f,; >... > 4, > ty, a Markov process has a probability density such that PUK n|Xpr Xz +» Xo) = (XulXn-1) (B.23) where the notation p(a|b) means “the probability density of @ given that b was true.” Thus, for a Markov process, the probability that x, = 5 depends only on what the value of x,_; was; it does not depend on what the values of x,..2, X,-3, etc. were. There are both discrete and continuous Markov processes. An example of a discrete Markov process is given by flipping a coin. A trivial example comes if we give each toss a value x(1,) = x, = + | for a toss of “heads” and a value xX, = — | for a toss of “tails.” Then x is clearly a Markov process, since p(x,) = Box, — 1) + 46(x, + 1) does not depend on x,.,, much less on Xp-2, Xp-31 ClC. A better example of a discrete Markov process is given by defining the random variableMarkov Processes and Fokker-Planck Equations 273 (1) Fig. B.6 Any function in nature can be drawn as a smooth curve as shown X(t) = X, = x (B.24) where the x; are given by the coin tosses of the previous paragraph. Now X is clearly a Markov process, whose probability density at f, very definitely depends on the value of X,_,, but on no previous value. EXERCISE Calculate p(X,|X,-1) for this example. To give an example of a continuous Markov process is more difficult, because a continuous Markov process cannot exist in nature. To see this, consider any random function that we can draw as a smooth curve, as in Fig. B.6. Now, on the time scale shown, it appears that x,,, not only depends on x,, but also on x,_). That is, X,4; not only depends on x,, but also on the derivative of the function dx(1)/dt},,, Which can be written dx(t) X, — Xp dt |=, Ar Thus, this function is not a Markov process. In fact, no function that is a continu- ous curve and, therefore no physical function, can be a Markov process. This does not mean that Markov proccsses cannot be a good approximation to a physical process. Consider the velocity function of the Brownian particle in the previous section (Fig. B.7). We have seen that the velocity consists of a rapid fluctuation due to each molecular collision, together with a slowing down or net friction force. Thus, on the time scale of molecular collisions, the process is not Markovian. However, on the much longer time scale of many collision times, the (B.25) (ey ' Fig. B.7 One realization of the velocity of a Brownian particle in a particular direction.274 Appendix B situation is very nearly Markovian. The Brownian particle is performing a random walk in velocity space, and soon forgets the details of its orbit near r ; it does, however, remember its velocity vp at t = 0. Thus, we consider the process to have three time scales (Fig. B.8): the collision time 7,, Which is the autocorrelation time of the force SF(t) in the Langevin equation; the time ¢ after which we may assume to good approximation that fhe process is Markovian; and the dissipation time v'', We must have At >> 7,; we shall further assume in this section that At << vy’. Let us develop some of the mathematical properties of Markov processes. This development will lead us to the Fokker-Planck equation. Consider the probability of a sequence of values of the random function x(1). This is P(X Xpty + + ++ Xo.X1,Xo) = {probability that, at time fo, the process x(t) has the value x, and at time t,, x(t) has the value x,, and... and at time ¢,, x(t) has the value x,} where, > t)-) > ty»... > 1, > to (B.26) By the definition (B.23) we can write P(X rXnery = = >» Xo) = P(X p|XnteXnae - - + + Xo) XK Xe tXyers oe 0) = PG Opt Xeas es + Xo) (B.27) The same procedure can now be applied to the last factor on the right of (B.27), so that PXnr%nds 6 Xo) S P(XnilXn-2)O( nas» «+ + Xo) (B.28) and so on until we have finally, for a Markov process, et) % t Hi Collision time, not Markovian [at Many collision times, almost Markovian |-___, _1 Dissipation time » B.8 Three time scales of Brownian motion.Markov Processes and Fokker-Planck Equations 275 PX Xmty%ay oe + 9 Xo) = (XI Xn |Xn-2) ++ o(%21%1)e(% 1X0)a(%o) (B.29) By elementary considerations it must also be true that PX nr XnevXnezs + +s XyXo) = MKa%eay e+» « XilXo)O(%o) (B.30) Comparing (B.30) and (B.29) we find PU XpXyiy 0 Xe1%o) = A XnlXn-1)O(%eal%r-2) + + - OC 1%0) (B31) In particular, we can choose n = 2 to obtain (2X1 1X0) = (21% )o(%1 1X0) (B.32) Let us row integrate this expression over all possible x, to obtain p(%2|Xo) =f de (%2,%1 1X9) (B.33) elaalre) = fd, ebrate Grn) (B34) which is the Chapman-Kolmogorov equation, or Smoluchowsky equation (8]. Suppose we identify x, with time ¢ and x, as x(t + Ar). Suppose we further assume that or AX) = plxst = to) = (x — Xo) (B.35) Then we can drop the references to xp in (B.34), and write P(xX2/Xo) = p(x.t + Ar) (B.36) that is, x, is now denoted by x, and (x, 1X0) = p(X) (B.37) We can also change the notation of p(x,|x;); with the definition Ax=x— x, (B.38) we can write Cx|%) = pl(x,t + Atlx — Ax,t) u(Ax,t + At|x — Ax,t) (B.39) where the transition probability is defined by (B.39); gives the probability that at time ¢ + At, the random process has made a jump of Ax from its previous value x — Ax at time J. With these notational! changes, we can rewrite (B.34) as p(x,t + At) = f danwars + At|x — Ax,Op(x — Ax,t) (B.40)276 Appendix B The value x appears on the right of (B.40) only in the combination x — Ax. Thus, if we assume that all of the important physics happens for small Sx, then we can make a Taylor series expansion on the right of (B.40), obtaining p(x,t + Ax) = faax s (rant oe | [wane + At|x — Ax,t)p(x — ax] HIE} or i = (=Ax) a! p(xt + An) = faax) = 7 X (W(dx,t + Arlx,n)p(x,1)) (B.41) If the infinite sum converges, and if we can interchange the summation and inte- gration, then we can write sev a p(x,t + dt) = = ested X (olen f danianiware + adlx0] (B.42) The quantity given by the Ax integration is just the expectation value or ensemble average of (Ax)', (4x) = f qanadiwars + Atlx,n) (B.43) which is itself a function of x,t through y. Equation (B.42) becomes 2 (1 a! , post + At) = > ar [olsX(O0)' 4] (B.44) = 7 Moving the / = 0 term to the left side, and dividing by As, we have p(x, + At) — plxt) _ ¥ Cy a at fe dt ax! [ele X(Axy1)] (B45) We next take the limit as At — “0”. This means that we Jet Ar become very small, much smaller than any macroscopic time scale (e.g., »"!). However, At cannot really go to zero, because this development has assumed that Ar is large enough to justify the Markovian assumption. Thus, the left side of (B.45) becomes plxt + A) = ox.) _ dolx.t) at ar (B.46) lim amo" where the time derivative refers to macroscopic time. Equation (B.45} becomesMarkov Processes and Fokker-Planck Equations 277 ae (dx) = 2 (-l! oy [um ne oss] (B.47) a = ! Defining the diffusion coefficients i (dx)! Dx.) = Min te a) (B.48) Equation (B.47) is Seed) Be 1 © De%s.npts, ) (B.49) If we keep only the first two a on the right of (B.49), we have So 2 ipo & pe BP Gy DOOMED] + Fs DMN] | (B50) which is the well-known Fokker-Planck equation [9] For Brownian motion, the random variable x is replaced by the particle velocity u(f). We shall leave it as an exercise to determine the diffusion coefficients D'!(v,t) and D?y,1). EXERCISE Use the results of the previous section to evaluate the coefficients in Eq. (B.50) for Brownian motion. Show that D'"(v,t) = — vu, and D?(u,1) = vT/M, so that the Fokker-Planck equation associated the Langevin equa- tion of Brownian motion is 8p(v,t) oT # at ay (P) + a aye (B.51) EXERCISE Use the results of the previous section to show that D"(v,t) ~ At and, thus, vanishes as Ar — “0” We can now understand why we are able to write the Lenard-Balescu equation in the form of a Fokker-Planck equation, ast) — 1 fe at WAN + 7 Vs, Vy BN (B.52) Because the derivation of Lenard-Balescu assumed g(1,2) << f,(1)f,(2), we have effectively limited ourselves to small angle two-body collisions. The quantity /(v,,t) may be thought of as the probability density of particles in velocity space. Thus, (¥,,1) is changing slowly on the time scale for a two-body collision. All of these features are precisely those assumed in the derivation of the Fokker-Planck equa- tien, It should come as no surprise to us that the Lenard-Balescu equation can be written in the form of the Fokker—Planck equation. The coefficient A in (B.52) is called the coefficient of dynamic friction, and plays the same role as vv in the Fokker-Planck equation (B.51) for Brownian motion. It represents the slowing278 Appendix B down of a particle due to many small angle Coulomb collisions, Likewise, the coefficient B in (B.52) is called the diffusion coefficient, and plays the same role as vT/M in (B.51). It represents the diffusion of the plasma particles in velocity space due to many small angle collisions. In the steady state, a typical particle is suffering dynamic friction plus diffusion; the net effect is to produce a Maxwellian. This is just as true in a plasma as itis for a Brownian particle. In addition to the stated references, sources for this appendix include the book by Stratonovich [10] and the ageless and excellent article by Chandrasekhar (11] REFERENCES [1] R. Brown, Phil. Mag., 4, 161 (1928). [2] R. Brown, Ann. Phys. Chem., 14, 294 (1928). [3] M. P. Langevin, C. R. Acad. Sci. Paris, 146, 530 (1908). [4] A. Einstein, Ann. Phys., 17, 549 (1905). [5] A. Einstein, Ann. Phys., 19, 289 (1906). [6] A. Einstein, Ann. Phys., 19, 371 (1906). [7] D. K. C. MacDonald, Noise and Fluctuations: An Introduction, Wiley, New York, 1962. [8] _M. von Smotuchowsky, Ann. Phys., 21, 756 (1906). [9] Lord Rayleigh, Nature, 72, 318 (1905). [10] R. L. Stratonovich, Topics in the Theory of Random Noise, Gordon and Breach, New York, 1963, Vol. 1. (11) S. Chandrasekhar, Rev. Mod. Phys., 15, 1 (1943)APPENDIX Pedestrian’s Guide to Complex Variables Many parts of this book make use of the basic results of the theory of complex variables [1, 2]. For the benefit of readers who have not yet studied this subject in detail, or who have studied it long enough ago to have forgotten it, these basic results are summarized here. The most useful result is the residue theorem, which states that the integral in a counterclockwise direction around a closed curve is 27/ times the sum of the residues. If the integrand is of the form f(z(z — 2)"!, the residue at the simple pole z = 2 is f(z9). For example, consider the integral aa dz where the integration is along the real z-axis, and @ > 0. The integration can be closed by a large semicircle at infinity, since the contribution from the semicircle is ~ lim (7R/R*) = lim (a/R) = 0 Roe Row The semicircle can occur in either the upper-half z-plane or the lower-half z-plane (Fig. C.1). Writing (C.1) as I -¢ Seas 2 EE ( + ia\z — ia) we close the contour downward, changing the sign on the result because this is in the clockwise direction, to obtain (only the pole at z = — ia is enclosed) (€.2) z— ia a2 (C3) ia I = (-2n/)280 Appendix C Fig. C.1 Integration contour in the z-plane. EXERCISE Obtain the same result by closing upward. The solution of many ordinary and partial differential equations is facilitated by Fourier and Laplace transformation. The Fourier uransform conventions used in this book, stated in Chapter 5, are for functions of one spatial dimension x, Ho =f" & seep cinn (C4) with inverse transform f(x) = J ” dk (ke) exp (ikx) (C5) The x and k integrations are along the real axes. The Laplace transform conven- tions for functions of time f are Say = [7 at f(0) exp (iat) (C6) F with inverse transform dw Ko =f FE fleyexp cian) cn where the Laptace contour L is a horizontal line in the complex w-plane that must pass above all singularities of f(w). The Laplace transform (C.6) can be considered as the Fourier transform of the function 7({#) such that 7(t) = f(t) fort > 0 and F(t) = 0 for t < 0. Then for + < 0 the inverse Laplace transform (C.7) can be closed upward [since exp (—iwt) ~ exp (—|t|w,) ~ 0 for w; + +], yielding F() = O fort < 0 since the Laplace contour passes above all singularities of f(w). Consider the solution of the differential equation q 4 = af (C.8)Pedestrian’s Guide to Complex Variables 281 with f(¢ = 0) = fo. The Laplace transform of the left side is £ a ane exp (fet) = f(1) exp (ia)|~ - taf a S(t) exp (dwt) = f(1) exp (iat)|” — iw f(o) (C9) ° Without knowing the function /(/), we can only say that this integral is defined for w; large enough, for only then is f(t) exp (‘t)| -.. equal to zero. How large is large enough remains to be seen. For large enough w,, Eq. (C.8) has the Laplace transform - ft = 0) ~ iwf(w) = a fla) (C.10) so that _ ft = 9 BAC) Beaerremnera (C1) The inverse transform is gy =f % = op cian (C.12) 2m wie where the contour must be placed high enough in the wplane so that f(w) is defined (the shaded region in Fig. C.2). Once the contour is drawn in the shaded region of Fig. C.2, it can be moved around only if f(a) is analytically continued to the remainder of the complex w-plane. An ana/ytic function is one that is differen- tiable (the derivative in the complex plane at the point z does not depend on which direction the point is approached from). The analytic continuation of a simple function like f(w) in (C.11) from the shaded region in Fig. C.2 (the only region where (C.10) is defined} to the entire orplane is easy; it is just the function so) = 9 (C.13) Fig. C.2- Inverse Laplace contour must be drawn initially high in the @-plane,282 Appendix C itself. This function is now analytic in the entire w-plane except at the point w = ia, The contour in (C.12) can now be closed by a large semicircle in the lower-half w-plane, since f(w) is defined everywhere. The result is 2rilé ray = - =O. Fa = oyexp (an 7 = f(t = 0) exp (at) (C.14) which is the desired result. In retrospect, we can now see that (C.9) converges for @, > a. This is why the inverse Laplace contour must be drawn above ail singular- ities of f(w) in the w-plane. Analytic continuation is not always quite as simple as in (C.13). Consider il dx r= f" ats (C)5) defined for z,; > Owitha > 0; the integration is along the real x-axis. Closing the contour either up or down (Fig. C.3), we find 1 a zti (C16) f(z) = for z, > 0. We canmot use (C.15) as the analytic continuation of f(z) for z; < 0, for then (C.15) yields _ 1 2n $0) Tia F+e@ ea for z; < 0. The function f(z) defined by (C.16) and (C.17) is discontinuous at z; = 0 and so is not analytic. In order to properly analytically continue (C.15), one must subtract the extra term in (C.17) that Jeads to the discontinuity, and write ft dx Qi sea) =f @tae@-a tte (C18) Fig. C.3_ Evaluation of f(z) for z, > 0 in (C.15).Pedestrian's Guide to Complex Verlebles 263 Fig. C.4 Integration contour for z, < 0 that gives the proper analytic continuation of (C.15). for z, < 0. The combination (C.15) for z; > 0 and (C.18) for z, < 0 is now an analytic function everywhere except at the pole of (C.16),z = — ia. Alternatively, one can deform the contour in (C.15) as shown in Fig. C.4 for z; < 0, and write ix @ F(a) =f @+a@—-_ (C.19) The contour C’ is as shown in Fig. C.4 for z, < 0, and is along the real x-axis for z, > 0. This gives the same result as the form (C.18) for z; < 0. A useful formula for Fourier transformation is f & exp (-ikx) = &(k) (C.20) This formula can be demonstrated by multiplying each side by an arbitrary func- tion f(k) and integrating over all k. The right side yields f(k = 0), while the left side is Ji ak p(k) f £ exp (“ikx) “f ZS dk f(k) exp (-ikx) =[" $1 = f(k = 0) (C.21) where it has been assumed that the order of integrations can be reversed. Since the right and left sides of (C.20) yield the same result, the identification (C.20) must be correct. Another useful formula concerns integrals of the form aa a 1 = lim Sie x-asim (C22)284 Appendix C atin Fig. C.5__ Integration contour leading to the Plemelj formulas where the integral is along the real x-axis, anda > 0. For the lower sign, the pole is at x = a + ilm|, and the integral can be performed by slightly deforming the contour as shown in Fig. C.5. This leads to r= Pf” ae 4 + wi (€.23) where the semicircle in Fig. C.5 contributes half of 2a/, and where Pf. aque et +S (C.24) Formally, one writes : ! es 1 ne lim soot 7? (gAg) + rte a) (C25) which when integrated over x yields (C.23). For the upper sign in (C.22), the pole approaches the integration contour from below, the integration contour must be deformed upward rather than downward, and the sign of the imaginary contri- bution changes. The general formula is finally 1 =p{—- 16x — praeaa 7? (gAq) = me 2) (C.26) lim. 0 Other properties of complex variables are explored throughout the book. REFERENCES (11. S. Sokolnikoff and R. M. Redheffer, Mazhematics of Physics and Modern Engineering, 2nd edition, McGraw-Hill, New York, 1966. (2], G. Arfken, Mathematical Methods for Physicists, Academic, New York, 1970.APPENDIX D Vector and Tensor Identities The following vector and tensor identities are useful in the study of plasma physics a). A+ (BX C) = (A X B)-C =B-(C X A) = (BX C)-A = C-+(A XB) = (CX A)-B (D.1) A x (BX C) = (A: OB — (A BIC (D.2) (A X B)+(C X D) = (A+ CB -D) — (A+ DYB-C) (D.3) VUE) = fVg + evs (D4) \ VGA) S=SVA FAVS (D.5) VX (fA) =fVKATUSXA (D.6) V-(A XB) =B+(V x A) — A+(V XB) (D.7) Vv x (A X B) = A(V-B) — B(V-A) + (B- VA —(A- VB (D.8) ViA-B) = A X (VX B) + BX (VX A) + (A> VB + (B- VA (D.9) VA = WV+A)— Vx (VX A) (D.10) vx vf=0 (D.11) V-(V KX A=0 (D.12) V+ (BA) = A(V-B) + (B+ VA (D.13) VisSH=V/A-T+ys/0-F (D.14) REFERENCE [1] D. L. Book, Plasma Formulary, Naval Research Laboratory, Washington, D.C., 1980.