2A, Afstorea Ieoduetion substance X formed by a chemical reaction of the form X = A.A quite well de fined deterministic motion is evident, and ths is reproducible, unlike the fuctuations fround this motion, which are not. 1.2 Some Historical Examples 1.2.1 Brownian Motion ‘The observation that, when suspended in water, small pollen grains are found to be in A very animated and iregula state of motion, was first systematically investigated by Robert Brown in 1827 [1.1], and the observed phenomenon took the name Brownian ‘Motion because of his fundamental pioneering work. Brown was botanis—indeed 4 very famous botanist—and he was examining pollen grains in order to elucidate the mechanism which by which the grains moved towards the ova when ferilising flowers. At first he thought this motion was a manifestation o life he was secking, but whem he found that this motion was present in apparently dead pollen, some over a century old, some even extracted from fossils, and then even in any suspension of Fine particles gas, minerals and even a fragment ofthe sphinx—he ruled out any specifically organi origin ofthis motion. The motion is ilustated in Fig. 1.2 ‘The riddle of Brownian motion was not quickly solved, and a satisfactory ex planation did not come until 1905, when Einstein published an explanation under the rather modest title “Uber die von der molekular-kinetischen Theorie der Ware geforderte Bewegung von in ruhenden Flissigketen suspendierten Teilchen” (con- cerning the motion, as required by the molecular-kinetic theory” heat of particles suspended in liquids at rest) [1.2]. The same explanation was independently devel ‘oped by Smoluchowski [1.3], who was responsible for much of the later systematic development and for much ofthe experimental verification of Brownian motion the- ory. ‘There were two major points in Einstein's solution to the problem of Brownian motion, | Fie. 1.2 Moron of a point undergoing Brownian 12 Some Histrical Examples 3 ') The motion is caused by the exceedingly frequent impacts on the pollen grain of the incessantly moving molecules of liquid in which it is suspended. {i) The motion of these molecules is so complicated that its effect on the pollen grain can only be described probabilstcally in terms of exceedingly frequent satistcally independent impact “The existence of fuctuations like these ones calls out fora statistical explanation of this kind of phenomenon. Statistics hod already been used by Maxwell and Boltz~ 1ann in their famous gas theories, but only as a description of possible states and the likelihood of theie achievement and not as an intrinsic part othe time evolution of the system, Rayleigh (14) was in fact the fst 0 considera statistical description inthis context, ut for one reason or another, very litle arse out of his work. For practical purposes Einstein's explanation of the nature of Brownian motion must be regarded asthe beginning of stochastic modelling of natural phenomena, Einstein's reasoning is very clear and elegant. I contains all the basic concepts ‘which will makeup the subject matter ofthis book. Rather than paraphrase a lassie piece of work shal simply give an extended excerpt fom Einsteins paper (author's translation: “T-must clearly be assumed that each individual particle executes a motion Which is independent of the motions of all other particles: it will also be considered thatthe movements of one and the same particle in different time imervals are independent processes, as long as these time intervals are not chosen too smal “We introduce a time interval + into consideration, which is very small com pared to the observable time intervals, but nevertheless so large that in wo successive time intervals r, the motions executed by the particle can be thought of as events which are independent ofeach other “Now let there be a total of m particles suspended in a liquid. Ina time intorval r the X-coordinates of the individual particles will nerease by an amount A, where for each particle A has a different (positive or negative) value. There will be a certain frequency law for A; the number dn of the particles which experience a shift which is between A and 4 + dA will be ‘expressible by an equation of the form dn = ngs. oan were Foca 0.2 and ¢ is only diferent from zero for very small values of A, and satisfies the condition a) = HA) 23) “We now investigate how the diffusion coeficiem depends on @. We shall ‘once more restrict ourselves tothe case where the number v of particles per ‘uit volume depends only on x and4 1. A Historical Irodction “Let y = f(t) be the number of particles per unit volume, We compute the istribution of particles atthe time r+ fom the distribution time f, From the definition of the function 4A). iti easy to find the number of particles ‘which at time 1 +7 are Found between two planes perpendicular to the x-axis nd passing through points x and x +d. One obtains flsrende= de F fore a.nbayda (24) But since ris very small, we can set fasten) = fleo+ 025) urthermore, we develop f(x + At) in powers of & fort A. flan SLED , LOLA , 126) eH Or ‘We can use this series under the integral, because only small values of A contribute to this equation, We obtain a are ay eat ft Foaaa + 3 F agawas Fs F Seana aan x) the second, fourth, et. terms on the right-hand side vanish, while out of the Ist, 3rd, th, etc. terms, each one is very small compared with the previous, We obtain from this equation, by taking into consideration Fanaa, (128) and sting (aula =D (129) and keeping only the Ist and third terms ofthe right af p@t _ p&t f= ott (1210) “This is already known as the differential equation of diffesion and it can be seen that Dis the diffusion coefficient. “The problem, which correspon to the problem of diffusion from a sin le point (neglecting the interaction between the diffusing particles), is now ‘completely determined mathematically: its solution is, Ho 124 Vaso ‘ 1.2 Some Historical Examples S “We now calculate, with the help of this equation, the displacement 2x in the direction of the X-axis that a particle experiences on the average oF ‘more exactly the square root of the arithmetic mean of the square of the displacement in the direction ofthe X-axis: iis a= VE = VEDI” 0.2.12) Einstein’s derivation is really based on a discrete time assumption, that impacts hap- pen only at times 0,1, 27,3... and his resulting equation (1.2.10) for the distribu tion function f(t) and its solution (1.2.11) are to be regarded as approximations, in ‘which ris considered so smal! that ¢ may be considered as being continuous. Nev’ eteless, his description contains very many ofthe major concepts which have been developed more and mote generally and rigorously since then, and which will be central to this book. For example: §) The Chapman-Kolmogorov Equation occurs as Einstein's equation (1.2.4). It states that the probability of the paticle being at point « at time 1+ ris given by the sum of the probability of all possible “pushes” 4 from positions + A, ‘mokipled by the probability of being atx+ 4 at ime &. This assumption is based ‘on the independence of the push A of any previous history of the motion: it is only necessary to know the intial position of the particle at time —not at any previous time, This is the Markov postulate and the Chapman Kolmogorov equa tion, of which (1.2.4) isa special form, is the central dynamical equation to all Markov processes. These willbe studied in detail in Chap. 3. ii) The Fokker: Planck Equation: Bq, (1.2.10) is the diffusion equation, a special ease ofthe Fokker Planck equation (also known as Kolmogorov's equation) which de: srs a large clas of very interesting stochastic processes in which the system hhas a continuous sample path. In this case. that means that the pollen grain's Position, if thought of as obeying a probabilistic law given by solving the diffu- ‘sion equation (1.2.10), in which time ris continuous (not discrete, as assumed by Einstein) can be written x(2), where x7) i a continuous function of time—but 8 ‘random function, This leads us to consider the possibility of describing the dy- ramics of the system in some direct probabilistic way, so that we would have a random or stochastic differential equation for the path. This procedute was ini- tiated by Langevin with the famous equation that to this day bears his name. We will discuss this in Sect. 1.2.2, and in detail in Chap. 4. ii) The Kramers-Moyal and similar expansions are essentially the same as that used by Einstein to go from (1.2.4) (the Chapman-Kolmogorov equation) to the diffu- sion equation (1.2.10) The use ofthis type of approximation, which effectively replaces a process whose sample paths need aot be continuous with one whose paths are continuous, is very common and convenient. Its use and validity will be discussed in Chap. 11.6 4, AHlistorcat troduction 1.2.2 Langevin’s Equation ‘Some time after Einstein's original derivation, Langevin [1.5] presented a new method which was quite different from Einstein’s and, according to him, “infinitely ‘more simple.” His reasoning was as follows. From statistical mechanics, i was known thatthe mean kinetic energy ofthe Brow= nian particle should, in equilibrium, reach a value (Sm) = ar. 213 (7; absolute temperature, &; Boltzmann's constant). (Both Einstein and Smolu- ‘chowski had used this fac), Acting on the particle, of mass m there should be two forces 4) A viscous drag: assuming this is given by the same formula as in macroscopic hydrodynamics, this is ~6rqadx/dt where 9 isthe viscosity and a the diameter of the particle, assumed spherical. 4i) Another fluctuating force X which represents the incessant impacts of the mo~ ecules of the liquid on the Brownian particle. All that is known about itis that fact, and that it should be positive and negative with equal probability. Thus, the ‘equation of motion for the position of the particle is given by Newton's law as ax dx mo = Gena + 1 oe = enna + x 24a and multiplying by x, this ean be written mde ae) Es) < mo = Sama + xx 2. ae) sana? + xx, 2.415) where v = dxdt. We now average over a large number of different particles and use (1.2.13) to obtain an equation for (2°): mac) der) eae ee raat where the term (xX) has been set equal to zero because (to quote Langevin) “of the irregularity ofthe quantity X". One then finds the general solution ae) dt where C is an arbitrary constant. Langevin estimated thatthe decaying exponential approaches zero with atime constant ofthe order of 10° s, whici Tor any practical observation at that time, was essentially immediately. Thus, for practical purposes, ‘we can neglect this term and integrate once more to get (2) = (x) = AT /Genayl. 218) This corresponds o (1.2.12) as deduced by Binstein, provided we identify D=AT HG), 1.2.19) (216, = KT (Gana) + Cexp(-Grnatin), a2iy 1.2 Some Historical Examples 7 a result which Einstein derived in the same paper but by independent means. LLangevin's equation was the first example ofthe stochastic differential equation— differential equation with a random term X and hence whose solution is, in some ‘sense, a random funetion, Each solution of Langevin’s equation represents a dtfer- cent random trajectory and, using only rather simple properties of X (his fluctuating force), measurable results can be derived. ‘One question arises: Einstein explicitly required that (on a sufficiently large time scale) the change A be completely independent of the preceding value of A. Langevin did not mention such a concept explicitly, but itis there, implicitly, when one sets (Kx) equal to zero. The concept that X is extremely irregular and (which is not men- tioned by Langevin, but is implicit) that X and x are independent ofeach other—that the irregularities in + asa function of time, do not somehow conspire to be always. inthe same direction as those of X, so that it would no be valid to set the average ‘of the product equal to ero. These are really equivalent to Einstein's independence assumption, The method of Langevin equations is cleanly very mich more direct, at least a first glance, and givesa very natural way of generalising a dynamical equation toaprobabilstc equation. An adequate mathematical grounding forthe approach of Langevin, however, was not available until more than 40 years later, when Ito for ‘mulated his concepts of stochastic differential equations. And in this formulation, 1 precise statement of the independence of X and « le tothe calculus of stochastic oo means X > 0, so the natural range (0,09) of prices is recovered. ‘There is also a certain human logic in the description. Prices move as a result of judgments by buyers and sellers, to whom the natural measure of a price change js not the absolute size of the change, but the fractional change. The improvement over Bachelicr’s results So significant, andthe resulting description in terms of the logarithm of the price and the fractional price change so simple, that ths i the pre- ferred model to this day. Samuelson termed the process geometric Brownian motion oralteratively economic Brownian motion. 1.32 Financial Derivatives In order to smooth the running of busines, itis often helpful to fx in advance the price of a commodity which will be needed in the future—for example, the price of Wheat which has not yet been grown and harvested 1s moderately uncertain. A baker ‘could choose to pay a fixed sum now for the future delivery of wheat. Rather than dal with an individual grower, the baker can buy the ungrown wheat from a dealer in wheat futures, who charges a premium and arranges appropriate contracts with rowers. However, the contract to deliver wheat ata certain price on a future date an itself become a tradable item. Having purchased such a contrat, the baker can sell itt another baker, or indeed, to anyone else, who may buy it with the view to selling it ata future date, without ever having had anything to do with any wheat at all Such a contract is known as a derivative security. The wheat future exists only because there is a market for real wheat, but nevertheless can develop an existence of its own, Another kind of derivative is an option, in which one buys the right to porchase something at a future date at a definite price. I the market price on the date at which the option is exercised is lager than the option price, one exercises the option. Ifthe market price turns out to be below the option price, one discards the option and pays the market price. Purchasing the option limits exposure to price101. A Historical Itroduetion rises, transferring the risk (othe seller ofthe option, who charges appropriately, and specializes in balancing risks. Options to purchase other securities, such as shares find stocks, are very common, and indeed there are options markets which trade under standardized conditions. 1.3.3 The Black-Scholes Formula [Although a desctiption of market processes in terms of stochastic processes was Well-known by the 1970s, it was not clear how it could be used as a tool for making investment decisions. The breakthrough came with the realization that a portfolio containing an appropriate mix of cash, stocks and options could be devised in which ‘the short erm fluctuations inthe various values could be cancelled, and that this gave a rclatively simple formola for valuing options—the Black-Scholes Formula—which ‘would be of very significant value in making investment decisions. This formula has truly revolutionized the practice of finance: to quote Samuelson {|-10] reat economist of an earlier generation said that, useful though eco- nomic theory is for understanding the world, no one would go to an eco- ‘homie theorist for advice on how to un a brewery oF prosuce a mousetrap. ‘Taday that sage would have to change his tune: economic principles really «do apply and woe the accountant or marketer who runs counter to economic Jaw, Paradoxically, one of our most elegant and complex secto’s of economic analysis—the modern theory of finance—is confirmed daily by millions of| statistical observations. When today’s associate professor of security analy- sis is asked "Young man, if you're so smart why ain't you rich”, he replies by laughing all the way to the bank or his appointment as a high-paid con- stltane to Wall Street” ‘The derivation was given fist in the paper of Black and Scholes [1.11], and a dif- ferent derivation was given by Merton (1.12). The formula depends critically on description of the returns on securities as a Brownian motion process, which is of mited accuracy. Nevertheless, the formula is sufficiently realistic to make investing in stocks and options a logical and rational process, justifying Samuelson’s perhaps. ‘over-dramatised view of modern financial theory 1.34 Heavy Tailed Distributions ‘There is, however, no doubt thatthe geometric Brownian motion mode of financial rmatkels is not exact, and even misses out very important features. One need only sudy the empirical values of the returns in stock market records (as well as other kinds of markets) and check what kinds of distributions are in practice observed. ‘The results are not really in agreement with a Gaussian distrittion of returns— the observed distribution of returns is usually approximately Gaussian for small values ofr, bat the probability of lage values of r is always observed to be significantly larger than the Gaussian prediction—the observed distributions are said tohave heavy tails. [4 Bith-Death Processes 11 “The field of Continuous Time Finance [1-10] is an impressive theoretical edifice built on this flawed foundation of Brownian motion, but so lar it appears to be the ‘most practical method of modelling financial markets. With modern electronic bank= ing and transfer of funds, itis possible to trade over very short ime intervals, during ‘which perhaps, in spite of the overall increase of trading activity which results, 3 Brownian description i valid I is certainly sufficiently valued for its practitioners to be highly valued, as Samuelson notes. However, every so often one of these practitioners makes aspectac~ ular lss, threatening financial institutions. While there is public alarm about billion dollar losses, those who acknowledge the significance of heavy tals are unsurprised. 1.4 Birth-Death Processes ‘A wide variety of phenomena can be modelled by a particular class of process called ‘a birth-death process. The name obviously stems from the modelling of human or ‘animal populations in which individuals are born, or die. One ofthe most entertaining ‘models is that of the prey- predator system consisting of two kinds of animal, one of ‘which preys on the other, which is itself supplied with an inexhaustible food supply. ‘Thus letting X symbolise the prey, Y the predator, and A the food of the prey, the process under consideration might be XA ox, (4a) x+y, aby yo, (ate) ‘which have the following naive, but charming interpretation. The frst equation sym: bolises the prey eating one unit of food, and reproducing immediately. The second equation symbolises a predator consuming a prey (who thereby dies—this is the only sath mechanism considered forthe prey) and immediately reproducing. The final ‘equation symbolises the death of the predator by natural causes. Iti easy (0 guess ‘model differential equations for x and y, the numbers of X and ¥. One might assume ‘that the first reaction symbolises arate of production of X proportional tothe proivct ‘of « and the amount of food: the second equation a production of ¥ (and an equal rate of consumption of X) proportional toy, and the lat equation a death rate of ¥, in which the rate of death of ¥ is simply proportional toy; thus we might write dx ax kay, a GF = hax hoy (1429) dy fe kyr ky 14.26) a> yy: (1.426) ‘The solutions of these equations, which were independently developed by Lorka 1.13) and Volterra [1.14] have very interesting oscillating solutions, as presented in Fig. 13a, These oscillations are qualitatively easily explicable. Inthe absence of significant numbers of predators, the prey population grows rapidly until the pres ence of so much prey for the predators to eat stimulates their rapid reproduction, a121. A istorica! Introduction 9) sr a Rico eo “Time (arbitrary units) Fig. L.a-e. Time development in prey-pedator systems, (a) Plot of scltons of the deter rinistic equations (4.28, 4.26) (x = sold line, y = dashed line. (b) Data fra real prey- predator system Hete the predator isa mite (Eottranyehus sexmaculatus—dashe Tine) which Feeds on ocanges. and the prey is another mite (Typhlodromus occidental). Data fom (1.15 1.16} @) Simulation of stochastic equations (1 4.3a-1.4.2)- the same time reducing the number of prey which get eaten, Becaase a large number cf prey have been eaten, there are no longer enough to maintain the population of predators, which then die out, returning us to our initial situation. The eyeles repeat indefinitely and are indeed, at least qualitatively feature of many real prey-predator systems. An example is given in Fig. 1.36 Of course the realist systems donot follow the solutions of differential equations cexacily—they fluctuate about such curves. One must include these fluctuations and 1.4 BinDeath Processes 13 the simplest way to do this is by means of a birthdeath master equation. We assume 1 probability distribution, PCy.) forthe number of individuals at a given time and ask for a probabilistic law corresponding to (1.4.2a, 1.4.2b). This is done by assuming that in an infinitesimal time Ar, the following transition probability laws. holds. Probar-rxttiy—y) = kasd, (1430) Prob(x—+x— try + y+ 1) = Anaya (436) Probix>xyoy-D = kwdr, (ae) Prob(x > sy > y) 1 (hia + kexy + kyr aad ‘Thus, we simply, for example, replace the simple rate laws by probability laws, We then employ what amounts to the same equation as Einstein and others used, ie. the CChapman-Kolmogorov equation, namely, we write the probability at + Aras a sum of terms, each of which represents the probability of a previous state multiplied by the probability ofa transition tothe state (x,y). Thus, we find by letting Ar —+ 0 2 atx — DPC 1 ys) + kate Dy BD x Plt Ly = La) +kaly + DPRK Y +1.) = (hart haty + ky) PC gD. 44) Tn writing the assumed probability laws (1.4.2a-1.4.30), we are assuming thatthe probability of each of the events occurring can be determined simply from the knowledge of x and y. This is again the Markov postulate which we mentioned in Sect. 1.2.1 Inthe case of Brownian motion, very convincing arguments can be made in favour ofthis Markov assumption. Here itis by no means clear. The concept of heredity, ie. thatthe behaviour of progeny is related to that of parents, clearly con- tradicls this assumption. How to include heredity is another matter; by no means «does a unique prescription exist. ‘The assumption of the Markov postulate in this context is valid to the extent that different individuals of the same species are similar; i i invalid to the extent that, nevertheless, perceptible inheritable differences do exist. ‘This type of model has a wide application—in fact to any system to which a popu lation of individuals may be attributed, for example systems of molecules of various chemical compounds, of electrons, of photons and similar physical particles as well, as biological systems. The particular choice of transition probabilities is made on various grounds determined by the degree to which details ofthe births and deaths. involved are known, The simple multiplicative laws, as illustrated in (.4.34-1.4.20, are the most elementary choice, ignoring, as they do, almost all details ofthe pro- cesses involved. In some ofthe physical processes we can derive the transition prob: abilities in much greater detail and with greater precision. Equation (1.44) has no simple solution, but one major property differentiates ‘equations like it fom an equation of Langevin’s type, in which the fluctuation term is simply added tothe differential equation, Solutions of (1.4.4) determine both the {gross deterministic motion and the fluctuations; the fluctuations are typically ofthe141, A Historica! Intoduetion same order of magnitude as the square rots ofthe numbers of individuals involved. [isnot difficult to simulate a sample time development ofthe process asin Fig. 1-3, ‘The figure does show the correct general features, but the model is so obviously sim- plied that exact agreement can never be expected. Thus, in conirast tothe situation in Brownian motion, we are not dealing here so much witha theory ofa phenomenon, as witha class of mathematical models, which are simple enough to have a very wide range of approximate validity. We will ee in Chap. 11 that a theory can be developed ‘which can deal with a wide range of models in this category. and tat there is indeed 4 close connection hetween this kind of theory and that of stochastic differential equations 1.5 Noise in Electronic Systems ‘The early days of radio with low transmission powers and primitive receivers, made it evident to every ear that there were a great numberof highly irregular electrical signals which occurred either inthe atmosphere, the receiver or the radio transmiter, and which were given the collective name of “noise”, since thsi certainly what they sounded like on a radio, Two principal sources of noise are shot oise and Johnson 15.1 Shot Noise In a vacuum tube (and in solid-state devices) we get a nonsteady electrical current, since i is generated by individual elecrons, which are accelerated across a distance and deposit their charge one ata time on the anode. The electric current arising from such a process can be writen n= ZF a~ 0). asp where F(r~ 1.) represents the contribution to tbe current ofan electron which arives at time f. Each electron is therefore assumed to give rise tothe same shaped pulse, but with an appropriate delay a in Fig. 14, ‘A statistical aspect arses immediately we consider what kind of choice must be made for 1. The simplest choice is that each electron arrives inéependently of the previous one—that is, the times “are randomly distributed with a certain average ‘number per unit time inthe range (—2, 0), or whatever time is under consideration, ‘The analysis of such noise was developed during the 1920's ard 1930"s and was summarised and laggely completed by Rice [1.17] It was first considered as early as 1918 by Schonky [1.18} We shall find that there is a close connection between shot cise and processes described by birth-death master equations. For, if we consider nthe number of elec~ trons which have arrived up to a time f, to be a statistical quantity deseribed by 1 probability P(n,1), then the assumption that the electrons arrive independently is clearly the Markov assumption. Then, assuming the probability that an electron will 4.5 Noise in Bleetronie Systems 15 | Pe Height Time Fig 14 Illustration of shot noise: identical electric pulses ave at random times arrive in the time interval between # and 1 + Aris completely independent of rand n, its only dependence can be on Ar. By choosing an appropriate constant 2, we may write Prob (n+ I,in time Ad) = Ar, «s2y so that Pine 0) = Penal = AN) + Pl L.Oaar sa) and taking the Fimit Ar —+ 0 aPIn.0) a Which isa pure birth process. By writing Gs.) = LS PIno. There, G(s.) is known as the generating funetion for Pin, 1), and the pati nique of solving (1.5.4) very widely used), we find = ALPin = 1.1) ~ POnD] asa) 20 eH. ase) oth Gta) = explo HG. assy By requiring at time ¢ = 0 that no electrons had arrived, it is clear that P(O,0) is t and Pln,0) is zero for all » > 1, so that Gts,0) = 1. Expanding the solution (1.5.7) in powers of s, we find Pn.t) = exp(-Aniary"/n, 58) Which is known as a Poisson distribution (Sect. 2.8.3). Let us introduce the variable _N(0), which isto be considered as the number of electrons which have arrived up to time and isa random quantity. Then, Pant) = Prob (N()= nl, 59) and N(t) can be called a Poisson process variable, Then cleaty, the quantity 44), formally defined by164. A Historica Itoduetion aaa (15.10) is zero, except when N(i) increases hy 1; at that stage i isa Dirac delta Function, arn), as. aH where the are the times of arrival of the individual electrons. We may write 4) = f dtr r ye) (15.2 A very reasonable restriction on F(¢~ ¢)is that it vanishes if < f and that for 1+ 6 it also vanishes, This simply means that no current arises from an electron before itarives, and thatthe effect of its arival eventually dies out. We assume then, for simplicity, the very commonly encountered form ae, «0, rw 1sa9 [eo Ga, so ta (15 12) an be rowriten as oo Kn) uae a (1.5.14) We cn derive sip deena equation. We diferente 106 otin AH) grgger-tNOD [gga dN) Mf aroge Hd [ponent sis sora ae) a 0516 ‘THis is a kind of stochastic differential equation, similar (o Langevin’s equation, in Which, however, the fluctuating fore is given by qu), where m(i) is the derivative ‘ol the Poisson process, as given by (1.5.11). However, the mean of) is nonzero, in fact, from (1.5.10) qutondty = (aN() = adr, asin dN) adr) = adr, (1s.18) from the properties of the Poisson distibution, for which the variance equals the ‘mean, Defining, then the fluctuation as the difference between the mean value and AN(0), we write dilt) = dN(W) ~ ade, 18.19) 0 that the stochastic differential equation (1.5.16) takes the form Nt) = Lag atid + gant) 820) 1.5 Noise in Electronic Systems 17 Now how does one solve such an equation? In this ease, we have an academie prob- Jem anyway since the solution is known, but one would like to have a technique. Suppose we try to follow the method used by Langevin—what will we get as an answer? The short reply to this question is: nonsense. For example, using ordinary caleulus and assuming (I()dr(@) = 0, we can derive $10) agony. as2n Laren _ oO Sage = enn) ao) (1522) Solving inthe limit ¢—» oo, where the mean values would reasonably be expected to bbe constant one finds (Ho0))_ = Agfa, (1523) (P(e) = (Aglay*. (1524) “The first answer is reasonable—it merely gives the average current through the $ys- {em in a reasonable equation, but the second implies thatthe mean square curentis the same as the square of the mean, ie, the current at 1 —r does not Huctuate! "This is rather unreasonable, and the Solution tothe problem will show tha stochastic differential equations are rather more subtle than we have sofa presented. Firstly. the notation in terms of differentials used in (1517-1520) has been cho- son deliberately. n deriving (1.5.22), one uses ordinary calculus, i.e. one writes Pye (+dl P= 21d! +a? (1525) and then one drops the (42) as being of second onder ind. But now look at (1.5.18): this is equivalent to ddn(a)?) = Adt, (15.26) 0 that a quantity of second order indy is actually of frst order in dt. Te reason 18 ot dffical to fnd, Cleary, dit) = ANC) ~ adr, sz but the curve of Nit) isa step function, discontinuous, and certainly not differen- tiable, atthe times of arrival ofthe individual electrons. In the ordinary sense, none of these calculus manipulations is permissible. But we can make sense ovt of them fs follows. Let us simply calculate (a2) using (1.5.20, 5.25, 15.26) Cath?) = 2UILag — atlas + qdn{n)) + lag alee +qdntoy?). (1.5.28) ‘We now assume again that ((n)) = O and expand, after taking averages using the fact that (dn()2) = Adi to Ist onder in dt, We obtain $dir) = [agin — at?) + qa] dt, 15.29) and this gives ; a (Poo) = (eo)? = £ (1530)181. A strca!Intodetion “Ths, there are fluctuations from this point of view, as ¢ —> co, The extra term in (1.5.29) as compared to (1.5.22) arises directly out of the statistical considerations implicit in N(e) being a discontinuous random function Thus we have discovered a somewhat deeper way of looking at Langevin’s kind fof equation—the treatment of which, from this point of view, now seems extremely naive, In Langevin’s method the fluctuating Force X is not specified, but it will be- come clear in this book that problems such at we have just considered are very ‘widespread in this subject. The mora is that random functions cannot normally be differentiated according to the usual laws of calculus: special rules have to be de- veloped, and a precise specification of what one means by differentiation becomes important. We will specify these problems and theit solutions in Chap.4 which will ‘concern itself with situations in which the uctuations are Gaussian, Autocorrelation Functions and Spectra ‘The measurements which one can carry out on fluctuating systems such as electric circuits are, in practice, not of unlimited variety. So far, we have considered the dis: Cibution functions which tel us, at any time, what the probability distribution of the values of a stochastic quantity are. If we are considering a measurable quantity xt) ‘which fluctuates with time, in practice we can sometimes determine the distribution of the values of x. though more usually, what is available at one time are the mean 40) and the variance varlx())- ‘The mean and the variance do not tell a great deal about the underlying dynamics ‘of what is happening. What would be of interest is some quantity which is a measure ‘of the influence ofa value of « a ime fon te value at time ¢ + r-Such a quantity is the autocorrelation function, which was apparently first introduced by Taylor (1.19) ur Ge) = fim F fdrsionr+ 1). (1531) This is the time average of a two-time product over an arbitrary large time ., which is then allowed to become infinite. Using modern computerized data collection tech- nology it is straightforward 10 construct an autocorrelation from sny stream of data, citer in realtime or from recorded data ‘A closely connected approach is to compute the spectrum ofthe quantity x(). This is defined in two stages. First, define yw) = fdre™*x(1), 5.32) then the spectrum is defined by 1 2, Stu) = i 5 Il a3) ‘The autocorrelation function and the spectrum are clesely connected. By a litle ma- nipulation one finds 1L5 Noise in Blecionic Systems 19 r te sw jim | Feoswondr® Fxom-+ it sap and aking the init 7 > eo (under suitable assumptions to ens the vay of ‘certain interchanges of order), one finds St) S Femwnctee. te ‘Thisis «fundamental result which relates the Fourier transform ofthe autocorrelation Fecon he rump sgh et one oe Gen= fin} 7 arave on 41336 swr= 2 Femorner, asa) vith coreg nee ote)= J staida sas) “This result is known asthe Wiener-Khinchin theorem 1.20, 1.21] and has widespread application Tt means that one may either directly measure the autocorrelation function of a signal, or the spectrum, and convert back and forth, which by means of the fast Fourier transform and computer is relatively straightforward 1.5.3 Fourier Analysis of Fluctuating Functions: Stationary Systems “The autocorrelation function has been defined so far as atime average of a signal, bot, we may also consider the ensemble average, in which we repeat the same measure- ‘ment many times, and compute averages, denoted by ¢ ) It will be shown that for very many systems, the time average is equal 1 the ensemble average; such systems are termed ergodio—see Sect. 3.7.1 If we have such a fluctuating quantity 2(7), then we can consider the average ofthe product of two time-values of x dadoxte+0) = GO). (1.529) "The fact thatthe result is independent ofthe absolute time £is a consequence of our ergodic assumption. ‘Now itis very natural to write a Fourier transform for the stochastic quantity x(f)201, Aistorcal Inteoduetion al and consequently, Sdwc(wye™, (540) L ctw) = 5 Farin «san Note that 2() real implies lw) = e(-w). (1842 If the system is ergodic, we must have a constant (x(), since the time average is clearly constant. The process is then stationary by which we mean that all time- dependent averages are functions only of time differences, ie., averages of functions Mls Xa) Ul) are equal 1 those of x(t) +A), a(f2 + Ay ose + A), For convenience, in what follows we assume (1x) = 0. Hence, tow) = sane =o asa “tat = ah f fanart (lancet = Gh fava" xt), 1 Ftlw—w') fare™Gtr), Bae Ww dre™"orn) Kw — WF )S(w) saa) Here we find not only a relationship between the mean square (J(w)?) and the spec: ‘rum, but also the result that staionarity alone implies that e() and ¢”(c') are un- correlated, since the term 6(w ~ «u') arses because (x(0)x(7)) isa function only of re 1.54 Johnson Noise and Nyquist’s Theorem ‘Two brief and elegant papers appeared in 1928 in which Johnson 1.22] demonstrated ‘experimentally that an electric resistor automatically generated fluctuations of elec- trie voluge, and Nyquist [1.23] demonstrated its theoretical derivation, in complete ‘accordance with Johnson's experiment. The principle involved was alteady known by ‘Schottky [1.18] and is the same as that used by Einstein and Langevin, Ths principle is that of thermal equilibrium. Ifa resistor R produces electric fluctuations, these will produce a current which will generate heat. The heat produced in the resistor must ‘exactly balance the energy taken out of the Mluctuations, The detailed working out ‘ofthis principe is not the subject of this section, but we will find that such results are common throughout the physics and chemistry of stochastic processes, where the principles of stalistical mechanics, whose basis is not essentially stochastic, are brought in to complement those of stochastic processes—such resulls are known a fuctuation-dissipation theorems. Nyquist’s experimental result was the following. We have an electric resistor of resistance R at absolute temperature T. Suppose by means of a suitable filter we measure F(w)duy the volge across the resistor with angular frequency in the range (w.+ da), Then, if kis Boltzmann's constant, 1.5 Noise in Electronic Systems 21 uT fa «1ssy ey) ‘This esul is known nowadays as Nyquist's theorem. Johnson remarked. “The effect is one ofthe causes of what is called ‘tube noise’ in vacuum tube amplifiers. Indeed, itis often by far te larger part ofthe ‘noise’ of a good amplifier” Johnson noise is easly described by the formalism of the previous subsection, “The mean noise voltage is zero across a resistor, and the system is arranged so that, itis in a steady state and is expected to be wel represented by a stationary process. Johnson's quantity is, in practice, a limit ofthe kind (1.5.33) and may be summarised by saying thatthe voltage spectrum 5 («) is given by S (a) = RET/ x, (1546) that i, the spectrum is fat, i. a constant function of w. In the ease of light, the ‘requencies correspond to different colours of light. If we perceive light to be white, iis found that in practice all colours are present in equal proportions—the optical spectrum of white light is thus flat—at least within the visible range. In analogy, the term white noise is applied toa noise voltage (or any other fluctuating quantity) ‘whose spectrum is fla ‘White noise cannot actually exist. The simplest demonstration is to note that the ‘mean power dissipated in the resistor inthe frequency range (11,02) is given by Faust so that the total power dissipated in all frequencies is infinite! Nyquist realised this, and noted tha, in practice, there would be quantum corrections which would, at room temperature, make the spectrum flat oaly up to 7 x 10! Hr, which is not detectable in practice, in a radio situation. The actual power dissipated inthe resistor would be somewhat les than ifinite—10-!® W in fact! And in practice there are other limiting factors such as the inductance ofthe system, which would limit the spectrum to even lover frequencies. From the definition ofthe spectrum in terms of the autocorrelation function given in Sect, 1.5, we have (et NE) = Gtr, «usa 2 Fave™omr, «39 AT ~ w2)/m, «san, = 2RETA(r), (1550) ‘which implis that no matter how small the time diference +, E(t +) and E(x) are not correlated. This i, of course, a direct result ofthe flatness of the spectrum. A {typical model of Sw) that is almost fat is Slo) = RET Htatw?eE + D1 ass ‘This is flat provided w <« 12!. The Fourier transform can be explicitly evaluated in this case to give (EU + ECO) = (RET Ire)expl-t/r0). (152)221, A Historical Ineoduetion Gu) Fig. 5, Correlation Functions (—) and corresponding spectra (==) fr (a) shor corre: lation time coresponding to an almost Nat spectrum: (b) long corelaton cme. giving a quite rapidly deereesing spectrum so thatthe autocorrelation function vanishes only fort >> rc, which isealled the cor relation time ofthe fluctuating voltage. Thus, the delta function correlation function appears as an ideatsation, only valid on a sufficiently long time scale. This is very reminiscent of Einstein's assumption regarding Brownian motion and of the behaviour of Langevin’s Nuctuating force. The idealised white noise will play aa highly important role inthis book but, in just the same way asthe fluctuation term that arises ina stochastic differential equation is not the same as an ordinary differ- ential, we will find that differential equations which include white noise as @ driving {erm have to be handled with great care. Such equations arse very naturally in any fluctuating system and it is possible to arrange by means of Stratonovich's rules for ordinary calculus rules to apply, bu a the cost of imprecise mathematical definition and some difficulties in stockastic manipulation. It turns ut to be far better to aban= don ordinary calculus and use the to calculus, which isnot very different (iis in fact, very similar to the calculus presented for shot noise) and t> preserve tractable statistical properties. All these matters will be discussed thoroughly in Chap. 4 ‘White noise, as we have noted above, does not exist as a physically realisable process and the rather singular behaviour it exhibits does not arse in any reaisable context. It is, however, fundamental in a mathematical, and irdeed in @ physical sense, in that it is an ideatisation of very many processes that do occur. The slightly sirange rules which we will devetop for the calculus of white noise are not really very difficult and are very much easier to handle than any method which always deals with areal noise. Furthermore, situations in which white noise is not a good approximation can very often be indirectly expressed quite simply in terms of white noise. In this sense, white noise is the starting point from which a wide range of stochastic descriptions can be derived, and is therefore fundamertal tothe subject of this book. Probability Concepts In the preceding chapter, we introduced probability notions without any definitions. In order to formulate essential concepts more precisely, it is necessary to have some ‘more precise expression of these concepts. The intention of this chapter isto provide some background, and to present a number of essential results. tis not a thorough ‘outline of mathematical probability, for which te reader is refered to standard math ematical texts such as those by Feller [2.1] and Papouts (2.2), 2.1 Events, and Sets of Events is convenient to use a notation which is as general as possible in order to describe those occurrences to which we might wish to assign probabilities. For example, we ‘may wish to tak about a situation in which there are 64x 10"* molecules in a certain region of space; of situation in which a Brownian panicle is at a certain point x in space; of possibly there are 10 mice and 3 owls in a certain region of a forest ‘These occurrences are all examples of practical realisations of events. More ab- siractly, an event is simply a member of a certain space, which in the cases most practically occurring can be characterised by a vector of integers (ns may ms Ds ey lo a vector of real numbers (x1, ea). 12 ‘The dimension of the vector is arbitrary. Wis convenient to use the language of set theory, introduce the concept of a ser of events, and use the notation wea, 13) to indicate thatthe event «is one of events contained in A. For example, one may ‘consider the set A(25) of events in the ecological population in which there are m0 ‘mote than 25 animals present; clearly the event 2 that there are 3 mice, a tiger, and ‘no other animals present satisfies Be AQ5) en ‘More significantly, suppose we define the set of events A(r, AV) that @ molecule is within a volume element AV centred on a point r In this ease, the practical sig- nificance of working in tems of sets of events becomes clear, because we should28 2, Probability Concepts normally be able to determine whether or not a molecule is within @ neighbourhood. AAV ofr, but to determine whether the particle is exactly tr is impossible. Thus, if we define the event ay) thatthe molecule is at pointy, it makes sense to ask whether cy) € ACr.AV), Qs) and to assign a certain probability to the set A(r, AV), which is to be inexpreted as the probability of the occurrence of (2.1.5). 2.2 Probabilities ‘Most people have an intuit ‘own experi tence, However, a precise formulation of intuitive concept is fraught with dificlties, and it has heen found most convenient to axiomatise probability theory as an essen- tially abstract science, in which a probability measure P(A) is assigned to every set ‘Ain the space of events, including ‘The set ofall events : 2, 21) ‘The set of no events : @, (2.22) in order to define probability. we need our sets of events to form a closed system (known by mathematicians as a o-algebra) under the set theorstic operations of union and intersection, 2.2.1 Probability Axioms We introduce the probability of A, P(A), as a function of A satisfying the follo probability axioms: i) PAYS forall A, 22) ii) PQ=1, (224) iil) IFA; (= 1,2,3,...) is a countable (but possibly infinite) collection of nonoverlapping sets Le. such that AOA =O forall i#j, (225) 226) ‘These are all the axioms needed. Consequentially, however, ve have: iv) if Ais the complement of A, ie. the set ofall events not contained in A, then PA) =1~ Pa), 227) » P) 228) 2.2.2 The Meaning of P(A) ‘There is no way of making probability theory correspond to reality without requir ing a certain degree of intuition. The probability P(A), as axiomatised above, isthe itive probability that an “arbitrary” event w. ie, an event w “chosen at ran- dom, will satisly w» € A. Or more explicitly, if we choose an event “ar random” from @ N times, the relative frequency thatthe particular event chosen will satisfy ‘WA approaches P(A) as the number of times, N we choose the event, approaches infinity, The number of choices NV can be visualised as being done one after the other (“independent” tosses of one die) or atthe same time (N dice are thrown at the same time “independent. All definitions ofthis kind must be intuitive, as we ‘can see by the way undefined terms (“arbitrary”, “at random”, “independent”) keep turning up. By eliminating what we now think of as intuitive ideas and axiomatis- ing probability, Kolmogorov [2.3] cleared the road for a rigorous development of ‘mathematical probability. But the circular definition problems posed by wanting an intuitive understanding remain. The simplest way of looking at axiomatic probabil- ity isas a formal method of manipulating probabilities using the axioms. In order 10 apply the theory, the probability space must be defined and the probability measure P assigned. These are a priori probabilities, which are simply assumed. Examples ‘of such a priori probabilities abound in applied disciplines. For example, in equilib- rium statistical mechanics one assigns equal probabiltes to equal volumes of phase space. Einstein's reasoning in Brownian motion assigned a probability ¢4A) to the probability of a “push” A from a position x atime t ‘The task of applying probability is ') To assume some set of a prior! probabilities which scem reasonable andl to de- ‘duce results from this and from the structure ofthe probability space, {i) To measure experimental results with some apparatus which is constructed to ‘measure quantities in accordance with these a priori probabilities. ‘The structure ofthe probability space is very important, especially when the space of events is compounded by the additional concept of time. This extension makes the effective probability space infnite-dimensional, since we can construct events such as “the particle was at points.ry at times fy for = 0,1,2,...,80". 2.2.3 The Meaning of the Axioms Any inuitive concept of probability gives rise to nonnegative probabilities, and the probability that an arbitrary event is contained in the set of all events must be 1 ‘no matier what our definition of the word arbitrary. Hence, axioms i) and ii) are understandable. The heart of the matter lies in axiom ii). Suppose we are dealing with only 2 sets A and B, and A.n B = ©. This means there are no events contained in both A and B. Therefore, the probability that © AU Bis the probability that either w € A ors € B. Intuitive considerations tellus this probability isthe sum of the individual probabilities, ic,26 2, Probability Concepes PYAU B) = Pl(w€ A) or a € B)) = P(A) PCB) 2.29) Notice this is aot proof—merely an explanation, ‘The extension now to any finite number of nonoverlapping sels is obvious, but the extension only to any countable number of nonoverlapping ses requires some ‘comment. ‘This extension must be made restrictive because of the existeace of sets labelled by a continuous index, for example, x, the position in space. The probability of a molecule being in the set whose only element in x is zero; but the probability of beeing in a region R of finite volume is nonzero. The region R is a union of sets of the form (}—but not a countable union. Thus axiom ii) is not applicable and the probability of being in Ris not equal to the sum ofthe probabilities of being in (x). 2.24 Random Variables ‘The concept ofa random variable is a notational convenience which is central to this book. Suppose we have an abstract probability space whose events can be written x. Then we can introduce the random variable F(x) which isa fimction of x, which ‘akes on certain values for each x. In particular, the identity function of x, written (2) is of interest: itis given by Xin) (2.2.10) We shall normally use capitals in this book to denote random variables and small letters to denote their values whenever its necessary to make adistinction, ‘Very often, we have some quite different underlying probability space 2 with val- tues w, and talk about X(w) which is some function of w, and then omit explicit sncation oF a, This eat be for either of two reasons i) we specify the events by the values of x anyway, ie. we identify x and w: fi) the underlying events w are too complicated to describe, oF sometimes, even to know. Forexample inthe case of the positon of a molecule in aliquid, we really should terret each was being capable of specifying all the postions, momenta, and ori- énlations ofeach melecule i that volume of igi; bu thsi simply too dificult to write down, and olten unnecessary One great advantage of inducing the concept ofa random variable i the sin- plicity with which one may handle functions of random variables eg. X2,sin(a-X), ‘c., and compute means and distributions of these. Further, by defining stochastic Aitferemial equations. one can also quite simply talk about time development of ran- ddom variables in way which suite analogous othe classical description by means of ditferemial equations of non-probebliste systems. 23 Join and Contional Probabilities: Independence 27 2.3 Joint and Conditional Probabilities: Independence 23,1 Joint Probabilities ‘We explained in Sect.2.2.3 how the occurrence of mutually exclusive events is re lated to the concept of nonintersecting sets. We now consider the concept P(A B). where ANBis nonempty. Anevent us which satisfies a» € A will only satisly w € ANB ifwe Bas wel ‘Thus, PCAN B) = Pllw € A) and (w € B)). 31) and P(A © B)is called the joint probability that the event « is contained in both lasses, of, alternatively, that both the events w € A and w € Boccur Join probabil- ities occur naturally in the context of this book in two ways: ') When the event is specified by a vector, ¢., mice and n igers. The probability ofthis event isthe joint probability of (m mice (and any number of tigers)} and In tigers (and any number af mice). All vector specifications are implicitly joint probabilities inthis sense. li) When more than one time is considered : what is the probability that (at time f there are my tigers and n; mice) and (at time fz there re ms tigers and n> mice) ‘To consider such a probability, we have effectively created out of the events at time f; and events at time f, joint events involving one event at each time. In essence, there ino difference between these two cases except forthe fundamental dynamical role of time. 23.2. Conditional Probabil ‘We may specify conditions on the events we are interested in and consider only these, eg. the probability of 21 buffaloes given that we know there are 10O lions. What does this mean? Clearly, we will be interested only in those events contained in the set B= [all events where exacily 100 lions occur). This means that we to define conditional Probabilities, which are defined only on the collection of al sets contained in B. we define the conditional probability as PLA|B) = PUAN BY PLB). 32) nd this satisfies our intuitive conception thatthe conditional probability that w € A (given that we know w € B),is given by dividing the probability of joint occurrence by the probability (w € B). ‘We can define in both directions, ie, we have PUAN B) = PALB)PB) = P(BLA)P(A). 33) ‘There is no particular conceptual difference between, say, the probability of {(21 buffaloes) given (100 lions)} and the reversed concept. However, when two times are involved, we do see a difference. For example, the probability that a particle is al position xy al time f, given that it was at x at the previous time 12, is a very natural thing to consider: indeed, it will turn out to be a central concept inthis book,28 2. Probability Cones “The converse looks to the past rather than the future; given that a particle is at xy at time ¢. what isthe probabfity that that atthe previous time fit was at position x2, The frst concep—the forward probability —looks at where the panicle will go, the second—the backward probability—at where it came from. ‘The forward probability has already occurred in this book, for example, the ‘4A)dA of Einstein (Sect. 1.2.1) is the probability that a particle at x at time f will be in the range [5+ A,x +4 + dA] at time 1 + 1, and similarly inthe other exam- ples. Our intuition tell us as it told Finstein (as can be seen by reading the extract from his paper) that this kind of conditional probability is direcly related wo the time development of a probabilistic system. 23.3 Relationship Between Joint Probabilities of Different Orders Suppose we have a collection of sets B; such that Bn By =O. 234) 233) 236) Using now the probability axiom ii), we see that A 7B, satisfy the conditions on the A; used there, so that EPANB) = RUAN BD), (237) PA). 238) and thus PAI B)PLB) = PLA) 2.39) ‘Thus, summing over afl mutually exclusive possibilities of B int eliminates that variable. Hence, in general, EPAN BAG... BNC 23.10) ‘The result (2.3.9) has very significant consequences in the developanent of the theory of stochastic processes, which depends heavily on joint probabilities. 23.4 Independence We need a probabilistic way of specifying what we mean by independent events ‘Two seis of events A and B should represeat independent sets of evens if the spec- ification that a particular event is contained in B has no influence on the probability ‘of that event belonging to A. Thus, the conditional probability P(A|B) should be independent of B, and hence 2.4 Mean Values and Probubiiy Density 29 POA B) = PlA)PLB). ai Inthe case of several events, we need a somewhat stronger specification. The events, (wo € AWM = 1,2,.--.n) will be considered to be independent if for any subset (tetas ath) of the 801, 2,.--4m, PLAY, OAs --- A) = PAG IPA) «os POG) 23.12) {tis important to require factorisation for all possible combinations, as in (2.3.12) For example, for three sets Ay, itis quite conceivable that PALA) = PAPA). 23.13) for al different i and j, but also that AL Ar=ArMAy= ANAL. (Soe Fig. 2.1) 314) ‘This requires POA, Az 1s) = Pld 9A 0143) = PAD VAS) POA2)PLAs) # PUA )P(A2)PUAS) 23:15) ‘We can sce that the occurrence of w € Az and w € As necessarily implies the occur- rence of w € Aj. In this sense the events are obviously not independent. Random variables X1, X2. Xs...» Will be said to be independent random variables, if forall sets of the form Ay = x such that a) < x :tc0, so we see that higher moments tell us only about the properties of unlikely large values of X. In practice we find that the most important quantities are related to the first and second moments. In particular, for a single variable X, the variance defined by varlX] = (o1XIP = (08 - OOF), 259) and as is well known, the variance var{X] ofits square root the standard deviation CX], is @ measure ofthe degree to which the values of X deviate from the mean value (0, In the case of several variables, we define the covariance matrix as (XG) = (X= XDI — OGM) = XD — XDXD. 54) Obviously (Xi) = varlX. 255) Ifthe variables are independent in pairs, the covariance matrix is diagonal 2.8.2 The Law of Large Numbers As an application ofthe previous concepts, let us investigate the following model of measurement. We assume that we measure the same quantity N times, obtaining sample values of the random variable X(n}; (1 = Iy2ese-e). Since theve are all ‘measurements of the same quanity at successive times, we assume that for every ‘n.X(n) has the same probability distribution but we do not assume the X(n) to be independent. However, provided the covariance matrix (Xin), X(m) vanishes sufi ciently rapidly as jx ~ mj +o, then defining yaa Sx, 036) wwe shall show im Ru = 0) 5 Nis clear that hy = a) 258) ‘We now calculate the variance of Xy and show that as N’ + oo it vanishes under certain conditions: 2.6 Characteristic Function 33 1 Sout A By) = a)® Provided (Xs, Xq) falls off suficiently rapidly as n— mj -> co, we find Jimvaty) =0, 25.10) so that fim Ry is 8 dterminsc variable equal 10. woods of (Xe Xa) canbe chosen, a) hnXe) ~ Ka ac), (25.11) for which one ids 2K (aN N(A= N= a) OK vetted = Fe (Pt) Foo, sz) b) Xa Xn)~ [n= mit, (nem), (2.5.13) and on fis approximately Diy 2 vari ~ Z log N- 7 +0 5.4) In both these cases, var[Xy] > 0, but the rate of convergence is very different. Inter. preting n,m asthe times at which the measurement is carried out, one sees than even very slowly decaying correlations are permissible. The law of large numbers comes jn many forms, which are nicely summarised by Papoulis (2.2). The central limit theorem is an even more precise result in which the limiting distribution function of Rw ~ (X) is determined (see Sect.2.8.2). 2.6 Characteristic Function (One would lke a condition where the variables are iuepeament, ot jus in pats, To this end (and others) we define the characteristic function. I isthe vector (51.83,---y84) and X = (X1,Xaye.e%y) is @ vector of random, variables, then the characteristic function (or moment generating function is defined by 6s) = Cexplis 29) = Jdx pleyexplis x) 26.1) “The characteristic function has the following properties ([2.1], Chap. XV) 60) ipo] ai) 4¢)is a uniormly continuous function ofits arguments forall inte reals (2.4) iv) I the moments (TT Xf") exist, then 262)342. Probab concep WA sequence of probability densities converges to limiting probability density if ‘and only i the corresponding characteristic functions converge tothe corespond- ‘ng characteristic function ofthe limiting probability density, vi) Fourier inversion formula lx) = (2ay"* fds (8) exp(-ix 8) 26.) Because of this inversion formula, (+) determines p(x) with probability 1 Hence, the characteristic function does trly characterise the probability density vii Independent random variables: from the definition of independent random vari ables in Sect. 2.3.4 it follows thatthe variables X,,Xo... are independent if and only if lta tha) = Puan PACED) Palins (264) in which case, 4545200005) = 665159) al) 265) viii) Sum of independent random variables: if Xi, X2,..., are independent random variables and if v= 3x, 266) and the characteristic function of ¥is 8) = (expieY, 267) then #460) = FL 64S) 268) “The characteristic function plays an important role in this book which arses from the convergence property (v), which allows us to perform limiting processes on the characteristic function rather than the probability distribution itself, and often makes proofs easier. Further, the fact thatthe characteristic function is truly characteristic, i.e, the inversion formula (vi), shows that different characteristic functions arse from different distributions. As well as ths, the straightforward derivation of the moments, bby (2.6.2) makes any determination of the characteristic function directly relevant to measurable quantities. 2.7 Cumulant Generating Function: Correlation Functions and Cumulants |A further important property ofthe characteristic function arses by considering its logarithm Os) = log 8) 74) 127 Comulant Generating Function: Corelation Functions and Cumulants 38 which is called the cumulane generating function. Let us assume that all moments ‘0 that g(s) and hence, (3), is expandable in a power series which can be ‘writen as 72) where the quantities QXP°X$2 AT) are called the cumulants ofthe variables X. “The notation chosen should not be taken to mean that the cumulants are functions of the particular product of powers of the X; it rther indicates the moment of highest ‘order which occurs in ther expression in terms of moments, Stratonovich (2.5 aso tecr the term correlation funerions, aterm which we shall reserve for cumulant ‘which involve more than one ;, For, if the X are all independent, the factrisation property (26.6) implies that @s) (Ihe cumblant generating function) is sum of frm, each of which is a function of only one s; and hence the coefficient of tnixed lemme. the correlarion functions (in our terminology) ar all zero and the ever sss tr. Thus, the magnitude of the coelation functions is a measure ofthe degree of corelation. “The cumulant and coreaton functions can be evaluated in erms of moments by expanding the characteristic function asa power series: Ps 4 one) = Ei Dag ay XE), a0 EE capa A et Oe dim) Oe 27» [Expanding the logarithm in a power series, and comparing it with (2.7.2) for (8). the relationship between the curmulants and the moments can be deduced. No simple formula can be given, but the first few cumulants can be exhibited: we find ap =X) 74) XP = HRY — KAD. 75) HGR Xd = OKA) — KRG) — IKE) ~ OAD + KOGA 076) Here, all formulae are also valid for any number of equal i .k./. An explicit gen- cral formula can be given as follows, Suppose we wish to calculate the cumulant AX XaXs...Xa)). The procedure is the following: i) Writea sequence of n dots .. ii). Divide ino p + | subsets by inserting angle brackets GMM de e277) i) Distribute the symbols X;...%q in place ofthe dots in such a way that all difer- ent expressions ofthis kind occur, ©... CX YKAKa) = (MYRDND) # OASYXIDDs 278) iv) Take the sum of all such terms for a given p. Call this Cy(Xi, X2,-.-+Xo)% Probability Concepts YOK Ky = EC Wp Kayo Xe 79 A drivation ofthis Formula was given by Meerom (2.6) The particular procedure is de to van Kampen (2.7) vi) Camutansin which there i one or more repeated element For example (X3X5X2}—simply evaluate (Xi XeXsXa) and set Xe = X; inthe resulting expression 2.7.1 Example: Cumulant of Order 4: (XiX2X3Xa) a) p=0 Only term is (Xi XoXsXe) = ColXiX2X0X0), b) pet Partition (.)...) ‘Ten (OX )MN Mad + a)(KKAN) + WAIN) +X A)N AGS) partition (96.2) ‘Term (X X2XsX) + (XIN ANNs) + (ML a)ONONG) Da Hence, Dy + Ds = CUR XXX). 2.7.10) Partition (3634-3) ‘Tear (X )¢Xa)(KaXa) + (X WGN NAKA) + KYKDNEXS) HKD) Mad + (Ka AUDEN) + KS) HDG AD) 5 CG KKK). ap Partition (424) ‘erm (Xi )(Xa)(XsMXs) = CNNNANA) Hence, AX XaXaXa) = Co ~ C14 203 ~ OC en. 2.7.2 Significance of Cumulants From (2.7.4, 2.7.5) we see that the first two cumulants are the means (Xj) and co- variances (XX;). Higher-order cumulants contain information of decreasing signif cance, unlike higher-order moments. We cannot set all moments higher than acertain order equal to zero since (X") > (X")? and thus, all moments contain information about lower moments. For cumulans, however, we can consistently set 28 Gaussian and Poissonian Probability Distebutions 37 @) =a, Wy =0?, WX) =0, (M>D, ‘and we can easily deduce by using the inversion formula forthe characteristic fune: tion that 27.1) that is, a Gaussian probability distribution. It does not, however, seem possible to give more than this intuitive justification. Indeed, the theorem of Marcinkiewicz [2.8 2,9) shows thatthe cumulant generating funetion cannot be s polynomial of degree ‘reater than 2 that is, either all but the first 2 cumulants vanish or there are an infinite fhumber of nonvanishing cumulants. The greatest significance of cumulant lies in the definition ofthe correlation functions of diferent variables in terms of them; this leads further to important approximation methods. 2.8 Gaussian and Poissonian Probability Distributions 2.8.1 The Gaussian Distribution By far the most important probability distribution is the Gaussian, or normal disti- bution, Here we collect together the most important facts about it EX is a vector of n Gaussian random variables, the corresponding multivariate probability density function can be written 1 ix) =§ exp [Hr 2-9), 2.8. 2) = aS p[-$r- 37 e-3)] 281) s0 that (X= fdex pox)= 5 082) 083) 084) ‘This particularly simple characteristic Function implies that all cumulant of higher order than 2 vanish, and hence means that all moments of order higher than 2 are expressible in terms of those of order 1 and 2. The celationship (2.8.3) means that is the covariance matrix (as defined in Sect. 2.5.1), Le. the matrix whose elements are the second-order correlation functions, OF course, is symmetric. “The precise relationship between the higher moments and the covariance matrix «can be written down straightforwardly by using the relationship between the mo- rents and the characters funtion [Set 2.6). The formula is only simple i 1 in which case the odd moments vanish and the even moments satisty3H 2. Probability Concepts OA Keo) = Oe youre lame 085) “where the subscript “sym” means the symmetrised form of the product of o"s, and 2N isthe order ofthe moment. For example, A Ae 2 ox = AS fot = 30h esr 2.82 Central Limit Theorem “The Gaussian distribution is important fora variety of reasons. Many variables are, in practice, empirically well approximated by Gaussians and the reason for this arises from the central limit theorem, which, roughly speaking, asserts that a random vat able composed ofthe sum of many pars, each independent bu arbitrarily distributed, is Gaussian. More precisely let Xy,%2,Xs,....y be independen’ random variables such that (%)= 0, varlX) = 5}, (28.8) and drain action of Hb Dene ix. eas) ond che vasal= £8 a9) ‘We vequine Further the Fl sn ofthe Lindeberg condition: i : sin |= Ef ave nto] =0. eat) for any fixed 1 > 0. Then, under these conditions, the distribution of the normalised sums S/o tends to the Gaussian with zero mean and unit varians The proof of the theorem can be found in [2.1]. It is worthwhile commenting fon the hypotheses, however. We fist note thatthe summands X} are required to be ‘independent. This condition is not absolutely necessary; for example, choose x LY, (28:12) where the ¥; are independent. Since the sum ofthe X's can be rewritten asa sum of ¥°s (with certain finite coefficients), the theorem is sill tue. Roughly speaking as long asthe correlation between X; and X, goes to ero suf- ciently rapidly as i- j| +c, conta limit theorem will be expected. The Lindeberg 28 Gaussian and Poissonian Probability Distributions 39 condition (2.8.11) is not an obviously understandable condition but is the weakest condition which expresses the requirement that the probability for [Xo be larg is ‘ery small. For example, if all the by are infinite or preater than some constant C, its clear that of diverges as n —» oo. The sum of integrals in (28.11) i the sum ‘of contributions to variances for al {Xj > 17x, and its clear that as 7 — 6 each contribution goes to zero. The Lindeberg condition requires the sum of all the con- tributions not to diverge as fast as 2. In practice it isa rather weak requirement; satisfied if [i] < C forall X, orf px) goto zero sufficiently rapidly'as.x ~» 4ce ‘An exception is a 2819 mera) the Cauchy, or Lorentsian distribution. The variance ofthis distribution is infinite and. in fac. the sum of al the X; has a distribution othe same form as (2.8.13) with 4; replaced by 5 a, Obviously, the Lindeberg condition isnot satisfied A related condition, also called the Lindeberg condition, will arise in Sect, 3.3.1, ‘where we discuss the replacement of a discrete process by one with continuous steps. pa 2.8.3 The Poisson Distribution {A distrbation which plays a central role in the study of random variables which take ‘on postive integer valves is the Poisson distribution. 1X isthe relevant variable the Poisson distribution i defined by PX =e Pe) (2814 and lea te facil moments, defined by = GED. Dy axis) sre givenby yea" ais) For variables whose range is nonnegative integral, we can very naturally define the generating function 60) = § P09 esi which related othe characteristic function by G05) = 6-i on). 28.18) “Te generating function has the sel property that or [(2) e9] ay For the Poisson distribution we havei= 2 eget) a2 ‘We may also define the Factorial cumulant generating function gs) by 918) = log G(s) 2821) and the factorial cumulants &X")¢ by way 7 2.822) as) = 36x: ‘We see that the Poisson distribution has all but the frst factorial cumulant zero. ‘The Poisson distribution arises naturally in very many contexts, for example, we have already met iin Sec. 1.5.1 a8 the solution ofa simple master equation. It plays similar cenral roe in the study of random variables which take on integer values to that occupied by the Gaussian distribution inthe study of variables with a continuous. range. However, the only simple multivariate generalisation ofthe Poisson is simply « produet of Poissons, ie. of the Form a ea)" Porm aye) 2823) There is no logical concept of a correlated mulipoissonian distibution, similar to ‘hat of a correlated multivariate Gaussian distribution. 2.9 Limits of Sequences of Random Variables “Much of computational work consists of determining approximations to random vari- ables, in which the concept of a limit of a sequence of random variables naturally arises. However, there is no unique way of defining sucha limit. For, suppose we have a probability space Q. and a sequence of random variables, X, defined on 2. Then by the limit ofthe sequence as e0| X= dim Xy 29.) ‘we mean a random variable X which, in some sense, is approached by the sequence cof random variables X,. The various possibilities arise when one considers thatthe probability space @ has elements w which have a probability density p(w). Then we ‘can choose te following definitions. 2.9.1 Almost Certain Limit X, converges almost certainly to X if, forall « except a set of probability eto X,(e) = Xu) (292) Thus each realisation of Xy converges to X and we write 2.93) 29 Limits of Sequences of Random Variables 2.9.2 Mean Square Limit (Limit in the Mean) Another possibility is to regard the Xq(w) as Functions of «and look for the mean square deviation of Xq(w) from X(w). Thus, we say that X, converges to X in the ‘mean square if lim f des ple){Xale) ~ X(w))? lim 4(Xy ~ X)°) = 0. 294) ‘Tiss the kind of iit which swell known in Hilbert space theory. We write mein X, =X 295) 29.3 Stochastic Limit, or Limit in Probability ‘We can consider the possibility that X,(w) approaches X because the probability of deviation from X approaches zero: precisely, this means that if for any z > 0 lim P(X, —X1> 6) = 0. 296) then the stochastic Himit of X is X. Inthis case, we write stlimX, =X. 299) 29.4 Limit in Distribution ‘Aneven weaker form of convergence occur if for any continuous hounded function 0) Jim F0%)) = FO) 098) In this case the convergence ofthe limit is said wo be in distribution. In particular, using exp(ixs) for /(), we find thatthe characteristic functions approach each other. and hence the probability density of Xy approaches that of X: 2.9.5 Relationship Between Limits ‘The following relations can be shown, 1) Almost certain convergence => stochastic convergence. 2) Convergence in mean square => stochastic convergence. 3) Stochastic convergence = convergence in distribution. All ofthese Fimits have uses in applications.3. Markov Processes 3.1 Stochastic Processes Ail othe examples given in Chap. 1 can be mathematically described as stochastic processes by which we mean, in a loose sense, systems which evolve probabils tically in time or more precisely, systems in which a certain time-dependent ra ddom variable X(1) exists. We can measure values x2 2, +.» € OF X(t) at times Fictostye... and we assume that a set of join probability densities exists Plxrenaas tastes ds LD which descr the system completely. In terms of these joint probability density functions, one can also define condi tional probability densities OEY Te) Plait tact os Imnet Weta) PON TYa Fa) 12 ‘These definitions are valid independently ofthe ordering ofthe tines. although it is ‘usual to consider only times which increase from right to left ie Neh Baer ene 1a ‘The concept of an evolution equation leads us 10 consider the conditional probabi ites as predictions of the future values of X(?) (ie X1.325-.. al times fs t3. given the knowledge ofthe past (values yy...» times 71.125 ..0) 3.1.1 Kinds of Stochastic Process ‘The concept of a general stochastic process is very loose. To defise the process we need to know at least all possible joint probabilities of the kind in (3.11). IF such Knowledge does define the process, it is known as a separable stochastic process. Al he processes considered inthis book will be assumed to be separable, 42) Complete Independence: This is the most simple kind of stochastic process it satisfies the property play nko tists.) = TL porns) G14) 32 Markov Process 43 Which means thatthe value of X a time ris completely independent of its values in the past (or future). ) Bernoolli Trials: An even more special case occurs when the p(t) are inde- ‘pendent of 7, so thatthe same probability law governs the process at all times. We then have the Bernoulli rials, in which a probabilistic process is repeated at succes- sive times, ©) Martingales: The conditional mean value of X(0) given that X(to) = i defined (Keollagtol) & fd x plastl sot) 1s) Ina martingale this has the simple property (XColLsa tol) = 0 G.L6) ‘The martingale property is rather strong property, and is associate with many similar and related processes, such as local martingales, sub-martingales, super-martingales ‘etc. which have come tobe extensively studied and used inthe past 25 years. Potter [3.1] has weitten the definitive book on their use in stochastic processes. 1d) Markov Processes: ‘The next most simple idea is that of the Markov process in which knowledge of only the present determines the future, and most of this book is. built around this concept 3.2 Markov Process “The Markov assumption is formulated in terms of the conditional probabilities. We require that ifthe times satisfy the ordering (3.1.3), the conditional probability is determined enttely by the knowledge ofthe most recent condition, ie. Play thsataicos | mpeTMae abe) = poetshionastaseo IWyTD) 20) This is simply a more precise statement of the assumptions made by Einstein, Smolu- chowski and others. Iti, even by itself, extremely powerful. For it means that we can. define everything in terms ofthe simple conditional probabilities p(x). 4.1). For ‘example, by definition ofthe conditional probability density p(x), 1; x2.f2l4.71) xy x2. f: Mie T1)PCR2falgy.71) and using the Markov assumption (3.2.1). we find leita fim. T1) = plate La. f)pCa2 tay. 7Ds 22) and it isnot dificult to see that an arbitrary joint probability can be expressed simply Pleats 229795) cits Lota) PCa tts peat tate) «PLE a tales 1 af) PCE nt provided HD DBD DM De 24)443, Markov Processes 3.2.1 Consistency—the Chapman-Kolmogorov Equation From Seot.2.3.3 we require that summing over all mutually exclusive events of one kind ina joint probability eliminates that vatiable, ie. ZPAnBC.. PARC...)s 25) «d when this is applied to stochastic processes, we get two deceptively similar equa- tions: poets Jes pry te2t2) = Fede> plea ts Laas h)planct 626) ‘This equation is an identity valid for all stochastic processes and is the first in @ hierarchy of equations, the second of which is lari st Lest) dx por. tisxta.f ts.) = Fede plat | ta t5x1.t))p Crate 83.0). 627) This equation is also always valid. We now introduce the Markov assumption. If 1 > af we en drop ther dependence in he dobly condoned probability and write lry.t Late) dry plot) | x3. 12)PCR fa Xt). G28) hich is the Chapman-Kolmogoray equation ‘What is the essential difference between (3.2.8) and (3.2.6)? The obvious answer is that (3.2.6) is for unconditioned probabilities, whereas (3.2.7) is for conditional Probabilities. Equation (3.2.8) is a rather complex nonlinear Functional equation re- lating all conditional probabilities p(x |.()) to each other, whereas (3.2.6) sim- ply constructs the one time probabilities inthe future tof ¢, given the conditional probability p(xy.t [2 f). The Chapman-Kolmogorov equation has many solutions. These are best under: food by deriving the differential form which is done in Sect. 34.1 under certain rather mild conditions 3.22 Discrete State Spaces In the ease where we have a discrete variable, we will use the symbol N= ().Na.Ni...), where the N; are random variables which take on integral values Cleary. we now replace fide 29) ‘and we can now write the Chapman-Kolmegoroy equation for such a process as Pant) macs) = BPs maa) Plt 1) 2.10) ‘This is now a matrix multiplication, with possibly infinite matrices. 43 Continuity in Stochastic Proceses 45 3.2.3 More General Measures, ‘A more general formulation would assume a measure d(x) instead of dx where a variety of choices can be made. For example, if (x) is a step function with steps at integral values of x, we eecover the discrete state space form, Most mathematica! ‘works attempt to be as general as possible. For applications such generality can lead to lack of clarity so, where possible, we will favour a more specific notation. 3.3 Continuity in Stochastic Processes Whether or not the random variable X()has a continuous range of possible values is a completely different question from whether the sample path of X(2) is acontinuous function of For example, ina gas composed of molecules with velocities V(0). itis clear that all possible values of V(0) are in principle realisable, so thatthe range of (is continuous. However, a model of collisions ina gas of hard spheres as occur. ring instantaneously is often considered, and in such a model the velocity before the collision, o, will change instantaneously at the time of impact to another value ry, 0 ‘the sample path of V(t) is not continuous. Nevertheless, in such a model, the pasition ‘fa gas molecule X(i) would be expected o change continuously. ‘A major question now arises. Do Markov processes with continuous sample paths actually exist in reality? Notice the combination of Markov and continuous. It is almost certainly the case that in a classical picture (i.e, not quantum mechanical), all variables with a contiquous range have continuous sample paths. Even the hard sphere gas mentioned above is an idealisation and more realistically, one should allow some potential to act which would continuously deflect the molecules during a collision, But it would also be the ease that, if we observe on such a fine time scale, the process will probably not be Markovian. The immediate history of the whole system will almost certainly be required to predict even the probabilistic fue. This is certainly born out inal altempis to derive Markovian probabilistic equations rom ‘mechanics. Equations which are derived are rarely troly Markovian—rather there is 1 certain characteristic memory time during which the previous history is important 132.331 ‘This means that inthe real world there is really no such thing as a Markov process; rather, there may be systems whose memory time is so small that, on the time seale ‘on which we carry out observation, itis fair to regard them as being well approx jmated by a Markov process. But in this case the question of whether the sample paths are actually continuous is not relevant. The sample paths ofthe approximating “Markov process certainly need not be continuous. Even if collisions of molecules are not accurately modelled by hard spheres, during the time taken for a collision, afi nite change of velocity takes place and this will appear in the approximating Markov process asa diseete step. On this time scale, even the position may change discon- tinuously ths giving the picture of Brownian motion as modelled by Einstein In chemical reactions, for example, the time taken for an individual reaction to proceed to completion—roughly of the same order of magnitude asthe collision time46-3. Markov Processes Fig. 31. Musation of sample paths of the ‘Cavehy process XU) (dashed line) and Brown- ian notion Wold tne) {for molecules—provides yet another minimum time, since during this time, sates ‘which cannot be described in terms of individual molecules exist, Here, therefore, the very description ofthe state in terms of individual molecules requires a certain minimum time scale to be considered However, Markov processes with continuous sample paths do exist mathemati ‘cally and are useful in deseribing reality, The model of the gas mentioned above provides a useful example, The position of the molecule is indeed probably best ‘modelled as changing discontinuously by discrete jumps. Compared tothe distances travelled, however, these jumps are infinitesimal and a continuous curve provides {200d approximation to the sample path. On the other hand, the velocities can change bby amounts which are ofthe same order of magnitude as typical values attained in practice. The average velocity of a molecule in a gas is about 100C m/s and during collision can easily reverse its sign. The velocities simply cannot reach (with any significant probability) values for which the changes of velocity can be regarded as very small. Hence, there is no sense in a continuous path description of velocities in a gas. 3.1 Mathematical Definition of a Continuous Markov Process For a Markov process, itcan be shown {3.4 that with probability one, the sample paths are continuous Functions oft if For any 5 > 0 we have 1 fim 2 f de plate ari. =0, a) My Bip hoe uniformly in 2. and Av ‘This means thatthe probability forthe final position x to be finitely different from £ goes 10 2er0 faster that Ar as AF goes to zer0. Equation (3.3.1) i sometimes called the Lindeberg condition. Examples i) Einstein's solution or his f(x,0) (Sect. 1.2.1) is really the conditional probability (x.110.0), Following his method we would find 34 Diflerental Chapman Kolmogorov Equation 47 post Atle) 32) and itis easy to check that (3.3.1) is satisfied inthis case. Thus, Brownian motion in Einstein's formulation has continuous sample paths. ii) Cauchy Process : Suppose a (xt Ang.) = a oa pe l= aT areaA ) “Then this does not satisfy (3.3.1) so the sample paths are discontinuous. However, in both cases, we have as requited for consistency im, rtst + Atha.) = 8x ~2), 34) and it is easy to show that in both cases, the Chapman-Kolomogoroy equation is satisfied. “The difference between the two processes just described is illustrated in Fig. 3.1 in which simulations of both processes are given, The difference between the two is striking. Notice, however, that even the Brownian motion curve is extremely ir regular, even though continuous—in fact itis nowhere differentiable, The Cauchy process curve, however, is only piecewise continuous. 3.4 Differential Chapman-Kolmogorov Equation Under appropriate assumptions, the Chapman-Kolmogorov equation can be reduced to differential equation. The assumptions made are closely connected with the con- tinuity properties of the process under consideration. Because of the form of the ‘continuity condition (3.3.1), one is led (0 consider a method of dividing the difer- entiability conditions into pats, one corresponding to continuous motion of arepre- sentative point and the other to discon ‘We require the following conditions forall e > 0: i) imple. + i070 = Weel. oan thifomly in x2, and ¢ for b= al >: a) fim bf acta — apple + Ariz.) = Az. + O10) 42) tide i pm, fais -sinreien= AYN #0. 43 the ast wo being uniform in , af ‘Notice that all higher oder coufficents ofthe form in (34.2, 3.4.3) must vanish, For example, consider the third-order quantity defined by Wim Lf datas antsy epee epee Arlee, Euan = Cale) + 060) aad483, Markov Processes Since Cx is symmetric in, jk, consider Sremicins. = Ciae.0, 45) so that 1a Clean = pe — Ce. (= Baca CE 646) Then, 1 eee. 2.o1< fim, xf la Ge ablla-ta—2iFptat+ Ala. nde+ Oe) Slee fim flex ~ 2) plat + Atl z, tds + OF6) = elarlieiaBy(e.1) + O16)) + O46), (6). Gan imilary, we can show that all corresponding higher-order quant so that C is ze, ties also vanish, ‘According tothe condition for continuity (3.3.1), the process ean only have contin- ‘vous paths if Wx | 2.1) vanishes forall x # 2. Thus, this function must in some way deseribe discontinuous motion, while the quanites A; and mst be connected ith continuous motion, 3.4.1. Derivation ofthe Differential Chapman-Kolmogorov Equation ‘We consider the time evolution of the expectation ofa function f(z) which is twice continuously differentiable. Thus, 4, fds fexplaty.t) =m, arf Fee sen e+ arian) pestle] (48) = Bay gy te fae fa) pie + arlene) ~ J dae sno}, 049) ‘where we have used the Chapman-Kolmogorov equation in the positive term of (34.8)10 produce the corresponding term in (3.4.9). ‘We now divide the integral over x into two regions |x ~ 2] > e and |x ~ 2] 35.4 General Processes In general, none of the quamtties in A(z,).B(z,f) and W(x 2.0) need vanish, and inthis case we obtain a process whose sample paths are as illustrated in Fig. 3.28, i.e. a piecewise continuous path made up of pieces which correspond to a diffusion process with a nonzero drift, onto whichis superimposed a fluctuating par. Ttis also possible that A(z, 1) is nonzero, but B(z,) is zero and here the sample paths are, asin Fig. 3.2b composed of pieces of smooth curve [solutions of (3.5.19)] With discontinuities superimposed. This is very like the picture one would expect ina dilute gas where the particles move freely between collisions which cause an ‘instantaneous change in momentum, though not positon. 3.6 Equations for Time Development in Initial Time—Backward Equations We can derive much more simpy than in Set. 34, some equations which give the time development with respect tothe initial variables yt of p(x.tly. 0). ‘We consider 1 im 2 fptastigae +f) ~ plz} 661) dim, gpleetlnt + at) - pe tye, SAL Ly. pletlyst + OP) ~ plate + O/)} (362) 1 BM ae 42 me “The second line follows by use of the Chapman-Kolmogorov equation in the second term and by noting that the fist term gives 1 x ptx,tly.t! + Af).563, Markov Processes ‘The assumptions that are necessary are now the existence of all relevant deriva- tives, and that p(x, ly.) is continuous and bounded in x,t,f for some range 11> 65> 0, Wemay then write (8.6.2) Jim, 3 fae pla’ + at Wot folate.) pati), G63) We now proceed using simitar techniques to those used in Seet.34.1 and finally derive 2 2 pF rketiy.t) PO y By. PE fee EST + fe Wolof ftastine) = patie} 364) “Wis wl be called the Boker dferetia Chapman Kolmogorov equation. In 2 naar sens, ster defn thn she Crespo fred eation (a2) The appropiate condo orb enon a ass) rspreseing the obvious fac tha i hepatic i a y time the rbebity pau oa ry “The frvar and ine acer equation recut! o cach ker Fo, sl von of the forvard equation subtotal conion (343) and ny ar prt boundary conditions, ye! Slaton of he Chapman Kolnogeror equ, pee Saree ier ser car eb rinner aati eet peel i in alien oat nt aia! quan dea wah n Sec S132) Tne ba iflrnce which et a ar a telnet ors > fost (363) nen nile fre ford eq ‘on arth bck equation, suo eit fer <0 tt sne he aca pa ete i mi rg en even anni Siace te are equivalent the frwarl an beckward equnions ae bath est ‘The forward equation gives more directly the values of measursble quantities as a finctinot he there ime tdtenstte wed erccommenl inappiaons The tciwardoqton fins os aplcaten nh sya fe pease me oF [Ee ele ee ny ie acre igen nada ir pete 3.7 Stationary and Homogeneous Markov Processes In Sect.1.5.3 we met the concept of a stationary process, which represents the stochastic motion of a system which has settled down to a steady state, and whose stochastic properties are independent of when they are measured. Stationarty can be ‘defined in various degrees, but we shall reserve the term “stationary process” for a strict definition, namely. a stochastic process X() is stationary if X() and the process 117 Stitionary and Homogeneous Markov Processes 57 X(1+ «) have the same statistics for any &. This is equivalent to saying that all joint probability densities satisy time translation invariance, ie, pox. tisaa iss fie inte) = plait + eines HeLa +E tes 2) and hence such probabilities are only functions of the time differences, «j= 1) In particular, the one-time probability is independent of time and can be simply writen pala, 672) and the two-time join probability as Palais — 552.0) G73 Finally, the conditional probability can also be written as ara) palxisth ~ 122.0) For a Markov process since all joint probabilities can be written as products of the two-time conditional probability andthe one-time probability, a necessary and suff- cient condition for stationarity is the ability to write the one and two-time probabil ties in the Forms given in (3.7.1-3.7.). 3.7.1. Engodie Properties I we havea stationary proces, it is reasonable 1 expec that average measurements could be constructed by taking values of the variable x at suocessive times, and av- feraging various functions of these. This is effectively a belief thatthe law of lage rumbers (as explained in Sect.2.5.2) applies to the variables defined by successive rmeasuremens ina stochastic process 2) Ergodic Property of the Mean: Let us define the variable 27) by Lt pal 875) R= 5p F dex, 675) where 31) isa stationary process, and consider the limit T —+ co. This represents @ possible model of measurement ofthe mean by averaging overall times. Clearly RO= Os. 678) We now caleuate the variance of X(T). Thws, ODP) = Ghy F J duacnteysend, 677) and ifthe process is stationary, Cateye) = Reh = 2) +P, 378) ‘where Ris the two-time correlation function, Hence,58 3, Markov Proceses OP) = shy Td RNA I, 0.79) where the ast ator follows by changing variables to renott 67.10) an integrating ‘The left-hand side is now the variance of (7) and we will show that under certain conditions, this vanishes as T — co, Most straightforwardly ll we require is that any Ure tt ting Lar(t- Fy) ne which isa litte obscure. However, it is clear that a sufficient condition for this mit to be zero is for Jariani ) Engodie Property ofthe Spectrum: We similarly fnd that the spectrum, given ty the Fourier eansform eGtndr, 7.19) as in Sect. 15.2, is also given by the procedure 7.20) 1 [Faroe 4) Engodic Property of the Distribution Function: Finally, the practical method cof measuring the distribution function isto consider an interval (xy, x2) and measure x(t) repeatedly to determine whether itis in this range or not. This gives a measure of [2 dx p.(x). Essentially, we are then measuring the time average value of the function x(x) defined by hoo n tp by Poet) = pilxstl ato). 128 pestle’) = pae'tlet), (37.28) and all other joint probabilities are obtained from these in the usual manner for a Markov process. Clearly. if 3.7.23) is satisfied, we find that as 1+ coor as fy —» —20, Plt) — pax) 726) and all other probabilities become stationary because the conditional probability is Stationary, Such a process is known as a homogeneous proces ‘The physical interpretation is rather obvious, We havea stochastic system whose variable x is by some extemal agency fined to have a vale al time fy. tthen evolves back to a stationary system wih the pasage of time. This # how many stationary systems are created in practic. From the point of view of the diferental Chapman-Kolmogorov equation, we ‘wil find that the stationary distibuion function pas) sa solution of the stationary diferent! Chapman-Kolmogorov equation, which ake the frm a Belem tla + 3 # O=-E Fes BME tl +f dx{Weel epee, tly.) — Weel 2)pte.tla Is 720 Where we have used the fact that the process is homogeneous to aoe that A, B and W, as defined in (3.4.1-3.4.3), are independent of r. This isan alternative definition, ‘of a homogeneous process. 3.7.3 Approach to a Stationary Process ‘A converse problem also exists. Suppose A.B and W are independent of time and pala) satisfies (3.7.27), Under what conditions does a solution of the differential ‘Chapman-Kolmogorov equation approach the stationary solution (2)? ‘There does not appear to be a complete answer to this problem. However, we can _givea reasonably good picture as follows. We define a Lyapunov functional K of any 'wo solutions p and pe ofthe differential Chapman-Kelgomorov equation by K = Sx prez ttoglpy(x.0/ p22.) 3.7.28) and assume for the moment that neither py nor pe are zero anywhere. We will now show that K is always positive and dK/d is always negative. Firstly, noting that both pax.) and py(x,1) are normalised to ose, we write 2.7 Stationary and Homogeneous Markov Processes 1 e)-1 Kipispatl= f dx pix. 0 logtp xs / poe.) + pala. 1px.) ~ 1} | (37.29) and use the inequality valid forall z> 0, ~loge+2-120, 67.0) to show that K > 0. Let us now show that dK/de < 0, We can write (using an abbreviated notation) = PPL tog py +1 tog pal ~ 221puI I} a7) 5 def tog ps +1 spa - FP ‘We now calculate one by one the contributions to dK /d from drift, diffusion, and jump terms inthe differential Chapman-Kolmogorov equation: i ond Zao} 0739 (B),,-Bte {-tostoun + gen + wuloag (a) which can be rearranged to give lesa ‘Similarly, we may calculate (FB) Bo egeeatosturenn. 0739 et) ae + 1 tBupn ~ (riled geee Burd} (GE), A Bi {totter Bud = (rn gr (Bud 0734) and after some rearranging we may write a a } =A Esde bs { om rron} {toto @ +40 fae By logtps/ po) - (3139) 4 See 5a [Bu toe | Finally, we may calculate the jump contribution similarly: (BE) = face {terterpra'9 —we'inpceo} he x loglps(.0)/pa6e.01+ 1) — [Wexix’yprta'.0- WO Levert.) ]n(x.0/p2.0}, G.7.36) and after some rearrangement, (B= saree westenmersnie eatee'i-o+e0, 737 where a738) = pile DI pales),62 3. Markow Processes and ¢' is similarly defined in terms of ‘We now consider the simplest case. Suppose a stationary solution py(x) exists which is nonzero everywhere, except at infinity, where it and its fist derivative van- ish. Then we may choose p2(x.1) = psx). The contributions to di/dr from (3.7.33) and the second term in (3.7.35) can be integrated to give surface terms which vanish at infinity so that we find ak a 67298) aK (Gi s, that HCOWCS) Loto}? = AWE — WESMIWES) + WLP) 824) Using the independence of increments, the fist average is zero and the second is given by (38.10) so that we have, in genera, HCW) {Wot} = mint ~ tp, 5~t0) + UG, 825) Which is corect for both ¢> sand r< s a he initil vale is 38.2 The Random Walk in One Dimension ‘A man moves along a lin, aking, at random, steps to the left or the right with equal probability. The steps are of length /so that his position can take en only the value nl, where nis integral. We want to know the probability that he reaches a given point ‘distance n! from the origin ater a given elapsed time. ‘The problem can be defined in two ways. The firs, which is more traditional, is to allow the walker to take steps at times Nr (N integral) at which times he must 38 Examples of Markov Processes 69 ‘step either left or right, with equal probability. The second isto allow the walker to take steps left or right with a probability per unit timed which means thatthe walker waits at each point for a variable time. The second method is describable by a master ‘equation 8) Continuous Time Random Walk: To do a master equation treatment of the problem, we consider thatthe transition probability per unit time is given by the form Wont Vin = Win Vin = 0. 3.826) otherwise, Wonlm,t) = 0 s0 that, according to Sect. 3.5.1, the master equation for the man to be atthe position nl, given that he started at nl, is BPlatin(f) = dlPln Lent) + Pa Leta). @827) by Disrte Time Random Walk: The more clase for of he random walk Aes nt atu ta the an makes his jump tothe eto igh according o 8 tse quton baht he upset cr gh with equal probability at times Ne. 0 that tine is adr variable In this case, we ca wee fetes Lei ney+ Pon tavedenD. ‘aa28) Pon N+ DyrlatN) Iris small, we can view (3.8.27) and (3.8.28) as approximations to each other by writing Poni + Ariat N's) = PlnNe lat N'r) + rOPCn.tln' A). 829) with ¢= Naf = N’randd = $1"! so that the transition probability per unit time in the master equation model corresponds to half of the inverse waiting time r in the discrete time model. © Solutions Using the Characteriatic Punction: Both systemscan be eaily solved by introducing the characteristic function (5.0 = (1) = E Plein 6.830) in which case the master equation gives AG(s,0) = dle" +e" -2G(8.0), G83 and the diserete time equation becomes G(s,(N + Dr) = He? +P IG(S,NA) 3.832) ‘Assuming the man starsat the origin n’ = Oa time r= 0, we find Gs,0) 3.833) in both cases, so tha the solution to (3.8.31) is xpliel” +e ~ 2d), 3.834) Gus.)703, Markov Processes and 10 (3.8.32) Gys,N2)=[He" +e)" 3.8.35) ‘The appropriate probability distributions can be obtained by expanding G\(s,NT) and G(5.1)n power of expis): we find Py(n.t}0.0) =e Fy(4td), 3.8.36) yt amy (Mea (mle HS MT eam “The discret time distribution is also known asthe Bernoulli distribution: it ives the probability of a total of» heads tossing an unbiased coin W times. {Continuous Space Limit: For both kinds of random walk, limit of continuous space—that is, very many steps of very small size—gives the Wiener process. If we set the distance travelled as. Pan,Ne10,0) xenl, 838) so thatthe characteristic function ofthe distribution of x is, O1(5,1) = Ce) = Gulls.) = expite” +e — 2yed 8.39) ‘Then the limit of infinitesimally small steps 1 > Ois (5,1) > exp-s"1D), 8.40) where D= lima. san, “This isthe characteristic function of a Gaussian (Sect. 2.8.1) ofthe form P(x, t10,0) = (4xD1)""expl—7/4D0, 842) sand is of course the distribution for the Wiener process (Sect 3.8.1) oF Brownian ‘motion, as mentioned in Sect. 1.2. Thus, the Wiener process can be regarded as the limit ofa continuous time random walk in the limit of infiitesimally small step size. The limit 1047-40, wih D=Hgt? 0 De Hig. sayy ‘of the discrete time random walk gives the same resul. From ths form, we see clearly the expression of D as the mean square distance travelled per unit time. ‘We can also sce more directly that expanding the right-hané side of (3.8.27) as a function of x upto second order in J gives G,p(xst10,0) = BddEPCx,t10,0). 84s) ‘The three processes are thus intimately connected with each other at two levels, namely, under the limits considered, the stochastic equations approach each other and under those same limits, the solutions to these equations approach each other, ‘These limits are exactly those used by Einstein. Comparison with Sect. 1.2 shows 38 Examples of Markov Processes 71 that he modelled Brownian motion by a discrete time and space random wal nevertheless, derived the Wiener process model by expanding the equations for time ‘development ofthe distribution Function. “The limit results of tis section are a slightly more rigorous version of Einst! method. There are generalisations of these results to less specialised situations and it isa fair statement to make tat almost any jump process has some kind of limit which isa diffusion process. However, the precise limits are not always so simple, and there are limits in which certain jump processes become deterministic and are governed by LLiowvile's equation (Sect. 35.3) rather than the full Pokker-Planck equation. These results are presented in Sect. 11.2 138.3 Poleson Process. ‘We have already noted the Poisson process in Sect. 1.5.1. The process in which elec- trons arrive at an anode or customers arrive ata shop with a probability per unit time “2 ofariving, is governed by the master equation for which Wont Winn = 2, 3.845) ‘otherwise, Worm =o. 6846) ‘This master equation becomes PCr th nf) = APC — Wot ett) = POL thn, @san and by comparison with (3.8.27) also represents a “one-sided” random walk, in ‘which the walker steps tothe right only with probability per unit time equal 10 2 “The characteristic Function equation i similar to (3.8.31): 8,G(5,0) = Alexptis) ~ NGLS., 3.8.48) with the solution G{s,t)= exp {arfexptis) — M. 849) for the initial condition that there are initially no customers (or electrons) at time = 0, yielding [eee 1) Continuous Space Limit: In contrast to the random walk, the only timit that exists is 1+ 0, with alee, (3852) held fixed, andthe limiting characteristic Function is3, Markov Processes Timfexplae!® — 1] = expttes), 8.8.53) ‘withthe solution p(x,110,0) (x a), 854) ‘We also see that in this limit the master equation (3.8.47) would become Liowvile’s ‘equation, whose solution would be the deterministic motion we have derived, ») Asymptotic Approximation: We can do a slightly more refined analysis. We ‘expand the characteristic function up to second order ins in the exponent and find Hs (ls.t) = expleios ~ #°D/2)]. 855) where, as inthe previous section, D=Pa 38.56) ‘This is the characteristic function of a Gaussian with variance Di and mean of, $0 that we now have i ap 8 st10,0)> Ghee o( = ) 857) Its als clear that this solution is the solution of A,p5110,0) = v8, 110,0) + SDAA. 110.0), 858) ‘which i obtained by expanding the master equation (3.8.47) to order P, by writing Pun = 1.110,0)= apt 1.110,0), = Aplx.110,0) = 128,p4.110.0) + $PADF 10.0). 3.8.59) However, this isan approximation or an expansion and not a lieit The limit + 0 «tives Liouvile's equation with the purely deterministic solution (3.8.54). Effectively, the limit + 0 with well-defined corresponds to D = 0. The kind of approximation {just mentioned is a special case of van Kampen's system size expansion which we {neat fully in Sect. 11.2.3. 38.4 The Ornstein-Uhlenbeck Process All the examples so far have had no stationary distibution, that is, as # + oo, the distribution at any finite point approaches zero and we see tht, with probability one, the point moves to infinity If-we add a linear drift term to the Wiener process, we have a Fokker-Planck equation ofthe form, ‘Ap = dutkxp) + $Dap, 6.860) 38 Examples of Markov Processes 73 where by p we mean p(x. 11 29,0). This is the Omstein-Uhlenbeck process (3.6) 8) Characteristic Function Solution: The equation for the characteristic function 43) = Fe ptt. 0dr, R61 1 dpa 38.62) sé “The method of characteristics can be used to solve this equation, namely, if ee ee 4.8. we o seas pac hf BED ghey rae anes ‘The particular integrals ae readily found by integrating the equation involving, dt and ds and that involving ds and dé; they are W(s,t.9) = sexp(-k). 8.66) 165,140) = dexp(Ds"/4k) 867) “The general solution can clearly be put inthe form v = g(u) with g(u) an arbitrary fonction of u, Thus, the general soli 8,0) = exp(-Ds* /4k\gl sexp(—ki)] . (3.8.68) petordey cot PAx,0 | 0,0) = 6x = 30). 3.8.69), a 8, 0) = explixos), (3.8.70) ocr @{s) = exp(Ds?/4k + ixos). G.8.71) = frm S028 arent 0 rein (X(0) = zoexpt-kt), (3.8.73) wate Bt -et-20) a774 3. Markov Processes b) Stationary Sotution: Clearly, as # > co, the mean and variance approach limits 0 and D/2k, respectively, which gives a limiting stationary solution. This solution cean also be obtained directly by requiring d,p = 0, so that p satisfies the stationary Fokker-Planck equation aaftsos 500.0] = 875) and integrating once, we nd ee |esvs S00] =0 876) ‘The requirement that p vanish at ~co together with its derivative, is necessary for normalisation. Hence, we have Spx 2 perp (38.77) sot 7 nin= V&en(-~ eam Ts a Gaon wih mean O nd vince DI, as pte tome ine {dependent solution Its clear that a stationary solution can always be obtained for a one variable system by this integration method if such a stationary solution exists. Ifa stationary solution does not exist, this method gives an unnormalisable soluion. ©) Time Correlation Functions: The time correlation function analogous to that ‘mentioned in connection with the Wiener process can be calculsied and is a mea- sutable piece of data in most stochastic systems. However, we have no easy way of ‘computing it other than by definition (XOX) | Leosol) = J Fade, x12 pCa te 428120640), 8.79) and using the Markov property S fdadanxy rapist x2. 8)pt22. 5120010). 8.8.80) ‘om the assumption that 1352 Gssn ‘The correlation function with a definite intial condition is not normally of as much interest as the stationary correlation function, which is obtained by allowing the system 10 approach the stationary distibution. It is achieved by putting the intial condition in the remote past, as pointed out in Sect.3.7.2. Letting -> —2e, we find tim Pra 5 |405t0) = ryan) = (D/A)? exp(—kx3/D) (3.882) and by straightforward substitution and integration and noting that the stationary ‘mean is zero, we pet 38 Examples of Markov Processes 75, croton = oat = Bessa «assy ‘This result demonstrates the general property of stationary processes: that the cor relation functions depend only on time differences. It is also a general result (3.7] thatthe process we have described in this section is the only stationary Gaussian, ‘Markov process in one real variable. ‘The results ofthis subsection are very easly obtained by the stochastic differential equation methods which will be developed in Chap. 4 ‘The Ornsiein-Uhlenbeck process is a simple, explicitly representable process, which has a stationary solution. In its stationary state, itis often used to model a realistic noise signal, in which X(2) and X(s) are only significantly coreated if bess Vker. asa) ‘More precisely, 7, known as the correlation time can be defined for arbitrary pro- cesses X(s) by Jaane, xy, ae. 3.8.85) ‘hich is independent ofthe precise functional form of the correlation function. 3.85 Random Telegraph Process ‘We considera signal X(t) which can have either of two values a and b and switches {rom one tothe other with certain probabilities per unit time. Thus, we have a master equation Plat 819) = ~APCa,t1 X40) + #PCD. FL 0), | asso A.P(b.t1 8510) = AP(a,t1 x10) ~ wP(b,t1 400). 18) Time-Dependent Solutions: These can simply be found by noting that Plast XQ) 4 P(D«tL Sto) = 1, (3.887) and using the intial condition POX tol ato) = Bex (2888) ‘A simple equation can then be derived for AP(a,11x,f0)— P(b, txt), Whose solu tion is PCa, x10) ~ HP(D.1t0) = exp| ~ (2+ HVE to] (Aba2— Hrs) (3.8.89) ‘The solution for the probabilities then takes the form rasa) gig eestern| ta (3.8.90) = -casynr-an {4 ee “The mean of X() is sraightforwardly computed:7h 3. Markow Processes (XCD Lao. tol) = E xP Koto) = EPA ctennn| wa ® 7 ‘The variance can also be computed bul is a very messy expression, 1b) Stationary Solutions: This process has the stationary solution chained by letting anh matt) 89) ” a Pa tpg 3. ere Py = = 3.8.92) Which is obvious from the master equation ‘The stationary mean and variance are ws ba wy, = EB 9 (X), ern 3893) (a=brna vary, = Coed IX), aw (38,94), «) Stationary Correlation Functions: To compute the stationary time correlation function Ie > sand write MONE = Ex’ Plt ha’, IPs), (3.8.95) LeU DAC 6895) Now ase .8.91-2.8.94) to obtain ORONO. = ODE engl + Kt IKE) — OD. eam fay + ba y (a-b¥pa expC + pone = 9 =e 3 (SEER!) tote Hence (XU Nas = (MDX), — C9? = HEHE soins 8.899) (ene Notice that this time correlation function is of exactly the same form as that of the ‘Ornstcin-Uhlenbeck process. Higher-order correlation Functions are not the same of ‘course, bat because ofthis simple correlation function and the simplicity of the two sate prooess, the random telegraph signal also finds wide applicatca in model build 4. The Ito Calculus and Stochastic Differential Equations 4.1 Motivation In Sect 1.2.2 we met for the first time the equation which isthe prototype of what is now known as a Langevin equation, which can be described heuristically san ‘ordinary differential equation in which a rapidly and irregularly lutuating random function of time [the term X() in Langevin’ original equation} occurs. The simplic- ity of Langevin’s derivation of Einstein's results isin itself sulficient motivation to attempt to pat the concept of such an equation on a reasonably precise footing. “The simple-minded Langevin equation that turns up most often can be written in the form dx BS aus.) bu nec 4.1 GF = ats. Be 6) at) where «isthe variable of interest, af) and D(x) re certain known functions and 40) is the rapidly fluctuating random term. An idealised mathematical formulation fof the concept ot “rapidly varying, highly iregular funetion” is that for ¢ # .£(0) and £(7’) are statistically independent, We also require (&()) = 0, ince any nonzero mean can be absorbed into the definition of af, 2), andl thus require that cena) = 1-1). (4.12) which satisfies the requirement of no conelaton at different times and furthermore, has the rather pathological result that &(¢) has infinite variance. From a realistic point fof view, we know that no quantity can have such an infinite variance, but the con- cept of white noise as an idealisation ofa realistic Auctuating signal des have some ‘meaning, and has already been mentioned in Sect 1.5.2 in connection with Johnson noise in electrical cireuits. We have already met two sources which might be consid- cred realistic versions of almost uncorrelated noise, namely. the Ornstein-Uienbeck process and the random telegraph signal. For both of these the second-order corrla- tion function can, up to a constant factor, be put in the form cana = Fem" a9) [Now the essential difference hetween these two is thatthe sample paths ofthe random telegraph signal are discontinuous, while those of the Ornstein-Uhlenbeck process fare not. If (4.11) isto be regarded as a real differential equation. in which €() is not white noise with a delta function correlation, but rather a noise with a finite7% 4, Thelto Coleus and Stocheste Ditferetil Equations ‘correlation time, then the choice ofa continuous function for &() seems essential 10 make this equation realistic: we do not expect dx/dt to change discontinuously. The limit as 7 — oo of the corelation function (4.1.3) is clearly the Dirac delta function ie w2 and fort. h Lay Yew sim Ze (415) ‘This means that a possible model ofthe £(7) could be obtained by taking some kind of limit as —» oof the Orsten-Ulenbeck process. This would corespond, inthe notation of Sect.3.8:4,t0 the limit k— co with D ‘This limit simply does not exist. Any such limit must clearly be taken after ca culating measurable quantities. Such a procedure is possible but co cumbersome to use asa caleulational to! ‘An alternative approach is called for. Since we write the differential equation (4.1.1, we must expect it tobe integrable and hence must expect tat a 15) ‘Suppose we now demand the ordinary property of an integra, tht u(t) is a contin tious function off This implies that u(7) is a Markov process since we can write vr) = fey face. wan = tgl Féren) «farce as anor nye» Othe (mht nega aiden of the i the cond integral. Hence, by continuity w#) and w(’) ~ ut) are statistically independent and Further. u(')~n(2) is independent of u(?”) forall 1” < 1 This means that u(t) fully {determined (probabilistically) from the knowledge ofthe value of u(t) and not by any past values. Hence. w() is a Markov process. ‘Since the sample functions of u(t) are continuous, we must be able to describe u(t) by a Fokker-Planck equation. We can compute the drift and diffusion coeficients for this proces by using the formula of Set 3.52. We cn tite (utr At) — ual uot) = ("P etsids) = 0, (419) and utr + 1) oto AD Pas'Paveioes. (4.1.10) fds’ Paves —sy= as au 42 Stochastic tegration 79 ‘This means that the deft and diffusion coefficients are (u(r + Af) ~ uot) Al.) = fi, x =o, arp shag. = fy Am iD Aa “The corresponding Fokker-Planck equation is that ofthe Wiener process and we ean Jett = ut) = We) (44.14) “Ths, we have the paradox thatthe integra of €() is (A), which i self not df= Ferentiable, as shown in Sect. 38.1, This means that mathematically speaking, the Langevin equation (4.1.1) does not exist. However, the corresponding integral equa tion = ¥0)= falotatde flat. sktads, Lis) ‘ean be interpreted consistently. ‘We make the replacement, which follows directly from the infepretation of the integral of €() asthe Wiener process WW(), that aaa) = Were dr) ~ We) = Eide 4116) and thus write the sond integral as frtx).sld Ms) (aay which isa kind of stochastic Stiles integral with respect to a sample function Wi) Such an integral can he defined and we will carry this out inthe next section Before doing so, it should be noted that the requirement that w(t) be continu cus, while very natural, can be relaxed to yield a way of defini fs stochastle differential equations. This has already been hinted atin the ff shot noise in Sect 5.1, However, it does not seem to be nearly so useful and will ‘ot be treated in this book. The interested reader is referred to [4.1] ‘As a final point, we should note that one normally assumes that £(1) is Gau sian, and satisfies the conditions (4.1.2) as well. The above did not require tl {Gaussian nature follows in fet from the assumed continuity of u(t). Which ofthese assumptions is made i, in a siict sense, a matte of tase. However, the continuity of 1) seems a much more natural assumption to make than the Gaussian nature of &(), ‘which involves in principle the determination of moments of arbitrarily high order 4.2 Stochastic Integration 4.2.1 Definition ofthe Stochastic Integral Suppose Gt) is an arbitrary function of time and Wr) isthe Wiener process. We define the stochastic integral Ge" \4W() asa kindof Riemann Sti neg80 4, The Mo Caleuls and Stochastic Dilfer Equations HALAL ig... Paitonng of the time interval used inthe defntion of stockasic negation Namely, we divide the interval tt] into subintervals by means of pattoning points (asin Fig. 4.1) WS Knee Se a2 and define intermediate points such that hi ens 422) ‘The stochastic integral {5 G(e)/¥ (is defined as iit ofthe partial sums S42 S Gerona) Woot 429) 10s heuristically quit easy to se that. in general, the ietegral defined asthe limit ‘Of S, depends on the particular choice of intermediate point . For example, if we take the choice of Gis) = Wir). Sa) =( i Wert) - Wee.) (424) = Eminent) ~ mints 425) Eev-no. 426) 1 forexample, we choose fr all Heat t(l-ainr 0. “The proof is quite straightforward. For N= 0, let us define84 4, Theo Calculus and Stochastic Differential Equations in ([56-s(aW2—a1))) 4228) Jim (3G cow? — an? +3 26,6 a ~a1p (aw? a} 423) ‘The horizontal braces indicate factors which are statistically independent of each ‘ther because ofthe properties of the Wiener process, and because the Gare values ‘of a nonanticipating function which are independent ofall AW; for j> i. Using this independence, we can factorise the means, and also using (AW) = an, fi) (AW? - An)? we find (from Gaussian nature of AW), 42 Stochastic megrtion 85 and although G(1) is nonanticipating, this is not suficent to guarantee the inde- pendence of AW, and Gj.» as thus defined. iv) By similar methods one can prove that im 3G, AW (4.2.38) fawyat awe a} and similarly for higher powers. The simplest way of characterising these reslls is to say that dW() isan infinitesimal order of $ and that in calculating diferental, infinitesimals of order higher than I are discarded. 4.2.7 Properties of the Ito Stochastic Integral 4) Existence: One can show that the Ito stochastic integral G(r}dW() exists ‘whenever the function Git’) is continuous and nonanticipating on the closed interval 1=24m[Zato.0")] 4220) ta) Cece eeepc ory b) Imegration of Polynomials: We ean formally us theres of See. 4.26 mi( 2.6.14? 26.184) =0, aan aqweor =p + ance - wo = $(")weor awe 6239) and since ee mim EG. = fare). 42m) = niveyr amine A Daveor ar, (4240) we have sothat flamorau = farce 239) fwerawey= oiwot wages 5fmerar, aan Comments 1) The proof {2 GintdWeP™™ = 0 for N > is similar and uses the explicit expressions for the higher moments of « Gaussian given in Sect.2.8.1 ii) dWV( only occurs in integrals so that when we restrict ourselves to nonantic pating functions, we can simply write aw? i, (4.2.34) aWi*™ 0, (N>0) (4235) ili) The results are only valid for the lo integra, since we have used the fact that AW, is independent of Gs. In the Stratonovich integral AW, = WHE) Ml. 4236) Gir = G(4ur+ 4-0). 4.237) <6) Two Kinds of Integral: We note that foreach G() there are wo kinds of integrals, namely, found’ and jouw), 4242) both of which occur in the previous equation. There is, in general, no connection between these two kinds of integral 4) General Differentiation Rules: In forming differentials, as in (b) above, one ‘must keep all terms upto second order in dWV(7. This means that, for example, dlexplW(s)}} = exp[W(n) + dW(1)] ~ explWi0)], (4.2.43) = exptwWen)lawen + 4awer'), (4248) = exptWonifaiven + $a] 4.245)86 4, Thelto Calculus and Stochastic Diferental Equations Foran arbitrary function 12 ae « Lanne LOL io. 12 Lane» Lawes 12 Liaw? anwo.n SEL an? + Lawn 3 2Laweoy * ZL awe 4.246) sd we use amor a, 247) draw) +0, [Set.426, comment (iv) (4.248) (ay? 0, (4.249) and all higher powers vanish, to atv at af VPS), apenas = (2 LBL) a i41We0 ( ¢ “i a+ Lawn 14250) ©) Mean Value Formula: For nonantcpaing G0, (jomane)=0 as) Proof: Since G(t) is nonanticipatin (26.1001) = E66) MAM) = 0. (4.252) in the definition of the stochastic integr ‘We know from Sect.2.9.5 that operations of ma-lim andl ¢ ) may be interchanged Hence, taking the limit of (4.2.52), we have the result. ‘This result is notte for Stratonovieh’s integral, since the value of G1 is chosen inthe middle ofthe interval, and may be correlated with AW;. eee funetions, (fowraweeyjmeyawer) 4253) Proof: Notice that (EGE H)10W,) = (SGeaaacamy?) + EGeaths +6108. pawyaw) 4254) 42 Stochastic Imepraion 87 In the second term, AW; is independent of all other tems since j < i. and G and ‘Hare nonanticipating. Hence, we may factorise out the term (AW) = 0 so that this term vanishes. Using (AW?) = As,, (4255) nd interchanging mean and limit operations, the result follows 4) Relation to Delta-Correlated White Noise: Formally, this is equivalent to the idea that Langevin terms €) are deta corelated and uncorrelated with F(?) and Git), For, rewriting awn) > end, (42.56) itis clear that if F(t) and G() ate nonanticipating, £1 is independent of them, and wwe get jar Jaeicuyms ee’ =f faras Gwyn es). = farGwne’y, (4257) which implies {E10 619) = 501-9) (4258) [An important point of definition arises here, however. In integrals involving dla funtion it frequently occurs inthe study of sochasti differential equations that the argorent ofthe delta function i equal to either the upper or the lower limit of the mcg thats, we ind integrals like 1 =farpnee-n), (4259) b= Farsnse~). (4.260) ‘Various conventions can be made concerning the value of such integrals. We will show that inthe present context, we must always make the iaterpretation ha ft), 4261) h=0, (4.2.62) corresponding to counting all the weight of a delta function atthe lower limit of an integral, and none ofthe weight atthe upper limit, To demonstrate this, note that (Jowawery| FAs yawis))) = 0. (42.6%) This follows, since the function defined by the integral inside the square bracket {s, by Sect.4.2.5 comment (v), a nonanticipating function and hence the complete88 4, The to Caleuus and Stochasti Dilferemial Equations integrand, {obtained by multiplying by G(r’) which is also nonantcipating} is itself ‘nonanticipating. Hence the average vanishes by the result of Sect. 4.2.7e. [Now using the formulation in terms of the Langevin source £(), we can rewrite (4.2.63) as fae fds Gurynes nse — (42.68) ‘which corresponds to not counting the weight ofthe delta function atthe upper limit. Consequently, the full weight must be counted atthe lower limit. This propery is a direct consequence of the definition of the Ito integral as in (4.2.10) in which the increment points “towards the future". Tha i, we can interpret awe) = Wier dd) ~ WO. (42.65) Inthe case of the Stratonovich integral, we get quite a different formula, which is by ‘no means as simple to prove asin the Ito case, but which amounts 0 choosing he hf), 4see) “his means that in both cases the delta function occurring atthe Hinit ofan inte al has half its weight counted. This formula, although intuitively more satistying than the Ito form, is more complieated to use especially in the perturbation theory of stochastic differential equations, where the Ito method makes very many term vanish, {Stratonovich) (42.66) 4.3 Stochastic Differential Equations (SDE) We canetnded in Sect. 4.1, that the most satisfactory interpretation ofthe Langevin equation Ho act) 4b x B= ax. + a.0€0). aa) Js a stochastic integrat equation = HO) = fala £1 fee Bete). 432) Unfortunately, the kind of stochastic integral tobe used isnot given by the reasoning of Sect. 4.1. The lo integral is mathematically and technically the most satisfactory, but it is not always the most natural choice physically, The Straonovich integral is the natural choice for an interpretation which assumes £(r) isa real noise (not a white noise) with finite corelation time, which is then allowed to ecome infnites- mally small after calculating measurable quantities. Furthermore, a Stratonovich interpretation enables us to use ordinary calculus, which is not possible for an Tio interpretation, 443 Stochastic Difernial Equations (SDE) 89 nee pee arora Fig. 42. lusaton ofthe Cauchy-Euler procedure for consiruting an approximate solution ofthe stochasti iferenil equation dx) = ax). + BL. tldW(0) From a mathematical point of view, the choice is made clear by the near impossi- bility of carrying out proofs using the Strtonovich integral. We will therefore define the Ito SDE, develop its equivalence with the Stratonovich SDE; and use ether form depending on circumstances, The relationship between white noise stochastic differ tential equations and the eal noise systems is explained in Sect. 8. 43.1 Ito Stochastic Differential Equation: Definition ‘A stochastic quantity x() obeys an Ko SDE written as a(t) = alx(, fdr + HEx(,1dWO), (43.3) if forall rand fo, att) + fala) fae + fole?)s"awe!) aaa) Before considering what conditions must be salisfied by the coethcients in (4.3.4), itis wise to consider what one means by a solution of such an equation and what ‘uniqueness of solution would mean in this context. For this purpose, we can consider a discretised version ofthe SDE obtained by taking a mesh of points (as illustrated in Fig. 4.2) such that ch encore Sar Stat 435) and writing the equation as wr = 314 aC NAN + Ba AW 436) Here, eats ANS teats 437) AW, = Wer) = WOH)904, The to Calculus and Stochasie Difeental Equations 18) Cauchy-Euler Construction ofthe Solution of an Ito SDE: We see from(4.3.6) that an approximate procedure for solving the equation isto calculate x; from the knowledge of x; by adding a deterministic term ast. (438) and a stochastic term Dai. AAW; (439) ‘The stochastic term contains an element AW;, which isthe increment of the Wiener process, but i statistically independent of 3 if {i xs itself independent of al W(r) ~ W(t) fort > ta (hus, the initial conditions if considered random, must be nonanticipating), and ii) a(x.0) is a nonanticipating function of ¢ for any fixed x. Constructing an approximate soltion iteratively by use of (4.2.6, we ee that xis always independent of AW, for j > i ‘The solution is then formally constructed by leiting the mesh size goto zero, To say that the solution is unigue means that fora given sample function ¥() of the random Wiener process W(1), the particular solution ofthe equation which arises is unique. To say thatthe solution exists means that with probability one, a soltion exists for any choice of sample fonction W() ofthe Wiener process WC. “This method of constructing a solutions called the Cauchy- Euler method, and can bbe used to generate simulations. However, there are significantly better algorithms, asis explained in Chap. 10 >) Existence and Uniqueness of Solutions of an Ito SDE: Existence and unique- sess wll not be proved here, The interested reader will find proofs in {4.3}. The conditions which are required for existence and usiquenes in a tne interval {7 ip Lipschitz condition: aK exists such that lax.) = ay.) + 16C4. 1) = bys < Kix al, (43.10) forall xand y, and all # inthe range 0.7. ii) Growth condition: a K exists such that forall inthe range (4,7), late, + Ibo. < 120+ [Pp au Under these conditions there will be a unique nonanticipating ‘olution x(t) in the range {to.T Almost every stochastic differential equation encountered in practice satisfies the Lipschitz condition since itis essentially a smoothness condition. However, the {growth condition is often violated. This does not mean that no solution exist; rather, ‘it means the solution may “explode” to infinity, that i, the value of x can become infinite ina finite time; in practice a finite random time. This phenomenon occurs in ‘ordinary differential equations, for example, 42 Stochastic Differential Equations (SDE) 91 [=har (43.12) ete asthe general solution with an inital condition x = x at = 0, a= ars tidy? ain I ais postive, this becomes infinite when xo = (a1)! but if a is negative, the solution never explodes, Failing wo satisy the Lipschitz condition does not guarantee the solution will explode, More precise stability results are required for one to be certain ofthat [4.3]. 4.32 Dependence on Initial Conditions and Parameters Inexactly the same way as in the case of deterministic differemial equations, ifthe functions which occur in a stochastic differential equation depend continuously on parameters, then the solution normally depends continuously on that parameter. Sim- italy, the solution depends continuously on the initial conditions. Let us formulate this more precisely. Consider a one-variable equation dx = aQd.x,Ddt + WA, )dHO, aay ‘with initial condition to) = (A), anus) where 2 isa parameter. Let the solution of (4.3.14) be x(1.1) Suppose (to). 4316) i) slime) ii) ForeveryN>0 sim { sop, [lcd ald.s.01+ Bea, x.0)~Bda.x08]} = 0. ve nt — iil) Thee exists aK independent of such that laa, x07 + 19 < AAC + be), 3.18) Then, slim {sup 0a. ~ to, ni} = 0 (63.19) ti sup ba.) ~ x.) 4319 Fora proof see [4.1] ‘Comments 5) Recalling the definition of stochastic Timi, the interpretation ofthe Himit (4.3.19) js that as > do, the probability that the maximum deviation over any finite inerval[f, T] between x(2,1) and xd, 1) is nonzero, goes to zero.92 4, The to Calculus and Stochastic Diferental Equations ii) Dependence on the initial condition is achieved by eting a andb be independent of a ii) The result will be very useful in justifying perturbation expansions iv) Condition (ji) is written in the most satural form for the case thatthe functions atx,t) and b(x,t) are not themselves stochastic. It often atises that a(x.) and Dx.) are themselves stochastic (nonanticipating) functions. n this case, condi tio (i) must be replaced by a probabilistic statement, Ii, in fact, sufficient to replace fim by slim 43.3 Markov Property of the Solution of an Ito SDE ‘We now show that (1), the solution to the stochastic differential equation (4.3.4). is a Markov Process. Heuristcally, the result is obvious, since with a given initial condition x), the future time development is uniquely (stochastically) determined, tha is, a(7) for ¢> fg is determined only by ') The particular sample path of W() fort >to; ii) The valve of s(t). Since x() is a nonantcipating function ofr, W(t) for # > ty is independent of xe) for 1 fg, Thus, the time development of x(t) for t > ta is independent of x(1) for t< tp provided .%p) is known. Hence, 1(f) is a Markov process. For a precise proof see (43. 43.4 Change of Variables: Ito's Formula Consider an arbitrary function of x1): f[2(2)). What stochastic differential equation does it obey? We use the resulls of Sect. 4.2.6 to expand d [x0] to second order in awn: Aflac) = flat + dat) ~ fl), (43.20) SUatoldat + LPL da? + (4321) =F Latnifalatn, ride + Lat), naw} + FECL Faw? 43.22) where all other terms have been discarded since they are of higher order, Now use dWin? = dt obvain flat) = (atx, AF Lx] + FLU AF FLX] de + BEC ALF LC CO, 3.3.23) This formula is known as Io's formula and shows that changing variables is not given by ordinary calculus unless fls(7)] is merely linear in x(). {42 Stochastic Differential Equations (SDF) 98 Many Variables: In practice, Io's formula becomes very complicate and the eas iest methodist simply use the multivariate form ofthe rle that dW() is an in fntesimal of order 4. By similar methods to those used in Sect. 4.2.6, we can show that forann dimensional Wiener process W(). AWA aWAD = 6d", (43.240) aw? = 0, (> 0, (43.240) awynd = 0, (43.240) an eo) (43.248) svhich imply that dW san infitesimal of order. Note however that 4.3.24) inaconsequence othe independence of dW?) and dW (). To develop o's feral for functions of an n dimensional vector xt) satisfying the stochastic differential equation dx = Acx pdt + Bex WO, 43.25) ‘we simply follow this procedure. The result is pee) = [EAs fx +} $1B(x 08% x00] +E Bylx NOs dW An), (4.3.26)| 4.35 Connection Between Fokker-Planck Equation and Stochastic Differential Equation a) Forward Fokker-Planck Equation: We now consider the time development of an arbitrary f(a(). Using to's formula daslsto _ (afi _ a, eon. ( iz TAF). = (ost. nh + $8L0.0°22f) 3.2m However hs condos probability density xt) an Sto = fdr fats to. = fase, + ps. 0PARp pts tt) 43.2% This is now of the same form as (3.4.16) Sect. 3.4.1. Under the same conditions as, there, we integrate by parts and discard surface terms to obtain Sasfisyayp = fdxflo| — delacx. op) + 4OLx pl}. (4.329) and hence, since f(x) is arbitrary,94 4 Thelto Calewas and Stochastic Diferental Equations 4px 11 Xoo) = Balas, pC%tx0.0)] + $02B(%.1% pCa #19. fo). 4.3.30) ‘We have thus a complete equivalence toa diffusion process defined by a drift coef cient a(x.) and a diffusion coeficient (x, ‘The results are precisely analogous to those of Sect. 3.5.2, in which it was shown that the diffusion process could be locally approximated by an equation resembling. an lo stochastic differentia} equation, ») Backward Fokker-Planck Equation—the Feynm: function g(x,2) obeys the backward Fokker-Planck equation Ag = ~atx,0.9 ~ 4614. 020. (43.30 with the final condition Ma.T) = GU) (43.32) Ifa() obeys the stochastic differential equation (4.3. then using Ho's rule (adapted appropriately to account for explicit time dependence), the Function glx), t} obeys the stochastie diferental equation dtd. = [Bg + alae dyglat e+ 4bLAC. AF Boleto.t dr + BLx.t1d,glx(st1 dW, 433) and using (4.3.31) this becomes dglxtt), | = BLx(t), 14,914, 1 dW0). (4.3.6) [Now integrate fom to T, and take the mean daaern.ry abana = (Folarneyaatarnflamn) =0. 4338) Let the initial condition of the stochastic differential equation fora) and 1” = 1 be xO) = (4336, ‘where « is a noa-stochastic value, so that dale) ) = 96.0) (aaa) At the other end ofthe interval, use the final condition (4.3.32) 10 write dalX7).T) = GUT at) = 2), (4.3.38) ‘where the notation on the ight hand side indicates the mean conditioned om the initial condition (4.3.36), Putting these two together the Feynman-Kac formula results: (GL TMI() = 2) = 91.0 (43.39) where g(x) isthe solution of the backward Fokker-Planck equation (4.3.31) with initial condition (4.3.32). ‘This formula is essentially equivalent tothe fact that p(x.t1s9,t) obeys the back- ward Fokker-Planck equation inthe arguments xo. as shown in Sect 3.6, since GGIs¢T Isto) = x0) = fdxGCIPCR,T L010). (4.3.40) 443 ‘Stochastic Dtferemial Equations (SDE) 98 4.36 Multivariable Systems ‘In gencral, many variable systems of stochastic differential equations can be defines for n variables by dx = Ate dt + BE. NAWO), (asaly where W(t) is an n variable Wiener process. as defined in Sect.38.1. The many variable version of the reasoning used in Sect. 4.3.5 shows thatthe Fokker-Planck equation forthe conditional probability density p(x. tl 30.0) =p is ap = ~ZalAde.opl+ $ A9l(@x. 9B". hip) aan Notice thatthe same Fokker-Planck equation arses from all matrices 8 such that BBB? is the same, This means that we can obtain the same Fokker-Planck equation by replacing B by BS where S is orthogonal, ic., SS" = 1. Notice that S may depend onx(0. “This can be seen directly from the stochastic differential equation. Suppose St) is an orthogonal matrix with an arbitrary nonanticipating dependence on t. Then define avin = Sinawe (4343) Now the vector €V(1 isa linear combination of Gaussian variables dW() with cocf- ficients S(t) which are independent of dW(7), since S(t) is nonanticipating. For any fixed valve of Sit, the dV(°) are thus Gaussian and their correlation matrix is LaVO) AVA) = ESAS pl (AWAD AWD) = ESS Ode = by dt, (4am) since S(0 is orthogonal. Hence, all the moments are independent of S() and are the same as those of dW'(), $0 d¥() is itself Gaussian with the same corelation matrix. as dW(0). Finally averages at different times factorise, for example, if > fin Fllawansyinr"tawceysuer. (43.45) we can factorise ot the averages of d'Vi() to various powers since dW, is inde- pendent ofall other terms. Evaluating these we will find thatthe orthogonal nature of St) gives after averaging over d1¥(t), simply Zalawyor cate sule a> (43.46) ‘which similarly gives [dWOI"{aW(7)1". Hence, the dV(0) are also increments of 4a Wiener process, The orthogonal transformation simply mixes up different sample paths of the process, without changing its stochastic nature. Hence, instead of (4.3.41) we can write (x, dr + Bex. NSTNSIDAWE, (aan Ales dt + Bla, NSO. (4.348) and since V(0) is itself simply a Wiener process, this equation is equivalent t096 4. The to Calculus and Stochasti Diferenil Equations dx = Atx rar + Box, NSTDAWi, (4.3.49) which has exacly the same Fokker-Planck equation (4.3.42). ‘We will return to some examples in which this identity is relevant in See. 4.5.5, 4.4 The Stratonovich Stochastic Integral ‘The Stratonovich stochastic integral is an alternative tothe Io definition, in whieh Nto's formols. developed in Sect.4.3.4, is replaced by the ordinsry chain rule for ‘change of variables. This apparent advantage does not come without cost, since in Siratonovich’s definition the independence of a non-anticipating integrand Git) and the increment d¥7() in a stochastic integral no longer holds. This means tht inere iment and the integrand are correlated, and therefore to give a full definition of the Stratonavich integral requires some way of specifying wha this cerrelation is This correlation is implicitly specified inthe situation of most interest, he ease in “which the integrand is a function whose stochastic nature arses fram its dependence ‘on a variable (0 which obeys a stochastic differential equation. Since the aim is 10 recover the chain rule for change of variables in a stochastic differential equation, sonable reteotion 44.1 Definition of the Stratonovich Stochastic Integral Stratomovich [4.21 d vurvand ry ined a stochastic integral ofan integrand which isa function of (5) fate 1aer im 8 GL (st + Mia))sfen}E Win) — WED) an “The Suratonovich imegal i clearly related to a mid-point choice of; inthe deti- sition of stochastic ion as given in Sect, bul eleary is nor necessarily ‘equivalent to that definition. Rather, instead of evaluating x atthe midpoint $i) the average of the values athe two time points s taken. Furhemrore its only the dependence ons) thai averaged inthis way. and no the explicit dependence on 1 However, if G(s differentiable inthe integral canbe shown t be independent ofthe patiular choice of value fort the range [ft 44.2 Stratonovich Stochastic Differential Equation Iis possible to write a stochastic diferential equation (SDE) using Stratonovich's integral Wi = says faded, 1468) Fawerstat?r) 442) 44 The Stamonavich Stochastic Ingral 48) Change of Variables for the Stratonovich SDE: The definition ofthe Straton= variables, This means, thal forthe Stratonovich integral, Ho's form replaced by the simple calculus rule (44) in} = Fat fat. rll + ato, 114We0 ‘This can be proves! quite simply (rom the definition (4.4.1). The essence ofthe proot ean he explained by using the simple SDE. (S)duin = Bian] dwen. aan In discreised form, this can be written tie) = 8 BLL ey + 89] Mien = WD. (445) ‘To ind the Stratonovich SDE for fLA(} we need only use the Taylor series expan- sion a a faction about a migpoat i he form cape 3 Emde! Ln Seen In expanding fC) we oly need to keep lerms up 10 second onde. so we drop all tut the fst wo tems and write jeeray 446) FUsined = fa) + Fea + x1 — 5+ aan = FL EGon + xd] BLS Csien + x0) Whos ~ Wh (448) “This means thatthe Stratonovich SDE for fLa(A] is (Spdflan} = fet] Banda, 449) ‘which is the ordinary caleulus rule, The extension to the general ease (4.4.3) is, straighitorward, 'b) Equivalent Ho SDE: We shall show thatthe Stratonovieh SDE isin fac equiva Tent to an appropriate Ho SDE. Let us assume that x7) isa solution of the lio SDE x(t) = af. fdr + Lx.) dW). (4.4.10) and deduce the and fora corresponding Stratonovich equation ofthe form (44.2) In both cases, the solution (7) isthe same Function, We first compute the connection between the Ito integral fdWW4r yiatr).} and the Stratonovich integral (S) dW" Blxtr). 71 (5) fawer tate). = Eal§ ata + ar0))tr]AW0.0. ain nA.) we wie ses) + Atti). ain and use the Ito SDE (4.4.10) to write xt68 4, TelloCaeus an Stott DieretlEgptons Ante) = aso. at tes + Bliss) ay “Then, applying Ho's formula, we can write [30 + athd)ea] = Batten + Sana. 1) + fle. tad $M] +} bl) BE AVAWE). (44.14) (For simplicy, we write) et, insted of fiat)] wherever posible). Pating all thse back inthe original equation (4.4.10) and dropping as usual di? d-dWor, and setting dW(0? = dr, we find (5) f= Satta WEI) = Wee M4 HY BO VB BANG = tea) Hence we derive (5) Feber 1dWir) = feted). }dWO+ $f 6L). 118, BLd yar (44.15) ‘This formula gives a connection between the Io and Stratonovich integrals of func- Lions flat") 1]. in which x(7’)is the solution ofthe Ito SDE (4.4.2. Itdoes not give a ‘general connection between the Io ané Stratonovich integrals of arbitrary functions. 1 wo mow make the choice (x,t) = atx.) XA bLN, A) Bn = x0 4416) we ee tha theo SDE dx=adi+ bdo). 517 is the same asthe Strafonovich SDE (S)de=(a~ 4o0,8)dr + b4W0), 41 and conversely. the Strtonovich SDE (Syax=aarspawe, (44.19 is the same asthe Ko SDE r+ 400,404 Bawey 4420 445 Some Examples and Solutions 99 ‘¢) Many Variables: Ifa many variable To equation is x dt + Bux.naWin, aan ralonovich equation can be shown similarly 1 be given by de then the corresponding S replacing A = A $5 BOB, B-8 (44.22) 1d) Fokker-Planck Equation: Corresponding to the Stratonovich SDE, (S)de = Ax nd + Bir. nd WO) (4.4.23) ‘we can, by use of (4.4.22) ané the known correspondence (Sect. 4.3.6) between the Ito stochastic differential equation and Fokker-Planck equation, show thatthe equiv alent Fokker-Planck equation is Op =~ Zalaiph + $F GABA yp. (4.4.24) ‘whieh is often knowin asthe “Stratonovich form” of the Fokker-Planck equation. Ia contrast to the two forms of the stochastic differential equation, the two forms of Fokker-Planck equation have a differen appearance but are (of course) interpreted ‘with the same rules—those of ordinary calculus. We will find later that the Stra e Fokker-Planck equation does arise very naturally in certain 4.5 Some Examples and Solutions 4.5.1 Coefficients without x Dependence The simple equation dx = athdr+ KOdWe, as.) ‘with a(t) and b(0) noaranddom functions of ime, is solved simply by integrating Mt) = so fattyaf + fe) de) 432) Fete, x» can be either nonrandom inital condition or may be random, but must be independent of (1) ~ Wht) for > fy: otherwise, x) is not nonanticipating ‘As constructed, s() is Gaussian, provided is either aonrandom or itself Gaus-1004, The ko Callus and Stochastic Ditfeweil Equations Soe awe), (453) is simply a linear combination of infinitesimal Gaussian variables. Further, (xi) = (aa) + fat ydr 454) (since the mean ofthe Ho integral vanishes) and ain ~ (LA) ~ GND # CD. 269) as = varlaol + ( Foie yaWer) {6(s.4M59)). 456) = arb) + Fou’ yP ar 450 where we have wed the result 4.2.53) with, however We), fet . we), es. “The process is thus completely determined. 4.5.2 Multiplicative Linear White Noise Process—Geometrie Browni Motion ‘The equation dx = cra) (45.10) is known as a multiplicative white noise process because it is Hinear in x but the ‘noise term” Wr) mauliplies x. Is lsu cossmwnly know as geomerie Brownian ‘We can solve this exacly by using Ito's formula, Let us define a new variable by (5.1) dW(t)~ $e dt 45.12) “This equation can now be directly inepated, so we obtain wht) = wlan) + ef Wit) — Wro)] — $c° = to), (4.5.13) and hence. A) = xinhexp etW¥e0 ~ Wea} = $e2~ m0} ss) 49) Mean value: We can calculate the mean by using the formula for any Gaussian variable with 2ero mean 4.5 Some Examples and Solutions 108 exo (423) 4515) so at (at) = anenp|J2U= = $20 1) = (x06) est “This resulis so obvious rom deft, since dein) = axn)= xed = 0 asin 1b) Autocorrelation Funetion: We can also calculate the autocorreation function atria = Cnt) (exp [el # Wis) = 2Weml— Jee s—2)]) Cara?) exp [LEU + WC) = 2WeayP) — 45 2s (ate) exp | $I = Aig + 2mings)= (r+ 6~ 20} = (alto? expe? min(~ io. = 0) 145.18) «© Stratonovich Interpretation: ‘The solution of this equation interpreted 0s a Stra- tonovich equation can also be obtained, but ordinary calculus would then be vali. ‘Thus insted of (4.5.12) we would obain Say = cay (45.19) and hence, st) = ntpoxpet WO) ~ Wee) (4520) Inthis case xo) = Gxonexn [$2 —0)} (3520 and dsc) = Cato) exp {Jet +s 2+ 2 mint ~ fp. Tol} (45.22) ‘One sees that there isa clea difference between these two answers Frequency 4.33 Complex Oscillator with No “This isa simplification of a model due 10 Kubo [4.4] and isa slight generalisation of the previous example for complex variables. We consider # ci(ur Wran)e, 14523) a ‘which formally represents a simple model of an oscillator with a mean frequency a perturbed by a noise term £0). Physically, this is best modelled by writing 8 Strtonovich equation (de =ifods Vrawa))e, which is equivalent tothe Ho equation (from Sect. 4.4)102 4, The fo Calelos and Stochastic Differential Equations Fig. 43. Mostration ofthe decay of the mean arp tude o complex oxcillatorof as «result of dephasing, [Gar yas i Vp amen (4.525) ‘Taking the mean value, we see immediatly that a 2 eo tiw- Ko (45.26) with the damped oscillatory solution (30) = expla ~ yketO}) 527) ‘We shall show fully in Seet.8.3, why the Stratonovich model is more appropriate ‘The most obvious way to see this isto note that £(0) would, in practice, be someshat smoother than a white noise and ordinary calculus would apply, ass the case in the Stratonovick interpretation [Now in this cas, the correlation function obtained from solving the eriginal Stra {onovich equation is (OP Yexpltias = ye + 5) ~ 2y mint, (45:28) Fo eiwith ere + netn) =0. (45.29) However. the correlation function of physical interest isthe complex correlation Mexpliads ~ 9) +i V21WE) ~ WoT) (COP yexptivg® ~ 8) ~ylr + s— 2mints 9), = (210) expiants — 9)— yr ~ sl (45.30) ‘Thus. the complex correlation function has a damping term which eises purely from the noise. I may be thought of as a noise induced dephasing effect, whereby the phases of an ensemble of initial states with identical phases diffue away from the value wf arising from the deterministic motion, as illustrated in Fig. 4.3. The mean of the ensemble consequently decays, aldhough the amplitude ir} of any mernber ‘of the ensemble is unchanged. For lage time differences, z(2) and 2s) become in- dependent 45 Some Examples and Solutions 19% [A realistic oscillator cannot be described by this model of a complex oscillatcr. ‘as discussed by van Kampen [4.5]. However the qualitative behaviour is very si Jar, and this model may be regarded as a prototype model of oscillators with noisy frequency, ‘Taking the FokkerPlanck equation given for the Omstein-Uenbeck process in Seot.3.84, we can immediately write down the SDE using the result of Sec. 4.35 de = ~kedt + VDAWin), (45.31) and solve this directly, Puting vase’, (4332) then dy = (dxjd(e") + (dx)e" + xd(e") = Lavar+ Daweotel drs [-ardrs WDawen] ky ot 45a) ‘We note that he fst product vanishes involving only a and dV Gin fh, an be seen hat tis wil always happen if we simply mulily shy # deterministic funetion of time). We get dy= Ve away, us) so that integrating and resubsttuing for y, we get xi we + VBS eH aWer) 143.35) vty IF the intial condition is deterministic or Gaussian distributed, then o() is eh Gaussian, with mean and variance (x) (oye, (4.5.36) varlatn) = ([Ex0)~ (ope + VBj eS? awe)’) 45.37) “Taking the inital condition to be nonanticipating, that is, independent of dW for 15-0, we can write using the result of Sect. 4.41 vari stn aixle™ + fe a = (varl(0)} = D/2K) e+ D/2K. (4538) “These equations are the same as those obtained directly by solving the Fokker-Planck. equation in Sect. 3.8.4. with the added generalisation of a nonantcipating random initial condition. Added to the fact that the solution is a Gaussian variable. we also have the correct conditional probability.1044, Theo Callus and Stochastic Differential Equations The time correlation function can also be ealeulated directly nd is, at.) = vary}! 4+ DE Fee™—Paweny fee-M awe), = var Oe DP tetra! ar) fratwo-2 ase) Notice that if & > 0, as ¢.5 -» 00 with finite { ~ s, the correlation Funetion becomes. Stationary and of the form deduced in Sect. 3.8.4 In actif we set the initial time at -s rather than 0, the solution 4.5.35) becomes t= NB f otra), san in which the corcelation function and the mean obviously asstene thei stationary values. Since the process is Gaussian, this makes it stationary, 48.5 Conversion from Cartesian to Polar Coordinates, ‘A. model often used to describe an optical fcld is given by & pair of Ornstein- Unlenbeck processes describing the real and imaginary components ofthe electri field i. MEW) = ~yE Md + ed (0, (Sata) dE) = ~yEAdt + edWy), (4s.41b) is of interest to convert to polar coordinates. We set Exit) = alnreoseXty, (45.42) auysin gt), (45.43) and for simplicity, also define w= (45.44) so that iO) + iC) = loglEs() + ECO) (assy We then use the ho ealeuls to deve UE) iE3) (dE + EP Ei vies HE +B? HEL +E) 4, sid) +idWoin] dW + iaWeen? Brie Evie) EE diy +i) = 45.46) and noting dW(0) dW) = 0, and dW; (1? = JWa(1)? = dt, it can be seen that the fast cerm vans 445 Some Examples and Solutions 108 iat) + i0)] = ~ydt + exxpl—ue ~ i@cNldW LC) + dW) (asany ‘We now take the real part, set at?) = expt?) and using the Ito ealeulus find dao =( at + 5) de oa yeosein +a} (4548 “Te imaginary pr yes ta = E(-amnsina + ayes) as) We now define dW.A1) = dWy{1)cos dt) + dWoi0) sin de}, } (45.50) AW) = -dY 0 sing) + WACO Cos ‘We note that his is an orthogonal transformation of the kind mentioned in Sect. 4.3.6, so that we may take dW,(?) and dWa(?) as increments of independent Wiener pro- wesses Wi) and Wel). Hence. the stochastic differential equations for phase and amplitude ane ein = oat (455i) dat) = (~ you + sla + ed Wait (4551b) ta) = (to 3a en ‘Comment. Using the rules given in Sect. 4.4 (i) itis possible fo convert both the Cartesian equation (4.5.41a, 4.5.41b) and the polar equations (4.5.51a, 4.5.51) to the Stratonovich form, and to find thet both are exactly the same as the lo form, Nevertheless, a direct conversion using ordinary calculus is not possible. Doing so wwe would get the same result until (4.5.47) where the term [e? 2a()] df would not be found. This must he compensated by an extra term which arises frm the fact that th ‘Stratonovich increments d¥(0) are correlated with fr) and thus, dW,(0) and dW.) ‘cannot simply be defined by (45.49). We se the advantage of the Io method which retains the statistical independence of dW(0) and variables evaluated at time the equations in Polar form are not soluble, as the corresponding Cartesian equations are, There isan advantage, however, in dealing with polar equa- tions in the laser, whose equations are similar, but have an aed term proportional toate} drin @SS1b). 45.6 Multivariate Ornstein-Ublenbeck Process we define the process by the stochastic diferental equation ax) (A and B are constant matrices) for which the solution is easily obtained (as in Sect. 4.5.4: Ax(dt + BAW), (45521064, The te Calculus and Stochastic Differential Equations 219) = expl-AnxiO) + Jexpl-Au~ IBEW) (45.53) The mean is (210) = expt-ANK0). 43s) The correlation function follows similarly Ate). xT) = (lett) ~ Cee Mlats) ~ Gets) expt-AratO). 20) e4ph-As) +f exk-Att— 1B expl-AT st 455s) ‘The integral can be explicitly evaluated in certain special cases, and for particular low-dimensional problems, its possible to simply multiply everything out term by term. In the remtainder we set (x10), (0) = 0), corresponding to a deterministic initial condition, and evaluate a few special eases a) The Case AAT = ATA: In this case (for seal A) we can find a unitary matrix S such that sst SAS* = SATS? = diag dy... 2s) (45.56) or simplicity. assume ¢ > s, Then tat.a(s) = S1GU.95 assy where BN [Gt shy fexpl-Ait — s) expla a9). (4.558) b) Stationary Variance: Ii has only eigenvalues with positive real par, a station- ary solution exists of the form ain f expr ryBawery (45.59) We have of course teln) <0. (45.60) (a). 163)) = "TF expl-Ale— C188" expl-ATs~ Pha? sen Let us define the stationary covariance mattis. by ia.aT 45.02) ‘This can be evaluated hy means of an algebraic equation thus: 45 Some Examples and Solutions INT Aer brat f Aexpl-A@— BB" expl-AT= 11d 4+ J expt-Atr— 19188 exph-aT AT at td Aaa t 45060 = f Lrexpl-a— #188" expla — hdr 56% CCarying out the integral, we find that the Tower limit vanishes by the assumed pos: tivity ofthe eigenvalues oF A and hence only the upper limit remains. gv Al = BBE (45068 as an algebraic equation forthe stationary covariance matrix. © Stationary Variance for Two Dimensions: We note that ifA isa 2% 2 mates. it ie equation satisfies the charac AP = (TeAA + (Det A) = 0, (45.65) ‘and from (45.60) and the fact that (4.5.65) implies exp(—A1) is a polynomial of degree | in A. we must be able 1 write /BBT + piABET + BBTAT) + yABBT AT 45.066) Using (45.65), we find (4.5.68) i satisfied if + (TRAY (Det Ay = 0, (467) 2ptDetA) +1 =0, (45.68) Beta o. (45.69) rom which we have = (DANA + [A = CTEANBBEA — CTA] er (TE ANDet Ay wary State: From the solution of 4.5.60) Matrix in the Stati 4) Time Correlation wwe see that if >; (.00,4148)) =expl-Ate~ 9) f expl-Ats PBB expl-As— PA. empl-Auasile, 1s 45.7ta and similarly. scenpl-AUs-n]. res casiby “This depends only on 5 ~ 1, as expected of a stationary solution. Defining then Gut 9) = (20,2119). asm©) Spectrum Matrix rather simple, We define similarly to Seo. 1.5.2: em 44 The lo Caeulus and Stochastic Ditferetial Equations wwe see (remembering o = 0) that Galt 9) = 1s =F (45.73) in Stationary State: ‘The spectrum marry jwens out to be 1 stay= Ef eG arn, (57) ite ° = sgl {ewl-os arrears f expla +A ea 4575) ! fy +t ~ ia! ENA tive eta iw) 4576) Hence (A+ iw)S KA" ~ io) 4637 and! using (4.5.64), we get t Ste) = 5214 + iw BBN AT — wy! (45.78) £ Regression Theorem: The result (4.5.71a) is also known as @egression theorem in that it states that te time development G7) is for T > O gover ‘of time development of the mean, as in (4.5.54). Iisa consequence of the Marko ian liner nature ofthe problem, The tie derivative ofthe stanary corelation by the same law d a : GlGlMde = FeadraTOar, = ((rAn (ride + BaWes)h a2 (0) (45.19) Since r > 0, the increment We) is uncoreated with 43 (0) this means that a FIG Un) = AG) (4.5.80) Thus, computation of G.(r) requires the knowledge of Gy(0) = o and the time de- velopment equation ofthe mean. This results similar to those of Sect. 3.7.4 45.7 The General Single Variable Linear Equation 8) Homogeneous Case: We consider firstly the homogeneous case dx = (bind +a) dWenx, assy and using the usual lo rules, write 445 Some Examples and Solutions 109 y= tog x. (15.82) so that ty 8 Ec nate gain Sarak, 458%) sing and inverting (4582). wea st) = oven} f r)— Satta! + foramen}. asap = W090 (45.88) which serves to define 0) ‘We note that using 4.5.15) Ate = eesoor (exp fed drat +» feutvawer)y = (ext xp [nf va + n= foo ar (45.86) ’) Inhomogeneous Case: Now consider lit) vx + LAO + gtOHT AO. (4587) and write (45388) st = atte. ‘with 0) as defined in (4.5.85) and a solution ofthe homogeneous & ‘Then we write ation (4.581), dz = dxigint! + xdlginy'] + dedloy] (4.5.89) Noting that d{gin]-! = -dednto]? + [doy HOT? and using Ho rules, we find d= = (late) — finginyldt + find wonnecy! (4.5.90) ‘which lire integrable, Hence, the solution is 11) = po) foe Mla) — Fear Lae + FPA) son ©) Moments and Autocorrelation: It is beter to derive equations for the moments from (45.87) rather than caleulate moments and autocorrelation ditecty from the solution (4.5.91). or we have ast) = nat" Mdatd) + nln Dato" EdstOP = airy date) + Jno = Hcy PLAC + ataCOP de (45.92)110 4, Theo Caleulus and Stochastic Difeental ay Renee Levy = (xeyInbe9 + $rtn~ gto"). + (ay [natn) + mtn DyfloMnT + da") ft = Df (45.93) ‘These equi fom a hearhy in which he ih equation inlet saa ons a ihe pons te, andcanbelepat sean 488 Multivariable Linear Equations 2) Homogeneous Case: The cation dni) = [Bindrs Soeoawso ec, 4594 where B().G,() are matrices. The equation i no, in general, soluble in closed form, ‘unles all the matrices (1) Ge) commute at all times with each other i. GANG Wr) = GArIGAD, BING.) = Gut'yBU (45.98) BiNBE) = BBL) In this ease, the solution is completely have alogous to the one variable case and we xu Bx0). (4596) with 1 = emp fa $ ZG. asm }) Inhomogeneous Case: We can reduce the inhomogeneous case to the homoge: ‘neous case in exactly the same way as in one dimension, Thus, we consider dx) = (AC) + Bln], + SIFU +Ganslawen, (4.598) and write m= oy 'xe), (45.99) where wis a matrix solution ofthe homogencous equ to evaluate dfs"), For any matrix M we have MM! onder, MdM™'] + dM"! + dMalM~) = 0, Hence, d[M-"] = =[M + dMf)""dMt Mand again to second order ar jon (4.5.64). We frst have |: $0, expanding to second MUM Ms Md aM, (45.100) 45 Some Examples and Solutions an thus, since i) satisfies the homogeneous equation, ayer) = wor Ben + LEAN] de EGAnaHan} 45.100 ana again aking diferenials ait = varia 3. Ganka + EReoamo) (5.102) Hence. x) o{o0 + four IA) = LGM EMA + srcwmeny : (45.103) when the solution forthe “This solution isnot very useful for practical purposes, even whe homogeneous equation is known, because ofthe difficulty in evaluating means and ‘correlation functions. 4.5.9 Time-Dependent Ornstein-Ulenbeck Process , hich is sob ‘This is a pacar ease ofthe provous gener iar eation whi Iisa gentsaton othe mutvarte Ornstin-Unenbeck process (Set. 4.5.6 1 includ time-dependent parameters, namely x(a) = Anetra + BNW) (45.106) “This is clearly ofthe same Form as (4.5.98) withthe replacements Ao =) 0: BO, + AW. me SRnawin + KNAW, Gan, SC ‘The corresponding homogeneous equation is simply the deterministic equation dx( = ~Aunxtna (45.106) which is soluble provided A(DA() = A()A(A) and has the solution x) = WO), 45.107) with (45.108) wo = exp| = facet Thus, applying (45.108) at) =exp|-fArdf]ntoy+ Flexo[ = facadslfacryawiry. (65.109)— 124 The to Coteus and Stohaate Dilfer Equations This very similar tothe solution of the time-independent Omstin-Uenbeck “1 i cess, a derived in See, 4.5.6, equation (45.53. . ee ae From this we have Grip mexp[- faurrae|ixioy. (45.110) ns) e0| ferrari, x09] fake at + faresp| - facnsfoerw ave|-fatons, cas.i1y The time-dependent Ornstein-Uhlenbeck process will arise very naturally in connec ‘ion with the development of asymptotic methods in low-noise systems, In the next two chapters, the theory of continuous Markov processes is developed from the point of view of the corresponding Fokker-Planck equation, which gives the lime evolution ofthe probability density function forthe system. This chapter is devoted mainly to single variable systems, since there are a large number of exact results for single variable systems, which makes the separate treatment of such syS- ‘ems appropriate, The next chapter deals with the more general multivariable aspests ‘of many ofthe same issues treated one-dimensionally inthis chapter. ‘The construction of appropriate boundary conditions is of fundamental impor tance, and is carried out in Sect.5.1 in a form applicable to both one-variable and ‘many-variable systems, A corresponding treatment for the boundary conditions on the backward Fokker Planck equation is given in Sect. 5.1.2. The remaining of the chapter is devoted to a range of exact results, on stationary distribution functions, properties of eigenfunctions, and exit problems, most of which can be explicitly solved in the one variable case, ‘We have already met the Fokker-Planck equation in several contexts, starting from Einstein's original derivation and uso ofthe diffusion equation (Sect. 1.2). again asa particular case of the differential Chapman-Kolmogoroy equation (Sect, 35.2), and Finally, in connection with stochastic differential equations (Sect. 4.3.5), There are many techniques associated with the use of Fokker Planck equations which lead 10 results more directly than by direct use ofthe corresponding stochastic differential ‘equation; the reverse is also true. To obtain a full picture of the nature of difusion processes, one must study both points of view. ‘The origin of the name “Fokker-Planck Equation” is from the work of Faker (1914) [5., 5.2] and Planck (1917) [5.2] where the Former investigated Brownian ‘motion in a radiation field and the latter attempted to build a complete theory of ed on it, Mathematically oriented works tend (0 use the term “Kol- ‘mogorov’s Equation” because of Kolmogorov's work in developing its rigorous basis [5.3]. Yet others use the tem “Smoluchowski Equation” because of Smoluchowshi ‘original use ofthis equation, Without in any way assessing the merits of this termi nology I shall use the term “Fokker-Planck equation” as that most commonly used by the audience to whom this book is addressed1145, The Fokkee Planck Equation 5.1 Probability Current and Boundary Conditions ‘The PPE is second-order parabolic partial differential equation, and for solutions ‘we need an initial condition such as (5.2.5) and boundary conditions atthe end ofthe Jnwerval inside which xis constrained. These take on a variety of forms. Ics simpler to derive the boundary conditions in general, than to restrict consi ion tothe one variable situation, We consider the forward equstion a a atest) = ~¥, EAdesple. + $Y 5 Ble. inte) = ~¥ Bade peat) + $B Poe Byte) 1) We note that this ean also be writen aptzt)| y @ 12 EY oy Fen =0 12) where we define the probability current a Adz. D2.) ~ FE Bile... 1a dies Equation (5.5) has the form of a focal conservation equation, and can be written in an integral form as follows. Consider some region R with a boundary S and define oe i eS PAE. asm sen. ne Where mis the outward pointing normal to S. Thus (5.1.5) indicates thatthe total loss ‘of probability is given by the surface imegral of J over the boundary of R. ss a Fig 8.1 Repions used to demonstrate tha! the probability current ithe Row of probability ‘We can show as well thatthe current J does have the somewhat stronger property, that a surface integral over any surface S gives the net flow of protability across that surlace, For consider two adjacent regions Ry and Ro, separated bya surface Sja. Let 'S; and S2 be the surfaces which, together with Siz, enclose respectively Ry, and Re (see Fig. 5.1) 51. Probbility Curent and Boundary Conditions 115 “Then the net flow of probability can be computed by noting that we are dealing here with a process with continuous sample paths, so that na sufficient! short time [Ar the probability of erossing Six from Ry t Ry isthe joint probability of being in Ry atime rand Ry, atime t+ Ai, = fide f dy poss + Ayo) (5.16) “The net flow of probability from Ry to Ry is obtained by subtracting from this the probability of erossing inthe ceverse direction, and dividing by At i. Bayh fe fairer aor) in ja fysy.)=0 18) since this isthe probability of being in Ry and R; simultaneously. Thus, we can write sara (5.19) AD) = fd f dy Oe post. — de Py, and using the Fokker-Planck equation in the form (5.5) D8 [EWE Ma tReds (6.1.10) where Ji. Re) formed from port: Rad) Ln) dy pt inthe same way as J(z,¢) is formed from plz, in (5.1.3) and Jy: Rais defined similery, We now convert the integrals to surface integrals. The integral over Sz vanishes, since it will ivolve p(x, 1:0), with x notin Rj or on its boundary (except fora set of measure zero.) Similarly the integral over Sy vanishes, but those over Siz donot, since here the integration is simply over pat of the boundaries of Ry and Re ‘Thus we find, the et flow from Rp t0 Ry is f dS mE RIs0) + Det Ray (1.12) s and we finally conclude, since x belongs the union of Ry and thatthe net flow of probability per unit ime from Rs to Ry tig © pdx f dota. + arn. ~ pt t=.) | wimp a 0 omy f dsm Hen),N65. The Fokker: ek Equation 5.1.1 Classification of Boundary Conditions ‘We can now consider the various kinds of boundary condition separately, 4) Reflecting Barrier: We can consider the situation where the particle cannot leave ‘a region R, hence there is zero net flow of probability across, the boundary of. Thus we require nS) =0, forzeS, = normal 1S. (Lay where Jz.) is given by (5.5.4) Since the particle eannot cross, it must be reflected there, and hence the name ‘reflecting barrier for this condition. b) Absorbing Barrier: Here, one assumes thal the moment the particle teaches, it is removed trom the system, thus the barrier absorbs. Consequertly the probability ‘of being on the boundary is 210, i. We=0, forzes. S15) ©) Boundary Conditions at a Discontinuity: It is possible for both the A; and Bi, ‘coefficients to be discontinuous ata surface S, but for there o be free motion across 'S. Consequently, the probability andthe normal component of the current must oth be continuous across S, n- Stas, = m- Feels (5.1.16) Pals, = ples. 6.117) ‘where SS. a8 subseripis. mean the limits ofthe quantities from the left and tight hhand sides of the surface. ‘The definition (5.1.3) of the curren necessarily continuous a. indicates tha the derivatives of p(z) are not 5.1.2 Boundary Condit ms for the Backward Fokker-Planck Equation ‘We suppose that p(x. 1x1’) obeys the forward Fokker-Planck equation fora set of 2:1 and x’, and that the process is confined to a region R with toundary 8. Then, ifs isa time between t and 7, a oye i = Frat) = 5 dy pari. np slx' 0), (5.1.18) where we have used the Chapman-Kolmogorov equation, We take the derivative 0/9 inside the integral, use the forward Fokker-Planck equation forthe second factor and the backward equation for the first factor. For brevity, lel us write DY.) = Ply.sla) PS) = platy.) Then, 6.119) 552 Fokker Planck Equation in One Dimension 117 5 5, Oe 1.20) “fool 458 Be and afer some mnpation 2 Apps ty |p2 coup) vt se (tay ove 2 {anos 33 [pho peStl} (5.1.22) eafolane peg} ap pene ma, ee orale oper creer tr ‘on substituting p= On that equation, we get 2p 45.123) = 3.450 22 Cachet 7 However if he boundary i absorbing. clearly patlys)=0, for ye boundary 124) since this merely states that the probability of X re-entering R from the boundary is, 220. b) Reflecting Boundaries: Here the condition on the forward equation makes the first integral vanish in (S.1-22). The final factor vanishes fr arbitrary p only if 8 (a(x. 9) = 0 (5.1.28) Embargo In one dimension this reduces to 5.1.26) a Gyre = unless vanishes. ©) Other Boundaries: We shall not consider these this section. For further details see [54] 5.2 Fokker-Planck Equation in One Dimension none dimension, the Fokker-Planck equation (FPE) takes the simple form ED Brace geno) Slow 0/000) 62 Th Seuis.24, 35 the Fokker-Plonck equation was shown to be valid fo the con tional probability, that i, the choice118 5. The Fokker Planck Equation Fa. plat toto), 622) for any initial o,o, and withthe initia condition PA fo fo) = OU ~ x4) (52.3) However. using the definition for the one time probability PLD) = fo pox 29,40) = fd pst to.t0)pl2aato)s 524) ‘we see that i is also vaid for pt.) with the initial condition PLD, = Plato), (525) hich is generally less singular than (5.2.3). From the result of Sect. 4.3.5, we know thatthe stochastic process described by a ‘conditional probability satisfying the FPE js equivalent tothe to stochastic differen. tial equation (SDE) dat) = Alstn.t]dr + BGR AaWe), 626) and that the two descriptions are to be regarded as complementary to each other. We will see that perurbation theories based on the FPE are very diferent from those based on the SDE and both have their uses. 52:1 Boundary Conditions in One Dimension ‘The genera formulation of boundary conkitions as given in Sect. 5.1.1 can be aug. ‘mented by some more specific results fr the one-dimensional ease, 4) Periodic Boundary Condition: We assume that the process takes place on an interval [a bin which the two end points are identified with each ober. this occur, for example, if the diffusion is on a circle). Then we impose boundary conditions derived from those for a discontinuity, ie. 1: dig poe = lim tx. (527) We isp dea) = fim J.) 628) Most frequently. periodic boundary conditions are imposed when the functions ‘A(s.1) and B4x.0) are periodic onthe same interval so that we have Alb.1) = Ata.t), Bb.) = Blan), ‘and this means that I and Il simply reduce to an equality of px.) and its derivatives atthe points and b, ») Prescribed Boundaries: If the diffusion coeficient vanishes at a boundary, we have a situation in which the kind of boundary may be automatically prescribed Suppose the motion occurs only for x > a. If a Lipschitz condition is obeyed by AU 1) and VEC T atx = a Sect. 4.3.16) and Btx. 1) is differentiable at x= a then 0, Bia,0) = 0. 52.10) (5.29 52 Fokker Planck Equation in One Dimension 119 “The SDE then has solutions, and se may write dalt) = Ato + VEE AWO) In this rather special case, the situation is determined by the sign of AUx.1. Three cases then occur, as follows, (S21) i) Exit boundary. In this case, we suppose Ma.) <9. a will certainly proceed out of region so that if the particle reaches the point to.x 0 ca © 1 sign of A(a.t) is such a8 10 In this case i the particle reaches the point a, the sir return it to x > a; thus a particle placed to the right of « can never leave the region. However, a particle introduced at x = a will certainly enter the region Hence the name, “entrance boundary”. i) Natural boundary. Finally consider (52.14) Ata.) veritcan be demon- “The pail once itreaches = , wll ain hte, How {rae ha cannot ever each hs point Ths sa ound fom which we em neither abo nor at which we can odie any pies. ©) Feller’s Classification of Boundaries: Feller [5.4] showed that in general the boundaries can be assigned to one of the foe types: regular. entrance. exit and natu ral. His general eritria forthe classification ofthese boundaries are as follows. Define fix) = exp|-2 fdsatsy a1). 62.15) 2IBL ft}. 62.16) So faisdds. (52.17) an f Foods. 52.18) Heres € (2,0) and is fixed. Denote by 52.19) Foun). the space of all functions integrable on the interval (x, 82)120.5. The FokkerPtanck Equation Then the boundary at a can be classified as Ee Regular: if f(x) ¢ 2(a.xo) and glx) e Llaso) Me Exit if gay € Lea.) and f(x) € ta. 10) Mk: Entrance: it g(x) © 2a.) and Mat) € Zea, x9) IV: Natural: all other eases It can be seen trom the results of Sect.5.3 that for an exit boundary there is m0 ormalisable stationary solution of the Fokker-Planck equation, and that the mean time to reach the boundary, (5.5.24, is finite, Similarly, ifthe boundary is exit, a stationary solution can exist, but the mean time to reach the boundary is infinite, In the case of a regular boundary, the mean time to reach the boundary is finite, but & stationary solution with a reflecting boundary at a does exist. The case of natural boundaries is harder to analyse. The reader is referred to [5.4] fer a more complete description. ©) Boundaries at Infinity: All of the above kinds of boundary can occur at infin ity, provided we can simultaneously guarantee the normalisation of the probability which. it p(2) is reasonably well behaved, requires ° 5.2.20) If Bxphx) is reasonably well behaved (i.e. does not oscillate infinitely rapidly as x9 9), lim 8,p0s.4) = 0. 7 (52.21) 50 that @ nonzero current at infinity will usually require either 4¢x,1) or Blx,4) to ‘become infinite there. Treatment of such cases is usually best carried out by changing (o.another variable which i finite at x = oo, Where there are boundaries at x = 40° and nonzero curtents at infinity are permit ted, we have two possibilities Whieh do not allow fon loss of proba DJs, =0. (5.2.22) ii) Je409,) = S09.) (5.2.28) These are the limits of reflecting and periodic boundary conditions, respectively, 5.3 Stationary Solutions for Homogeneous Fokker-Planck ‘Equations We recall (Sect. 3.7.2) tha in a homogeneous process, the drift and diffusion coefi- cients are time independent. In such a case, the equation satishied by the stationary distribution is: a 1@ enn 3S tape a) 5.3 Stationery Solutions for Homogeneous Fokker-Planck Equations 121 which can also be written simply in terms of the curent (as defined in Sect. 5.1) we 832) ae ‘which clearly has the solution JU) = constant (ay ‘Suppose the process takes place on an interval (a,b), Then we must have Sta) = He) = 0) 8 J. 634) and i one ofthe boundary conditions is reflecting, this means that both are reflecting, and J : the boundaries are not reflecting (5.3.4) requires them to be periodic. We then use the boundary conditions given by (5.2.7) and (5.2.8). 2) Zero Current—Potential Solution: Seting J = 0, we write (5.34) as td Ava) = 5 H1Bla)p.t0)] = 0, (535) {for which the solution is exp|2f dx’ Aar)/B1x’) (536) pals) ae where isa normalisation constant such that an) Jasna 7 rical reasons, but sath estan known a pret slo, ovine aay te te sony ton tn! ty = gle megan el “Toc atom rien ine 622 Wea Boundary Conn: He We have za sont J ws ete hi Atnpala) = FBP = I 638) However, J is not arbitrary, but is determined by normalisation and the periodic boundary condition 7 recone ‘Then we can easily integrate (5.3.8) to get PAs) _ padBla) 9 F a in vay) wee)—— 1225. The Fokker Prank Equation By imposing the boundary condition (5.3.9) we find that 22) _ By i) wa |” I ae (83.13) Loa) ota jf BO) pa Boy Laceyea *! aero 10) = nly] ED oan) coho ©) Infinite Range and Singular Boundaries: In ether of these cases, one or the sither ofthe above possibilities may turn out to be forbidden because of divergences, ele. A full enumeration ofthe possibilities is, in general, very complicated. We shal demonstrate these hy means ofthe examples given in the next section. 5.3.1 Examples of Stationary Solutions 4) Diffusion in a Gravitational Field: A strongly damped Brownian particle mov ing in 2 Constant gravitational field is olten described by the stochastic diferential equation (8.2.15) dx = -gdt+ VDdwin, (5.3.15) for which he Fokker Planck equation i ® _ Bon, \poe 2. Diop tn@e (5316 (On the interval (a.) with reflecting boundary conditions, the stationary solution is given by (5.3), ie. Paix) = AF expl-29x/D 1, AI? ‘where we have absorbed constant factors into the definition of 1 Clearly this solution is normatisable on (a,b) only if aa is finite, though b may be infinite The result is no more profound than to say that particles diffusing in a ‘beaker of fuid will lall down, and if the beaker is infinitely deep, they will never stop falling! Diffusion upwards against gravity is possible for any distance bat with exponentially small probability. ‘Now assume periodic boundary conditions on (a,b), Substitution into (5.3.14) yields Pad pata), (53.18) ‘constant distribution, ‘The interpretation is that the particles pas frely from a to band back. 1) Ornstein Ublenbeck Process: We use the notation of Sect.3.8.4 where the Fokker Planck equation was 453 Stationary Solution for Homogeneous FokkerPlanck Equations 123 ras) Fig, $2, Nontomatsale “stoma p06 the ee ¥ thos Ae 2k 9 _ Beary s to® 6319) Pt Pa whose stationary slain on he intra ab) with reflecting ares x pals) = expl-kx?/D) (5.3.20) Provided & > 0, this is normalisable on (~0, 00). IK <0, one can only make sense of ton a finite interval. In this case suppose <0. 9320, 0 that from (5.3.11). Kigaa sore tee and if we consider the periodic boundary condition on this interval, by Ha) = wa) a wwe find that a3) wef Fue al. 532 patsy = tS ~ puarero |p ion as in the ease of reflecting barirs so thatthe symmetey yields the same solu the Letting a + 0, we see that we still ave the same solution. The result is also true ifa—+ ov independently of b + ce, provided k > 0. ©) A Chemical Reaction Models Although chemical reactions are normally best modelled by a birth-death master equation formalism, as in Chap. 11, approximate treatments are often given by means of a Fokker-Planck equation. The reaction X+Ae oN (53.25) = 0 (where xis the number of is of interest sine it possesses an exit boundary at x = 0 ( rat molecules of X}. Clearly if there is no X, a collision between X and A cannot occur s0 no more X is produced “The Fokker-Planck equation is derived in Sect. 1.6.1 and is apts) = ~8, (ax—P) piso] + $02 |(ax+ A) pes.9} oe1245. The Fokker Planck Equation We introduce reflesting houndaries at x solution is and x = B. In this case, the stationary Paap eMart 63.27) whieh is not normalisable if ¢ = 0. The pole at x = 0 is a resul ofthe absorption ‘there. Infact. comparing with (5.2.18), we see that BO.) = (art hg =0. AO.1) = (ax ¥)a0 = 0, (5.3.28) 8,810.1) = (04 28)e0g > 0. 80 we indeed have an exit boundary. The stationary solution has relevance only if, > 0 since itis otherwise not normalisable. The physical meaning of a reflecting ‘barrier is quite simple: whenever a molecule of X disappears, we simply add another ‘one immediately. A plot of psx) is given in Fig. 5.2. The time for all x to disappear 's in practice extraordinarily long, and the stationary solution (5.3.27) i, in practice, 8 good representation ofthe distribution except near x = 0. 5.4 Eigenfunction Methods for Homogeneous Processes ‘We shall now show how. in the case of homogeneous processes, slutions can most naturally be expressed in terms of eigenfunctions. We consider refecting and absorb ing boundacies, 5.4.1 Bigenfunctions for Reflecting Boundaries We considera Fokker-Planck equation fora process on a interval (a,b) with refle ing boundaries. We suppose the Fokker-Planck equation to have a sationary solution Pao) and the from O,p13.8) = ~ILACPL D+ YRLBDPL. A 64 We define a function gts.) by PED = BadgCH.D 542) and, by direct substitution, find that 42 satisfies the Backward equation Aglet) = Aas.) + $BEHAAC.1 (543) ‘We now wish to consider solations ofthe Form Pian = Pune’, (544) at) = Qaaye*, (645) which obey the eigenfunction equations ANACIPAAN] + $LBYP AL] = ~AP AX) (8.46) AGH, Qi(2) + $BOIBO eC) = =2" yx). 647) 5.4 Eigenfunction Methods for Homogeneous Processes 128 rom (514.2) and (5.4.3) i follows that (548) i) Relationship berween Py and Qu Paka) = pals}Qutx) “This simple result does not generalise completely to many dimensional situs tions, which are treated in Sect. 65. 4i) Orthogonality of eigenfumctions: We can straighttorwardly show by partial inte ration that ura farryoQ.09 [oren{-aunrions ga.tanr ceo} ~ $BL9PsC99. Geta 649) Using the reflecting boundary condition on the cocficient of Q(x), we see that, this ceficien vanishes, Further, sing the definition of g(t) a tems ofthe Stationary solution (5.4.2), its simple to show that $298, O08) = ALP) + $LBEOP (0. (54.10) 50 that term vanishes also, Hence, the Qu(x) and P(x) form & bi-orthogonal system falxPusvQr4a)= bie. 4.1) ‘Threats reo alte ohana systems, J dx pssyOua)Qn(0) = Sax (54.12) i (53) fasip.cor PAP eC) = bas onary i should be noted that setting 4 = 2° = O gives the normalisation ofthe solution py(x) since 44) Pa = pad. ou = 1 415) assuming comple: it) Expansion neigefntons Using this orthogonality and r ess) ne can write any ston interme of egenncions Foc if pint = EA, Pane, (5.4.16) then Faroe. (6417)1265. The Fokker Planck Equation ‘v) Conditional probability: For example, the conditional probability p(x] s0.0) is piven by the inital condition P2,0).%9,0) = dt x0), (54.18) so that Ay = Fede Q,2)6~ x9) = Orban). (54.19) and hence, Pact 10.0) = E PriQataaye" (54.20) Vi Autocorrelation function: We can write the autocorrelation fonction quite ele- any as ((OM(O)) = Fade Fro x20 plxt149, 09.00), 42 Z[Jacmcnfe® 42 where we have used the definition of Q4(x) by (5.4.5) SA2 igenfunctions for Absorbing Boundaries These are treated similarly. We define P and Q; a8 above, except that p(x is lll he stationary solution of the Fokker-Planck equation with reflecting boundary conditions, With his definition, We find that we must have Plo) = Qx(a) = Px(b) = Q41b) = 0, (5.4.23) {and the orthogonality proof stil follows through. Eigenfunctions are then computed using this condition andthe eigenfunction equations (5.4.6) and (5,7) and all other results look the same. However the same of 4 does not Include 1 = 0, and hence Lt 49,0) 0 as 1 > 60 54.3 Examples 2) Wiener Process with Absorbing Boundaries The Fokker Planck equation ap = ip. (5424) is treated on the itera (0,1). The absorbing boundary conitoe requires 100.9 = and the appropri 0. (5.4.25) igenfunctions are sin(nzx) so we expand in a Fourier sine series pian = ¥ bansininas, 6.426) ‘hich automaticaly saisies (5.4.25) The initial condition is chosen so that 54 Ejgentunetion Methods for Homogeneous Processes 127 (5427) for which the Fourier coeticients are 2,0) = 2 Favotx ~ xp)sin(unx) = 2 sin(nmay) (54.28) Substituting the Fourier expansion (5.4.26) into (5.4.24) gives ald = a,b, (5429) i =~). with Ae = (5430) and the solution lO) = beO) expla) aan) So we have the Solution [which by the initial condition (5.4.27) is for the conditional probability px.t1x.0)) Us t1a9,0) = 2 5, expl—aut)sin(mm) sin(as). (5.4.32) 'b) Wiener Process with Reflecting Boundaries: Here the boundary condition re duces to fon the interval (0, 1)] A,pA0.1) = BPC, 8) =O, (5433) and the eigenfunctions are now cos(nur), so we make a Fourier cosine expansion 140 +E ayltheosns), te ‘withthe same initial condition our a0) (54.35) (0) = 2d costnnsi(x— ag) = 20s. (5436) Inthe same way as before, we find 440) = a0) EXPK-2ats (saan) with ener (538) sothat plx.t\,0) = 1 +2 F costs) costo expl—2yt) (54.29) [As #0, the process becomes stationary, with stationary distribation128 5S. The Fokker-Planck Equation (5.40) an compute the stationary autocorrelation function by P.O), = ff dds plet a. 0100. (saat) and carrying out the integrals explicitly, las (nO = 5+ FE exmetananQn + 4 (5.442) We see that asr + c, all the exponentials vanish and l (ONO. > 5 = CO aaa and as t+ 0, Ae (tna, > 3+ 5 Zane ty (544d) when one takes sesount of the entity (rom the theory of te Riemann zeta a ae Eats % 5.485) ©) Ornstein-Unlenbeck Process: As in Sect. 3.8.4 the Fokker Panck equation is 4 plot) = AlkxpOO + $D8%px.0, (534.46) ‘The eigenfunction equation for Q is ene Qn tO + FOr 0, (5.4.47) tnd his becomes he diferent equation for Hermite peynomis ion or Hert pobmomils Ha) (5.560 making the replacement y = x VE7D 7 ee 4} Q1 ~ 24d, 01 + 211K), = 0. 5.4.48) Weean write Q) = tute '?H, (x VAT) « (5.4.49) where Aan. 6450) lutions are normalised as in (5.4.11-5.4.13), jonary solution is, as previously found, (ky) exp—kx?/D) 6.451) tnd a general solution can be written as Psst = Z VRE ARDH exr-k (DIK, (x YTB), (5452) 154 igentunction Methods for Homogeneous Processes 129 with Ag= J deploy ( #). Since the system is time homogeneous, we ean write plast13,0)= plO1x-0. 654) and the backward Fokker Planck equation ean be written ‘Aypla ts 0) = ALA PLA t}%0) + § BLD 12,0), (535) and hence, G0) obeys the equation B,GLx,1) = ALIGL.) + $BUIRGCO.1) 656) Initial condition: Clearly that plu’.O1x,0) = r=, (659) and hence, G0) eae 5.5.8) 0. elsewhere. iit) Boundary conditions: Wx = a oF b, the panicle is absorbed immediately, so ProbiT > #) =O when x= aorx = bie, Gla,t) = Gib, =0. (5.5.9) in) Moments of the exit time: Since Gl.) is the probability that Tf. the mean ‘of any function of Tis gery =F sasexn iss10 Thus heme eine oe ras ine) Tix) = (T) (5.5.11) beens ria) = =feaceena sin = Focund (5.13) stringy pte Sma ie Tax) = (T"), (5.5.14) wn ra0)= foun, 519 ¥) Differential equation for the mean exit ime: We can derive a simple ordinary dif ferenial equation for T(x) by using (5.5.13) and integrating (5.5.6) over (0,0) Noting that1525. The Fokker Planck Equation Ja, Gts.nd = G1x,00)~ 614.0 1 516) wederive AcoaATEN + $BUNETUA) (65.17) with he howndary condition To) = 10 6518) Simiay we se that ay) = ACT) + EBL) 6519) i Solin of he Equation: Bquton (517 an be volved on (S517 can be solved diel by ner ‘on The slo ater some mariplaion cane writen interns 419-09 {f artes} 5.20) eg _(f ta few Bo awit du5t me a 521) fe Jey 55.2 One Absorbing Barrier We omit motion sil inthe interval.) tat eral (0) bsp ihe air ber flecting. The boundary conditions then become a a Gla.) = 0. Gib.t) i ame (55.226) sich flow fom the candiiony om he backward Fokker Patch svar Fokker Manc eatin dived Seo 5.12, Weselve (3.17) wih the comespondng boundary candi amd in Falage sors Similarly, one finds reflecting, a absorbing, ae a © Po 7 wo \y \ no ; I oo ; | a 2 Zs Fig. 5.3. (a) Double well potential U(2;¢b Stationary dstiation p10; () Mean fs pas sage lime from at0 x, T(a +9) 5.53 Application—Escape Over a Potential Barrier ‘We suppose that a point moves according 10 the Fokker-Planck equation A, peat) = Bef". + DPC A (55.25) “The potentiat has maxima and minima, as shown in Fig. 5.3. We suppose that motion json an infinite range, which means the stationary solution is pala) =H expl-UO/D1, (5.526) which is bimodal (as shown in Fig. 5.3) so that there is a relatively high probability, Of being onthe left or the ight of b, But not near b. What is the mean escape time from the let hand well? By this we mean, what is the mean first passage time from ‘ato. where xis in the vicinity of B? We use (5.5.23) with the substitutions es ssa sot ro 3s fares f ex -Ute ND sam If the central maximum of U(x) is large and D is small, then exp [U()/D) is sharply peaked at x = b, while expl-U@)/D] is very small near z = b. Therefore, ‘Pha expl-U(2)/ Diz isa very slowly varying function of y near y = b. This means that the value ofthe integral f, expl—U(2)/ dz will be approximately constant for those values of y which yield a value of exp{U(y)/D] which is significantly eliferent from zero. Hence, inthe inner integral, we can set y = B and remove the resulting constant factor from inside the integral with respect to y. Thus, we ean approximate (5.5.28) by reo {5 f ayerot-uieDI}jdvenstu/} (5529) [Notice that by the definition of py(x) in (5.5.26), we can say that nal (5.5.20) F dyexp(-Uey0)IMS. The FotkerPlnck Equation Which means that isthe probability thatthe particle is o the left of & when the system is stationary. A plot of Ta ~> 49) against pis shown in Fig, 5.3 and shows thatthe mean frst, Passage time to xy is quite small for xp in the let well and quite large for tp in the ‘ight well. This means that the particle, in going over the bartier to the right wel {takes most ofthe time in actually surmounting the barrier. Its quite meaningful 10 alk of the escape time as that time for the particle, initially ata, to reach a point rear c since this time is quite insensitive tothe exact location ofthe inital and final Points. We can evaluate this by further assuming that near b we ean write voy vos 4(254) esa antmara tes vers (2) 3529 “The constant factor in (5.5.29) is evaluated as {dcext-vooroi f deorp[-4O =a"), say =a VixDexp|-U(a)/D), (5.5.34) And the inner factor becomes, on assuming ty is well to the right ofthe central point b. Fevevanor= j aver] 4 SP], 6538) VIxDexplU(b)/D1 (5.5.36) Ping tah oF een 55.29) Tea 1) 2abrexlUd)~ Ul0/D) osm ‘This is the classical Arrhenius formula of chemical reaction theory. In a chemical reaction, we can model the reaction by introducing a coordinate such that x = a is species A and x = ¢ is species C, The reaction is modelled by the above diffusion Process and the two distinet chemical species are separated by the potential barrier 1b inthe chemical reaction, statistical mechanics gives the value D=it (55.38) here &is Boltzmann's constant and T is the absolute temperature. We see thatthe ‘ms important dependence on temperature comes from the exponential factor which is often writen explAE AT), (55.39) and predicts a very characteristic dependence on temperature. Intuitively, the answer Js obvious. The exponential factor represents the probability that the energy will ‘exceed tha ofthe barrier when the system sin thermal equilibrium, Those molecules that reach this energy then react, witha certain finite probability ‘We will come back to problems like this in great detail in Chap 14 5.5 Fis Passage Tines for Homogencous Processes 135 5.54 Probability of Exit Through a Particular End of the Interval ‘What isthe probability thatthe particle, intially atx in (a,b) exits through a. and ‘hat i the mean exit ime?” iL “The total probability thatthe particle exits through a after time 1 is given by time integral of the probability current ata. We thus define this probability by ans.) =~ fal Jat (50), (53.40) Fat [-Alaypta,r’ | 2.0) + $341 BCa)pa.t [2,00]} 540) the nopative sign being chosen since we nee the euren pointing othe Ill, Simi lary we define ‘gulsat) = J at (ACbyplb4 F250) ~ LOBOYPLE.1 2,001} (5.5.42) “These two quantities give the probabilities thatthe particle exits through a or b after time ¢ respectively. The probability that (given that it exits through a) it exits after time ris Prob, > 1) = aals.19/996.0) 543) We now find an equation for gu(x.0). using the fact that p(a,1\x,0) satisfies 3 backward Fokker-Planck equation. Thus, Fdt ae Ha,t \x.0), Aena,gtsi+ aCe) at), = Bal.0) 65.49) “The mean exit time, given that exit is through ais gost) 98.22) (5.845) Ta.s) =~ r5Pu(Ty > Na Simply integrating (5.5.44) with respect to 1, we get Atala s9T(a.2)] + $BUVOLR ITAA) = Ales (55.46) where we define _m(2) = (probability oF exit through a) = 94(%,0). (35.47) “The boundary conditions on (5.5.46) are quite straightforward since they follow from those for the backward Fokker Planck equation, namely, (S548) ea)T (aya) = na(b)P(ab) is zero) and In the first of these clearly Ta,a) is zero (the time 10 reach a from a is zero) in the second, x,(b) is zero (the probability of exiting through a, starting from b, is 2210)(365. The Fokker Planck Equation By leting ¢ ~» 0 in (5.5.44), we see that J(a,Olx.0) must varish ifa + x, since, a.0x.0) = 6x ~a), Hence the Fight-hand side tends o zero and we get ACI RL) + $ BLY (5.549) the boundary condition this time being mala 1, mb) =0. (5.5.50) ‘The solution of is boundary condition and the condition (551) 5.49) subject to (8) 4 m2) = 1 Rasy = {i ay 0) l J dy wo. (5552) mtn) = [fave | vv. (55.5%) with WO) as defined in 55.20. “These formulae ind application inthe problem of relaxation of a distribution ini- sally concentrated at an unstable stationary point (Set 141.2). Example—Dilfusive Traversal Time of a One-Dimensional medium: A particle tse ina one-aimension according the difesion equation ip = {dp What ste mean time forthe pate to cifse frm bo a under the condition tha it does. sot eae the interval (a,b) before reaching a? [S6) In the case that the particle starsat within (2,6), we find rom (5.552) and (5546) bes a) a 6554 SDARRAITAbO) = =) (55.55) Using the boundary conditions (5.5.48) the second equation is easy integrated 10 ive Ex bia = aXe +a-2b) As)Pela) = 4 (oT) oD) (55:56) and hence Tx) = FeO so) (3550) 3D In the limiting case x ~» b the probability of exit through a as given by (5.5.54), is zero. Nevertheless, inthe limit that xis approaches b, the mean time to make the exit ven that the ext is at ais quite well defined and is -aF Tb) = (55.58) 5.5 First Passage Times for Homogeneous Processes 137 “This is also clearly the time to exit ata without ever leaving the interval (a,b) before Somtntnte ineral: Notice thai we fix x ant be, we find 1 (5559) Ta) 0 (ssa “The eesull is thatthe particle is certain fo escape a a, but the average time to escape is infinite This arises esas the particle an spend ret dea fine exper the infinite half of the interval, giving rise to an escape time distribution which is normalisabl, but decays so slowly that all moments diverge.6. The Fokker-Planck Equation in Several Dimensions ln many variable situations, Fokker-Planck equations take on an essentially more complex range of behaviour than is possible inthe case of one variable, Boundaries are no longer simple end points ofa line but rather curves or surfaces, and the nature ‘of the boundary can change from place to place. Stationary soluions even with re- Aecting boundaries can correspond to nonzero probability currentsand eigenfunction ‘methods are no longer so simple. [Newertheles, the analogies between one and many dimensions are useful, and this chapter wil follow the same general outline as that on one-variable situations. ‘The subject matter of this chapter covers @ number of exact Fokker Planck equa sion results for many variable systems. These results are not as explicit as forthe one variable ease. An extra feature which is included is the concept of detailed balance in multivariable systems, whichis slmost trivial in one variable systems, but leads to very interesting conclusions in multivariable systems, “The chapter concludes with a treatment of exact results in ext problems for mul- tivariable Fokker-Planck equations. 6.1 Change of Variables Suppose we have a Fokker-Planck equation in variable. A,ple,1) =~ LAAAAx plat) + $F ASByx ple, (61.0) ‘and we want to know the corresponding equation for the variables = fe) (6.12) where fare certain differentiable independent functions. Let us denote by p(y.) the probability density for the new variable, which is given by dus. x2 Bysya- ‘The simplest way o effect the change of variables isto use Ito's formula on a corre- sponding stochastic differential equation iy.) = pla.t (6.13) ax where (ard + benyawin, (6.14) 6.1 Change of Variables 139 bixy"btx) = BOD, (6.15) ‘and then recompute the corresponding Fokker-Planck equation for piy.t) trom the resulting stochastic differential equation as derived in Sect. 4.3.5. "The result is rather complicated. In specific situations direct implementation (6.1.3) may be preferable, There is no way of avoiding a rather messy calculation Lunes full use of symmettes and simplifications is made Example—Cartesian to Polar Coordinates: As an example, one can consider the transformation to polar coordinates of the Rayleigh process, previously done by the stochastic differential equation method in Sect. 4.5.5. Thus, the FokkerPlanck equation is fp, & wo ce). (6.1.6) EF * 362 ‘and we want to find the Fokker-Planck equation for a and @ defined by a apts = Gbps Gen be By = acosd, Ep =asind 17 “The Jacobian is cos sind |_ Be sing acosd ie We use the polar form of the Laplacian to write 18 apa 2 tZ (oe), 61.9) rales) i) and inverting (6.1.7) a= JET+E}, p= tan (Es/Ey) (61:10) We nove oe Et cos Cet at ant ane un fe. -§ Fs = sing, OB: fet and (Ob iaan tanya: te B+ @ a6 _F:___ sing a, H+e ae (6.1.12) Hence,M06, The Fokker Planck Equation in Several Dimensions a a aE ape. ra (ES 5 228), 2 eae ae, 0 1 8a p+ oe = Etat, 6.13) Let us use the symbol pa) For the density Function in terms of 2 and &, The Jaco- ban formula (6.1.8) us that AE. Bo 7) = |AEE ct,£) =a EE ons Pating opt (6.1.6 (6.1.9) (6.1.13) and (6.1.1, we Be 2 Mya) oe ( 188 2 Elo) $ (GZ 22). (15) “which (of course) isthe Fokker Planck equation, corresponding tthe two stochas- ic differential equations in Sect.4.5.5, which were derived by changing variables according to to's Formula, 6.2 Stationary Solutions of Many Variable Fokker-Planck Equations 62.1 Boundary Conditions We have already touched on boundary conditions in general ia Sect. 5.1 where they were considered in terms of probability current. The full range 0! boundary condi tions for an arbitrary moltidimensional Fokker-Planck equation is very much broader than for the one dimensional case, and probably has never heen conpletely specified. In this book we shall therefore consider mostly reflecting barrier boundary condi tions at a surface S, namely. mJ =0 forse, (62 where mis the normal othe surface and Hla = Adastpta) = BE, aye Deke. 622) and absorbing barrier boundary conitions Pa.) =0 fore 23) 1 practice, some par ofthe surfce may be reflecting and anotieabsorbing. Ata surface S on which 4; or Bj are discontinuous, we enforce JO), = 2 Ha, may Epa | ms is2 62 Siationary Solutions of Many Variable FokkerPlanck Equations 141 “The tangential current component is permitted to be discontinuous. “The boundary conditions on the backward equation have already been derived in ‘Sect 5.1.2 For completeness, they are Absorbing Boundary: pox. ty.) = 0 yes. 625) a 5 Reflecting Boundary: Em Bywrgrrie t=O yes. (626) 6.22 Potential Conditions {A large class of interesting systems is described by Fokker-Planek equations which permit a stationary distribution for which the probability current vanishes forall in BR. Assuming this tobe the ease, by rearranging the definition of J (6.2.2), we obtain ‘a completely equivalent equation ap.tx) 3x, $B Btn Las she matrix B(x) has an inverse forall x, we can write (6.2.7) a : a , Lvoeinoon = Zaiten|emey-2 20,40). (628) ZAA,Bl 629) “This equation cannot be satisied for arbitrary By(x) and A(x) since the leltchand side is explicitly a gradient. Hence, Z; must also be a gradient, and a necessary and {condition for that isthe vanishing ofthe curl i.e. Hs (6.2.10) If this condition is satsted, the stationary solution can be obtained by simple inte- ration of (6.2.8) oy-vo jae.) amy tities Z; from derivatives of log [p,(2)], which, therefore, is offen thought of as a potential ~@x) so that more precisely. pala) = exp (6212) and lx) =~ fade ZA, Bea) (62.13)1426 The Fokker Planck Equation in Several Dimensions Example—Raylefgh Process in Polar Coordinates: From (6.1.15) we find ya + e/a" ae ("e% - 215) from which 2 Bah (6.2.16) so that cap ta = (~2mle? + Ha’ ermia (De t19) oe an ceay yy a eg 218 ‘The satonary soleton shen aad) = exp| Fab doze] 62.19 ye 1 o9( 2 sg. oxy pid ve 2) 220 6.3 Detailed Balance 6.3.1 Definition of Detailed Balance ‘The fact that the stationary solution of certain Fokker-Planck equations corresponds ‘o2 vanishing probability current isa particular version of the physical phenomenon of detailed balance. A Markov process satisfies detailed balance if. roughly speak- ing, in the stationary situation each possible transition balances with the reversed lrnnsition. The concept of detailed balance comes from physies, so let us explain ‘more precisely with # physical example, We consider a gas of particles with pos tions r and velocities v. Then a transition corresponds to a particle at some time £ with position velocity (r.6) having acquired by a later time t+ + pestion and veloc- ity of). Te probabil deny ofthis easton ihe on poe deny (OH T: FB) ‘We may symbolically we this transition as (0) 4.0.02), a4) {63 Detaled Balinve 143 ned and unprimed ‘The reversed transition is not given simply by interchanging pr quantities Rather, itis Fo) rer) (6.32) i corresponds to the time reversed transition and requires the velocities to be re versed because the motion from z” to ri in the opposite direction from tha from r tor. “The probability density forthe reversed transition thus th joint probability den sity DAR B48. (633) The principle of detailed balance requires the equality ofthese tw joint probabilities when the system isin a stationary state, Ths, we may write (634) pe 7:758,0) palrs-v,t5 0-0 “The principle can be derived under certain conditions from the laws of physies—see [6.1] and Seet.6.4.2 6.3.2 Detailed Balance for a Markov Process When the probabilities correspond 10 those of a Markov process we ean rewrite (63) as pire] -,0, 002.8) rie. 0 pur). 635) or, where the conditional probabilities now apply to the corresponding homogeneous Markov process (iF the process was not Markoy, the conditional probabilities would be forthe stationary system only). Inits general form, detailed balance is formulated in terms of arbitrary variables x; whieh, under time reversal, (ransform (othe reversed variables according tothe rule mee. 636) ee al 37) depending on whether the variable is odd or even under time reversal Inthe above, riseven, vis odd, ‘Then by detailed balance we require ex 4 7880) (63.8) port as0 By ex, we mean (e) 81.4282.) "Notice that setting + = 0 in (6.3.8) we obtain ox ~ xpi!) = dlex — 8x’ )p.ler) 639 The two delta functions are equal since only sign changes are involved, Hence, palex). (63.10) pals)(errr 1446. ‘The Fokker Planck Equation in Severat Dimensions isa consequence of the formulation of detated balance by (6.3.8), Rewriting now in ‘terms of conditional probabilities, we have | nsrie.0on0e)= nee. ren pin ain 6.33 Consequences of Detailed Balance fr Stationary Mean, Autocorrelation Function and Spectrum nl An important consequence of (63.10) that (x), = eCx)s (6.3.12) hence all odd variables have zero stationary mean, and for the auiocorretation func: tion Gr) = (x(a, (63.13) we have GO) = eM, (63.4) hence, Gir) = eG" ine, (63.15) and setting + = O and noting that the covariance matrix o satisfies = o", o (6.3.16) For the spectrum matrix net Sw)= 3 FeMom, 3.17) we find from (6.3.15) that Sew) = 68 whe (6.3.18) led Balance must be Generalised Possible that there exist several stationary solutions to a Markov process, and in this situation, a weaker form of detailed balance may hold, namely, instead of (6.38), we have ; plier x0) = plex ners ext. 63.19) where the superscripts 1 and 2 refer to two different stationary solutions. Such a situation can exist if one of the variables is odd under time reversal, but does not change with time; for example, ina centrifuge the total angular mementurs has this property. constant magnetic field acts the same way. ‘Mostly, one writes the detailed balance conditions in such situations as, 63 Dewiled Blane 1S poet eri att) = plex 4s en). 63.20) where 1 is a vector of such constant quantities, which change to 6A under time re- versal. According to one point of view, such a situation does not represent detailed balance: since ina given stationary situation, the transitions do not balance in detail Wis perhaps better to cal the property (6.3.20) fime reversal invariance. Ta the remainder of our considerations, we shall mean by detailed balance the situation (6.3.10), since no strong consequences arise from the form (6.3.20) 6.35 Implementation of Detailed Balance in the Differential Chapman-Kolmogorov Equation ‘The formulation of detsiled balance for the Fokker-Planck equation was done by vat Kampen [6], and Graham and Haken {6.2}. We will formulate the conditions in a slightly more direct and more general way. We want necessary ang suficient conditions on the drift and diflusion coefficients and the jump probabilities for a homogeneous Markov process to have stationary solutions which satisfy detailed balance. We shall show that necessary and sufficient conditions are given by @ Werle ype) = Weer’ ienp, a i eAKeRPA) = “AKIO +E FIBULA, Git) ee jByfox) = Bylxy “The specialisation to a Fokker Planck equation is simply done by sting the jump probabilities Wex 2’) equal to zero. 2) Necessary Conditions: 11s simpler to formulate conditions for the cifeentil CChapman- Kolmogorov equation than to restit ourselves to the Fokker-Planck equa- tion. According to Sect. 3, which defines the quanities Wox|2).Ai¢x) and By) {allo course feng tine iependent. since we are considering homoseneous po cess), we have the trivial result that detailed balance require. from (63.11) Wer x’)pse’) = Wier’ lexdpan). (3.21) Consider now the dif coeficient. For simplicity write 63.22) x46 ‘Then from (63.11) we have spl Se + ob ae, Opa) =f aB Sra dole +40) +8), 6323) ‘we use K instead of ein the range of integration to avoid confusion with «divide by Ar and take the limit Ar > O, and the lef-hand side yields eA) pala) + O1K)- (63.28)1466. The Fokker Planck Equation in Several Dimensions ‘On the right-hand side we write Wa 8-8.At x +8,0)p,0x +8) per 6.0014, 0)p.(4) a +E) 2 [pla 8, dr) x. O)pia)] + 06), (6.325) so that the right-hand side is Ln, Fe {2 4,012,090) +560 grote a + 01K), “Adapted 5 (Bite + O18), (6326) free etre es ed ered In Se 34 des rms involving higher powers of than 6 are of order K. tating K 0, we find sAlexipAla) = ~Adadp da) + EI tBu (6327) ‘The condition on B(x) is obtained sina, but in this case no term like the second ‘on the right of (6.3.27) arises, since the principal term is O48"). We find 598 B (6x) = Bix) (6.3.28) A thitd condition i, of course, that p(x) bea stationary solution of the differential (Chapman-Kolmogorov equation. This is not a trivial condition, and is, in general independent ofthe others. by) Sufficient Conditions: We now show that (6.3.21) are sufficient. Assume that these conditions are satisfied, that py(x) is a stationary solution of the differential CChapman-Kolmogorov equation, and that p(x, rx’, 0) is solutior of the differential CChaprman-Kolmogorov equation. We now considera quantity Phest1x'.0) = plex’ }ox,O)p.(x)ps2") (3.29) Clearly x.) 2,0) = dtx =x) = pox. 0] 2.0) 63:30) ‘We substitute p into the differential Chapman-Kolmogorov equation and show that because plx’,t]x,0) obeys the backward differential Chapman-Kolmogorov equa: dion in the variable x, the quantity pis a Solution of the forward differential Chapman Kolmogorov equation. We do this explicitly. The notation is abbreviated for clarity, so ‘hat we write PD for pletle.0), pe for pals), for pe’), pes) for ple'thx.0) amy 63 Daaled Blane 7 We proceed erm by tc. i) Dat Term -E fama -2 Zeamenniy. 6330) = 3 [Flames ngs 2 mex) 0% inno 4 a [Bypsdplex) Lege mum 45 oo mare We 339 Be # +22 (Bp) 2 ples) + Birig amen aera ge wee) * Boas PA it) Jump Term: J del Wx 2)D(z.t 12,0) ~ Wea x) px tha’, = fade Wx] z)psz)plex’,tez,0) ~ Wiel x)p.Cx) plex’. t1ex,O)IPs (63:30) ‘We now use the fact that p(x) is solution ofthe stationary differential Chapman- Kolmogorov equation to write 2 ® -s|2earoe bE scion Jaz Weelsype) ose Werle. a3) and sing the detailed balance condom (6.210) for W = - fide Wezlex)pdz) (6.3.36) 50) Combining all terms : Now subsite age (633) and all up ll thre contains taking cae of (6.338), (63.36% {-pencennen| Zp] + Zeng enn ge | + JdsiW eu |2)p.(c)pty' 162.0) +4 Zee,B: steno | aoa ~ Woey| ps2) ty. 00} /p.) 63.28) We ow satiate died talance conttons (62.20) =(saunatin.o Bsns mK + [dete] yip12.0)~ Welwrpty tiv. 0o}psayiedy’) (6229)1486. The Fokkes anck Equation in Sever! Dimensions ‘The term in the lage curly brackets is now recognisable asthe backward differ ential Chapman-Koimogorov operator (Sect. 3.6, (3.6.4). Nete that the process is homogeneous, so that ra’ t1v.0)= ta0i9.-0) (340) We se that (63.9)= iy’ Opsyvinay | = Bokate’.0) (6341) Siow Osrtea= 2p. hich means that tx, 1x’, 0). defined in (6.3.29), satisfies the Forward differen: tial Chapman-Kolmogorov equation. Since the intial condition of plx,t|x’.0) and lx.t|x',0) at 1 = 0 are the same (6.3.30) and the solutions are unique, we hhave shown that provided that detailed balance conditions (6.3.21) are satisfied ‘detailed balance i satisfied, Hence, suficiency is shown, ©) Comments: ') Even variables only: the conditions are considerably simpler if all ¢, are +1. In this case, the conditions reduce 10 Woxe')p) = Wex'lx) pax), (6.3.42) Aiea) = 4, Fyltitennte (6.3.43) ys) Be (6344) the last of which is trivial. The condition (6.3.43) is exactly the same as the potential condition (6.2.7) which expresses the vanishing of J. the probability ‘current inthe stationary state ‘The conditions (6.3.42) and(6.3.43) taken together imply that pa(x) satisfies {he stationary differential Chapman-Kolmogorov equation, which isnot the ease for the general conditions (6.3.21), i) Fotker Planck equations: The cone of reversible and ireversile dit pats ss intodced by va Kampen 1), and Graham and Haken [6.21. The ie, versie dit is D4) = HA) +eAKon. 6345) and the reversible dit Na) = Asa) ever). (6346) Using again the potenil defined by Pax) = oxpl-#l0. (6347 ‘we sce that in the case of « Fokker Planck equation, we ean write the conditions {or detailed balance as nck Equations 149 64 Examples of Det Bex) = By). 348) Day 15 Layton : (6349) Yay 4 240 (9350) p[ Zo -no%@2]=0, where the last equation is simply the stationary Fokker-Planck equation for psx) alter substituting (6.3.21 i). AS was the ease forthe potential conditions, i bre seen that (6.3.49) gives an equation for 69/2x; which can only be satisfied provided certain conditions on D(x) and By(x) are satisfied. 1 B(x) has an inverse, these take the form a, ash Bs,” OR” where i a (6352) 4 io] 2De ~ BBY and we have p= expat) = exp (Fae ) oe ‘Thus, asim the case of a vanishing probability current, p,(x) can be determined explicitly as an integral. iii} Connection between backward and forward operators of differential Chapman Kolmogorov equations is provided by the detailed balance, The proof of suf cient conditions amouots to showing that if fx, is a solution of the forward tiferential Chapman-Kolmogoroy equation, then Fiat) = flex pe. 16.354) isa solution of the backward differential Chapman-Kolmogorov equation. This telationship will be used in Sect. 6.5 for the construction of eigenfunctions. 6.4 Examples of Detailed Balance in Fokker-Planck Equations 6.4.1. Kramers? Equation for Brownian Motion in a Potential “This problem was given its definitive formulation by Kramers 6.3}. when consider ing a mode! of molecular dissociation. : 'We take the motion of a particle in a fluctuating environment, The motion isin one dimension and the state ofthe particle is described by its postion x and velocity 6 ‘This gives the differential equations1506. The Fok nck Equation in Several Dimensions ds é (64) le = v9) pos VERTED, 42) which resenilly Langevin equations (1.2.14) n which, for brevity, we write ana = 8, (43) and V(x) isa potential whose gradient V'(x) gives rise o a force on the particle. By ‘making the assumption thatthe physical uctuating force g(t) is io be interpreted as eindr= awe), (644) «s explained in Sect. 41, we obtain stochastic diferenta! equations ds=odt, (645) indo = -1V'G) + polars V3BRTaWin, (646) for which the comesponding Fokker Planck equation is T Py (Ve +A4p) + AEE 647) The equation canbe slightly simplified by introducing new scaled variables v=sVaiT. (648) = 0ymiT, (649) uw) = Yeayer, (6.4.10) y=BIm, (64.11) so thatthe Fokker Planck equation takes the form &® a * - Gun Svs vo) +e (oo 2). (64.12) which we shall eall Kramers” equation Here, y (the position) isan even variable and u(the velocity) n odd variable, as explained in Seet.6.3, The drift and diffusion can be written Aya) = |-vsi-m (oan) ewn=[5 3] 64.19) ‘The detailed balance transformation is lcl-[4] wy We can check the conditions one by one, ‘The condition (6.3.21 ii) is trivially satisfied. The condition (6.421) is somewhat degenerate, since B is not invertible. It can be writen OO 464 Examples of Desiled Balance in Fokker Planck Equations (St o ‘or, more fully “1 - ¢ (6.7 oe P= era vou] ny See ne “The fist line isan identity and the second states pay.) = Be, (64.18) ala) = exp(—$1") fly) (64.19) “This means that if ps(y.u) is written in the form (6.4.19), then the detailed bal- ance conditions are satisfied, One must now check whether (6.4.19) indeed gives, stationary solution of Kramers’ equation (6.4.12) by substitution. The final bracket vanishes, eaving of =u ~ vrs, 64.20) By TOMS « which means W expl-UWI. 420, se] (64.2) In terms of the original (x,2) variables, Plxt)= er [2 me (6.4.23) 2KT |* “which is the familiar Boltzmann distribution of statistical mechanics, Notice that, the denominators AT arise from the assumed coefficient /2BKT of the flectwating force in (6.4.2). Thus, we take the macroscopic equations and add a fluctuating force, whose magnitude is fixed by the requirement that the solution be the Boltzmann distribution corresponding to the temperature T. But we have also achieved exactly the right distribution function. This means that the assumption that Brovsnian motion is deseribed by a Markov process of the form (64.1), (6.4.2) must have considerable validity.152.6 ‘The Fokker Planck Equation in Several Dimensions 6.4.2 Deterministic Motion Here we have B(x) and Werle’) equal to zero, so the detailed balance conditions are simply GiAKeR) = ~Asx) (6.4.24) Since we are now dealing with a Liouville equation (Sect,3.5.3), the motion of a Point whose ecordinates are « is described by the ordinary differemial equation a Ga = Aleta (6.4.25) Suppose a solution of (6.4.25) which passes through the poin al aia. 6426) Which therefore satisfies (6420) ion 6.4.24) implies thatthe reversed solution eq(-1.00), (6.428) is also a solution of (64.25), and since eq(0.cy) = sey =y, (6.4.29) initial conditions are the same, these solutions must be identical, eq\-t.0y) = qt.9) 6.4.30) Now the joint probability in the stationary state can be wei pasta’) = dy paeestix! 139.0) = Lely t= qe” — quae). (aay an pilex’ 1 sex.) = {dy dex — gl-ryilex’ ~ q(-1 wlpau) 64.32) Change the vias from y to ey and note that i) = pe nd ey = dy, so (64.32) = Jdyolx~ eq(-t.cy)olx’ ~ eq(-1 ewIp.ty), (64.33) and using (64.30), = Sey Sx — qt etx! ~ gt pala) (64.34) paestx'0) (6.4.38) Using the stationarity property, that py depends only onthe tha nds only on the time difference, we see that detailed balance is satistied. This direct proof is, of course, unnecessary since the original general proof is valid for this determi {64 Examples of Detailed Balance in Fokker Planck Equations 188 6.43 Detailed Balance in Markovian Physical Systems tn physical systems, which are where detailed balance is important, we often have an unbelievably larze numberof variables, ofthe order of 10" at least. These variables (Gay, momentum and velocity ofthe particles ina gas) are those which occur in the disribation function which obeys a Liouville equation for they follow determinist {equations of mation, like Newton's laws of motion. can be shown directly that, for appropriate forms of interaction, Newton's laws ‘obey the principle of microscopic reversibility which means that they can be pot in the form (6.4.25), where A(x) obeys the reversibility condition (64.24). The macro- scopically observable quantities in uch a system are functions of these variables (for example, pressure, temperature, density of particles) and, by appropriate changes of variable. can be represented by the frst few components of the vector x. “Thus, we assume x can be weitten x=(as) Cen where the veetor a represents the macroscopically observable quantities and # is all the others. Then in practice, we are interested in Bay, tsag.f343.t5--) So fbaodity parr tian tik tsio) (64.37) From the mieroscopic reversibility, it follows from our reasoning above that p, and thus also, both obey the detailed balance conditions but, of course, ? does not obey ‘a Liouville equation. [fi turns out or can be proven that B abeys, to some degree “approximation, a Markov equation of morion, then we must preserve the detailed bal- lance property. which takes the same form for pas for p. In this sense, the condition (6.38) for detailed balance may be said tobe derived from microscopic reversibility ‘of the equations of motion. 64.4 Ornstein-Uhlenbeck Process Most systems in which detailed balance is of interest can be approximated by an Omnstein-Uhlenbeck process, i. this means we assume Ada) = BAG (6.4.38) By (6.4.39) Bye) “The detailed balance conditions are not trivial, even for the case of such a linear system, They take the form a Beery + Aude = EB Yo mt, (64.40) and on wash,S46 The Fokker Planck Equation in Several Dimensions Equation (6.4.40) has the qualitative implication that pa (x) a Gaussian since deriva- tive of log ps(x) logis linear in x. Furthermore, since the leli-hand side contains no ‘constant term, this Gaussian must have zero mean, hence, we can write ¥exp(-baton's) (6.442) One ean ow substitute (64.42) in the stationary Fokker-Planck equation and re- arrange to obtain pa) : (6.4.43) ~ BAG HE Bye! + B(Beutas +4507! Beat) Where we have used the symmetry of the matrix. The quadratic term vanishes if the symmetric part of its coefficient is zero, This condition may be written in matrix form as oat ato = Bo, (64.48) Ao tea" =~B. (6.445) The constant term also vanishes i(6.4.44) is satisfied, Equation (6.4.45) i, of course, exactly that derived by stochastic differential equation technigues in Sect. 4.5.6 (4.5.64) withthe substitutions ed (6.446) Bes Bs, 447) ‘We can now write the detailed balance conditions in their most elegant form. We define the matrix by = diagley,e0.65....) (6.4.48) and clearly ea (6.449) ‘Then the conditions (6.4.40) and (6.4.41) become in matrix notation bAB+A = Bo (6.4.50) Be = B (6451) ‘The potential condition (6.3.51) is simply equivalent to the symmetry of ‘As noted in Sect.6.3 (6.3.16), detailed balance requires Lo = oe. (64.52) Bearing this in mind, we take (6.4.45) Aotod™ =—B, (6.453) and from (6.4.50) her + Aa = (458) 464 Examples ofDesiled Balance in Fokker Planck Equations 185 which yield eas (6455) and with (6.4.52) atAo) = Ao". Nee ‘These are equivalent to the celebrated Onsager relations; Onsager, (64; Casimir, [6.5]. The derivation closely follows van Kampen’s [6.1] work GAS The Onsager Relations “The physical form of the Onsager relations arises when we introduce phenomeno- logical forces defined as the gradient of the potential ¢ = logl p(x) Fix) = Vox) = ox, 457) (in physics, 4/47 isthe entropy of the system). Because of the Tinear form of the ‘A (6.4.38), the exact equations of mation for (x) are as) dr MD) ~ stay = Ae FH) (6.4.58) “Thus, ifthe flares d(x) are related linearly to the forces FU(x)) by a mateix L defined by L=Ae. (6459) then (6.4.56) says as (6.460) and e, of the same sign, ny 6, and e} of different slg. Notice also that the relations eoe=o. (6462) imply that fy and cry vanish if e; and ¢; have opposite signs. Tn the special ease that all the have the same sign, we find tht bent 64.63) and noting that, since ois symmetric and positive definite it has a real square root (?, we find that ebe deo tac? (6468) is symmetric, that Ais similar. asymmetric matrix. Hence, al the eigenvalues of Aa real1566 TheFokkecPlanck Equation in Several Dimensions L e SO — t é avn ES Foe dn Fig. 6.1. Bec ict used inthe dorivation of Nyguist's formal 6.6 Significance of the Onsager Relations—Fluctuation-Dissipation Theorem The Onsager relations are for a set of macroscopically observable quantities. and thus provide an easily observed consequence of detailed balance, which is itself @ consequence of the reversibility of microscopic equations of motion, as outlined in Sect. 6.4.2 above. However, to check the validity in a given siuation requires a knowledge of the covariance matt c. Fortunately. in such situations statistical mechanics gives us the form of the stationary distribution, provided this is thermodynamic equilibria in which de- tailed balance is always satisied. The principle is similar to that used by Langevin (Sect. 1.22), Example—Derivation of Nyquists’s Formula: As an example of the use of the ‘Onsager relations, we ean give a derivation of Nyquist’s formula ‘We assume an electric circuit asin Fig. 6.1 in which there is assumed to be a fluc- ‘wating voltage and a fluctuating charge. both of which arise from te system having ‘nonzero temperature, and which we will show have their origin in the resistors R andr. The elecrical equations aise from conservation of an elecrie charge q and Kirchoft’s voltage law. The charge equation takes the form ag ra in which we equate the rate of gain of charge on the capacitor to the current i less leakage term yq = g/rC plus a possible fluctuating term A/(2), which arises from Jeakage through the capacitor, and whose magnitude will be shortly calculated, Kirchofl’s voltage law is obtained by adding up all the voltages around the circuit, including 2 possible Huctuatng voltage AV( di a We now assume that A/(1) and AV(A) ate white noise. We can write in the most general ease, AK = bugil) + bre), (6.4.67) ave = ine. (sass) wtauy, —y ire. (6.465) rbE- save) 6s) in which €(0) and (0) are uncorrelated Langevin sources, i, Fokker Planck Equations 157 (64 Examples of Detailed Bal andr = am. waa eude = a0. (e470) ze (and Ws) ar independent Wiener processes Ths hve 4 |: : (am ie The tal ery inthe 5)tem i BP + te 47) Batt at From statistical mechanics, we know that p,(q.i) is piven by the Bolremann Distri- ution at temperature T. ie. lg.) = AV expl-E/kT) eae a7) =o sada): where ks the Boltzmann constant, so that the covariance matrix is, arc 0 on elo eri! ‘The Onsager symmetry can now be checked: ear (6.4.78) =" VeryL RUTH? For this system g the total charge is even under time inversion and i the current, is ‘odd Thus, ' Tee (6.4.78) pesaevotent['S 2) ‘ o lane an eo h suggests that and we see that Bp = By; = 0, a8 required by physical inition, which suzgests the two sources of fluctuations arise from different causes and should be independent.= i 158 6. The Fokker-Planck Equation in Several Dimensions ‘These results are fluctuation dissipation results, The magnitudes ofthe fluctuations ‘hy ace determined by the dissipative terms r and R. In fact, the result (6.4.80) is precisely Nyquist’s theorem, which we discussed in Sect. 15.4. The noise voltage and current inthe circuit ae given by AV) = VETR EM), aly = VT en, (6.482) so that {AVr+ DAV(N) = ATR, (Are AND) = 2T/r6(2), (6.4.83) (avirw nanny = 0. ‘The first ofthese is Nyquist” theorem in the form quoted in (1.5.48-1.5.50), the second isthe corresponding result fr the noise regarded as arising from a fluctuation current Irom the resistor r, and these fuetuations are independent ‘The terms r and B are called dissipative because they give rise to energy dissipa- ‘ion: in fact, dleterministically (ie. setting noise equal to 0), caer Ea ee, (6.484) ‘which explicitly exhibits the dissipation, 6.5 Bigenfunction Methods in M: ny Variables ere we shall proceed similarly to Sect.5.4, weating only homogeneous processes. We assume the existence of a complete set of eigenfunctions P(x) of the forward Fokker-Planck and a set Q,(x) of the backward Fokker Planck equstion, Thus, = ESUACNPARD + § FAB yCeyPie) = APC), (65.1) TAMA +L = = vs) 652) Whether Que) and Py(x) satisfy absorbing or reflecting boundary conditions, one ean show, in a manner very similar fo that used in Sect. 5.1.2, that AA-AD fds Pye) =O. 653) 0 that the P(x) and Q(x} from a bi-orihogonal set, which we normalise as Idx Py Qu(2) = by. 54) if the spectrum of eigenvalues 1 is discrete. Ifthe spectrum is continuous, the Kro- recker 8, isto be replaced by (4~ 2’), except where we have reflecting boundary ‘conditions, and thus 4 = 0 corresponds to the stationary state, The normalisation of ala) then gives J dx Pox )Qolx) = fidx Pols) = 1 (655) ‘0 that there is also a discrete point with zero eigenvalue inthe specteum then, (65 Eigenfunetion Methods in Many Variables 159 65.1 Relationship between Forward and Backward Eigenfunctions ‘The functional relationship (5.8) between the P, and Q), which is always tue in ‘one dimension, only pertains in many dimensions if det ‘To show this, note that if detailed balance is valid, we have already seen in Sect. 6.3 Selii) that plex, —2/px(x) is a solution of the backward Fokker-Planck ‘equation so tha, from the uniqueness of solutions, we ean say Ou(x) = mPaen pn. 656) Here, my =

You might also like

• Numerical Analysis of Concrete Crash Barriers: J. Kala, P. Hradil, and V. Salajka

  

  Document4 pages

  Numerical Analysis of Concrete Crash Barriers: J. Kala, P. Hradil, and V. Salajka

  Barry Liberty

  No ratings yet
• D10 Handbook of Stormwater Drainage Design

  

  Document59 pages

  D10 Handbook of Stormwater Drainage Design

  Anonymous WqYvpVd

  100% (3)
• 5.3.2 Design Member Moment Capacity: Asi: Design Capacity Tables For Structural Steel Volume 2: Hollow Sections

  

  Document5 pages

  5.3.2 Design Member Moment Capacity: Asi: Design Capacity Tables For Structural Steel Volume 2: Hollow Sections

  Barry Liberty

  No ratings yet
• • Magazines
  • Podcasts
  • Sheet music
• Thermal Oxidation Plants: February 2016

  

  Document29 pages

  Thermal Oxidation Plants: February 2016

  Barry Liberty

  No ratings yet
• Sulzberger Method

  

  Document8 pages

  Sulzberger Method

  Barry Liberty

  No ratings yet
• Standard For Australian Survey Control Network - v2.1

  

  Document13 pages

  Standard For Australian Survey Control Network - v2.1

  Barry Liberty

  No ratings yet
• Geotechnical Characterization of A Residual Soil Profile: The ISC'2 Experimental Site, FEUP

  

  Document9 pages

  Geotechnical Characterization of A Residual Soil Profile: The ISC'2 Experimental Site, FEUP

  Barry Liberty

  No ratings yet
• Mast Foundation Software PDF

  

  Document10 pages

  Mast Foundation Software PDF

  Barry Liberty

  No ratings yet
• Code of Practice RC 50022 Selection and Design of Pavement and Surfacings December 2018

  

  Document33 pages

  Code of Practice RC 50022 Selection and Design of Pavement and Surfacings December 2018

  Barry Liberty

  No ratings yet
• Seismic Hazard Analysis

  

  Document36 pages

  Seismic Hazard Analysis

  Barry Liberty

  No ratings yet
• Residual Soils

  

  Document18 pages

  Residual Soils

  Barry Liberty

  No ratings yet
• Welding Calculations: Return To Main Menu Basics

  

  Document27 pages

  Welding Calculations: Return To Main Menu Basics

  Barry Liberty

  No ratings yet
• Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  

  From Everand

  Hidden Figures: The American Dream and the Untold Story of the Black Women Mathematicians Who Helped Win the Space Race

  Margot Lee Shetterly

  Rating: 4 out of 5 stars

  4/5 (897)
• The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  

  From Everand

  The Subtle Art of Not Giving a F*ck: A Counterintuitive Approach to Living a Good Life

  Mark Manson

  Rating: 4 out of 5 stars

  4/5 (5807)
• Shoe Dog: A Memoir by the Creator of Nike

  

  From Everand

  Shoe Dog: A Memoir by the Creator of Nike

  Phil Knight

  Rating: 4.5 out of 5 stars

  4.5/5 (537)
• Team of Rivals: The Political Genius of Abraham Lincoln

  

  From Everand

  Team of Rivals: The Political Genius of Abraham Lincoln

  Doris Kearns Goodwin

  Rating: 4.5 out of 5 stars

  4.5/5 (234)
• Never Split the Difference: Negotiating As If Your Life Depended On It

  

  From Everand

  Never Split the Difference: Negotiating As If Your Life Depended On It

  Chris Voss

  Rating: 4.5 out of 5 stars

  4.5/5 (842)
• The Yellow House: A Memoir (2019 National Book Award Winner)

  

  From Everand

  The Yellow House: A Memoir (2019 National Book Award Winner)

  Sarah M. Broom

  Rating: 4 out of 5 stars

  4/5 (98)
• Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  

  From Everand

  Elon Musk: Tesla, SpaceX, and the Quest for a Fantastic Future

  Ashlee Vance

  Rating: 4.5 out of 5 stars

  4.5/5 (474)
• Yes Please

  

  From Everand

  Yes Please

  Amy Poehler

  Rating: 4 out of 5 stars

  4/5 (1898)
• The Little Book of Hygge: Danish Secrets to Happy Living

  

  From Everand

  The Little Book of Hygge: Danish Secrets to Happy Living

  Meik Wiking

  Rating: 3.5 out of 5 stars

  3.5/5 (401)
• A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  

  From Everand

  A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story

  Dave Eggers

  Rating: 3.5 out of 5 stars

  3.5/5 (231)
• On Fire: The (Burning) Case for a Green New Deal

  

  From Everand

  On Fire: The (Burning) Case for a Green New Deal

  Naomi Klein

  Rating: 4 out of 5 stars

  4/5 (74)
• Grit: The Power of Passion and Perseverance

  

  From Everand

  Grit: The Power of Passion and Perseverance

  Angela Duckworth

  Rating: 4 out of 5 stars

  4/5 (590)
• The Emperor of All Maladies: A Biography of Cancer

  

  From Everand

  The Emperor of All Maladies: A Biography of Cancer

  Siddhartha Mukherjee

  Rating: 4.5 out of 5 stars

  4.5/5 (271)
• The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  

  From Everand

  The Hard Thing About Hard Things: Building a Business When There Are No Easy Answers

  Ben Horowitz

  Rating: 4.5 out of 5 stars

  4.5/5 (345)
• Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  

  From Everand

  Devil in the Grove: Thurgood Marshall, the Groveland Boys, and the Dawn of a New America

  Gilbert King

  Rating: 4.5 out of 5 stars

  4.5/5 (266)
• Steve Jobs

  

  From Everand

  Steve Jobs

  Walter Isaacson

  Rating: 4.5 out of 5 stars

  4.5/5 (806)
• Fear: Trump in the White House

  

  From Everand

  Fear: Trump in the White House

  Bob Woodward

  Rating: 3.5 out of 5 stars

  3.5/5 (738)
• Principles: Life and Work

  

  From Everand

  Principles: Life and Work

  Ray Dalio

  Rating: 4 out of 5 stars

  4/5 (599)
• Rise of ISIS: A Threat We Can't Ignore

  

  From Everand

  Rise of ISIS: A Threat We Can't Ignore

  Jay Sekulow

  Rating: 3.5 out of 5 stars

  3.5/5 (137)
• The World Is Flat 3.0: A Brief History of the Twenty-first Century

  

  From Everand

  The World Is Flat 3.0: A Brief History of the Twenty-first Century

  Thomas L. Friedman

  Rating: 3.5 out of 5 stars

  3.5/5 (2259)
• The Unwinding: An Inner History of the New America

  

  From Everand

  The Unwinding: An Inner History of the New America

  George Packer

  Rating: 4 out of 5 stars

  4/5 (45)
• The Glass Castle: A Memoir

  

  From Everand

  The Glass Castle: A Memoir

  Jeannette Walls

  Rating: 4.5 out of 5 stars

  4.5/5 (1716)
• Angela's Ashes: A Memoir

  

  From Everand

  Angela's Ashes: A Memoir

  Frank McCourt

  Rating: 4.5 out of 5 stars

  4.5/5 (440)
• The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  

  From Everand

  The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are

  Brene Brown

  Rating: 4 out of 5 stars

  4/5 (1091)
• John Adams

  

  From Everand

  John Adams

  David McCullough

  Rating: 4.5 out of 5 stars

  4.5/5 (2409)
• Bad Feminist: Essays

  

  From Everand

  Bad Feminist: Essays

  Roxane Gay

  Rating: 4 out of 5 stars

  4/5 (1016)
• The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  

  From Everand

  The Sympathizer: A Novel (Pulitzer Prize for Fiction)

  Viet Thanh Nguyen

  Rating: 4.5 out of 5 stars

  4.5/5 (122)
• The Outsider: A Novel

  

  From Everand

  The Outsider: A Novel

  Stephen King

  Rating: 4 out of 5 stars

  4/5 (1849)
• A Man Called Ove: A Novel

  

  From Everand

  A Man Called Ove: A Novel

  Fredrik Backman

  Rating: 4.5 out of 5 stars

  4.5/5 (4610)
• The Woman in Cabin 10

  

  From Everand

  The Woman in Cabin 10

  Ruth Ware

  Rating: 3.5 out of 5 stars

  3.5/5 (2520)
• The Light Between Oceans: A Novel

  

  From Everand

  The Light Between Oceans: A Novel

  M.L. Stedman

  Rating: 4.5 out of 5 stars

  4.5/5 (789)
• Brooklyn: A Novel

  

  From Everand

  Brooklyn: A Novel

  Colm Tóibín

  Rating: 3.5 out of 5 stars

  3.5/5 (1946)
• The Perks of Being a Wallflower

  

  From Everand

  The Perks of Being a Wallflower

  Stephen Chbosky

  Rating: 4.5 out of 5 stars

  4.5/5 (2104)
• Wolf Hall: A Novel

  

  From Everand

  Wolf Hall: A Novel

  Hilary Mantel

  Rating: 4 out of 5 stars

  4/5 (3811)
• Manhattan Beach: A Novel

  

  From Everand

  Manhattan Beach: A Novel

  Jennifer Egan

  Rating: 3.5 out of 5 stars

  3.5/5 (792)
• Little Women

  

  From Everand

  Little Women

  Louisa May Alcott

  Rating: 4 out of 5 stars

  4/5 (104)
• The Art of Racing in the Rain: A Novel

  

  From Everand

  The Art of Racing in the Rain: A Novel

  Garth Stein

  Rating: 4 out of 5 stars

  4/5 (4203)
• Sing, Unburied, Sing: A Novel

  

  From Everand

  Sing, Unburied, Sing: A Novel

  Jesmyn Ward

  Rating: 4 out of 5 stars

  4/5 (1104)
• A Tree Grows in Brooklyn

  

  From Everand

  A Tree Grows in Brooklyn

  Betty Smith

  Rating: 4.5 out of 5 stars

  4.5/5 (1929)
• Her Body and Other Parties: Stories

  

  From Everand

  Her Body and Other Parties: Stories

  Carmen Maria Machado

  Rating: 4 out of 5 stars

  4/5 (821)
• The Constant Gardener: A Novel

  

  From Everand

  The Constant Gardener: A Novel

  John Le Carré

  Rating: 3.5 out of 5 stars

  3.5/5 (104)