Diploma Thesis
On Brownian motion of asymmetric particles An application as molecular motor
MARTIN REICHELSDORFER
May 2010 Institut f¨urTheoretische Physik I Prof. Dr. Klaus Mecke Friedrich-Alexander Universit¨atErlangen-N¨urnberg Zusammenfassung
Asymmetrische Brownsche Teilchen k¨onnenunter Nichtgleichgewichtsbedingungen eine im Mittel ge- richtete Bewegung ausf¨uhren. Dies kann zur Konstruktion molekularer Motoren genutzt werden. Die vorliegende Arbeit befasst sich mit der Erweiterung und Analyse eines biologisch inspirierten Modells, welches von Susan Sporer et al. [42] eingef¨uhrt wurde. Aufgrund der oft sehr allgemeinen und abstrakten Darstellung sind die Betrachtungen jedoch auch im weiter gefassten Kontext der generellen Brownschen Bewegung asymmetrischer Objekte relevant. Das Modell besteht aus einem asymmetrischen zweidimensionalen K¨orper, der sich entlang einer Schiene in einem zweidimensionalen, unendlich ausgedehnten idealen Gas bewegt. Durch vor¨uber- gehendes Verlangsamen an Stoppstellen wird der Motor bzw. das zugeh¨origestatistische Ensemble aus dem Gleichgewicht gebracht. Anhand des Beispiels eines parabolischen Potentials mit exponentiell abfallender Federkonstante wird gezeigt, dass sich diese Bindungspl¨atzedurch zeitlich ver¨anderliche Potentialmulden darstellen lassen. Ahnlich¨ der K¨uhlung eines Gases durch Expansion, f¨uhrtdieser Ausschaltevorgang zu einer Verschm¨alerungder Geschwindigkeitsverteilung. Haupts¨achlich w¨ahrend der darauffolgenden Relaxation erh¨altdie zuf¨alligeBewegung eine Vorzugsrichtung bis sich erneut die Gleichgewichtsverteilung einstellt. Ein einmaliges L¨osendes Motors aus einer Stoppstelle erweist sich im betrachteten Beispiel als eher ineffizient, in dem Sinne dass die Breite der Ortsverteilung stets deutlich gr¨oßerals die Verschiebung ihres Mittelwertes ist. Durch periodische Anordnung der Bindungspl¨atzejedoch kann eine effektive Gleichrichtung der Bewegung erreicht werden, da eine Equilibrierung des Systems verhindert wird und sich stattdessen ein station¨arerNichtgleichgewichts- zustand einstellt. Da das Modell, bedingt durch seine Komplexit¨at,nur n¨aherungweise gel¨ostwerden kann, werden ferner eine Reihe reduzierter Modelle entwickelt. Diese weisen die gleichen charakteristischen Eigen- schaften wie das allgemeinere Modell auf, sind jedoch analytisch besser zug¨anglich. Hervorzuheben ist insbesondere ein System aus nur drei diskreten Geschwindigkeitszust¨anden,welches als Minimalmo- dell eines asymmetrischen Brownschen Teilchens in einer Dimension angesehen werden kann. Es zeigt sich, dass sich die Bewegung von asymmetrischen und symmetrischen Teilchen nicht nur im Nichtgle- ichgewicht (wie vielleicht aus den vorangegangen Beobachtungen zu vermuten gewesen w¨are)sondern auch im Gleichgewicht unterscheidet. Durch Erg¨anzungum diskrete Ortszust¨aendeerlaubt dieses Modell dar¨uber hinaus die bereits angesprochene Untersuchung sich wiederholender Stoppstellen. Hierbei lassen sich qualitative Ubereinstimmungen¨ zu experimentellen Daten eines monomerischen Kinesin KIF1A Motors feststellen. Zu ¨uberpr¨ufen, ob sich die Ergebnisse auch quantitativ repro- duzieren lassen, w¨areAufgabe weiterer Untersuchungen. Verh¨altnism¨aßiggroße thermische Fluktuationen sind die treibende Kraft des Motors, weshalb neue Methoden zu seiner thermodynamischen Beschreibung angewandt werden m¨ussen.Im Rahmen der sogenannten stochastischen Thermodynamik k¨onneneinzelnen Phasenraumtrajektorien Gr¨oßen wie Energie, Arbeit, W¨armeoder sogar Entropie zugeordnet werden, die nunmehr als stochastische Variablen mit ausgedehnten Wahrscheinlichkeitsverteilungen zu verstehen sind. Hinsichtlich dieser Verteilungen l¨asstsich eine Reihe von Relationen, genannt Fluktuationstheoreme, finden, die oft eine detailliertere Fassung bekannter thermodynamischer Aussagen, wie etwa des zweiten Hauptsatzes, darstellen und bereits bei der Analyse experimenteller Daten, z.B. bei der Messung von freien En- ergien, Anwendung gefunden haben. Die G¨ultigkeit dieser Theoreme f¨urdas vorliegende Modell kann durch direkte Rechnung best¨atigtwerden. Abstract
Asymmetric Brownian particles are able to conduct directed average motion under non-equilibrium conditions. This can be used for the construction of molecular motors. The present paper is concerned with the extension and analysis of a biologically inspired models, which was introduced by Susan Sporer et al. [42]. Due to the often quite abstract formulation, the considerations are also relevant in the broader context of general Brownian motion of asymmetric objects. The model consists of an asymmetric two-dimensional body, which moves along a rail through a two-dimensional, infinite ideal gas. By temporarily slowing it down at stopping sites, the motor or the corresponding statistical ensemble, respectively, is pushed away from equilibrium. The example of a parabolic potential with exponentially decreasing spring constant shows that these binding sites can be modelled by time-dependent potential wells. Similar to the cooling of a gas by expansion, this release process leads to a narrowing of the velocity distribution. Particularly during the subsequent relaxation the random motion is biased until the equilibrium distribution is established again. A singular release of the motor is found to be rather inefficient in the considered case, in the sense that the width of the distribution of positions is always profoundly larger than the drift of its mean value. By periodically distributing the binding sites, however, an effective rectification of the motion can be achieved, since the equilibration of the system is prevented and a stationary non-equilibrium sate establishes instead. As the model can only be solved approximately due to its complexity, a series of reduced models is developed. These systems exhibit the same characteristic features as the more general model, but are easier to access analytically. In particular, a system composed of only three discrete velocity levels can be regarded as a minimal model of an asymmetric Brownian particle in one dimension. It turns out that the motion of asymmetric and symmetric particles differ not only in non-equilibrium (as might have been anticipated from prior observations), but also in equilibrium. Moreover, augmenting the model with discrete position states allows for the already mentioned investigation of repeated stopping sites. Hereat, qualitative accordances to experimental data from a monomeric kinesin KIF1A motor are observed. To test, whether the results can also be reproduced quantitatively could be the subject of further studies. Comparatively large thermal fluctuations are the driving force of the motor, wherefore new methods have to be applied for its thermodynamic description. In the framework of so-called stochastic thermodynamics, quantities like energy, work, heat or even entropy may be attributed to individual phase space trajectories. Hence, they are now to be understood as stochastic variables with wide- stretching probability distributions. With respect to these distributions a multitude of relations, called fluctuation theorems, can be found, which often constitute a more detailed version of common thermodynamic statements like the second law, and which have already been applied in the analysis of experimental data (e.g. in the determination of free energies). The validity of these theorems for the present model can be confirmed by direct calculation. 4 Contents
1 Why Brownian motion is still interesting 7
2 Outline 9
3 Molecular motors in theory and reality – the context of this work 11 3.1 Two examples of theoretical models ...... 11 3.2 Kinesin ...... 11 3.3 A kinesin inspired model ...... 14
4 Brownian motion as stochastic process - the mathematical framework 17 4.1 Selected foundations of stochastic processes ...... 17 4.2 Description of the Model ...... 21 4.3 Modelling the stochastic process ...... 21 4.4 Time evolution of the phase space probability density ...... 25 4.5 Time evolution of moments of the phase space probability density ...... 29
5 Using potential wells as stopping sites 31 5.1 Relaxation without potential ...... 32 5.2 Potential-change without collisions: time-dependent undamped harmonic oscillator . . 37 5.3 Time evolution with potential and collisions ...... 41 5.4 A remark on re-opening and multiple stopping sites ...... 50
6 Minimal Models 53 6.1 Two-state model ...... 53 6.2 Three-state generic model of an asymmetric Brownian particle in one dimension . . . 57 6.3 Augmented three-state model ...... 64 6.4 A step towards bridging the gap back to reality ...... 70
7 Stochastic thermodynamics 73 7.1 Energy, work and heat ...... 73 7.2 Entropy ...... 75 7.3 Fluctuation theorems ...... 76
8 Conclusion and Outlook 81
A Derivations and Proofs 85 A.1 Coefficients of the differential Chapman-Kolmogorov equation ...... 85 A.2 Relation of dissipated heat and transition rates ...... 88 A.3 Expansion of the Kramers-Moyal coefficients ...... 89 A.4 Limit of the mean squared velocity of the time-dependent undamped harmonic oscillator 90 A.5 Moments of the phase space distribution of an harmonic oscillator in equilibrium . . . 91 A.6 Properties of the spectrum of the transition matrix ...... 91
5 Contents
6 1 Why Brownian motion is still interesting
For thou wilt mark here many a speck, impelled By viewless blows, to change its little course, And beaten backwards to return again, Hither and thither in all directions round.
Titus Lucretius Carus (ca. 60 BC) [5]
Actually, Brownian motion is a fairly old story. Observations date back as far as to ancient times, where, for instance, in the first century BC Titus Lucretius Carus [5] describes the jittering motion of small specks of dust in sunlight and – probably inspired by Epicurean atomistic physics – originates it in the underlying dynamics of the “primeval atoms”. In 1827, the Scottish botanist Robert Brown [4] finds erratically moving floating particles during his observations of pollen and succeeds in drawing the scientific community’s attention back to this phenomenon, bringing him the honour of coining its name. Some 80 years later, Albert Einstein [12] and Marian Smoluchowski [44] contribute greatly to clarify the physical background and present a quantitative analytical description, serving as well as a way of indirectly confirming the existence of atoms and molecules, being still disputed at that time. Their models successfully explain the characteristic observations of Brownian motion, namely a linearly growing mean squared displacement1 of the particle and and unchanged average position:
hXi = 0, (1.1) hX2i = 2dDt. (1.2)
The parameter d gives the number of dimensions and D is the diffusion constant. Moreover, Einstein and Smoluchowski are able to relate friction and diffusion to the same underlying mechanism, the collisions of fluid molecules with the Brownian particle. This fact is expressed by the so-called Einstein-Smoluchowski relation (cf. Reference [12] or [30] p. 120)
k T D = B , (1.3) γ which relates the diffusion constant D to the friction coefficient γ. As factor of proportionality serves the thermal energy, where kB is Boltzmann’s constant and T the absolute temperature. Now, again more then a hundred years later, still many physicists are working on this topic. One of the reason is an increasing interest in small (often also non-equilibrium) systems, where thermal fluctuations are comparably large and play a fundamental role. This is in particular true in the research field of biological physics, where biological processes are physically investigated down to molecular scales. The biological background of the present paper are so-called molecular motors,
1In practise, this linear growth law of hX2i does not hold for arbitrary small times, as Einstein already remarks in Reference [13]. Instead, one typically observes a transition from ballistic (hX2i ∝ t2) to diffusive (hX2i ∝ t) motion.
7 1 Why Brownian motion is still interesting which are known to carry out active transport by, for instance, dragging along loads or pumping ions through membranes. Even force generation in muscle cells is due to the co-operation of many individual motor molecules (for instance, see References [3, 30, 23, 22, 2]). Remarkably, many models for these system are based on the seemingly undirected Brownian motion, but at a closer look one finds that, indeed, under non-equilibrium conditions and in presence of a spatial asymmetry net motion may be obtained. This class of molecular motors is therefore called Brownian motors.
8 2 Outline
The focus of the present work lies on the analysis and the extension of a model of a Brownian motor, which was introduced by Susan Sporer et al. (cf. References [42, 41]), and which makes use of an intrinsic asymmetry of the Brownian particle itself. Non-equilibrium is established by stopping and releasing the motor at stopping sites. In course of the generalisation, the approach to the problem resembles in places the steps taken in Reference [41], and results therein obtained are retrieved as specialisations. As the treatment is often quite abstract, it has also relevance for the general issue of Brownian motion of asymmetric particles, especially in non-equilibrium states. The paper is organized as follows: Chapter 3 introduces the conceptional and the biological background of the present model. The major results of Reference [42] are outlined. In Chapter 4, the model is described in more detail and the mathematical framework is developed. The original description, which was based on a master equation for the time evolution of the velocity probability density, is extended to include positions and time-dependent potentials. Moreover, it is derived from a Poisson type stochastic process, which models individual collisions of particles of the medium with the motor. The probability of a specific realisation of this stochastic process is given, which may in principal serve as a starting point for path integral techniques an which is applied to directly verify so-called fluctuation theorems in Chapter 7. The inclusion of potentials allows to show in Chapter 5 how stopping sites may be realised by potential wells – a point completely left open in Reference [42]. Interestingly, not the trapping, but the release from the well by gradually turning it off stops the motor in the sense that the velocity distribution is narrowed. The efficiency of the process and different influences are studied. In general, a singular release from a potential well is found to have a rather weak rectifying effect. Since a complete solution of the “full” model is not feasible, a set of minimal models is developed in Chapter 6. These systems have the advantage of allowing for better analytical access, while still exhibiting the same characteristic behaviour. In particular, a system with three velocity states is devised as a minimal model for an asymmetric Brownian particle in one dimension. The effect of asymmetry, being perceivable even in equilibrium, and the role of detailed balance are examined. Additionally, the implementation of a discrete position space makes it possible to study the influence of (multiple) stopping sites that are able to bind a released particle again. This setup is capable of preventing the equilibration of the system successfully and the motion of the particle is effectively rectified. Finally, in Chapter 7, the model is viewed from the point of stochastic thermodynamics. As thermal fluctuations are the driving force of the motor, they have to be accounted for in thermodynamic considerations, leading to trajectory dependent formulations of quantities like work, heat or entropy, which – instead of single deterministic values – posses probability distributions, which obey the fluctuation theorems mentioned above. Using the path integral weights derived in Chapter 4, it is possible to verify these relations by direct calculation.
9 2 Outline
10 3 Molecular motors in theory and reality – the context of this work
In this chapter, two theoretical models are very briefly discussed. The first one is a very prominent example, showing very instructively how the combination of asymmetry and non-equilibrium can generate net motion. The second one has great similarities to the present system and served as an inspiration in fact. Additionally, also the biological background, a molecular motor called kinesin, is introduced. Eventually, the model from References [42, 41], which is the basis of the present work, is briefly sketched and relevant results are summarised.
3.1 Two examples of theoretical models
There are two basic ingredients, needed to rectify Brownian motion: asymmetry and non-equilibrium1 (cf., for instance, Reference [30] p. 422). In the last few decades, a multitude of models have been brought forward, all sporting different implementations of this concept. Reference [32] gives a very thorough overview of this topic. Some of these models have already successfully been used to explain experimental data of biological molecular motors. One of the first models that has been proposed features a periodic sawtooth potential that is turned off and on repeatedly (cf. Reference [33]). See Figure 3.1 for an illustration. The particles gather in the minima and start diffusing symmetrically when the potential is turned off. Since the distances to the neighbouring maxima on the left and on the right are different, more particles will have crossed the closer one when the potential is turned on again and slide into the next minimum. Thus, a net current is established. Another – not so intuitive – model is sketched in Figure 3.2. Two asymmetric objects, which are stiffly linked, are placed each in a separate heat bath, consisting of ideal gases. Different variants of this setup are studied in References [10, 29, 11, 46, 45, 9], including various object shapes, two- dimensional, three-dimensional, translational and rotational realisations, and even the application as heat pump or refrigerator, respectively. If the temperatures are different, the system is not in equilibrium and the motor performs a net motion while transferring heat (by collisions with gas particles) from the warmer to the cooler reservoir. In contrast to the first system, which uses an asymmetric external potential, this model possesses an intrinsic asymmetry. The same principle is used by the motor under consideration in the present paper, but another mechanism for perturbing thermal equilibrium is applied.
3.2 Kinesin
According to Reference [30] p. 58, the family of single-protein motors called kinesins are, for instance, involved in the transport of proteins to axon terminals and of the ingredients from which synaptic vesicles will be built. The following is based on Reference [30] p. 437 ff. A kinesin molecule usually possesses two heads or motor domains, with which according to common theory it “walks” down microtubules by binding and unbinding the heads at subsequent binding sites. During one cycle the energy from
1Throughout this paper, when speaking of the motor being in or out of equilibrium, or having a certain distribution, this, of course, actually refers to the whole corresponding statistical ensemble.
11 3 Molecular motors in theory and reality – the context of this work
(a) U
tential U
(b) concentration po
(c) concentration
Figure 3.1 Model of a Brownian motor using a sawtooth potential that is periodi- cally switched on and off (a). (b) In the “on”-state particles gather in the potential minima. (c) In the “off”-state particles diffuse freely. Due to the asymmetry of the po- tential, more particles cross the right-hand maximum (shaded area) than the left-hand maximum. (adapted from Reference [33])
Figure 3.2 Model of a thermal Brownian motor, making use of intrinsic asymmetry. Two (in this particular case identical) rigidly linked objects are placed in separated heat baths with different Temperatures T1 and T2. (adapted from Reference [29])
12 3.2 Kinesin
Figure 3.3 Details of the model for the stepping of two-headed kinesin. Each of the steps of this cyclic reaction is described in the text of Section 3.2. The steps form a loop, to be read clockwise from upper left. The grey symbols represent a microtubule, which is slightly polar, indicated by the “+” and the “-” sign. Strong physical bonds are denoted by multiple lines, weak ones by single lines. The symbols T, D and P denote ATP, ADP and inorganic phosphate, respectively. (from Reference [30]) the hydrolysis of one adenosine triphosphate molecule (ATP) is transduced and the motor crosses a distance of approximately 8 nm. In Figure 3.3 a proposed mechanism, some elements of which are still under debate, is graphically summarised. The cycle is not meant to be taken literally; it just shows some of the distinct steps in the enzymatic pathway and follows. Although taken from Reference [30], it originally follows an analysis in Reference [36]. Initially (top left panel of the figure), a kinesin dimer approaches the microtubule from solution and binds one head, releasing one of its adenosine diphosphates (ADP), meanwhile the other head cannot reach any binding site because its 0 tether is too short (E). Then, the bound head binds an ATP molecule from solution (ES1, ES1), which causes its so-called neck linker to dock on the head, biasing the other head’s random motion in the forward direction (see also Reference [26]). Being thrown forward by the bound head’s neck linker greatly increases the probability that the elastic tethers will momentarily stretch far enough for the free head to reach the next binding site. It may bind weakly, then detach, many times (ES2). Eventually, instead of detaching, the forward head releases its ADP and by that binds strongly to the microtubule (ES3). Meanwhile, the rear head splits its ATP and releases the resulting phosphate (EP). This reaction weakens its binding to the microtubule. The strain induced by the binding of the forward head then biases the rear head to unbind from the microtubule (rather than releasing its ADP). Obviously, kinesin has been cunningly designed to coordinate the action of its two heads, and one might wonder how such a complex motor could have evolved from something simpler. Or, putting the
13 3 Molecular motors in theory and reality – the context of this work
Figure 3.4 Model for single-headed kinesin motility. Bound ATP is denoted by T; ADP and P molecules are not shown. Other symbols are as in Figure 3.3. (a) Initially, the kinesin monomer is strongly bound to site n on the microtubule. (b) In the weakly bound state, the kinesin wanders freely along the microtubule. (c) When the kinesin reenters the strongly bound state, it is most likely to rebind to its original site, somewhat likely to rebind to the next site and least likely to bind to the previous site. Relative probabilities are represented by shading. (adapted from Reference [30]) matter differently, one could ask whether there is a simpler force-generating mechanism, perhaps not as efficient or as powerful as two-headed kinesin, which could have been its evolutionary precursor. In fact, a single-headed (monomeric) form of kinesin, called KIF1A, has been found to have single- molecule motor activity (cf. References [31, 30]). For this motor, experiments showed strongly stochastic trajectories with linearly increasing average displacement and variance. The authors of Reference [31] successfully modelled this system as a Brownian motor in an asymmetric ratchet potential (cf. Figure 3.1), which is assumed to originate from the polar structure of the microtubule. The chemical cycle of ATP binding, ATP hydrolysis and ADP release is comprised in two states (s and w), one corresponding to strong binding (s) (i.e, the motor feels the potential of the binding sites) and the other corresponding to weak binding (w) (i.e., the sawtooth potential is essentially turned off). After entering a state, the motor waits an average time ts or tw, respectively, before snapping back to the other state. This cycle is illustrated in Figure 3.4.
3.3 A kinesin inspired model
Inspired by the single headed kinesin Susan Sporer et al. [42, 41] brought forward a model of a Brownian motor, which picks up characteristic features of the biological system, but uses a different approach than Reference [31]. Instead of focussing on the slight asymmetry in the binding potentials, they based their model on the asymmetry of the Brownian particle itself. The microtubule is repre- sented by a rail with periodic stopping sites, where the motor is stopped and subsequently released, whenever it reaches one of them. This serves to sustain a non-equilibrium state. Figure 3.5 depicts a triangular two-dimensional motor in a Lennars-Jones fluid, which was used for simulations. As well by simulations as by analytical calculations it was possible to show that indeed an average motion could be generated (cf. Figures 3.6 and 3.7). Although it is unclear to what extent this mechanism plays a role in the motility of the kinesin KIF1A, the model constitutes an elegant concept to realise rectified Brownian motion in a completely symmetric environment and in a single heat bath.
14 3.3 A kinesin inspired model
Figure 3.5 An asymmetrically shaped motor (here a triangle) is built from fluid par- ticles and placed in a Lennard-Jones fluid. Its motion is restricted to a one-dimensional track with periodically spaced binding sites along the X axis. If the motors center of mass crosses a binding site, the velocity of the motor is set to zero. (adapted from Reference [42])
Figure 3.6 Average position (400 runs) of an asymmetrically oriented, isosceles triangle in a Lennard-Jones fluid for different distances δs between stopping sites (σ is the Lennard-Jones radius of a particle). Directed motion is possible if the triangle is oriented asymmetrically with respect to the direction of motion. Otherwise the motor performs symmetric fluctuations around its starting position. Inset: trajectories of individual motors. (from Reference [42])
15 3 Molecular motors in theory and reality – the context of this work
Figure 3.7 Relaxation of a motor to thermal equilibrium: the molecular dynamics simulations of the LJ-fluid (symbols; averages over 400000 runs) can be well described by the analytic results in the ideal gas limit (lines) if the data are normalized by the maximum velocity Vmax and the relaxation time. (from Reference [42])
16 4 Brownian motion as stochastic process - the mathematical framework
Due to the enormous amount of particles in the medium into which the Brownian particle is embedded, it is almost impossible to treat the system deterministically and a statistical approach is appropriate. This chapter is concerned with how the motion of the motor may be modelled as a stochastic process. While Section 4.1 provides some selected theoretical background, the remaining sections apply the latter to the specific model under consideration in a yet very general fashion, that is for still arbitrary potentials and motor shapes. Starting from the description as a Poisson type process in path integral representation, a time evolution equation of the phase space probability density function (phase space pdf) is derived. Additionally, techniques for obtaining approximate solutions of this equation are introduced, including a Kramers-Moyal expansion, an expansion in the mass ratio of motor and medium particles, and equations for the time evolution of the moments of the phase space pdf.
4.1 Selected foundations of stochastic processes
According to Reference [19] p. 42 ff., systems which evolve probabilistically in time or more precisely, systems in which a certain time-dependent random variable X(t) exists, can mathematically be described as stochastic processes. One can measure values x1, x2, x3, ..., etc., of X(t) at times t1, t2, t3, ... and it is assumed that a set of joint probability densities P (x1, t1; x2, t2; x3, t3; ...) exists which describe the system completely. In terms of these joint probability density functions, one can also define conditional probability densities
P (x1, t1; x2, t2; ...|y1, τ1; y2, τ2; ...) := P (x1, t1; x2, t2; ...; y1, τ1; y2, τ2; ...)/P (y1, τ1; y2, τ2; ...), (4.1) which may be used for the prediction of future values of X(t) (i.e., x1, x2, ... at times t1, t2, ...), given the knowledge of the past (values y1, y2, ..., at times τ1, τ2, ...).
4.1.1 Markov processes A stochastic process is called a Markov process if the conditional probability is determined entirely by the knowledge of the most recent condition, that is
P (x1, t1; x2, t2; ...|y1, τ1; y2, τ2; ...) = P (x1, t1; x2, t2; ...|y1, τ1), (4.2) given the times satisfy the ordering t1 ≥ t2 ≥ t3 ≥ ... ≥ τ1 ≥ τ2 ≥ ... . As a consequence, any arbitrary joint probability density may be decomposed into a product of simple conditional probabilities of the type P (xi, ti|yi, τi). For example, by definition of the conditional probability P (x1, t1; x2, t2|y1, τ1) = P (x1, t1|x2, t2; y1, τ1)P (x2, t2|y1, τ1) and using the Markov assumption (4.2), one finds P (x1, t1; x2, t2|y1, τ1) = P (x1, t1|x2, t2)P (x2, t2|y1, τ1). (4.3) Integrating over all events of one kind in a joint probability density eliminates that variable, that is Z P (x1, t1|x3, t3) = dx2 P (x1, t1; x2, t2|x3, t3) Z = dx2 P (x1, t1|x2, t2; x3, t3)P (x2, t2|x3, t3). (4.4)
17 4 Brownian motion as stochastic process - the mathematical framework
Application of the Markov assumption (4.2) then yields Z P (x1, t1|x3, t3) = dx2 P (x1, t1|x2, t2)P (x2, t2|x3, t3), (4.5) which is the Chapman-Kolmogorov equation.
4.1.2 The differential Chapman-Kolmogorov equation of a Markov process In this paper, the formalism developed here is going to be applied to a Brownian particle with one degree of freedom, X(t), and corresponding velocity V (t) (see Section 4.2 for the detailed definition of the system). The phase space position of the motor, Γ(t) := (X(t),V (t)), thus takes over the role of the random variable. The sample paths of X(t) are expected to be continuous, whereas V (t) will posses discontinuities due to collisions with instantaneous hard body interaction. According to Reference [19] p. 46, it can be shown that with probability one the sample paths of a Markov process are continuous functions of t, if for any > 0 1 Z lim dx P (x, t + ∆t|z, t) = 0 (4.6) ∆t→0 ∆t |x−z|> uniformly in z and t. On page 47 it is stated that the Chapman-Kolmogorov equation can be reduced to a differential equation under appropriate assumptions, which are closely connected with the continuity properties of the process under consideration. Because of the form of the continuity condition (4.6), one is led to consider a method of dividing the differentiability conditions into parts, one corresponding to continuous motion of a representative point and the other to discontinuous motion. The following conditions are required for all > 0:
lim P (x, t + ∆t|z, t)/∆t = W (x|z, t) (4.7) ∆t→0 uniformly in x, z and t for |x − z| ≥ ; 1 Z lim dx (xi − zi)P (x, t + ∆t|z, t) = Ai(z, t) + O(); (4.8) ∆t→0 ∆t |x−z|< 1 Z lim dx (xi − zi)(xj − zj)P (x, t + ∆t|z, t) = Bij(z, t) + O(); (4.9) ∆t→0 ∆t |x−z|< the last two being uniform in z, and t. It can be shown that all higher-order coefficients of the above form must vanish. According to the condition for continuity (4.6), the process can only have continuous paths if W (x|z, t) vanishes for all x 6= z. Thus, this function must in some way describe discontinuous motion, while the quantities Ai and Bij must be connected with continuous motion. The derivation of the differential Chapman-Kolmogorov equation, which is to be found in detail in Reference [19] p. 48 ff., is only sketched here. It starts with considering the time evolution of the expectation value of a function f(z) which is twice continuously differentiable. Thus, Z Z 0 1 0 0 ∂t dx f(x)P (x, t|y, t ) = lim dx f(x)[P (x, t + ∆t|y, t ) − P (x, t|y, t )] ∆t→0 ∆t 1 Z Z = lim dx dz f(x)P (x, t + ∆t|z, t)P (z, t|y, t0)− ∆t→0 ∆t Z − dz f(z)P (z, t|y, t0) , (4.10)
18 4.1 Selected foundations of stochastic processes where the Chapman-Kolmogorov equation has been used in the second line. The next steps are to divide the integral over x into two regions |x − z| ≥ and |x − z| < , to expand f(x) in the terms where |x − z| < to second order around z with a remainder that vanishes as |x − z| → 0, to introduce a one in the form of R dx P (x, t + ∆t|z, t) into the last term of Equation (4.10), to identify the quantities W , Ai and Bij, and finally to take the limit → 0. Equation (4.10) now reads Z Z X ∂f(z) 1 X ∂2f(z) ∂ dz f(z)P (z, t|y, t0) = dz A (z, t) + B (z) P (z, t|y, t0)+ t i ∂z 2 ij ∂z ∂z i i i,j i j Z Z + dz f(z) dx [W (x|z, t)P (x, t|y, t0) − W (x|z, t)P (z, t|y, t0)] . (4.11)
Notice that the integral over x actually is a principal value integral lim R dx. The last thing →0 |x−z|> to do, is to integrate by parts, where an apt limitation of the stochastic process and f to some finite regions lets surface terms vanish, and which yields the differential form of the Chapman-Kolmogorov equation,
X ∂ ∂ P (z, t|y, t0) = − [A (z, t)P (z, t|y, t0)] t ∂z i i i X 1 ∂2 + [B (z, t)P (z, t|y, t0)] 2 ∂z ∂z ij i,j i j Z + dx [W (z|x, t)P (x, t|y, t0) − W (x|z, t)P (z, t|y, t0)], (4.12) onto which large parts of the analysis in this paper are based.
4.1.3 Detailed balance According to Reference [19] p. 148 f., a Markov process satisfies detailed balance if, roughly speak- ing, in the stationary situation each possible transition balances with the reversed transition. The mathematical formulation of this statement reads
0 0 0 P (x, t|x , 0)Ps(x ) = P (x , t|x, 0)Ps(x), (4.13) where x := (1x1, 2x2,...) represents the behaviour of x under time reversal and Ps(x) signifies the stationary distribution. The variable xi transforms to the reversed variable according to the rule
xi → ixi (4.14)
i = ±1 (4.15) depending on whether the variable is odd or even under time reversal. Positions X, for instance, are even, whereas velocities V are odd. This concept is motivated by physical considerations. For exam- ple, take a one-dimensional system in classical mechanics. A transition (X, V, t) → (X0,V 0, t + ∆t) is not simply reversed by interchanging primed and unprimed quantities. Rather, it is (X0, −V 0, t) → (X, −V, t + ∆t), which corresponds to the time reversed transition and requires the velocities to be reversed because the motion from X0 to X is in the opposite direction from that from X to X0. It is noteworthy that for t = 0 Equation (4.13) implies
Ps(x) = Ps(x), (4.16) since the conditional probabilities reduce to delta functions.
19 4 Brownian motion as stochastic process - the mathematical framework
On page 151 ff. of Reference [19] it is shown that for the differential Chapman-Kolmogorov equation necessary and sufficient conditions of detailed balance are given by 0 0 0 W (x|x )Ps(x ) =W (x |x)Ps(x), (4.17) X ∂ A (x)P (x) = − A (x)P (x) + [B (x)P (x)], (4.18) i i s i s ∂x ij s j j
ijBij(x) =Bij(x). (4.19) These conditions are in fact fulfilled by the present model as demonstrated in Section 4.4.2.
4.1.4 The Poisson process A key assumption of the model will be that collisions of particles of the medium with the motor occur independently of each other. The rates of these events will depend strongly on the current motor velocity, the velocity of the incoming particle and the location of the point of impact on the motor surface. For the derivation of basic properties, however, it is convenient to study the elementary process with only one sort of event and one corresponding rate. The generalisation and adaptation to the model is left to Section 4.3. The following is largely based on Reference [19] p. 12 f. Consider a stochastic process that consists of some kind of events occurring independently of each other and at constant rate λ. If N is the number of events counted up to a time t, then the probability that N increases by one in an infinitesimal interval ∆t is given by P (N + 1, t + ∆t|N, t) = λ∆t, (4.20) and one has for the total probability of having N counts at t + ∆t P (N, t + ∆t) =(1 − P (N + 1, t + ∆t|N, t))P (N, t) + P (N, t + ∆t|N − 1, t)P (N − 1, t) =(1 − λ∆t)P (N, t) + λ∆tP (N − 1, t), (4.21) for N > 0. The validity can be extended to N ≥ 0 by setting P (−1, t) = 0. Taking the limit ∆t → 0 yields ∂ P (N, t) = λ[P (N − 1, t) − P (N, t)], (4.22) ∂t which is solved by the Poisson distribution (λt)N P (N, t) = e−λt. (4.23) N! It is very instructive to reconsider this last formula in a slightly different way, as it allows for the assignment of illustrative interpretations to its individual components and for setting up detailed expressions for individual series of events. For instance, one finds that the probability that nothing happens during an interval t is given by
−λt 1 Pvoid(t) := P (0, t) = e . (4.24) Conversely, the probability of an event to occur in the infinitesimal interval dt is λdt. Having these expressions at hand, one can now construct the joint probability of a series of events taking place at times 0 < t1 < t2 < . . . < tN < t,
−λt1 −λ(t2−t1) −λ(tN −tN−1) −λ(t−tN ) dP ({ti}) = e λdt1 e λdt2 ... e λdtN e = N −λt = λ e dt1 ... dtN . (4.25)
By integrating out t1 to tN one eventually recovers Equation (4.23). 1An alternative way to arrive at this result is to divide t into m parts of length t/m. Then P (0, t) = (1 − λt/m)m, which goes to exp(−λt) as m → ∞.
20 4.2 Description of the Model
U(X,t) m ê M ê θ X motor V SF(θ)dθ X T,�ρ
Figure 4.1 Sketch of the system: The two-dimensional Figure 4.2 A line element of convex motor particle is immersed in an ideal gas and length SF (θ)dθ on the motor sur- free to move along the X axis with velocity V . M is face, parametrised by the angle θ the mass of the motor, whereas the gas particles have between the tangent vector eˆ|| and masses of m. Their density is ρ and the temperature of the X axis. the gas is T . In addition to collisions with gas particles the motor experiences forces due to a time-dependent potential U(X, t).
4.2 Description of the Model
The system under consideration resembles largely the one studied in Reference [42] and is sketched in Figure 4.1. A two-dimensional convex body, called the motor, of mass M, having one translational degree of freedom X, is placed inside an infinite two-dimensional ideal gas at constant temperature T with density ρ and particle masses m. The gas is assumed to be in thermal equilibrium at any time. Concerning interactions, motor and gas particles behave as fully elastic hard objects. The limitation to convex bodies has two reasons. First, Minkowski functionals may be applied to describe the shape of the motor, which was carried out in detail in Reference [41], and second, multiple collisions of gas particles with the body are mostly omitted. In combination with the ideal behaviour of the medium, impacts of particles on the motor may thus be treated as independent random events, which is be the basis of the stochastic process describing its motion. The model is extended by allowing the body to additionally experience forces due to a time-dependent potential U(X, t).
4.3 Modelling the stochastic process
As impacts of gas particles on the body are regarded independent and random, the chronology of these events may be modelled as a Poisson-type stochastic process with the phase space position Γ(t) := (X(t),V (t)) being the random variable. However, the rates of velocity-changes will be highly dependent on the current motor velocity. Generally, the faster the motor is travelling into one direction, the likelier a collision which gives it a push in the opposite direction. This is the microscopic realisation of friction and ensures a stable equilibrium distribution. Moreover, the shape and especially its asymmetry translate into these rates, which is in fact the crucial detail enabling asymmetric motion in non-equilibrium. Section 4.3.1 shows how the transitions rates explicitly derive from the shape of the motor. After- wards they are used to construct expressions for the probabilities of individual trajectories and the phase space density in Section 4.3.2.
21 4 Brownian motion as stochastic process - the mathematical framework
4.3.1 Collision rates An infinitesimal line element of the motor’s surface is characterised by an angle θ and a length factor F (θ) (cf. Figure 4.2). θ is defined as the angle between the tangent unit vector eˆ|| := (cos θ, sin θ) and the X axis. Hence, the corresponding normal vector is given by eˆ⊥ = (sin θ, − cos θ). The length of the line element SF (θ)dθ, where S is the total perimeter, implicitly defines the function F (θ) representing the fraction of the surface with tangent angle θ. Let v˜ be the velocity vector of a gas particle with respect to the motor’s rest frame. The current density of particles having a velocity between v˜ and v˜ + dv˜ heading towards the body’s surface is then given by
j(˜v) d˜v = ˜vρΦ(˜v + V)Θ(−˜veˆ⊥) d˜v, (4.26) where V := V eˆx, Φ(v) is the velocity probability density of the gas and the Heaviside Θ ensures that no particles are coming from inside the body. Accordingly, the flux or impact rate on a line element with tangent angle θ is
λi(θ, ˜v,V ) d˜vdθ = −SF (θ)eˆ⊥j(˜v) d˜vdθ. (4.27)
If one demands conservation of total energy, total momentum in X direction and conservation of the particle’s momentum tangential to the body’s surface, the velocity-change caused by a collision takes the form
∆V = αv˜⊥, (4.28)
m 2 M sin θ withv ˜⊥ := ˜veˆ⊥ and α := m 2 (cf. References [41, 11]). Usage of the two dimensional Maxwell- 1+ M sin θ Boltzmann velocity distribution
m mv2 Φ(v) = exp − , (4.29) 2πkBT 2kBT eventually leads to an expression for the rate of events inducing a velocity-change between ∆V and ∆V + d∆V at motor velocity V
2π ∞ Z Z d∆V λ(∆V,V ) d∆V := dθ d˜v λ (θ, (˜v , v˜ = ∆V/α),V ) || i || ⊥ |α| 0 −∞ 2π Z ∆V ∆V = dθ a Θ − exp −b(∆V c + V )2 d∆V (4.30) α|α| α 0 q with a := −ρSF (θ) m , b := m sin2 θ and c := 1 . An exemplary plot of λ(∆V,V ) is 2πkBT 2kBT α sin θ presented in Figure 4.3.
4.3.2 The phase space probability density
∗ Consider an interval t0 ≤ t ≤ t during which the motor experiences N collisions at times {ti|i = ∗ 1, ..., N; t0 ≤ t1... ≤ t } causing velocity-changes {∆Vi|i = 1, ..., N}. As a shorthand notation, this trajectory of velocity-jumps will be denoted by the symbol ∆V (t). In this section, expressions for three probabilities will be developed: first, the conditional proba- bility density for the system to take a specific path ∆V (t), P [∆V (t)|Γ0, t0], given the starting point Γ0 at t0, second, the conditional probability density of finding it a the phase space position Γ at time ∗ ∗ ∗ t , P (Γ , t |Γ0, t0), given the same starting point, and third, the phase space probability density ∗ ∗ function (pdf), P (Γ , t ), given a certain initial distribution P0(Γ) at t0.
22 4.3 Modelling the stochastic process
q 1 λ(∆V,V )/ ρS 2π /100 30 1.2
20 1
10 0.8 T B M k 0 0.6 q V/ -10 0.4
-20 0.2
-30 0 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1 q kBT ∆V/ M
Figure 4.3 The rate λ(∆V,V ) of velocity-changes ∆V at current motor velocity V (cf. Equation (4.30)). The plotted data correspond to the test motor of Chapter 5 (cf. Figure 5.1), an equilateral triangle with one side perpendicular to the X axis and the opposite corner pointing in positive X direction. Moreover, δ := pm/M = 0.1.
23 4 Brownian motion as stochastic process - the mathematical framework
Similar to Equation (4.24), the void probability of no collision to occur between ti and tj is defined as tj ∞ Z Z Pvoid(tj, ti) := exp − dt d∆V λ(∆V,V (t)) . (4.31)
ti −∞ R ∞ The quantity −∞ d∆V λ(∆V,V ) is the combined rate of all possible collisions at motor velocity V . The integration over t becomes necessary as V (t) may vary in between two collisions due to the potential U(X, t) and can be motivated by considering the continuum limit of the product of many void probabilities for subsequent small time intervals with constant V or λ, respectively. With the definition of Vi, the velocity just before collision i,
Vi := lim V (t), (4.32) t%ti the probability density of the velocity-change ∆Vi to happen at ti becomes λ(∆Vi,Vi). It is now straightforward to write down the first sought conditional probability density (cf. Equation (4.25)):
N Y ∗ P [∆V (t)|Γ0, t0] := [Pvoid(ti, ti−1)λ(∆Vi,Vi)] Pvoid(t , tN ) i=1 ∗ t ∞ N Z Z Y = exp − dt d∆V λ(∆V,V (t)) λ(∆Vi,Vi). (4.33) i=1 t0 −∞
∗ For integrations over all possible collision histories between t0 and t , it is convenient to define another symbol:
∗ ∗ ∗ ∞ t t t ∞ ∞ ∞ Z X Z Z Z Z Z Z D[∆V (t)] := dt1 dt2... dtN d∆V1 d∆V2... d∆VN . (4.34) N=0 t0 t1 tN−1 −∞ −∞ −∞
By this means, the normalisation condition of P [∆V (t)|Γ0, t0] reads Z D[∆V (t)] P [∆V (t)|Γ0, t0] = 1, (4.35) but turns out to be difficult to prove. In Section 6.1.2 it is shown for a simpler, but similarly constructed process, which shall serve as a motivation for the assumption of the validity of the above equation. The second desired quantity is easily obtained by integrating over all paths ∆V (t) weighted by their probabilities and keeping only those which end at Γ∗: Z ∗ ∗ ∗ ∗ P (Γ , t |Γ0, t0) := D[∆V (t)] P [∆V (t)|Γ0, t0]δ(Γ − Γ(t )). (4.36)
Finally, integrating out Γ0 yields the phase space pdf: Z ∗ ∗ dΓ0 ∗ ∗ P (Γ , t ) := P (Γ , t |Γ0, t0)P0(Γ0) 2π~/M Z Z dΓ0 ∗ ∗ = D[∆V (t)] P [∆V (t)|Γ0, t0]P0(Γ0)δ(Γ − Γ(t )), (4.37) 2π~/M where the factor M/(2π~) is needed to render the expression dimensionless and assure consistency with quantum mechanics.
24 4.4 Time evolution of the phase space probability density
4.4 Time evolution of the phase space probability density 4.4.1 The time evolution equation The collision probability densities or rates λ(∆V,V ) only depend on the current phase space position Γ. Obviously, the same is true for the deterministic motion between collisions. Thus, the condi- tional probability density P (Γ, t|Γ0, t0) is generated by a Markov process and consequently obeys the Chapman-Kolmogorov equation Z P (Γ, t|Γ0, t0) = dΓ1P (Γ, t|Γ1, t1)P (Γ1, t1|Γ0, t0). (4.38)
Following Section 4.1.2, the latter can be transformed into a differential time evolution equation. 0 If the uniform convergence of the limits is assumed, the quantities W (Γ|Γ , t), Ai(Γ, t) and Bij(Γ, t) are found to be W (Γ|Γ0, t) = W (V |V 0)δ(X − X0); W (V |V 0) := λ(V − V 0,V 0), (4.39)
A1(Γ, t) = V, (4.40) 1 ∂U(X, t) A (Γ, t) = − and (4.41) 2 M ∂X Bij(Γ, t) = 0, (4.42) as demonstrated in Appendix A.1. In fact, the vector A := (A1,A2) simply corresponds to Γ˙ if dis- continuous changes due to collisions are excluded. With these coefficients, the differential Chapman- Kolmogorov equation (4.12) takes the form ∂P (Γ, t|Γ , t ) ∂ 1 ∂U(X, t) ∂ 0 0 = −V + P (Γ, t|Γ , t )+ ∂t ∂X M ∂X ∂V 0 0 ∞ Z 0 0 0 0 + dV [W (V |V )P ((X,V ), t|Γ0, t0) − W (V |V )P (Γ, t|Γ0, t0)] . (4.43) −∞
R dΓ0 By means of the definition P (Γ, t) := P (Γ, t|Γ0, t0)P0(Γ0), the evolution equation of the 2π~/M phase space pdf is obtained: ∞ ∂P (Γ, t) ∂ 1 ∂U(X, t) ∂ Z = −V + P (Γ, t)+ dV 0 [W (V |V 0)P ((X,V 0), t) − W (V 0|V )P (Γ, t)] . ∂t ∂X M ∂X ∂V −∞ (4.44) The first term of the right hand side corresponds to the Liouville equation of a classical particle in the potential U(X, t) and accounts for the deterministic motion between collisions. The second term has the form of a so-called master equation and represents the stochastic part of the evolution of P (Γ, t) due to the random collisions. By setting U(X, t) = const and integrating out X, the master equation for velocities, which was used in References [42, 41, 10], can be recovered.
4.4.2 Stationary solution and detailed balance For a time invariant potential U(X) the canonical distribution 1 1 1 2 Peq(Γ) := exp − MV + U(X) (4.45) Z kBT 2 with the partition function Z dΓ 1 1 Z := exp − MV 2 + U(X) (4.46) 2π~/M kBT 2
25 4 Brownian motion as stochastic process - the mathematical framework is a stationary solution of the time evolution equation (4.44). One convinces oneself easily that
∂ 1 ∂U(X) ∂ −V + P (Γ) = 0. (4.47) ∂X M ∂X ∂V eq
To show that the integral part of Equation (4.44) vanishes as well, it is written as
∞ Z d∆V [W (V |V + ∆V )Peq((X,V + ∆V )) − W (V + ∆V |V )Peq((X,V ))]. (4.48) −∞
By noticing that
∆q Peq((X,V + ∆V )) = Peq((X,V ) exp , where (4.49) kBT 1 ∆q := − M (V + ∆V )2 − V 2 , (4.50) 2 which can be associated with the heat dissipated during a collision (cf. Section 7.1), and
∆q W (V |V + ∆V ) exp = W (−(V + ∆V )| − V ), (4.51) kBT proved in Appendix A.2, one finds (using W (a|b) = λ(a − b|b))
∞ Z (4.48) = d∆V [λ(−∆V | − V ) − λ(∆V |V )]Peq((X,V )) = −∞ 2π Z = −ρV Peq((X,V )) dθ SF (θ) sin(θ) = 0, (4.52) 0 where Equation (4.30) was used for the second equality. The integral over θ vanishes because dθSF (θ) sin θ is the projection perpendicular to the X axis of the line element dθSF (θ) weighted with a sign that reflects the notion of an inside and an outside of the body, or in other words, that indicates whether the surface element is oriented to the right or to the left. As the cross section of the body is the same no matter whether it is looked at from positive or negative X direction, the sum of the projections – labelled with appropriate signs – along the complete outline of the body adds up to zero. It is now easy to show that this stationary state exhibits detailed balance (cf. Section 4.1.3). With 1 = 1, 2 = −1, A = (V, −∂X U(X)/M) (cf. Equations (4.40) and (4.41)) and Bij(Γ) = 0 (cf. Equation (4.42)), one readily finds that conditions (4.19) and (4.18) hold, that is
ijBij(Γ) =Bij(Γ) and (4.53)
iAi(Γ) = − Ai(Γ). (4.54)
Condition (4.17) reads
0 0 0 W (V |V )Peq((X,V )) = W (−V | − V )Peq((X,V )). (4.55)
By substituting V 0 = V + ∆V and applying identity (4.49), one finds that the above equation transforms into a relation similar to Equation (4.51), also proved in Appendix A.2.
26 4.4 Time evolution of the phase space probability density
4.4.3 Kramers-Moyal expansion of the time evolution equation
A common way of tackling master equations is to perform a so-called Kramers-Moyal expansion. For instance, expanding a master equation generated by a simple undirected random walk produces the diffusion equation and almost all expansion coefficients vanish. In case of the master equation like integral part of the time evolution equation, none of the coefficients disappear in general. Still, this method proves to be useful since for example only certain coefficients play a role if moments of the phase space pdf are considered (cf. Section 4.5). By means of a smart substitution and the help of an auxiliary function, f, the second part of Equation (4.44) can be expanded in powers of the step size V 0 − V :
∞ Z dV 0 [W (V |V 0)P ((X,V 0), t) − W (V 0|V )P (Γ, t)] =
−∞ ∞ ∞ Z Z = dy W (V |V − y)P ((X,V − y), t) − dy W (V + y|V )P (Γ, t) = | {z } | {z } −∞ f(y,V −y) −∞ =:f(y,V ) ∞ Z = dy [f(y, V − y)P ((X,V − y), t) − f(y, V )P (Γ, t)] =
−∞ ∞ ∞ Z X (−y)n ∂n = dy [f(y, V )P (Γ, t)] (4.56) n! ∂V n −∞ n=1
Re-substitution of V 0 − V for y yields the Kramers-Moyal expansion of the time evolution equation:
∞ ∂P (Γ, t) ∂ 1 ∂U(X, t) ∂ X (−1)n ∂n = −V + P (Γ, t) + [α (V )P (Γ, t)] (4.57) ∂t ∂X M ∂X ∂V n! ∂V n n n=1 with the n’th Kramers-Moyal coefficient being defined by
∞ Z 0 0 n 0 αn(V ) := dV (V − V ) W (V |V ). (4.58) −∞
q If one inserts Equation (4.30), Equation (4.39), a = −ρSF (θ) m , b = m sin2 θ, c = 1 2πkBT 2kBT α sin θ m 2 M sin θ q m 0 and α = m 2 into Equation (4.58) and substitutes R := (V − V )/α, one readily finds 1+ M sin θ 2kBT
r n+1 3n+1 kBT ρS n−1 α (V ) = − 2 2 √ δ × n M π 2π 0 !2 Z sin θ n Z r M δ × dθ F (θ) dRRn+1 exp − R + √ V sin θ , (4.59) 2 2 1 + δ sin θ kBT 2 0 −∞
p m where δ := M .
27 4 Brownian motion as stochastic process - the mathematical framework
4 6 exact 2 5 O(δ1) 3 ) 0 2
T O(δ )
T 4 B
M 3 B M k O(δ )
k -2
( 3 q /
-4 / 1 2 g 2 -6 exact g -8 O(δ1) 1 2 -10 O(δ ) 0 O(δ3) -12 -1 -20 -15 -10 -5 0 5 10 15 20 -20 -15 -10 -5 0 5 10 15 20 q q kBT kBT V/ M V/ M
Figure 4.4 Different approximations of the integrands g1(V, θ = π/2) and g2(V, θ = π/2) (cf. Equation (4.61)) of the first two Kramers-Moyal coeffiencts for δ = 0.1. For θ = π/2 deviations are maximal. Guided by thermal equilibrium, typical values of V q kBT are expected to be of a few M .
4.4.4 Series expansion of the Kramers-Moyal coefficients
The integral over R in Equation (4.59) can be solved analytically (cf. Reference [41] p. 87), but results in a rather complicated V dependence involving the Gauss error function. Luckily, a series expansion in the square root of the mass ratio, δ, which usually will be a small quantity, turns αn(V ) into a polynomial in V . In detail one has
2π r n+1 Z n 3n−1 kBT ρS n−1 n α (V ) = (−1) 2 2 √ δ dθ F (θ) sin θ× n M π 0 ∞ " l k l−k−m k−2m n r 2k X X X (−1) 2 (n − 1 + l − k)!Γ( + 1 + k − m) M × δ2l sin2l θ 2 V 2k+ (l − k)!(n − 1)!(2k − 2m)!m! kBT l=0 k=0 m=0 l k l−k−m k−2m+ 1 n+3 r 2k+1 # X X (−1) 2 2 (n − 1 + l − k)!Γ( + k − m) M +δ2l+1 sin2l+1 θ 2 V 2k+1 . (l − k)!(n − 1)!(2k − 2m + 1)!m! kBT k=0 m=0 (4.60)
The derivation of this formula is sketched in Appendix A.3. Truncating the series then yields approximate expressions for the Kramers-Moyal coefficients. To give an impression of its performance, the quantity gn(V, θ), implicitly defined via
Z 2π αn(V ) = ρS dθ F (θ)gn(V, θ), (4.61) 0 is considered. As the order of sin θ increases with the order of δ, the approximation to gn(V, θ) will be worst if sin θ = 1. In Figure 4.4.4, expansions up to different orders of δ of g1(V, π/2) and g2(V, π/2) are compared with the exact solutions. The focus is set on the first two coefficients as they will reappear in the analysis in Section 5.3. It is noteworthy that it is possible to express the integrals over θ in terms of tensorial Minkowski
28 4.5 Time evolution of moments of the phase space probability density functionals, namely (cf. References [42, 41]) Z 1 n (0,n) dθSF (θ) sin θ = W1 (K) . (4.62) 2 ∂K X...X | {z } n times K signifies the body the functional is applied to, being the motor in this case. The n-fold subscript (0,n) X denotes the X...X entry of W1 (K), that is the element of the rank n tensor solely associated with the X direction. In general, the first tensorial Minkowski functional of a convex body K is defined as n times Z m times (m,n) 1 z }| { z }| { W1 (K) := dO r ⊗ ... ⊗ r ⊗ eˆ⊥ ⊗ ... ⊗ eˆ⊥, (4.63) d ∂K where d is the dimension, dO denotes a surface integral, r is the position and eˆ⊥ the normal vector on the surface (see References [42, 28, 27]).
4.5 Time evolution of moments of the phase space probability density
As already hinted in Section 4.4.3, one of the virtues of the Kramers-Moyal expansion is that the time evolution equations of the moments of the phase space pdf, Z dΓ hXiV ji(t) := XiV jP (Γ, t), (4.64) 2π~/M may be expressed in terms of the expansion coefficients αn(V ). By taking the time derivative of Equation (4.64) and applying Equation (4.44), one finds Z i j dΓ i j ∂thX V i(t) := X V ∂tP (Γ, t) 2π~/M Z Z dΓ i j+1 1 dΓ i j =− X V ∂X P (Γ, t) + X V ∂X U(X, t)∂V P (Γ, t)+ 2π~/M M 2π~/M | {z } | {z } (i) (ii) Z dΓ Z Z dΓ Z + dV 0 XiV jW (V |V 0)P ((X,V 0), t) − dV 0 XiV jW (V 0|V )P (Γ, t). 2π~/M 2π~/M | {z } (iii) (4.65) Term (i) and (ii) are easily integrated by parts, where surface terms are assumed to vanish. If V and V 0 are swapped in the first integral of part (iii), the latter can be written as j Z Z X j hXi dV 0 (V 0j − V j)W (V 0|V )i = hXi dV 0 V j−n(V 0 − V )nW (V 0|V )i. (4.66) n n=1 The last equality is readily appreciated by writing V 0j = ((V 0 − V ) + V )j and applying the binomial identity. Now it is possible to identify the expression for αn(V ) in (iii), which yields in combination with (i) and (ii): j j X j ∂ hXiV ji = ihXi−1V j+1i − hXiV j−1∂ U(X, t)i + hXiV j−nα (V )i. (4.67) t M X n n n=1
With the expansion of αn(V ) in δ or V , respectively, (cf. Equation (4.60)) approximate equations for the moments may be obtained (see Section 5.3 for a concrete example).
29 4 Brownian motion as stochastic process - the mathematical framework
30 5 Using potential wells as stopping sites
In the original model (cf. References [42, 41]) the velocity of the motor is set to zero, whenever it reaches a stopping site. In terms of the statistical ensemble this means turning the velocity distribution into a delta peak, but, striving to make a different point, the author did not dwell on how this may actually be achieved. In fact, this turns out to be very hard if not impossible. Thinking of binding sites, one usually pictures some sort of potential well, where strong attractive forces keep the bound particle in place. Leaving aside quantum effects, one can by that means in principle confine the motor to a more or less arbitrary small region. However, the equilibrium velocity distribution of such a trapped particle is independent of the potential. It is simply given by the Gaussian shaped Maxwell-Boltzmann distribution, the width of which is determined by temperature alone. Nevertheless, potential wells can still be used to narrow the velocity distribution – namely by turning them off. If the motor is gradually released and given enough time to perform work against the “walls” of the potential well, the effect is quite similar to the cooling of a gas by expansion. With respect to real systems such a change in binding energy may probably be induced by ATP hydrolysis or some other biochemical reaction. To get deeper insight into the release process, a case study with a parabolic potential 1 U(X, t) := k(t)X2 (5.1) 2 is carried out in this chapter. The release from the well is realised by an exponentially decreasing spring constant −t/τ k(t) := k0e (5.2) with typical decay time τ. Moreover, the analysis is mostly restricted to moments of the phase space pdf, making use of the approximations introduced in Section 4.4.4. Any quantitative data presented in this chapter is obtained by using the test motor particle depicted in Figure 5.1. It is an equilateral triangle of circumference S with one side perpendicular to the X axis and the opposite corner pointing in positive X direction. For this setup, the elements of the Minkowski tensors (see Equation (4.62) for a definition) adopt values of ! S 1n−1 W (0,n)(K) = + (−1)n . (5.3) 1 X...X 6 2 | {z } n times
The shape dependence of the motor’s motion was studied in Reference [41] and is therefore not subject of the analysis. All numerical solutions to differential equations were obtained using the NDSolve algorithm of Wolfram Mathematica 6.0 with default options.
S/3
30° X
Figure 5.1 The test motor particle, used for any quantitative calculations.
31 5 Using potential wells as stopping sites
The following considerations are divided into three parts. Initially, the process is studied either without potential or without collisions, allowing for better analytical access, before both are combined. In Section 5.1, the free relaxation towards equilibrium out of a given initial distribution at zero potential is investigated, yielding that the difference of the starting value of hV 2i from the equilibrium value plays a crucial role in the efficiency of the process. It is therefore desirable to bring the velocity distribution as close to the delta peak as possible. This is important information for the design of the release process studied under negligence of collisions in Section 5.2. There, the importance of the correct choice of time scales is pointed out. Finally, Section 5.3 comprises both aspects. A method of deriving approximate sets of equations for the moments of the phase space pdf is proposed, and different approximation levels are numerically solved and compared. It turns out that many features of the combined system (i.e., collisions and potential) can already be understood from the reduced systems of the first two parts. Moreover, if the time scales of potential-change and relaxation in the medium are well separated, an analytical approximation may be obtained by treating both processes as being independent of each other and occurring subsequently, restoring to some extend the assumptions used in References [42, 41]. Additionally, a short outlook comments on the binding process, where in contrast to releasing some extra dissipation mechanism might be needed, and a way to incorporate multiple or re-opening stopping sites is outlined.
5.1 Relaxation without potential 5.1.1 Evolution of moments of the phase space probability density To investigate the collision driven evolution of the system in the case of zero potential, the following set of equations is studied:
∂thXi = hV i, (5.4) k T ∂ hV i = AδhV i − Bδ2 B − hV 2i , (5.5) t M 2 ∂thX i = 2hXV i, (5.6) k T ∂ hV 2i = −2Aδ B − hV 2i , (5.7) t M 2 ∂thXV i = hV i + AδhXV i, (5.8) with r r 2 kBT (0,2) A := − 4 ρ W1 (K) , (5.9) π M XX (0,3) B := − 2ρ W1 (K) . (5.10) XXX It is obtained by use of the δ expansion of the Kramers-Moyal coefficients (4.60) and constitutes a good approximation for small values of the mass ratio (δ . 0.1). It shall not be deduced here, since, in fact, it is the ω(t) = 0 case of Equations (5.86) to (5.90) in Section 5.3. For their derivation, however, it is necessary to include a potential, as it is reasonable to demand of any approximate equations to posses stationary solutions that match the corresponding equilibrium distribution for a non-zero time-invariant potential. To zeroth order in δ the above equations are those of a free particle. With δ1 merely linear damping terms proportional to Aδ are introduced and one can identify
γ := −AδM (5.11)
32 5.1 Relaxation without potential as the first order approximation of the linear friction coefficient. Due to this negative feedback, one expects exponential decays of initial values with a typical relaxation time of M −1 τ := = . (5.12) r γ Aδ Note that A is insensitive to asymmetry due to the sin2 θ terms. Asymmetry comes into play in the form of B when going to the second order in δ. In this case, an additional force term arises in the equation of hV i, which, however, vanishes in case of a symmetric object (B = 0) or if hV i has reached kBT its equilibrium value of M . Solutions to Equations (5.4) to (5.8) are found using the method of variation of constants:
2 hV i(t0) B kBT hXi(t) = hXi(t ) − 1 − eAδ(t−t0) − − hV 2i(t ) 1 − eAδ(t−t0) , (5.13) 0 Aδ 2A2 M 0 | {z } translation due to asymmetry and non-equilibrium Bδ k T Aδ(t−t0) B 2 Aδ(t−t0) Aδ(t−t0) hV i(t) = hV i(t0)e + − hV i(t0) 1 − e e , (5.14) | {z } A M exp. dissipation of initial mean vel. | {z } velocity boost
kBT 2 2 2hXV i(t0) − hV i(t0) hX2i(t) = hX2i(t ) − 1 − eAδ(t−t0) − M 1 − eAδ(t−t0) − 0 Aδ A2δ2 2 kBT kBT − M 1 − eAδ(t−t0) + 2 M (t − t ), (5.15) A2δ2 −Aδ 0 | {z } =:D kBT kBT hV 2i(t) = − − hV 2i(t ) e2Aδ(t−t0), (5.16) M M 0 | {z } | {z } equil. value exp. relaxation back to equilibrium
kBT 2 kBT − hV i(t0) hXV i(t) = hXV i(t )eAδ(t−t0) + M 1 − eAδ(t−t0) eAδ(t−t0) − M 1 − eAδ(t−t0) . 0 Aδ Aδ | {z } lim hXV i t→∞ (5.17) Exemplary plots of these functions are shown in Figure 5.2. There are several aspects that are worth a remark.
• First of all, net motion is possible even from a symmetric initial distribution (hV i(t0) = 0) if B 6= 0. This is true for most asymmetric objects, due to the sin3 θ term in the integral. However, this effect dies away as the velocity distribution approaches equilibrium as already anticipated from the differential equations.
• The typical time scale of the process is indeed set by τr. For instance, this quantity appears in the exponential dissipation of the initial mean velocity as well as the exponential relaxation of hV 2i back to its equilibrium value. • For large times (i.e., when the equilibrium distribution of velocities has been re-established) the mean square displacement goes linear with t, which corresponds to normal diffusion. The quantity k T D := B (5.18) −AδM may thus be identified with the first order approximation of the diffusion constant, and one recovers the Einstein-Smoluchowski relation (1.3) k T k T D = B τ = B (5.19) M r γ
33 5 Using potential wells as stopping sites
0 0 -2 -1 -4 -2 -6 1000 100 -8 × -3 × T
-10 B -4 M k ρS
i -12 -5 q X -14 / h i -6
-16 V -18 h -7 -20 -8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
t/τr t/τr
6 1 5 4 0.8 3 T B ρS 2 M 0.6 k i slope 2 1 3 / X i h
0 2 0.4
-1 2 V lg -2 h 0.2 -3 1 -4 0 -2 -1 0 1 2 3 0 0.5 1 1.5 2 2.5 3 3.5 4
lg(t/τr) t/τr
14 12 T B M 10 k
q 8
ρS/ 6 i 4 XV h 2 0 0 1 2 3 4 5 6 7 8
t/τr
Figure 5.2 Evolution of the first five moments of the phase space pdf in absence of a potential U(X, t), as given by Equations (5.13) to (5.17). Parameter values q q 2 kBT 1 are A = − π ρS M and B = 4 ρS, corresponding to the triangular test motor (cf. Figure 5.1). Moreover, δ = 0.1 and t0 = 0. Initially, all moments are set to zero, which constitutes an ideal “stopped” state and matches the case discussed in References [42, 41].
34 5.1 Relaxation without potential
1.1 2.5 η × 100 t /τ 1 max r approximation 2 f(σ, t = tmax) 0.9 0.8 1.5 0.7 0.6 1 0.5 0.5 0.4 0.3 0 0 1 2 3 4 5 6 7 8 -2 -1.5 -1 -0.5 0 0.5 1
t/τr σ
Figure 5.3 Left: The efficiency as defined in Equation (5.20) for the same set of parameters as in Figure 5.2. The approximation (5.24) is in perfect agreement with the original curve. Right: Dependence of the temporal position tmax (given by Equa- tion (5.26)) and hight (∝ f(σ, tmax); cf. Equation (5.25)) of the efficiency maximum on the initial deviation σ of hV 2i from the equilibrium value.
within the limits of the given approximation.
• The intensity of the velocity boost and consequently the final value of hXi, too, are – among 2 kBT others – controlled by the initial difference of hV i from its equilibrium value M . Therefore, it is desirable to drive the system away from equilibrium as far as possible to obtain maximum net motion. Moreover, the magnitude of the boost and its duration vary with δ reciprocally to each other, which is why the contribution of the integrated boost to the long time limit of hXi does not depend on δ itself.
• Last, one might notice that hXV i does not approach its equilibrium value of zero. This is explained by the fact that for an unbounded system there indeed exists no equilibrium distri- bution for the positions, or at least none that can be reached in finite time. However, this does not hinder the equilibration of the velocities.
5.1.2 Efficiency
As a measure of randomness or “directedness” one may regard the quantity
|hXi(t)| η(t) := , (5.20) phX2i(t) − hXi(t)2 which will be further referred to as efficiency. An exemplary plot is presented in Figure 5.3 (a), which shows that after having reached a maximal value after a comparatively short time, η declines towards zero, as random diffusive motion begins to dominate. In terms of optimizing the process it is desirable to estimate the position and height of η’s maximum. The analysis is be restricted to cases where hXi(t0) = 0 and hV i(t0) = 0, to focus solely on the effect of the asymmetric relaxation process. Moreover, t0 = 0. With the definition of M 2 σ := 1 − hV i(t0) , (5.21) kBT
35 5 Using potential wells as stopping sites
Equations (5.13) and (5.13) take the form:
2 B kBT hXi(t) = − σ 1 − e−t/τr , (5.22) 2A2 M neglect
z }| { 2 2 2 2 −t/τr kBT −t/τr −t/τr A δ hX i(t0) − 2AδhXV i(t0) 1 − e − M σ 1 − e + 2 1 − e − t/τr hX2i(t) = . A2δ2 (5.23)
Compared to hX2i, hXi2 can be neglected. If higher orders of δ in the expression for hX2i are left out, too, one finds that η separates nicely into a product of terms with different dependences:
r δ B kBT η(t) ≈ f(σ, t), (5.24) 2 A M where
2 |σ| 1 − e−t/τr f(σ, t) := q . (5.25) −t/τ 2 −t/τ −σ 1 − e r − 2 1 − e r − t/τr
This approximation is plotted in Figure 5.3 (a) as well and matches the exact curve excellently for the given set of parameters. Setting its first derivative with respect to t equal to zero, shows that f(σ, t) posses a minimum at t = 0 and is maximal when
3e−t/τr − et/τr + 4t/τ − 2 σ = r . (5.26) 2 1 − e−t/τr
Being a transcendental equation, this cannot be solved for t analytically to find the location of the maximum tmax. However, one may simply plot σ versus tmax and swap the axes, which is how Figure 5.3 (b) was obtained. Similarly, the plot of f(σ, t = tmax) (also contained in Figure 5.3 (b)) was generated. Notice that σ cannot be greater than one, as hV 2i ≥ 0, which is why the corresponding axis terminates at that value. If the initial distribution is narrower than the equilibrium pdf, then 0 ≤ σ ≤ 1, if it is broader, then σ ≤ 0. In the latter case f(σ, t = tmax) grows unbounded with decreasing σ which allows the conclusion that in this sense pushing the motor randomly (to increase its average kinetic energy) has the potential of being more efficient than trying to slow it down or stop it. Unfortunately tmax will increase as well. In summary on finds that, within this approximation, the efficiency may be optimised by choosing a motor design that maximises |B/A| and by developing a mechanism that initially pushes hV 2i away from equilibrium as far as possible. Concerning the mass ratio δ, larger values seem preferable, but one has to be aware that there is a trade off with σ as a higher damping makes it generally more difficult to obtain highly non-equilibrium states (cf. Section 5.3.2). Last, the temperature T , density p ρ and motor size S are found to have no direct influence (since A ∝ ρS kBT/M and B ∝ ρS), but may enter through σ.
36 5.2 Potential-change without collisions: time-dependent undamped harmonic oscillator
5.2 Potential-change without collisions: time-dependent undamped harmonic oscillator
The trajectory of the motor is composed of parts of deterministic motion disrupted by discrete jumps caused by collisions. To describe the deterministic segments, it is necessary to study the time- dependent undamped harmonic oscillator, which is the subject of this section. Moreover, this will give an impression of how the system is pushed away from its equilibrium state during the release process.
5.2.1 Linear dynamics
1 2 The motion of an undamped oscillator in the potential U(X, t) := 2 k(t)X is governed by the equation
Γ˙ = A(t)Γ, (5.27) with 0 1 A(t) := −k(t) . (5.28) M 0
Due to the linearity of Equation (5.27), its solution depends linearly on the initial value Γ(t0) and can be written as
Γ(t) = M(t, t0)Γ(t0), (5.29)
R t 0 0 where M(t, t0) is a 2 × 2 matrix. It might by worth noticing that M(t, t0) 6= exp dt A(t ) t0 in general, as R t dt0 A(t0) and A(t) not necessarily commute. Nonetheless, it is always possible to t0 write the solution in the form (5.29), which is readily verified by direct insertion into Equation (5.27), yealding also a conditional equation for M(t, t0):
∂tM(t, t0) = A(t)M(t, t0). (5.30)
The matrices M(t0, t) have the following properties:
M(t, t) = I2, (5.31) (M(t0, t))−1 = M(t, t0), (5.32) M(t00, t0)M(t0, t) = M(t00, t), (5.33) det (M(t0, t)) = 1, (5.34) where I2 denotes the 2 × 2 identity matrix. Equation (5.34) is no other than conservation of phase ∂ 0 space and can be obtained by noticing that ∂t0 det (M(t , t)) = 0, using Equation (5.30), and that det (M(t, t)) = 1, obviously. It also leads to the useful relations
0 0 M11(t , t) =M22(t, t ), (5.35) 0 0 M12(t , t) = − M12(t, t ), (5.36) 0 0 M21(t , t) = − M21(t, t ), (5.37)
0 0 where Mij(t , t) denotes the ij’th matrix element of M(t , t).
37 5 Using potential wells as stopping sites
5.2.2 Exponentially decreasing spring constant Let −t/τ k(t) = k0e . (5.38) With k(t) ω(t)2 := (5.39) M one easily obtains from Equation (5.27)
X¨ + ω(t)2X = 0. (5.40)
Substituting y := 2τω(t), Equation (5.40) is found to be equal to the Bessel differential equation (cf. Reference [1] p. 102) d2 d y2 + y + y2 − n2 X = 0 (5.41) dy2 dy with n = 0, which is solved by X(y) = C1J0(y) + C2Y0(y). (5.42)
J0 and Y0 are Bessel functions of the first and second kind, respectively, and C1 and C2 are integration constants. Re-substitution for y and specialisation to X˙ (t0) = V (t0) eventually lead to