Statistical Mechanics Lecture Notes (2006), L2
Statistical Mechanics Lecture Notes (2006), L2
Statistical Mechanics Lecture Notes (2006), L2
The most famous example of how entropy increases in a real system is the dilute classical gas. The rst assumption we will make is that = na3 1, (1)
or that the gaseousness parameter is small. Here n is the density of particles per unit volume and a is the range of the interparticle interaction: is just the number of particles within an interaction range of a xed particle. For hard spheres, a is just the sphere radius. For Van der Waals-type forces between neutral atoms or molecules, the interaction potential falls o sucently rapidly (usually V (r) r6 ) that our approximations are still well justied. One reason why plasma physics is dicult is that there the interparticle potential is given by the rather strong Coulomb force, V (r) r1 , so that the physics is quite complicated and the Debye screening must be taken into account. The second assumption is that the particles lack internal rotational or vibrational structure. At room temperatures, this is a reasonable assumption for the vibrational levels of many gases. Rotations may be quite low energy, but in practice molecules often rotate so fast (i.e., the rotational levels are at such low energy) that they are simply averaged over on the intercollision time scale. Now we need a model for how the particles in the gas collide. Let 1 , 1 , 2 , and 2 be the momentum variables of two incoming and two outgoing particles from a collision: = (p). We write instead of p so that the formulas can be generalized to cases where internal angular momenta, for example, are conserved in collisions as well as normal momenta; then = (p, L, . . .). We consider collisions that take 1 , 2 to 1 , 2 . Then dene w(1 , 2 ; 1 , 2 ) so that the rate per spatial volume at which pairs of particles from 1 , 2 get transferred via collisions to 1 , 2 is R = (f (t, r, 1 ) d1 )(f (t, r, 2 ) d2 )w(1 , 2 ; 1 , 2 )(d1 d2 ). (2)
In the above we ignored the position variables in w and are assuming that the collision process is translation invariant. There will be a bit more said about the assumptions which go into the denition of f below. The picture is that we are obtaining the collision function w from some microscopic physics like quantum mechanics. That physics will probably be time-reversal invariant, in which case w(1 , 2 ; 1 , 2 ) = w((1 )T , (2 )T ; (1 )T , (2 )T ). (3)
For the monatomic gas to be considered below, which has no additional angular momentum or other internal variables so T = (p) and also has parity invariance, (3) becomes just w(1 , 2 ; 1 , 2 ) = w(1 , 2 ; 1 , 2 ) = w(1 , 2 ; 1 , 2 ). (4)
In order to understand the meaning of w in more familiar terms, note that in three dimensions f d has units of 1/volume, so w d1 d2 must have units of volume/time. Now w(1 , 1 ; 2 , 2 ) d1 d2 = d. |v1 v2 | 1 (5)
The left side is a quantity with dimensions of area, and is equal to the dierential cross section to send particles into the momentum ranges (d1 , d2 ), which you may be familiar with in singleparticle scattering. Classically, this is just the target area in the relative coordinate system of the two particles. Consider the time evolution of a gas of particles with no collisions or external forces. Each particle moves with constant velocity, and the distribution function along phase space paths is constant: df = t f + v x f = 0. (6) dt The convective derivative above follows the distribution function along the paths of a particle. For an example, consider one particle moving on a line. If at t = 0 there is a -function for the one-particle distribution function to be at position x0 with momentum p0 , f (t = 0, x, p) = (x x0 )(p p0 ), then after a time t the new distribution function is f (t, x, p) = (x (x0 + pt/m))(p p0 ), which is easily shown to satisfy (6). When the particles are not moving freely but under the action of an external one-body force, (6) becomes just df = t f + v x f + F p f = 0. (9) dt Here F comes from some force law that acts on each particle independently. Such forces will still not increase the entropy: the distribution function f is still convected along particle trajectories, even though the particle trajectories are more complicated. There is no approach to equilibrium yet. Now lets get rid of the one-body potential for brevity and add a two-body collision term: df = t f + v dt
xf
(7)
(8)
= C(f ).
(10)
What is C(f )? Well, the variation it induces in f (t, x, p) will contain two parts. The rst part will be losses because of particles which are scattered from p to some other momentum p . The second part will be gains because of particles which are scattered to p from some other momentum p . From our denition of w above, we know that the total number of collisions per time in a volume dV that take particles out of a momentum range dp1 is dV dp1 w(p1 , p2 ; p1 , p2 )f (t, r, p1 )f (t, r, p2 ) dp2 dp1 dp2 . (11)
Likewise, assuming I have the indices and signs in the right place, the rate of collisions that bring particles into the momentum range dp1 is dV dp1 w(p1 , p2 ; p1 , p2 )f (t, r, p1 )f (t, r, p2 ) dp2 dp1 dp2 . (12)
For the simple monatomic gas, w = w from (4), but postpone using that for now. Now we can write out the collision term as C(f ) = (w f1 f2 wf1 f2 )dp2 dp1 dp2 . (14)
Here for brevity f1 f (t, r, p1 ) and so forth for p2 , p1 , p2 . Later we will discuss the collision term in an explicit case. For the monatomic gas, we can use w = w so C(f ) = w (f1 f2 f1 f2 )dp2 dp1 dp2 . (15)
It is a nontrivial exercise to derive (15) from (14) without time-reversal invariance, so that w = w no longer holds. Finally, pulling everything together, we obtain t f1 + v
x f1
(16)
This is the celebrated Boltzmann equation, derived in 1872 (!). Now lets look at a rather fundamental prediction of this equation. The entropy for a oneparticle distribution function f is, as shown below, S= f log f dd x dd p. (17)
Boltzmann actually used a denition diering by a constant and a sign: his H = N S, which you may see in some textbooks. Assume for simplicity a monatomic gas, so that w = w , although this assumption is not fundamental. Our goal is to show that the Boltzmann equation predicts dS 0, dt or that entropy cannot decrease. Let me pause to justify the entropy denition (17) of a probability distribution. We will show that S = i pi log pi for a discrete distribution of probabilities pi ; then (17) is simply the generalization to a continuous distribution of probabilities. Consider, as in the previous lecture, a large number N of identical systems which can be in any of k dierent states. Let i be the number of systems in state i, with i = N. (19)
i
(18)
The number of microstates of the N systems which correspond to a given set (1 , . . . , k ) is P N! . 1 !2 ! . . . k ! 3 (20)
The fundamental assumption here is that every microstate is equally likely. Now using Stirlings approximation log P = N (log N 1) i (log i 1). (21)
i
In using Stirlings approximation we are assuming that N is large but not that k is large: k may be as small as 2, as for a coin ip, and the above is still valid. To be more precise, we are assuming that N is so large that any congurations in which one of the i is small are very rare. Then Stirlings approximation is justied not just for N ! but also for i !. We can immediately cancel two terms because the sum of the i is just N (19). Doing so and introducing the probabilities pi i /N gives log P = N log N
i
N pi log(N pi ) = N (log N
i
(22)
Once again two terms can be cancelled via the sum condition (19). Finally, the entropy for one system is dened to be log P S pi log pi , (23) = N i the promised formula. The only subtleties in the generalization of this to continuous functions f arises for singular functions like the -function. For nonsingular functions on phase space, just divide phase space into a large number of small boxes dV and interpret f dV /N as the probability that a given particle is in dV to use the above. The time derivative of the entropy is then dS dt = = t (f log f ) dx dp (1 + log f )t f dx dp. (24)
The only change in the entropy occurs from the collisional term, not the convective terms (mathematical verication of this fact is left to the reader). So the entropy is now dS dt = = (1 + log f1 )C(f1 ) dp1 dx (1 + log f1 )(w(f1 f2 f1 f2 )dp2 dp1 dp2 ) dp1 dx. (25)
Now imagine shuing the variables p1 p2 , p1 p2 on the right-hand-side, without changing its value, and adding that to the above. Then 2 dS = dt (2 + log f1 f2 )w(f1 f2 f1 f2 )dp2 dp1 dp2 dp1 dx. (26)
Now do another shue, of incoming and outgoing momenta (i.e. primed and unprimed variables) and add the result to the above: 4 dS dt = (2 + log f1 f2 )w(f1 f2 f1 f2 )dp2 dp1 dp2 dp1 dx
(27)
Canceling terms in the above leaves simply 1 dS = dt 4 w(log f1 f2 log f1 f2 )(f1 f2 f1 f2 ) dp1 dp2 dp1 dp2 dx. (28)
The last step is just showing that (log f1 f2 log f1 f2 )(f1 f2 f1 f2 ) cannot be negative. This is a function of two variables x = f1 f2 and y = f1 f2 , which simplies to g(x, y) = (x y) log(x/y). All the f functions are probability distributions and hence nonnegative, so x and y are nonnegative and g(x, y) = y(x/y 1) log(x/y) will be nonnegative if g (x/y) = (x/y 1) log(x/y) is nonnegative. Now write u = x/y, so g (u) = (u 1) log u. Finally, the function (u 1) log u is always nonnegative for positive u, as the two parts u 1 and log u always have the same sign. What property was assumed about the two-particle distribution function in the form of w in order to obtain the Boltzmann equation, which is closed? Answer: that it was a product distribution, i.e., particle momenta were independent except for collisions: F (t, r, p1 , r, p2 ) = f (t, r, p1 )f (t, r, p2 ). (29)
This is sometimes known as the assumption of molecular chaos. The rst part of the next lecture will show how to relax this assumption and obtain the BBGKY hierarchy of evolution equations. Another question to be addressed in the next lecture is how the above is consistent with what we know about the deterministic evolution of Hamiltonian systems.