising that the same, apparently time-local, equation (4.26) holds also for non-
Markovian processes [82, 83]. This can be understood by realizing that the complete
information about processes with memory is contained in the transition probability rather
than in the single-point distribution ψ(x, t). We denote with Û(t|t 0 ; t 0 ) the operator that
propagates the non-Markovian system from time t 0 < t to the later time t. These prop-
agators do not form a semi-group (2.6) and the fact that the propagator depends on the
whole history in principle back to the time of preparation is made explicit through the
dependence on t 0 . The propagator actually depends on the protocol Γ up to t since any
change of the protocol will have consequences for the following evolution.
Using the propagator as the starting point, we define an evolution operator

L̂NM (t; t 0 ) ≡ ∂ t Û(t|t 0 ; t 0 ) t 0 =t

(4.27)
2
Non-ergodic systems result in non-Markovian processes which then do not relax towards a stationary
state, for a discussion of such processes see, e.g., [81].

58
 4.4 Extension to non-Markovian processes

describing a “substitute”, non-stationary Markov process leading to the same single-point

distribution ψ(x, t) but to a different transition probability than the non-Markovian pro-
cess [82]. In particular, knowledge of the operator (4.27) is not sufficient to calculate
correlation functions. In contrast to the Markov case, the dependence on the control
parameter γ of the operator (4.27) is implicit.
We restrict our proof to dynamics with a unique steady state, i.e., for fixed γ the system
will relax towards a unique probability distribution ψs (x; γ) depending on the control
parameter,

lim ψ(x, t) → ψs (x; γ). (4.28)

t→∞

This statement is equivalent to assuming an ergodic process with or without memory. The
second ingredient to the proof is the existence of a stationary solution

L̂(t; t 0 )ψs (x; γ(t)) = 0 (4.29)

for the corresponding value γ = γ(t) of the control parameter. Whereas this is evident
in the case of a Markovian operator, due to the implicit dependence on γ it is not so
obvious in the non-Markovian case and we give a proof by contradiction. First, we note
that for a proper Markovian substitute process, the operator (4.27) must have a stationary
solution. Now suppose that at time t 0 we stop the process and hold the parameter fixed
with value γ = γ(t 0 ). Under very general conditions, which are fulfilled by any transition
probability, the Perron-Frobenius theorem ensures that the propagator Û(t|t 0 ; t 0 ) has an
eigenstate ψ1 (x; t, t 0 ) corresponding to the eigenvalue 1 depending on t 0 and in principle
also depending on t, i.e.,

Û(t|t 0 ; t 0 )ψ1 (t, t 0 ) = ψ1 (t, t 0 ). (4.30)

Furthermore, this eigenstate ψ1 (x; t, t 0 ) is ensured to be a normalized, nonnegative prob-

ability distribution. From the definition (4.27), we calculate
1
L̂NM (t 0 ; t 0 )ψ1 (t, t 0 ) = lim Û(t 0 + "|t 0 ; t 0 )ψ1 (t, t 0 ) − ψ1 (t, t 0 ) 6= 0

(4.31)
"→0 "
which is nonzero for both arbitrary functions ψ1 and for the eigenfunction ψ1 (t, t 0 ) of
the propagator if the latter would depend on t since t then does not match the leading
time argument of the propagator. This would mean that the substitute operator (4.27)
has no stationary solution. This contradiction is resolved only if the eigenfunction ψ1 (t 0 )
is independent of t. Moreover, taking then the limit t → ∞ in (4.30), we find from
the ergodicity condition (4.28) that ψ1 (x, t 0 ) = ψs (x; γ). Finally, we note that due to
causality, we do not have to actually stop the process at a t 0 since the system cannot
depend on the future protocol and (4.29) must hold for all times t.
With these ingredients, the generalized proof of nonequilibrium fluctuation relations
becomes easy. Inspecting the expression for the function (4.15), we see that its instan-
taneous change ṙ(x, ẋ; γ) only depends on the actual state x the system is in. Hence,

59
4 Probability Distributions of Work, Heat, and Entropy

the time-local operator L̂(t; t 0 ) is all we need and the generalized proof goes along the
same line as in section 4.3.1, but with the Smoluchowski operator L̂SM replaced by the
substitute operator L̂NM .

4.4.1 Substitute operator for a moving trap

The Markovian substitute process is defined through the operator (4.27). Unlike the
Markov case, explicit expressions for this operator are hard to come by. One exception
are Gaussian processes for which these substitute operators can be calculated since we
know the explicit transition probability. As an illustration, we calculate the substitute op-
erator in case of a particle moving in one dimension with position x which is trapped in a
harmonic potential U(x; γ) = (k/2)(x − γ)2 . The generalized Langevin equation
Z t
dt 0 β(t − t 0 )ẋ(t 0 ) = −∂ x U(x(t); γ(t)) + ζ(t)
0

with friction kernel β(t) then becomes linear and can be solved by Laplace transformation
as
Z t
x(t) = G1 (t)x 0 + dt 0 G2 (t − t 0 )[kγ(t 0 ) + ζ(t 0 )],
0

where the two kernels are given as the inverse Laplace transform of Ĝ1 (s) = β̂(s)Ĝ2 (s) and
Ĝ2 (s) = [sβ̂(s) + k]−1 , respectively. The system is prepared at time t 0 in equilibrium with
initial position x 0 drawn from ψeq (x; γ(t 0 )). Due to the change of the external parameter
γ, the mean
Z t
m(t; Γ) ≡ 〈x(t)〉 = dt kG2 (t − t 0 )γ(t 0 )
0

is a functional of the protocol Γ. Without loss of generality, we have set γ(0) = 0 and
hence 〈x 0 〉 = 0.
The substitute operator for one-dimensional Gaussian processes has been worked out
explicitly in [82] reading in general
1
 
L̂NM (t) = −∂ x χ̇(t)x + µ̇(t) − σ̇(t)∂ x .
2
The functions µ̇(t) and σ̇(t) are determined through the differential equations

ṁ(t) = µ̇(t) + χ̇(t)m(t), v̇(t) = σ̇(t) + 2χ̇(t)v(t)

with time-dependent mean m(t) and variance v(t). The correlation function
〈[x(t) − m(t)][x(t 0 ) − m(t 0 )]〉
χ(t, t ) ≡
0
(4.32)
v(t 0 )

60
 4.5 Conclusions

with χ(t, t) = 1 determines χ̇(t) ≡ ∂ t χ(t, t 0 )| t 0 =t .

To be more specific, we choose an exponential friction kernel

κ δ(t)
β(t) = κe−κt ⇒ β̂(s) = ⇒ G2 (t) = (κ̄/k)2 e−κ̄t +
s+κ κ+k
with inverse time scale κ̄ ≡ κk/(κ + k). In the Markov limit, κ → ∞ yields κ̄ → k as
expected. Using the explicit expression for the kernel G2 (t), we calculate the mean
0 0
m(t) = e−κ̄(t−t ) m(t 0 ) + γ 1 − e−κ̄(t−t )
” —

where we have stopped the process at t 0 with parameter γ = γ(t 0 ). This equation shows
the basic features of ergodic non-Markovian processes. For fixed γ, the mean m(t → ∞) →
γ relaxes towards the value of the control parameter. It is a functional of the protocol Γ
up to t 0 and afterwards depends on the time difference t − t 0 only. The time derivative
yields ṁ(t) = −κ̄m(t) + κ̄γ and indeed a straightforward calculation of (4.32) confirms
χ̇ = −κ̄. Therefore, we have µ̇ = κ̄γ and since we do not change the strength of the trap,
the variance is v = T /k leading to σ̇ = 2κ̄T /k. Hence, the substitute operator for fixed γ
becomes

L̂NM (γ) = κ̄∂ x (x − γ) + (T /k)∂ x

 

with stationary solution ψeq (x; γ) ∝ exp[−k/(2T )(x − γ)2 ] for all times t ¾ t 0 .

4.5 Conclusions
The nonequilibrium fluctuation relations compromise a class of exact results holding ar-
bitrarily far from equilibrium. They arise from the behavior of the path action for a time
reversed process and/or reversed dynamics of the system. Mathematically, any skew-
symmetric form on the space of trajectories will generate a similar structure. How-
ever, without an explicit link to physical quantities, the value of the resulting relations
is doubtable. It therefore seems that time reversal is exceptional since it is connected to
the entropy production. Then a wealth of relations involving real physical quantities like
the transition functional and the housekeeping heat can be derived. Nevertheless, future
studies of other trajectory transformations might yield surprising results.

61
5 Discrete Systems

5.1 Introduction
The concepts introduced and discussed in the previous chapters for mechanically driven
systems can be extended to systems with a discrete configuration space in a very straight-
forward manner. This has been pioneered in [84] for death–birth processes. In this chap-
ter, we will study the extension of the Jarzynski relation and the entropy production along
a single stochastic trajectory for the specific system of a single defect center in diamond.
Moreover, the defect center is an athermal system, i.e., it is not coupled to a heat bath
with a well-defined temperature providing thermal noise. However, since the relations
we will study essentially derive from the behavior of the system under time-reversal, they
can be defined for other dynamics as well. We will restrict our discussion to the transition
functional Y and the concept of a nonequilibrium entropy. A more thorough extension to
discrete systems driven by chemical gradients has been given in [85].

5.2 The system: A single defect center in diamond

The simple energy level scheme 5.1 describes the observed optical properties of the defect
center in diamond as two coupled two-level systems (TLS). In the first TLS, the center
can be excited with red light of wavelength 680 nm, responding with a Stokes shifted
fluorescence of rate k−1
b
= 5.5 ns. The second TLS is excited with green light of wavelength
514 nm and decays non-radiatively with rate kd . Since these nano-second transitions are
not resolved, the first TLS appears as bright whereas the second TLS is dark. Depending
on the intensity of the red excitation light, the bright TLS decays with another rate b into

Figure 5.1: A single defect center in diamond de-

a
scribed as an effective two-level system with a dark
b kd and a detectable, fluorescent bright state. It con-
green

sists of two coupled two-level systems, where the

red

kb
fluorescent state is driven by a red laser and the
dark state is driven by a tuneable green laser. The
system switches between these two states with rates
bright state dark state a and b.

63
5 Discrete Systems

the dark TLS, from which it can be pumped back with rate a using the green laser. The
transition rates a and b between the two TLSs are several orders of magnitude smaller
than k b and kd and depend linearly on the intensities of the green (rate a) and red (rate
b) laser, respectively. Hence, it is sufficient to consider the whole system as one effective
two-level system with a dark and a bright state,
a
0 (dark) ‹ 1 (bright), (5.1)
b

with transition rates a and b. The stochastic trajectory X ≡ {n(t) : 0 ¶ t ¶ t 1 } of the state
occupied by the system at time t consists of N consecutive jumps at times t i between state
n = 0 and state n = 1, where t 0 = 0 and t N +1 = t. In the i th interval t i ¶ t ¶ t i+1 , the
state is denoted by ni .
The system is driven out of the initial equilibrium by modulating the intensity of the
green laser with a sinusoidal protocol γ(t) with modulation period t m . This leads to the
time-dependent rate
a(t) = a0 [1 + δγ(t)] (5.2)
with protocol
γ(t) ≡ sin(2πt/t m ), (5.3)
where 0 < δ < 1 is the strength of the modulation. The intensity of the red laser is
constant and therefore b = b0 . The state of the system is the vector ψ = (ψ0 , ψ1 ). For
fixed γ, the system relaxes towards the steady state
b0
ψ0 (γ) = = 1 − ψ1 (γ), (5.4)
a(γ) + b0
which, for a two-level system, is necessarily an equilibrium state.

5.3 The transition functional

In an athermal system, due to the lack of a bath, the notions heat and work lose their
immediate physical meaning. We therefore elude to more abstract concepts like the tran-
sition functional and the entropy production. The transition functional (3.19) reads
Z t1
∂ Φn(t) (γ(t))
Y (X ; Γ) = dt γ̇(t) , (5.5)
0
∂γ
where Φn (γ) ≡ − ln ψn (γ) defined from (5.4) corresponds to a “pseudo” energy. Along
the trajectory X , the transition functional can be rewritten as
N
X a(t i+1 ) a(t) + b0
Y (X ; Γ) = − ni ln + ln (5.6)
i=0
a(t i ) a0 + b0

64
 5.3 The transition functional

exploiting the discrete nature of the system. The second term vanishes if we start and end
the driving at the same laser intensity.
We distinguish moderate driving from strong driving by comparing the intrinsic relax-
ation time of the unmodulated system

τ = (a0 + b0 )−1 (5.7)

with the modulation period t m . For the first two experiments, τ = 14 ms compared to
t m = 50 ms means that the system is only moderately driven into non-equilibrium. For the
third run, the modulation period is reduced to t m = 20 ms compared to a relaxation time
τ = 20 ms. In this case modulation period and intrinsic relaxation time are approximately
equal, which corresponds to a strongly driven system.

5.3.1 Probability distribution

Insight into the statistical properties is gained by looking directly at the probability dis-
tribution p(Y ). For the calculation of this distribution, it is convenient to introduce the
joint probability ρn ≡ ρn (Y, t), which is the probability of the system to be in state n at
time t and to have accumulated an amount Y up to this time. Starting from the master
equation (2.3), the time evolution of ρn is then governed by the differential Chapman-
Kolmogorov equation [75, 76]
1
X ∂ Φn ∂ ρn
∂ t ρn = L nm (t)ρm − γ̇ , (5.8)
m=0
∂γ ∂Y

where L10 = −L00 = a(t) and L01 = −L11 = b0 . Since we start out of equilibrium,
the initial condition is ρn (Y, 0) = ψn (0)δ(Y ). In general, equation (5.8) must be solved
numerically. The distribution

p(Y, t) = ρ0 (Y, t) + ρ1 (Y, t)

can then be obtained by adding the contributions of the two possible final states.

9 2 Figure 5.2: Comparison of the

8 a) b)
7
numerically calculated probabil-
1.5
6 ity distribution P(Y ) with the
5 experimentally obtained nor-
1
4 malized histogram at times a)
3
2 0.5 t = 60 ms and b) t = 200 ms.
1 (In both cases: a0−1 = 64 ms,
0 0
-1.5 -1 -0.5 0 0.5 1 1.5 b0 = 30 ms, t m = 20 ms, and
−1
-0.6 -0.3 0 0.3 0.6
Y Y δ = 0.23.)

65
5 Discrete Systems

Figure 5.3: Test of the moment re-

1.3 1.3 lation (5.10) for the second (k =
a) k =2 b) k =2
1.2 k =4 1.2 k =4 2) and forth (k = 4) moment. The
1.1 1.1 plots show the ratio 〈Y k e−Y 〉/〈Y k 〉
1 1 over trajectory length t for t m =
20 ms. At the vertical dashed lines
0.9 0.9
the protocol becomes symmetric,
0.8 0.8
see Eq. (5.9). a) Experimental
0.7 0.7 data; b) Numerical data. (Param-
10 30 50 70 90 110 10 30 50 70 90 110
t [ms] t [ms] eters are as in Fig. 5.2.)

For sufficiently slow driving, i.e., if the relaxation time of the system is much smaller
than the modulation period t m , the distribution p(Y ) is expected to be a Gaussian. For
large t, our experimental results and numerical calculations indicate that p(Y ) again is a
Gaussian. In the intermediate regime of short trajectories and fast driving, the distribu-
tion p(R) shows distinctly non-Gaussian behavior with a pronounced peak structure, as
shown in Fig. 5.2. Here, we compare experimentally obtained histograms for two different
trajectory lengths to the basically exact numerical solution.
The numerically obtained center peak and the four narrow side peaks can be resolved
partially by the experimental histograms. These peaks can only be observed for short tra-
jectories, where there are at most a few jumps. The center peak derives from trajectories
which do not jump within t. The positions of the other four peaks are at Y = ± ln(1 ± δ),
independent of the driving frequency, which demonstrates that this is not a resonance
phenomena. Rather the explanation is as follows. Independent of the probability density
p(t i ) to jump at t i , the most probable value of a is either near the maximum a = a0 (1 + δ)
or the minimum a = a0 (1 − δ), as can be seen by inverting p(a)da = p(a(t i ))dt i . For a
jump at those values of a, Y in (5.6) picks up a contribution ± ln(1 ± δ) corresponding to
the location of the peaks. Therefore, these peaks are a consequence of the discrete nature
of the system and the shape of the protocol γ(t).
For particular protocols obeying the symmetry relation

γ(t 1 − t) = γ(t), (5.9)

adaption of the arguments developed by Crooks [13] show that then the distribution P(Y )
obeys even the stronger fluctuation theorem (4.6). This relation implies in particular an
intriguing condition on the k-th moment

〈Y k e−Y 〉 = (−1)k 〈Y k 〉. (5.10)

In Fig. 5.3, we show the ratio between the two sides of this relation as a function of the
length t of the trajectory. The theoretically calculated curves show clearly that the moment
relation is valid for symmetric protocols, i.e., for t = l t m where l = 1/2, 3/2, 5/2, . . . . For

66
 5.4 Entropy production

other values of t the relations (4.6) and (5.10) do not hold. The oscillations of the ratio
are damped and hence the moment relation will become valid for all t in the limit t → ∞.
Even though the experimental data are somewhat noisy, they also illustrate this particular
feature of a symmetric protocol which is a consequence of the fluctuation theorem.

5.4 Entropy production

For an extension of the concept of a nonequilibrium entropy to discrete systems, we define

s(t) ≡ − ln pn(t) (t)

analog to (3.10), where now pn (t) is the time-dependent solution of the master equa-
tion (2.3) to be distinguished from the equilibrium solution ψ(γ). The measured proba-
bility pn is evaluated at the actual state n(t) at time t. Fig. 5.4a shows the protocol a(t)
together with the probability p1 (t) to dwell in the bright state. Fig. 5.4b displays a sam-
ple binary trajectory n(t) jumping between the two states. In Fig. 5.4c we see that the
evolution of s(t) is governed by two effects. First, the time-dependent driving of the rates
leads to an evolving probability resulting in a continuous contribution. Second, a jump
between the two states gives rise to a contribution − ln[p+ /p− ], where p− and p+ are the
probabilities of the states immediately before and after the jump, respectively.
Beside the entropy of the system itself, energy exchange and dissipation lead, in general,
to a change in medium entropy. For an athermal system, this change in medium entropy
∆sm can not be inferred from the exchanged heat. Rather it has to be defined. In Ref. [16],
the choice
w(i → j)
∆sm = ln (5.11)
w( j → i)

for a jump from state i to state j with instantaneous rate w(i → j) [w( j → i) being the
backward rate] has been motivated in analogy to the thermal case. In our case it becomes
∆sm = − ln[a(t)/b] for a jump 1 7→ 0 and ∆sm = − ln[b/a(t)] for a jump 0 7→ 1. As
demonstrated in Fig. 5.4d, the medium entropy changes only when the system jumps,
thereby balancing to some degree the change of s(t).

67
5 Discrete Systems

a 4 e
a(t) / Hz

80 0.6

p (t)
1
60 2

Dsm ,Ds / k B
40 0.5 0

b 1.0 2
f
n(t)

0.5 0

0.0 -2

c 0 250 500 750 1000

0.8
s / kB

t / ms
600
g 400 h 400 i
# trajectories
0.3
d 1.0
400
300 300
Dsm / kB

0.5 200 200

0.0 200
100 100
-0.5
0 0 0
0 50 100 150 200 -1.0 -0.5 0.0 0.5 1.0 -6 -4 -2 0 2 4 6 8 -6 -4 -2 0 2 4 6 8

t / ms Ds / k B Dsm / kB Dstot / kB

Figure 5.4: Entropy production in a single two-level system with parameters a0 =

(15.6 ms)−1 , b = (21.8 ms)−1 , t m = 50 ms, and δ = 0.46. (a) Transition rate a(t) (green
line) and probability of the bright state p1 (t) (red circles – measured, red line – theoret-
ical prediction) over 4 periods. (b) Single trajectory n(t). (c) Evolution of the system
entropy. The gray lines correspond to jumps (vertical dotted lines) of the system whereas
the dark lines show the continuous evolution due to the driving. (d) Entropy change
of the medium, where only jumps contribute. (e,f) Examples of (e) entropy producing
and (f) entropy annihilating trajectories. The change of system entropy ∆s = s(t) − s(0)
(black) fluctuates around zero without effective entropy production, whereas in (e) ∆sm
(red) produces a net entropy over time. In (f), ∆sm consumes an entropy of about 1 after
20 periods. (g-i) Histograms taken from 2,000 trajectories of the system (g), medium (h),
and total entropy change (i). The system entropy shows four peaks corresponding to four
possibilities for the trajectory to start and end (0 7→ 1, 1 7→ 0, 0 7→ 0, and 1 7→ 1). The
distribution (h) of the medium entropy change has mean 〈∆sm 〉 = 1.7 and width σ = 3.7,
on this scale it differs only slightly from the distribution of the total entropy change (i).

68
6 Linear Response Theory

6.1 Introduction
The behavior of systems in response to small perturbations lies at the heart of many dis-
ciplines in physics, e.g., the electrodynamics of macroscopic media, and in particular at
the heart of statistical mechanics. The linear response of equilibrium systems defines the
linear response regime. In this regime, loosely speaking following Onsager, the decay of
a fluctuation is independent of whether it has been created spontaneously due to thermal
noise or whether is has been induced by a small applied force. The physical picture be-
hind this regression principle is manifested in the fluctuation-dissipation theorem relating
a response function with the equilibrium correlations of the thermodynamically conju-
gated observable [86, 87]. Beyond the linear response regime, a generalized fluctuation-
dissipation theorem has been found, which, however, does not allow for an interpretation
in terms of thermodynamic observables anymore [88, 83]. The question we attempt to
answer in this final chapter is whether the thermodynamics for small driven systems we
have constructed so far may point to a generalization of Onsager’s regression principle for
nonequilibrium steady states.

6.2 Fluctuation-dissipation theorem

6.2.1 Equilibrium
In its most general form for thermal systems, the equilibrium fluctuation-dissipation the-
orem
eq
T RA,h(t − t 0 ) = −∂ t 〈A(t)B(t 0 )〉0 (6.1)
eq
relates the response function RA,h with an equilibrium correlation function through the
temperature T . The observable B(x) appearing in (6.1) is not arbitrary but fixed through
the external perturbation h. In equilibrium, a perturbation causes a shift in energy of the
type

U 7→ U − Bh, (6.2)

where B is the observable conjugated to h in the energy.

69
6 Linear Response Theory

The response function is defined as

δ〈A(t)〉"
RA,h(t − t 0 ) ≡ . (t ¾ t 0 ) (6.3)
δ["h(t 0 )] "=0

It quantifies the change of the average of an arbitrary observable A(x) at time t in response
to a small perturbation h(t 0 ) of order " applied to the system at an earlier time t 0 , where
the average is measured in the perturbed system. The response function follows from
Taylor expanding the perturbed mean,
Z t
〈A(t)〉" = 〈A〉0 + " dt 0 RA,h(t − t 0 )h(t 0 ) + O(" 2 ).
−∞

Such a response function can be defined not only for a perturbation out of equilibrium
but also for a perturbation out of a nonequilibrium steady state. In both cases, due to
stationarity, the response function depends only on the time difference. Causality requires
RA,h(t) = 0 for times t < 0. In the following, we always assume nonnegative times for
response functions.

6.2.2 General Markov processes

We will now summarize two long-known relations for the linear response of Markovian
stochastic processes. The first relation is based on an operator approach [83], whereas
the second relation quantifies the response of the noise [89, 90].
Smoluchowski equation
In appendix A.5, we have derived the Smoluchowski operator (A.26) as

L̂SM (t) = −∇k · v̂k (t).

Let us assume that the system is in a nonequilibrium steady state and that it is then
responding to a small external perturbation. We make this explicit by splitting the operator

L̂SM (t) = L̂s + " L̂p (h(t))

into a constant part corresponding to the steady state L̂s ψs = 0 and a part of order " which
is time dependent through its dependence on the external field h(t). We further assume
that the distribution function can be expanded in powers of ",

ψ" (t) = ψs + "ψp (t) + O(" 2 ). (6.4)

The equation of motion for the deviation from the steady state up to first order then
becomes

∂ t ψp (t) = L̂s ψp (t) + L̂p (h(t))ψs

70
 6.2 Fluctuation-dissipation theorem

with formal solution

Z t
0
ψp (t) = dt 0 e L̂s (t−t ) L̂p (h(t 0 ))ψs .
−∞

The initial condition is ψp (t → −∞) = 0.

To obtain the response function (6.3), we write out the average of an observable A(x)
Z
〈A(t)〉" = dx A(x)ψ" (x, t).

The time dependence is completely contained in the perturbed distribution function ψ" (t).
Inserting the expansion (6.4), we obtain

δ〈A(t)〉" δψp (x, t) ∂ L̂p

Z Z
L̂s (t−t 0 )
= dx A(x) = dx A(x)e ψs (x).
δ["h(t 0 )] δh(t 0 ) ∂h

We can write the last expression in the form of a fluctuation-dissipation theorem

T RA,h(t − t 0 ) = 〈A(t)B(t 0 )〉0 (6.5)

similar to (6.1) but with the “conjugate observable” B(x) now given as

∂ L̂p
B ≡ T ψ−1 ψs . (6.6)
s
∂h
Moreover, correlations 〈·〉0 are now measured in a nonequilibrium steady state.
In the linear response regime near equilibrium, both expressions (6.1) and (6.5) should
of course become equal. To this end, we calculate the time derivative in (6.1) and obtain

∂
Z Z
L̂eq (t−t 0 ) 0
− dx A(x)e B(x)ψeq (x) = dx A(x)e L̂eq (t−t ) [−ψ−1 L̂ Bψeq ]ψeq (x).
eq eq
∂t

The Smoluchowski operator in equilibrium reads

L̂eq = −∇k · µkl [−(∇l U) − T ∇l ] (6.7)

and the term in square brackets can be worked out to give

−ψ−1 L̂ Bψeq = −T ψ−1

eq eq eq
∇k · µkl (∇l B)ψeq + ψ−1
eq
∇k · µkl B[−(∇l U) − T ∇l ]ψeq
= −T ψ−1
eq
∇k · µkl (∇l B)ψeq .

On the other hand, with (6.2) we obtain from (6.7) the perturbation operator

L̂p = −∇k · µkl [∇l (Bh)]

71
6 Linear Response Theory

with

B = −T ψ−1
eq
∇k · µkl (∇l B)ψeq ,

which shows the equivalence of both fluctuation-dissipation theorems in the linear re-
sponse regime.
Noise response
For a second relation involving explicitly the noise, we realize that the current microstate
x(t) is a functional x(ζ; t) of the noise trajectory ζ up to time t. The probability of a
given noise history in the case of Gaussian noise is
¨ Z «
1
P(ζ) = exp − dt ζk (t) · µ−1 ζ (t) .
kl l
4T

This probability is normalized according to the path integral over all noise histories
Z
Dζ P(ζ) = 1,

which determines the functional measure Dζ.

An external perturbation now leads to different trajectories and we calculate the re-
sponse with respect to a change of the noise trajectory as

δA(t) δA(x(ζ; t)) δP(ζ)

  Z Z
= Dζ P(ζ) = − Dζ A(x(ζ; t))
δζk (t )
0
δζk (t )0
δζk (t 0 )
Z
1
(6.8)
” —
= Dζ A(x(ζ; t)) µ−1 kl
(x(ζ; t 0
)) ζ l (t 0
) P(ζ)
2T
1
= 〈A(t)µ−1
kl
(t 0 )ζl (t 0 )〉.
2T
To arrive at the second line, we perform a functional integration by parts.

6.3 The nonequilibrium velocity fluctuation-dissipation

theorem

6.3.1 Derivation
We consider a system in a nonequilibrium steady state which is perturbed through a small,
spatially homogeneous change of the nonconservative force

h 7→ δfm

72
 6.3 The nonequilibrium velocity fluctuation-dissipation theorem

exerted on the mth particle. The response function can be rewritten as

δA(t) δA(t) ∂ ζk
   
T RA,δfm (t − t ) = T
0
=T ·
δf (t 0 ) δζk (t 0 ) ∂ fm
 m (6.9)
δA(t)

=T · µkm (t )
0
δζk (t 0 )

employing the chain rule. To obtain the last expression, we use that within the brackets,
the trajectory is fixed. The Langevin equation (2.13) solved for the noise is then a function
of the velocities and, in particular, of the nonconservative forces. Partial differentiation
with respect to fm then leads to the mobility matrices µkm . The response function is now
a vectorial quantity where the components are given by the functional derivative with
respect to the components of fm . We follow the steps outlined in the end of the previous
section and perform a functional integration by parts of (6.9), leading to

δµkm (x(ζ; t 0 ))P(ζ)

Z
T RA,δfm (t − t ) = −T Dζ A(x(ζ; t))
0
.
δζk (t 0 )

However, here we are confronted with the treatment of multiplicative noise due to the
spatial dependence of the mobility tensors.1 For the sake of simplicity, we set α = 0 cor-
responding to the Itô calculus. Then 〈rk (t)ζlT (t)〉 = 0 holds at same times which implies
that microstate and noise are independent variables leading to the response function

1
T RA,δfm (t − t 0 ) = 〈A(t)ζm (t 0 )〉. (6.10)
2
With this intermediate result, we can now calculate the observable conjugate to the non-
conservative force in a nonequilibrium steady state.
The conjugate observable in general is given by (6.6). The perturbation operator is
found to be

L̂p (δfm ) = −∇k · µkm δfm

and we obtain the vector

B = −∇k · Dmk − Dmk (∇k ln ψs )

employing the Einstein relation (2.15). This is certainly correct but seems not very useful.
However, combining the Langevin equation (2.13) with the local mean velocity (2.9), we
find the Langevin equation (remember α = 0)

ṙm = vs m + Dmk (∇k ln ψs ) + ∇k · Dmk + ζm

1
See appendix A.4.

73
6 Linear Response Theory

and hence

B = ζm − (ṙm − vs m ).

The fluctuation-dissipation theorem (6.5) amounts to

T RA,δfm (t − t 0 ) = 〈A(t)B(t 0 )〉0 = −〈A(t)[ṙm (t 0 ) − vs m (t 0 )]〉0 + 〈A(t)ζm (t 0 )〉

and finally, using (6.10), we arrive at the velocity fluctuation-dissipation theorem

T RA,δfm (t − t 0 ) = 〈A(t)δvm (t 0 )〉0 (6.11)

involving the relative velocity

δvm (t) ≡ ṙm (t) − vs m (x(t)). (6.12)

Hence, through combining the two known relations introduced in the previous section,
we could show that the abstract expression (6.6) equals the relative velocity (6.12).

6.3.2 Restoring the equilibrium form and a violation function

In equilibrium, the velocity fluctuation-dissipation theorem (6.1) reads

eq ∂
T Rṙn ,δfm (t − t 0 ) = 〈ṙn (t)rm
T
(t 0 )〉0 = 〈ṙn (t)ṙm
T
(t 0 )〉0 ,
∂ t0
where the response function now is a response matrix. The conjugate variable of the force
is the position rm . We have shifted the derivative to the earlier time t 0 and due to linearity,
we can pull the time derivative within the brackets. From (6.11), the same form holds in
nonequilibrium steady states but for the relative velocity,

T Rδvn ,δfm (t − t 0 ) = 〈δvn (t)δvm

T
(t 0 )〉0 . (6.13)

This is not the equilibrium fluctuation-dissipation theorem since the response is defined
with respect to a perturbation out of a nonequilibrium steady state but it strongly resem-
bles the form of the fluctuation-dissipation theorem in equilibrium.
Equation (6.13) demonstrates the role of the three velocities involved: the actual ve-
locity ṙk , the local mean velocity vs k , and the global mean velocity 〈ṙk 〉. Surprisingly, in
order to recover the equilibrium form of the fluctuation-dissipation theorem, the actual
velocity has to be measured with respect to the local mean velocity which depends on the
microstate of the system in contrast to the global mean velocity. At least for the velocity,
this implies a modification of Onsager’s principle: forced fluctuations out of a nonequilib-
rium steady state cannot be distinguished from spontaneous fluctuations with respect to
the local mean.

74
 6.3 The nonequilibrium velocity fluctuation-dissipation theorem

The fact that the equilibrium form is contained in (6.11) also for a general observable
A suggests that we introduce a “violation function”

IA,δfm (t − t 0 ) ≡ 〈[A(t) − 〈A〉][vs m (t 0 ) − 〈ṙm 〉]〉0 (6.14)

measuring the excess. The so defined violation function vanishes in equilibrium. In prin-
ciple, the fluctuation-dissipation theorem is not “violated” in the sense that a prerequisite,
namely detailed balance, is no longer met. It is, however, a convenient and commonly
used term. Violations of the fluctuation-dissipation theorem have been studied mostly in
the context of glassy systems [91, 92, 93, 90, 94], where time-translational invariance is
broken but the entropy production rate is small. In contrast to our result leading to an ad-
ditive violation (6.14), the concept of an effective temperature [95, 93, 90] aims to keep
the simple form (6.1) at the expense of a temperature different from that of the fluid. The
concept of an effective temperature has been applied also to driven systems [96, 97].

6.3.3 Generalized Einstein relation

In chapter 2 we already met the Einstein relation (2.15) connecting the short-time diffu-
sion tensors to the mobility tensors through the temperature T . This relation holds strictly
only in equilibrium but we have required it to be valid even when we drive the system. In
principle it corresponds to the assumption that driving the small system of interest does
not affect the strength of fluctuations determined by the larger heat bath, i.e., the fluid.
In the following discussion of a generalized Einstein relation, we consider the one-
dimensional case with particle position x. The particle is moving in a periodic potential
U(x + L) = U(x) with periodicity L. We introduce a second definition for the diffusion
coefficient as
1 ” 2 —
D ≡ lim 〈x (t)〉 − 〈x(t)〉2 (6.15)
t→∞ 2t

with particle position x. Whereas (A.9) defines the short-time diffusion coefficient D0
determining also the noise strength, D from (6.15) defines the effective long-time diffu-
sion coefficient. The free diffusion coefficient D0 is a property of the bath alone and does
not depend on external potentials nor particle–particle interactions. For a better under-
standing, consider a deep potential where the particle is trapped. Then clearly D = 0 but
D0 6= 0 as the particle still diffuses within the potential minimum. In equilibrium, D ¶ D0
since interactions hinder the diffusion of the particle, where the equal sign holds for a
free particle only. It has been shown that in a periodic potential under nonequilibrium
conditions, i.e., a nonconservative force f is driving the particle, the effective diffusion
is enhanced [98, 99] with D  D0 under certain conditions. We will now show that the
excess of the diffusion coefficient is quantitatively given by the violation function.
In one dimension, the violation function (6.14) becomes

I(t) = 〈[ẋ(t 0 + t) − 〈ẋ〉][v s (x(t 0 )) − 〈ẋ〉]〉, (6.16)

75
6 Linear Response Theory

which is a function of the time difference t. The offset t 0 is arbitrary because of time-
translational invariance in a steady state, and in the following we set t 0 = 0. The diffusion
coefficient (6.15) can be rewritten through the Green-Kubo formula
Z∞
D= dt 〈[ẋ(t) − 〈ẋ〉][ẋ(0) − 〈ẋ〉]〉. (6.17)
0

This suggests to integrate the fluctuation-dissipation theorem over time, leading to

Z∞ Z∞
T dt R ẋ, f (t) = D − dt I(t).
0 0

However, the left hand side is the effective mobility of the particle defined as the response
with respect to a small variation of the driving force,
Z∞
∂ 〈ẋ〉
µ≡ = dt R ẋ, f (t). (6.18)
∂f 0

Therefore, the integrated version of the fluctuation-dissipation theorem (6.13) holding in

nonequilibrium implies a generalized Einstein relation
Z∞
D = Tµ + dt I(t), (6.19)
0

which reduces to D = T µ in equilibrium. Hence, in equilibrium the Einstein relations

for bare and effective diffusion coincide. Driving the system into a nonequilibrium steady
state, the “bath” Einstein relation for the bare diffusion coefficient still holds but the effec-
tive diffusion attains an additional contribution given by the integrated violation function.
For the one-dimensional case, explicit formulas for both the global mean velocity
Z L
1 − exp(−L f /T ) 1
〈ẋ〉 = R L , I± (x) ≡ dz exp {±[V (x) − V (x ∓ z)]/T }
dx I+ (x)/L D 0 0
0

1.1

1.05
Figure 6.1: The effective diffusion con-
1
stant D over the force f for the periodic
D

0.95 model potential V (x) = V0 cos 2x (with

V0 = 12 , D = T = 1). The solid line shows
0.9 the analytical solution (6.20) whereas the
closed circles are obtained numerically.
0.85
0 0.5 1 1.5 2 2.5 3 3.5 4 The dotted line shows the effective mobil-
f ity (6.18).

76
 6.4 Outlook

1.5 violation integral

mobility
D
D [µm /s]

1.0
2

Figure 6.2: Experimental test of

0.5
Eq. (6.19) for different driving forces f .
The open bars show the measured dif-
0.0 fusion coefficients D. The stacked bars
0.04 0.06 0.08 0.10 0.12 0.14 0.16 are mobility (gray bar) and integrated
driving force [pN] violation (hatched bar), respectively.

and the diffusion coefficient

Z −3 Z
L L

D = D0 L 2
dx I+ (x) dx I+2 (x)I− (x) (6.20)
0 0

with V (x) ≡ U(x) − f x exist [98]. With these exact results, we compare the generalized
Einstein relation versus numerical data in figure 6.1.
We have also tested the Einstein relation experimentally, see figures 6.2 and 6.3. To
this end, a single colloidal particle is moving in a toroidal optical trap which restricts
the motion of the particle effectively to one dimension. The diffusion coefficient and the
violation function can be calculated from a single trajectory. First, global mean velocity
and the stationary probability distribution ψs (x) are determined. From these quantities,
the local mean velocity follows as v s (x) = 〈ẋ〉/[Lψs (x)]. The correlation function (6.16)
is then calculated along the trajectory, resulting in the violation function which is finally
integrated. The diffusion coefficient is obtained from (6.15). For the effective mobility,
the force is varied experimentally by a small amount and a short trajectory is recorded.
The difference of the global mean velocity finally leads to the mobility (6.18).

6.4 Outlook
These final remarks concerning the extension of the fluctuation-dissipation theorem to
nonequilibrium steady states conclude our efforts to build a thermodynamics of small
systems. The method described in this chapter yields a promising approach, which hope-
fully can be extended to other quantities than the velocity. Especially the extension of
the Onsager principle seems to be of greater generality than just concerning velocities.
The remaining task of finding the correct local mean for other observables, however, will
certainly be challenging.

77
 6 Linear Response Theory

3 .0

2 .5

2 .0 tim e c o n s t. local mean velocity

2 .3 s
1 .5
/s ]
2

1 .0
(3)
2

mean velocity
I ( t ) [ µm

0 .5 (2 ) (3 ) (4 ) (1)

0 .0

-0 .5 5 .8 s
11111111
00000000
00000000
11111111 (2,4)
-1 .0
drift velocity
0 5 1 0 1 5 2 0
t [s ]
0 1 2 3 4 5 6
angular position

Figure 6.3: Left: Experimentally measured violation function I(t) (solid line). Right:
Comparison of the velocities involved in the violation function I(t). For an ideal cosine
potential, we sketch the probability distribution ψs (x) (solid gray line), the local mean
velocity v s (x) together with the drift velocity and their mean 〈ẋ〉 versus the angular par-
ticle position. The drift velocity is the deterministic part of the actual velocity ẋ. The sign
change in I(t) at (2), (3), and (4) can be understood as follows. In a steady state, a single
particle trajectory will start with highest probability in the shaded region and, for an il-
lustration, we choose its maximum as starting point (1) determining the value v s (x(t)) in
Eq. (6.16). Neglecting thermal fluctuations, the particle would follow the dashed line and
during a small time step τ the product of drift velocity time v s (x(t 0 )) is positive. If the
particle passes (2), the product will become negative. The sign changes again if the parti-
cle passes (3) and then (4) and so on due to the periodic nature of the potential. Thermal
noise and averaging over all trajectories does not change this behavior responsible for the
oscillations of I(t).

78
A Appendix

A.1 Projected processes

There are in general two kinds of projected processes. The Zwanzig-Mori formalism allows
to derive exact equations of motion for configuration space functions including the mean
of an observable, correlation functions, etc [91]. This has found applications in the mode-
coupling theory of glasses [100] but also driven colloidal suspensions [101]. The second
technique, also called adiabatic elimination [55, 102], reduces the configuration space by
means of a time scale separation which allows to project out unwanted degrees of free-
dom. We will use the latter method to get rid of the particle momenta in the overdamped
regime.
We assume that the time evolution
∂ t Ψ(t) = L̂Ψ(t) (A.1)
of a state Ψ(x, x 0 , t) is governed by the linear, constant operator L̂. Let us denote 〈·, ·〉
the linear functionals with respect to the space {x 0 }. We aim to describe the evolution
of the process depending on x alone, i.e., we look for the function ψ(x, t) given by the
projection
P̂Ψ(t) = ψ(x, t)ϕ0 (x 0 ). (A.2)
The projector is defined as
P̂· ≡ 〈ϕ̄0 , ·〉ϕ0 ,
where ϕ0 (x 0 ) and ϕ̄0 (x 0 ) are two yet arbitrary functions which must fulfill 〈ϕ̄0 , ϕ0 〉 = 1 to
ensure idempotency P̂ 2 = P̂.
The specific application we have in mind is that of a process ∂ t ϕ(x 0 , t) = L̂0 ϕ(x 0 , t) to
which the evolution of another process is coupled. We model this through L̂ = L̂0 + L̂1
and choose to project onto the null space of L̂0 . Hence, L̂0 ϕ0 = 0 and L̂0† ϕ̄0 = 0 lead to
L̂0 P̂ = P̂ L̂0 = 0. (A.3)
Applying the projector to (A.1) and using (A.3), we can rewrite the evolution equation as
the two coupled equations
∂ t Ψ0 (t) = P̂ L̂1 Ψ0 (t) + P̂ L̂1 Ψ1 (t), (A.4)
∂ t Ψ1 (t) = ( L̂0 + P̂ L̂1 )Ψ1 (t) + P̂ L̂1 Ψ0 (t),
0 0
(A.5)

79
A Appendix

where Ψ0 ≡ P̂Ψ = ψϕ0 is the projected state and Ψ1 ≡ P̂ 0 Ψ ≡ (1 − P̂)Ψ is the deviation
from the projected state. The formal solution of (A.5) for time independent operators L̂0
and L̂1 is
Z t
Ψ1 (t) = exp[( L̂0 + P̂ 0 L̂1 )t]Ψ1 (0) + dt 0 exp[( L̂0 + P̂ 0 L̂1 )(t − t 0 )] P̂ 0 L̂1 Ψ0 (t 0 ).
0

The generalization to time-dependent operators is straightforward through a time-ordered

exponential. Putting this solution back into (A.4) would lead to an integro-differential
equation for ψ(x, t) but to little avail since the memory kernel can only be expressed
formally.
We can proceed further assuming that the time scales on which the two processes evolve
are very different. The first time scale is set by the relaxation time τ0 = k L̂0−1 k. We define

" = k L̂0−1 (∂ t − P̂ 0 L̂1 )k ¶ k L̂0−1 k · k∂ t − P̂ 0 L̂1 k = τ0 k∂ t − P̂ 0 L̂1 k  1

as the ratio of the relaxation time and the typical rate of change of the coupled process.
Equation (A.5) is solved as

Ψ1 (t) = [∂ t − ( L̂0 + P̂ 0 L̂1 )]−1 P̂ 0 L̂1 Ψ0 (t).

Assuming " to be small, we can expand

[(∂ t − P̂ 0 L̂1 ) − L̂0 ]−1 = −[1 − L̂0−1 (∂ t − P̂ 0 L̂1 )]−1 L̂0−1 ' − L̂0−1 + O("),

leading to a deviation

Ψ1 (t) ' − L̂0−1 P̂ 0 L̂1 Ψ0 (t).

The equation of motion for the reduced process thus reads

∂ t ψ(t) = L̂red ψ(t), L̂red ≡ 〈ϕ̄0 , L̂1 ϕ0 〉 − 〈ϕ̄0 , L̂1 L̂0−1 P̂ 0 L̂1 ϕ0 〉.

Through discarding the component of the coupled process in the nullspace of L̂0 , the
projector P̂ 0 makes sure that we can always apply the inverse L̂0−1 . If P̂ L̂1 ϕ0 = 0 then the
reduced evolution operator simplifies to

L̂red = −〈ϕ̄0 , L̂1 L̂0−1 L̂1 ϕ0 〉. (A.6)

A.2 Derivation of the Fokker-Planck equation

The Fokker-Planck equation is the equation of motion for the probability distribution of a
general Markov process [55]. In this section, we basically apply the methods of stochastic
processes as introduced in textbooks [55, 103] to many-particle systems while keeping

80
 A.2 Derivation of the Fokker-Planck equation

in mind Lau and Lubensky’s more recent discussion of the correct interpretation of multi-
plicative noise [60] following in the next section.
Derivations usually start from the propagation of the probability distribution (2.5)
Z
ψ(x, t + ") = dx 0 P(x, t + "|x 0 , t)ψ(x 0 , t) (A.7)

using the transition probability. The transition probability is normalized according to

Z
dx P(x, t + "|x 0 , t) = 1.

The probability for the transition x 0 → x during the small time step " is defined as
P(x, t + "|x 0 , t) = 〈δ(x − x(t + "))〉| x(t)=x 0 ,
where the mean is over the noise for trajectories starting in x 0 . Taylor expanding the
δ-function around x 0 yields

δ(x − x(t + ")) = δ(x − x 0 ) − [rk (t + ") − r0k ] · ∇k δ(x − x 0 )

1
+ [rk (t + ") − r0k ][rl (t + ") − r0l ] T : ∆kl δ(x − x 0 ) + . . .
2
where ∆kl ≡ ∇k ∇lT is the symmetric matrix of second-order derivatives. The minus sign
in the second term arises because we shift the derivative to the first argument. We expand
the Taylor series up to second order since for normal diffusion the first two moments,
〈[rk (t + ") − r0k ]〉 ≡ mk (x 0 )", (A.8)
〈[rk (t + ") − r0k ][rl (t + ") − r0l ] T 〉 ≡ 2Dkl (x )",
0
(A.9)
are of order " and higher moments are at least of order " 2 [55]. These moments are the
mean displacements mk and the short-time diffusion matrices Dkl , respectively.
We put the resulting transition probability
P(x, t + "|x 0 , t) = δ(x − x 0 ) − " mk (x 0 ) · ∇k δ(x − x 0 ) − Dkl (x 0 ) : ∆kl δ(x − x 0 )
 

back into (A.7) and pull the derivatives in front of the integral. With the time derivative
∂ψ ψ(x, t + ") − ψ(x, t)
= lim
∂t "→0 "
and using the properties of the δ-function, we arrive at the Fokker-Planck equation
∂ t ψ + ∇k · (mk ψ) − ∆kl : (Dkl ψ) = 0.
 
(A.10)
The Fokker-Planck equation is a partial differential equation of second order. The explicit
expressions of the mean displacements mk and short-time diffusion matrices Dkl depend
on our interpretation of the actual stochastic processes as defined by the Langevin equa-
tion.

81
A Appendix

A.3 Time evolution operator: The harmonic oscillator

The simplest Fokker-Planck operator is

L̂ = τ−1 ∂ x [x + ∂ x ] (A.11)

which describes motion in one dimension with a linear restoring force. The eigenvalue
equation becomes
” —
−λk ψk (x) = L̂ψk (x) = τ−1 ψk (x) + xψ0k (x) + ψ00k (x) ,
2
/2
where the prime denotes derivation with respect to x. With the ansatz ψk (x) = e−x ϕk (x)
we obtain

ϕk00 (x) − xϕk0 (x) + λk τϕk (x) = 0,

the generating differential equation – replace x by 2x – for the Hermite polynomials

H k (x). The normalized eigenfunctions of the operator (A.11) are therefore
2
e−x /2 p
ψk (x) = p H k (x/ 2)
2π
with nonnegative eigenvalues λk = k/τ. For the definition of the inverse operator we
exclude the null space (k = 0) and since there is a gap between zero and the smallest
nonzero eigenvalue 1/τ the inverse is a bounded operator with norm

k L̂ −1 k = τ,

i.e., all its eigenvalues are positive and smaller than τ.

A.4 Multiplicative noise

In this section, we write the Langevin equation as

ṙk = dk + Gkl ξl (A.12)

with arbitrary drift vectors dk . The noise has zero mean and correlations

〈ξk (t)ξlT (t 0 )〉 = 21δkl δ(t − t 0 ). (A.13)

If the strength of the noise Gkl (x) depends on the microstate itself, we are confronted with
the problem of multiplicative noise

G(x[ξ(t)]) ξ(t), (A.14)

82
 A.4 Multiplicative noise

which carries an ambiguity due to the δ-correlated nature of the noise. To understand the
problem, consider the integral
Z t+" Z t+"
dt 0 G(x(t 0 ))ξ(t 0 ) → G(x( t̄)) dt 0 ξ(t 0 )
t t

over a short time interval with the uniquely determined time t < t̄ < t + ". The right
hand side conclusion follows from the first integral mean-value theorem which only holds
for continuous functions. However, the noise ξ(t) is a discontinuous function. Stochastic
products of the kind (A.14) have therefore to be augmented by the answer to the question
of how the function G(x) is to be evaluated. In [60], the following rule
x( t̄) 7→ x̄ ≡ (1 − α)x(t) + αx(t + ") (A.15)
involving a parameter 0 ¶ α ¶ 1 has been proposed. In particular, it reduces to the Itô
convention for α = 0 and to the Stratonovich convention for α = 1/2 [55].
In the previous section, we have derived the Fokker-Planck equation (A.10). Using the
rule (A.15), we can calculate the involved moments from the particle’s displacement
Z t+" Z t+"
rk (t + ") − rk (t) = dt 0 ṙk (t 0 ) = dk (x̄)" + Gkl (x̄) dt 0 ξl (t 0 ) (A.16)
t t

using the Langevin equation (A.12). Taylor expanding a function depending on x̄ around
x = x(t) up to first order in " yields
dG
G(x̄) ' G(x) + " = G(x) + α[ṙk · ∇k G(x)]".
d" "=0

Putting this expression back into (A.16) and taking the mean over the noise leads to an
additional contribution
〈ṙk (t)ξlT (t 0 )〉 = dk (x)〈ξlT 〉 + Gkm (x)〈ξm (t)ξlT (t 0 )〉 = 2Gkl (x)δ(t − t 0 )
to the mean displacement mk besides the drift dk . Integrals over the δ-function are eval-
uated according to
Z t+" Z t+" Z t+"
1
dt δ(t − t ) = ,
0 0
dt 0
dt 00 δ(t 0 − t 00 ) = ".
t
2 t t

The moments as defined in (A.8) and (A.9) then become

mk = dk + αGl m : ∇m Gkl ,
Dkl = Gkm Gml .
Here, we have assumed that the Gkl are symmetric matrices. We can assume this because
the Dkl must be symmetric matrices (and also symmetric with respect to k ↔ l), which
does not determine the Gkl uniquely.

83
A Appendix

A.5 Derivation of the Smoluchowski equation

In section 2.4.2, the Smoluchowski equation has been introduced phenomenologically.
However, it can also be derived from the Fokker-Planck equation (A.10) if we consider
inertial particles and then project out their momenta {pk }.
For the derivation of the Fokker-Planck equation, we start with the coupled equations
of motion

ṙk = pk /m, ṗk = −∇k U + fk − µ−1 pl /m − u(rl ) + ξk ,

 
kl

where m is the mass of the particles. The noise has correlations

〈ξk (t)ξlT (t 0 )〉 = 2T µ−1

kl
δ(t − t 0 ).

The equations of motion for the momenta are linear with additive noise since the mobility
tensors µkl (x) do not depend on the momenta. It therefore does not suffer from the
ambiguities we discussed in section A.4.
Besides the positions x we gather the momenta x p ≡ (p1 , . . . , pN ) such that the config-
uration space turns into phase space with microstate (x, x p ). The state Ψ(x, x p , t) is then
the joint probability distribution of both particle positions and momenta. The Fokker-
Planck equation is obtained from (A.10) and we write it in the form

∂ t Ψ(t) = [ L̂0 + L̂1 ]Ψ(t)

with operators
µ−1 ∂ ∂
 
kl
L̂0 ≡ · pl + mT , (A.17)
m ∂ pk ∂ pl
pk ∂ ∂
L̂1 ≡ − · ∇k − [−∇k U + fk ] · − µ−1 u(r l ) · . (A.18)
m ∂ pk kl
∂ pk
The benefit of splitting the operator in this way is that L̂0 acts and depends on the mo-
menta only, which are the variables we want to get rid of. The splitting therefore is
amenable to the projecting scheme we developed in section A.1. The relaxation time τ0
is the maximal eigenvalue of mµkl . For the projecting scheme to hold, τ0 must be larger
than the typical time scale on which we observe the positions of the particles.

A.5.1 Reduced time evolution operator

For the explicit calculation of the reduced time evolution operator, we need the stationary
solution of L̂0 ϕ0 = 0 which is of course the Maxwell-Boltzmann distribution (in three
dimensions)
 3N (
N
)
1 p2k

2 X
ϕ0 (x p ) = ϕMB (x p ) ≡ exp − . (A.19)
2πmT k=1
2mT

84
 A.5 Derivation of the Smoluchowski equation

Its gradient becomes

∂ ϕMB pk
=− ϕMB . (A.20)
∂ pk mT

We define the brackets as the integral

Z
〈·, ·〉 ≡ dx p · (x p ) · (x p )

and accordingly, ϕ̄0 = 1.

The first step is to calculate
pk” — pk
L̂1 ϕ0 = ϕ0 · −T ∇k + (−∇k U + fk ) + µ−1
kl
u(r l ) ≡ ϕ 0 · µ−1 v̂ ,
kl l
mT mT
where we have introduced the velocity operator

v̂k ≡ u(rk ) + µkl −∇l U + fl − T ∇l .

 
(A.21)

Applying this operator to the probability distribution ψ(x, t) leads to the local mean ve-
locity (2.9),

v̂k ψ = vk ψ.

Applying the projector (A.2) shows that

P̂ L̂1 ϕ0 = 〈ϕ̄0 , L̂1 ϕ0 〉ϕ0 = ϕ0 〈1, pk /(mT )ϕMB 〉 · µ−1 v̂ = 0

kl l
R
vanishes because of the Gaussian integral dx p pk ϕMB = 0. Therefore, the Smoluchowski
operator is given by the expression (A.6),

L̂SM = −〈ϕ̄0 , L̂1 L̂0−1 L̂1 ϕ0 〉. (A.22)

In the next step, we have to calculate L̂0−1 L̂1 ϕ0 involving

L̂0−1 pk ϕ0 = −mµkl pl ϕ0 . (A.23)

Hence,
1 pk
L0−1 L̂1 ϕ0 = ( L̂0−1 pk ϕ0 ) · µ−1 v̂ = −ϕ0
kl l
· v̂k . (A.24)
mT T
The last step is to apply the operator L̂1 again. It involves the derivative

∂ pk plT
(pl ϕ0 ) = 1δkl ϕ0 − ϕ0
∂ pk mT

85
A Appendix

which is, however, zero if we integrate over the momenta. As the result, most terms cancel
in the integral
Z
dx p L̂1 pk ϕ0 = −T ∇k . (A.25)

Finally, putting everything together, we obtain the Smoluchowski operator

L̂SM = 〈ϕ̄0 , L̂1 (pk ϕ0 )/T 〉 · v̂k = −∇k · v̂k (A.26)

as the reduced time evolution operator in the overdamped limit.

A.6 Evolution of time-local functionals

The derivation of the evolution equation for time-local functionals goes along the same
lines as in the derivation of the Smoluchowski equation in the previous section. We start
from the equations

ṙk = pk /m,
ṗk = −∇k U + fk − µ−1 pl /m − u(rl ) + ξk ,
 
kl
ṙ = [pk /m − u(rk )] · ak + β,

where the first two are the same as in the previous section and the last equation is an
ansatz with arbitrary vectors ak (x; γ) and a scalar function β(x; γ) justified in section 4.3
in the main text.
The state is Ψ(x, x p , r, t) obeying

∂ Ψ(t) = [ L̂0 + L̂1 + L̂ r ]Ψ(t).

The operators L̂0 and L̂1 are given in (A.17) and (A.18), respectively. In addition, there is
an operator

L̂ r ≡ − [pk /m − u(rk )] · ak + β ∂ r

(A.27)

acting on r. First, we note that due to the additional operator now there is a component
in the null space of L̂0−1 ,

P̂( L̂1 + L̂ r )ϕ0 = P̂ L̂ r ϕ0 = ϕ0 [u(rk ) · ak − β]∂ r ,

and hence

P̂ 0 ( L̂1 + L̂ r )ϕ0 = [ L̂1 − (pk /m) · ak ∂ r ]ϕ0 .

86
 A.6 Evolution of time-local functionals

The time evolution operator for the reduced joint probability ρ(x, r, t) without the mo-
menta then becomes

L̂ = [u(rk ) · ak − β]∂ r − 〈ϕ̄0 , ( L̂1 + L̂ r ) L̂0−1 [ L̂1 − (pk /m) · ak ∂ r ]ϕ0 〉.

This expression contains the Smoluchowski operator (A.26). The first part of the remain-
ing operator reads
Z Z
dx p L̂1 L̂0−1 (pk /m) · ak ∂ r ϕ0 = − dx p L̂1 pk · µkl al ∂ r ϕ0 = T ∇k · µkl al ∂ r

using (A.23) and (A.25). The second part is slightly more complicated and we proceed in
two steps. First,
•p ˜
k
L̂0 [ L̂1 − (pk /m) · ak ∂ r ]ϕ0 = −
−1
· v̂k − pk · µkl al ∂ r ϕ0
T
employing (A.24) and again (A.23). This operator is proportional to pk . Now we apply
the operator (A.27) but only the first term containing the momentum will survive the
following integration, leading to the result

ak · v̂k ∂ r − T ak · µkl al ∂ r2 .

Collecting everything together, we obtain the evolution operator

L̂ = L̂SM − ak · [v̂k − u(rk )] − ∇k · Dkl al + β ∂ r + ak · Dkl al ∂ r2 .


(A.28)

87
Bibliography
[1] S. Carnot, “Réflexions sur la Puissance Motrice du Feu” (1824).

[2] D. A. McQuarrie, Statistical Mechanics (University Science Books, Sausalito, 2000).

[3] D. J. Evans and D. J. Searles, “The Fluctuation Theorem”, Adv. Phys. 51, 1529 –
1585 (2002).

[4] C. Bustamante, J. Liphardt, and F. Ritort, “The nonequilibrium thermodynamics of

small systems”, Physics Today 58(7), 43 (2005).

[5] F. Ritort, “Single-molecule experiments in biological physics: methods and applica-

tions”, J. Phys.: Condens. Matter 18, R531 (2006).

[6] R. J. Harris and G. M. Schütz, “Fluctuation theorems for stochastic dynamics”, J.

Stat. Mech.: Theor. Exp.,P07020 (2007).

[7] D. J. Evans, E. G. D. Cohen, and G. P. Morriss, “Probability of second law violations

in shearing steady states”, Phys. Rev. Lett. 71, 2401 (1993).

[8] G. Gallavotti and E. G. D. Cohen, “Dynamical Ensembles in Nonequilibrium Statis-

tical Mechanics”, Phys. Rev. Lett. 74, 2694 (1995).

[9] J. Kurchan, “Fluctuation theorem for stochastic dynamics”, J. Phys. A: Math. Gen.
31, 3719 (1998).

[10] J. L. Lebowitz and H. Spohn, “A Gallavotti-Cohen-Type Symmetry in the Large De-

viation Functional for Stochastic Dynamics”, J. Stat. Phys. 95, 333 (1999).

[11] C. Jarzynski, “Nonequilibrium Equality for Free Energy Differences”, Phys. Rev.
Lett. 78, 2690 (1997).

[12] G. E. Crooks, “Entropy production fluctuation theorem and the nonequilibrium

work relation for free energy differences”, Phys. Rev. E 60, 2721 (1999).

89
Bibliography

[13] G. E. Crooks, “Path-ensemble averages in systems driven far from equilibrium”,

Phys. Rev. E 61, 2361 (2000).

[14] R. van Zon and E. G. D. Cohen, “Extension of the fluctuation theorem”, Phys. Rev.
Lett. 91, 110601 (2003).

[15] P. Gaspard, “Fluctuation theorem for nonequilibrium reactions”, J. Chem. Phys.

120, 8898 (2004).

[16] U. Seifert, “Entropy production along a stochastic trajectory and an integral fluctu-
ation theorem.”, Phys. Rev. Lett. 95, 040602 (2005).

[17] A. Einstein, “Über die von der molekularkinetischen Theorie der Wärme geforderte
Bewegung von in ruhenden Flüssigkeiten suspendierten Teilchen”, Ann. Phys. 17,
549 (1905).

[18] H. Nyquist, “Thermal Agitation of Electric Charge in Conductors”, Phys. Rev. 32,
110 (1928).

[19] L. Onsager, “Reciprocal Relations in Irreversible Processes. I.”, Phys. Rev. 37, 405
(1931).

[20] L. Onsager, “Reciprocal Relations in Irreversible Processes. II.”, Phys. Rev. 38, 2265
(1931).

[21] D. J. Evans and G. P. Morriss, Statistical Mechanics of Nonequilibrium Liquids (Aca-

demic Press, London, 1990).

[22] G. Gallavotti, “Extension of Onsager’s Reciprocity to Large Fields and the Chaotic
Hypothesis”, Phys. Rev. Lett. 77, 4334 (1996).

[23] G. N. Bochkov and Y. E. Kuzovlev, “Nonlinear fluctuation-dissipation relations and

stochastic models in nonequilibrium thermodynamics I. Generalized fluctuation-
dissipation theorem”, Physica A 106, 443–479 (1981).

[24] G. N. Bochkov and Y. E. Kuzovlev, “Nonlinear fluctuation-dissipation relations and

stochastic models in nonequilibrium thermodynamics II. Kinetic potential and vari-
ational principles for nonlinear irreversible processes”, Physica A 106, 480–520
(1981).

[25] C. Jarzynski, “Comparison of far-from-equilibrium work relations”, C. R. Physique

8, 495–506 (2007).

90
 Bibliography

[26] G. Hummer and A. Szabo, “Free energy reconstruction from nonequilibrium single-
molecule pulling experiments”, Proc. Natl. Acad. Sci. U.S.A. 98, 3658 (2001).

[27] F. M. Ytreberg and D. M. Zuckerman, “Single-ensemble nonequilibrium path-

sampling estimates of free energy differences”, J. Chem. Phys. 120, 10876 (2004).

[28] O. Braun, A. Hanke, and U. Seifert, “Probing Molecular Free Energy Landscapes by
Periodic Loading”, Phys. Rev. Lett. 93, 158105 (2004).

[29] H. Oberhofer, C. Dellago, and P. L. Geissler, “Biased sampling of nonequilibrium

trajectories: Can fast switching simulations outperform conventional free energy
calculation methods?”, J. Phys. Chem. B 109, 6902 (2005).

[30] D. Collin, F. Ritort, C. Jarzynski, S. Smith, I. Tinoco, and C. Bustamante, “Verifica-

tion of the Crooks fluctuation theorem and recovery of RNA folding free energies”,
Nature 437, 231 (2005).

[31] F. M. Ytreberg, R. H. Swendsen, and D. M. Zuckerman, “Comparison of free energy

methods for molecular systems”, J. Chem. Phys. 125, 184114 (2006).

[32] W. Lechner, H. Oberhofer, and C. Dellago, “Equilibrium free energies from fast-
switching trajectories with large time steps”, J. Chem. Phys. 124, 044113 (2006).

[33] D. M. Zuckerman and T. B. Woolf, “Theory of a Systematic Computational Error in

Free Energy Differences”, Phys. Rev. Lett. 89, 180602 (2002).

[34] J. Gore, F. Ritort, and C. Bustamante, “Bias and error in estimates of equilibrium
free-energy differences from nonequilibrium measurements”, Proc. Natl. Acad. Sci.
U.S.A. 100, 12564 (2003).

[35] T. Hatano and S. Sasa, “Steady-State Thermodynamics of Langevin Systems”, Phys.

Rev. Lett. 86, 3463 (2001).

[36] Y. Oono and M. Paniconi, Prog. Theor. Phys. Suppl. 130, 29 (1998).

[37] U. Seifert, “Stochastic thermodynamics: Principles and perspectives”,

arXiv:0710.1187 (2007).

[38] F. Ritort, “Nonequilibrium fluctuations in small systems: From physics to biology”,

arXiv:0705.0455 (2007).

91
Bibliography

[39] F. Douarche, S. Ciliberto, A. Petrosyan, and I. Rabbiosi, “An experimental test of the
Jarzynski equality in a mechanical experiment”, Europhys. Lett. 70, 593 (2005).

[40] G. Binnig, C. F. Quate, and C. Gerber, “Atomic Force Microscope”, Phys. Rev. Lett.
56, 930 (1986).

[41] A. Ashkin, “Acceleration and Trapping of Particles by Radiation Pressure”, Phys.

Rev. Lett. 24, 156 (1970).

[42] A. Ashkin, “Optical trapping and manipulation of neutral particles using lasers”,
Proc. Natl. Acad. Sci. U.S.A. 94, 4853 (1997).

[43] G. M. Wang, E. M. Sevick, E. Mittag, D. J. Searles, and D. J. Evans, “Experimen-

tal Demonstration of Violations of the Second Law of Thermodynamics for Small
Systems and Short Time Scales”, Phys. Rev. Lett. 89, 050601 (2002).

[44] D. M. Carberry, J. C. Reid, G. M. Wang, E. M. Sevick, D. J. Searles, and D. J. Evans,

“Fluctuations and Irreversibility: An Experimental Demonstration of a Second-Law-
Like Theorem Using a Colloidal Particle Held in an Optical Trap”, Phys. Rev. Lett.
92, 140601 (2004).

[45] E. H. Trepagnier, C. Jarzynski, F. Ritort, G. E. Crooks, C. J. Bustamante, and J.

Liphardt, “Experimental test of Hatano and Sasa’s nonequilibrium steady-state
equality”, Proc. Natl. Acad. Sci. U.S.A. 101, 15038 (2004).

[46] G. M. Wang, J. C. Reid, D. M. Carberry, D. R. M. Williams, E. M. Sevick, and D. J.

Evans, “Experimental study of the fluctuation theorem in a nonequilibrium steady
state”, Phys. Rev. E 71, 046142 (2005).

[47] J. Liphardt, S. Dumont, S. B. Smith, I. Tinoco Jr, and C. Bustamante, “Equilib-

rium Information from Nonequilibrium Measurements in an Experimental Test of
Jarzynski’s Equality”, Science 296, 1832 (2002).

[48] H. Kamegawa, T. Hondou, and F. Takagi, “Energetics of a forced thermal ratchet”,

Phys. Rev. Lett. 80, 5251–5254 (1998).

[49] C. Jarzynski and O. Mazonka, “Feynman’s ratchet and pawl: An exactly solvable
model”, Phys. Rev. E 59, 6448–6459 (1999).

[50] I. M. Kulic, R. Thaokar, and H. Schiessel, “A DNA ring acting as a thermal ratchet”,
J. Phys.: Condens. Matter 17, S3965–S3978 (2005).

92
 Bibliography

[51] J. C. M. Gebhardt, A. E. M. Clemen, J. Jaud, and M. Rief, “Myosin-V is a mechanical

ratchet”, Proc. Natl. Acad. Sci. U.S.A. 103, 8680–8685 (2006).

[52] R. Lipowsky, “Universal Aspects of the Chemomechanical Coupling for Molecular

Motors”, Phys. Rev. Lett. 85, 4401 (2000).

[53] P. Reimann, “Brownian motors: noisy transport far from equilibrium”, Phys. Rep.
361, 57 (2002).

[54] P. Szekeres, A Course in Modern Mathematical Physics (Cambridge University

Press, Cambridge, 2004).

[55] H. Risken, The Fokker-Planck Equation, 2nd ed. (Springer-Verlag, Berlin, 1989).

[56] J. K. G. Dhont, An Introduction to Dynamics of Colloids (Elsevier, Amsterdam,

1996).

[57] M. Doi and S. F. Edwards, The Theory of Polymer Dynamics (Clarendon Press,
Oxford, 1986).

[58] A. Berman and R. J. Plemmons, Nonnegative Matrices in the Mathematical Sciences

(SIAM, Philadelphia, 1994).

[59] M. P. Langevin, “Sur la théorie du mouvement brownien”, Comptes Rend. Acad.

Sci. (Paris) 146, 530 (1908).

[60] A. W. C. Lau and T. C. Lubensky, “State-dependent diffusion: Thermodynamic con-

sistency and its path integral formulation”, Phys. Rev. E 76, 011123 (2007).

[61] D. S. Dean, “Langevin equation for the density of a system of interacting Langevin
processes”, J. Phys. A: Math. Gen. 29, L613 (1996).

[62] A. Andreanov, G. Biroli, J.-P. Bouchaud, and A. Lefevre, “Field theories and ex-
act stochastic equations for interacting particle systems”, Phys. Rev. E 74, 030101
(2006).

[63] K. Sekimoto, “Kinetic Characterisation of Heat Bath and the Energetics of Thermal
Ratchet Models”, J. Phys. Soc. Jpn. 66, 1234–1237 (1997).

[64] K. Sekimoto, “Langevin Equation and Thermodynamics”, Prog. Theor. Phys. Supp.
130, 17 (1998).

[65] D. Prieve, “Measurement of Colloidal Forces with TIRM”, Adv. Colloid Interface Sci.
82, 93 (1999).

93
Bibliography

[66] V. Blickle, D. Babic, and C. Bechinger, “Evanescent light scattering with magnetic
colloids”, Appl. Phys. Lett. 87, 101102 (2005).

[67] S. ichi Sasa and H. Tasaki, “Steady State Thermodynamics”, J. Stat. Phys. 125,
125–224 (2006).

[68] J. Johnson, “Dynamics of a flexible polymer chain in steady shear flow: the Rouse
model”, Macromolecules 20, 103 (1987).

[69] J. Zinn-Justin, Quantum Field Theory and Critical Phenomena, fourth ed. (Oxford
University Press, New York, 2002).

[70] M. Chaichian and A. Demichev, Path integrals in physics (Institute of Physics pub-
lishing, Bristol and Philadelphia, 2001).

[71] V. Y. Chernyak, M. Chertkov, and C. Jarzynski, “Path-integral analysis of fluctua-

tion theorems for general Langevin processes”, J. Stat. Mech.: Theor. Exp.,P08001
(2006).

[72] C. Maes, “On the Origin and Use of Fluctuation Relations for Entropy”, Sém.
Poincaré 2, 29 (2003).

[73] C. Maes and K. Netocný, “Time-reversal and entropy”, J. Stat. Phys. 110, 269
(2003).

[74] C. Jarzynski, “Equilibrium free-energy differences from nonequilibrium measure-

ments: A master-equation approach”, Phys. Rev. E 56, 5018 (1997).

[75] O. Mazonka and C. Jarzynski, cond-mat/9912121 (unpublished).

[76] A. Imparato and L. Peliti, “Work probability distribution in single-molecule experi-

ments”, Europhys. Lett. 69, 643 (2005).

[77] A. Imparato and L. Peliti, “Work distribution and path integrals in general mean-
field systems”, Europhys. Lett. 70, 740 (2005).

[78] A. Imparato and L. Peliti, “The distribution function of entropy flow in stochastic
systems”, J. Stat. Mech.: Theor. Exp.,L02001 (2007).

[79] H. Brenner, “The slow motion of a sphere through a viscous fluid towards a plane
surface”, Chem. Eng. Sci. 16, 242 (1961).

[80] Handbook of Mathematical Functions, 9th ed., edited by M. Abramowitz and I. A.

Stegun (Dover, New York, 1972).

94
 Bibliography

[81] J.-D. Bao, P. Hänggi, and Y.-Z. Zhuo, “Non-Markovian Brownian dynamics and non-
ergodicity”, Phys. Rev. E 72, 061107 (2005).

[82] P. Hänggi and H. Thomas, “Time Evolution, Correlations, and Linear Response of
Non-Markov Processes”, Z. Physik B 26, 85 (1977).

[83] P. Hänggi and H. Thomas, “Stochastic processes: Time evolution, symmetries and
linear response”, Phys. Rep. 88, 207 (1982).

[84] U. Seifert, “Fluctuation theorem for birth-death or chemical master equations with
time-dependent rates”, J. Phys. A: Math. Gen. 37, L517 (2004).

[85] T. Schmiedl and U. Seifert, “Stochastic thermodynamics of chemical reaction net-

works”, J. Chem. Phys. 126, 044101 (2007).

[86] R. Kubo, “Fluctuation-Dissipation Theorem”, Rept. Progr. Phys. 29, 255 (1966).

[87] R. Kubo, M. Toda, and N. Hashitsume, Statistical Physics II, 2nd ed. (Springer-
Verlag, Berlin, 1991).

[88] G. S. Agarwal, “Fluctuation-dissipation theorems for systems in non-thermal equi-

librium and applications”, Z. Physik 252, 25 (1972).

[89] L. F. Cugliandolo, J. Kurchan, and G. Parisi, “Off equilibrium dynamics and aging
in unfrustrated systems”, J. Phys. I 4, 1641 (1994).

[90] P. Calabrese and A. Gambassi, “Ageing properties of critical systems”, J. Phys. A:

Math. Gen. 38, R133 (2005).

[91] K. Binder and W. Kob, Glassy Materials and Disordered Solids (World Scientific,
Singapore, 2005).

[92] S. C. L. Bellon and C. Laroche, “Violation of the fluctuation-dissipation relation

during the formation of a colloidal glass”, Europhys. Lett. 53, 511 (2001).

[93] A. Crisanti and F. Ritort, “Violation of the fluctuation-dissipation theorem in glassy

systems: basic notions and the numerical evidence”, J. Phys. A: Math. Gen. 36,
R181 (2003).

[94] R. Mauri and D. Leporini, “Violation of the fluctuation-dissipation theorem in con-

fined driven colloids”, Europhys. Lett. 76, 1022 (2006).

95
Bibliography

[95] L. F. Cugliandolo, J. Kurchan, and L. Peliti, “Energy flow, partial equilibration,

and effective temperatures in systems with slow dynamics”, Phys. Rev. E 55, 3898
(1997).

[96] J. L. Barrat and L. Berthier, “Fluctuation-dissipation relation in a sheared fluid”,

Phys. Rev. E 63, 012503 (2001).

[97] L. Berthier and J.-L. Barrat, “Nonequilibrium dynamics and fluctuation-dissipation

relation in a sheared fluid”, J. Chem. Phys. 116, 6228 (2002).

[98] P. Reimann, C. van den Broeck, H. Linke, P. Hänggi, M. Rubi, and A. Pérez-Madrid,
“Giant Acceleration of Free Diffusion by Use of Tilted Periodic Potentials”, Phys.
Rev. Lett. 87, 010602 (2001).

[99] P. Reimann, C. V. den Broeck, H. Linke, P. Hanggi, J. M. Rubi, and A. Perez-Madrid,

“Diffusion in tilted periodic potentials: Enhancement, universality, and scaling”,
Phys. Rev. E 65(3), 031104 (2002).

[100] W. Götze, “Recent tests of the mode-coupling theory for glassy dynamics”, J. Phys.:
Condens. Matter 11, A1–A45 (1999).

[101] M. Fuchs and M. E. Cates, “Integration through transients for Brownian particles
under steady shear”, J. Phys.: Condens. Matter 17, S1681 (2005).

[102] N. G. van Kampen, “Elimination of fast variables”, Phys. Rep. 124, 69 (1985).

[103] C. W. Gardiner, Handbook of Stochastic Methods, 3rd ed. (Springer-Verlag, Berlin,

2004).

96
Danksagung
Ich danke allen, die zum Gelingen dieser Arbeit beigetragen haben:

• Herrn Prof. Dr. Seifert für die Ermöglichung dieser Dissertation, die wissenschaft-
liche Betreuung und Förderung, sowie die anregenden Diskussionen,

• Herrn Prof. Dr. Dietrich für die Übernahme und schnelle Anfertigung des Mitberichtes,

• Herrn Prof. Dr. Bechinger für die gelungene Zusammenarbeit in experimentellen

Fragen,

• Valentin Blickle, dessen experimentelle Arbeit wichtige Impulse und Ergebnisse ge-
liefert hat,

• Sebastian Schuler und Carsten Tietz für die gelungene Zusammenarbeit im Rahmen
der Experimente an einer Fehlstelle in Diamant,

• Frau Meyer-Haake und Frau Steinhauser für die freundliche Hilfe in organisato-
rischen Belangen,

• dem gesamten Institut für die schöne Zeit und die stets angenehme Atmosphäre,

• insbesondere Reimar Finken für sein unerschöpfliches mathematisches Wissen sowie

Jakob Mehl für die Korrekturlesung,

• meiner Familie und Assja Ivanova für ihre Unterstützung während der gesamten
Zeit.

97