Statistical Mechanics Lecture Notes (2006), L3
Statistical Mechanics Lecture Notes (2006), L3
Let us begin by nding what one-particle distributions f are preserved by the Boltzmann equation. The collision term derived in Lecture II for a monatomic gas at one spatial point r is C(f ) = w(f1 f2 f1 f2 )dp2 dp1 dp2 . (1)
This collision term will vanish identically at a given point in space if for all incoming and outgoing momenta with p1 + p2 = p1 + p2 , (2) we have f1 f2 = f1 f2 , or log f (p1 ) + log f (p2 ) = log f (p1 ) + log f (p2 ). (3)
The algebra we went through in proving the H theorem shows that this is actually a necessary condition to obtain dS/dt = 0, assuming that w is nonzero. Now (3) will hold as long as log f is a sum of quantities which are conserved in the collision process. The conservation of momentum and energy in collisions forces some constraints on the collision term C(f ): we have C(f ) dp1 = C(f )p1i dp1 = C(f ) p1 2 dp1 = 0. 2m (4)
Assuming that the only conserved quantities are energy, momentum, and particle number, then we take log f (p) = c0 + c1 p + c2 p2 = c2 (p p0 )2 + C. Here p0 = c1 /2c2 and C = c0 c1 2 /4c2 . Now we can start to recognize the familiar Maxwellian distribution of velocities in a gas. The integral over all momenta should give the local number density, so c2
d/2
(5)
eC = n(r),
(6)
which determines eC = n(r)(c2 /2)d/2 in d dimensions. The local momentum density is n(r)p0 , and the local energy density is n(r) p0 2 d p2 f (p) dp = n(r) + . 2m 2m 4c2 m (7)
It remains to derive c2 , which is best done by constructing the equation of state (cf. Huang). Taking a shortcut and invoking the answer from an undergraduate course or equipartition, we know the mean energy for a gas at rest is dkT /2, so c2 = (2mKT )1 . Finally, the equilibrium distribution given by the collision term is 2 n(r)e(pp0 ) /2mkT f0 (p, r) = . (8) (2mkT )d/2 1
There are still at this point multiple possibilities for equilibrium: it could be that at every point the collision integral vanishes because the distribution is Maxwellian, but the local mean energy and velocity vary. It remains to show that any gradient in local density, momentum, or energy will lead to a current directed so as to eliminate the gradient. Such gradients can survive for quite a long time relative to the collisional time scale, and are best studied using the Navier-Stokes equation to be derived below. We now change gears and detour into classical dynamics to understand the physical content of the Boltzmann equation and how it can be reconciled with the microscopic equations of motion. Recall that the Hamiltonian H = T + V of a classical system (here T is the kinetic energy and V the potential energy) generates the equations of motion for the position and momentum (p, q) for each of N particles through Hamiltons equations: dqi dt dpi dt H pi H = . qi =
(9)
A simple example is the motion of a single 1D particle in a quadratic potential, the harmonic oscillator: p2 kq 2 H = + 2m 2 p = kq q = p/m
(10)
The trajectories in the phase space (p, q) are ellipses of constant energy, around which the particles move clockwise. If one tracks the evolution in time of a region of initial conditions for the harmonic oscillator described above, it is fairly simple to show that the original region is rotated but preserves its volume. We are going to show that this is a general property of Hamiltonian systems, starting with a geometric example in 2D phase space and then formalizing the result as Liouvilles theorem. We want to show that the area element in phase space, dp dq, is preserved by the Hamiltonian ow. Consider the evolution of a region that starts as a rectangle at time t with corners (p0 , q0 ), (p0 + dp, q0 ), (p0 + dp, q0 + dq), (p0 , q0 + dq). At time t + dt, the corners have evolved under Hamiltons equations. The rst corner evolves to (p0 dt H , q0 + dt H ) (p , q ). The second corner evolves q p to (p + dp dt dp p H , q + dt dp p H ). The rectangle becomes a parallelogram to leading order q p in dt, and the fourth corner evolves to (p dt dq q H , q + dq + dt dq q H ). q p The change in area of this parallelogram to order dt can be computed by the following argument: you can check geometrically that the only terms of order dt in the area change are (x0 dy 4 + y0 dx2 ), where x0 = dp and y0 = dq are the original dimensions of the rectangle, dx2 is the change in the x separation between corner 1 and corner 2 as dened above, and dy 4 is the change in the y separation between corner 1 and corner 4 as dened above. This gives dA = dpdqdt( H H dV + ) + O(t2 ) = 0 + O(t2 ) = 0. p q q p dt (11)
This argument suggests correctly that the minus sign in Hamiltons equations is crucial for the volume conservation. Now we will give an alternate derivation of the same result for Hamiltonian systems in any dimension. 2
The above geometric argument is formalized in Liouvilles theorem for the distribution function g(p1 , . . . , pN , q1 , . . . , qN ) on the phase space of N particles moving in d dimensions, dg g N d g g = + qi + pi dt t i=1 qi pi = 0. (12)
Proof: Systems are locally conserved (never appear or disappear), so for any phase space volume V the change in the integral of g satises
V
g dV = t
n vg dS.
V
(13)
Here V is a bounding surface for V in 2dN -dimensional phase space and n is an outward-directed normal. The phase space velocity vector (not the normal velocity) v is 2dN -dimensional, with components v = (p1 , . . . , pN , q1 , . . . , qN ) (14) Equation (13) can be transformed into an integral of the divergence over V ,
V
g dV = t
(vg) dV.
V
(15)
=
i=1 Nd
=
i=1
Now we note than the second derivatives of H cancel since partial derivatives commute, leaving g N d g g + pi + qi = 0, t i=1 pi qi which proves the theorem. Comment I: this theorem justies the intuitive argument given before that df /dt = 0 for singleparticle distribution functions f (p, q) in the absence of collisions, as Liouvilles theorem can be rewritten for a single particle as p f + t m
xf
(17)
+F
pf
= 0.
(18)
Comment II: Note that generalizations of (12) hold for other conserved quantities, such as energy and momentum, that travel with the system in phase space. We will use this fact repeatedly below. Comment III: It should be claried that the conservation of volume in phase space is much more than just conservation of probability. Liouvilles theorem implies, for instance, that a distribution function f which starts at value a for volume V1 and value 0 for the rest of phase space will always have value a for a volume of the same size as V1 , and value 0 elsewhere. Conservation of probability would imply only that the integral of f would be V1 a. 3
At rst glance this seems dicult to reconcile with the Boltzmann equation, which we expect by the H theorem to describe ow to an equilibrium conguration. Before resolving the paradox, lets work out one surprising consequence of Liouvilles theorem. Consider a region of phase space at time 0 and its image at time t. The image must have the same volume as the original region. Now advance to time 2t; the image at time 2t must still have the same volume, and so on. If phase space is of nite volume, as for a spatially bounded system of particles with nite energy (so that momentum is bounded), then eventually two of these images must overlap, which means that some point of phase space must, when mapped under time evolution by some time nt, return very nearly to itself (to an arbitrarily small neighborhood of the initial point). We can actually make this statement much stronger. Let 0 be the union of all the images of for t 0, and let be the union of all the images of for t . Clearly 0 . But think of 0 itself as a set of initial points. Then evolving 0 to time clearly yields , which with Liouvilles theorem implies that 0 and have the same volume. A set which is contained in another set of the same volume must dier from the larger set by a set of volume 0, so almost every point of 0 is also in . In particular, almost every point of is also in (assuming is of nonzero volume). So almost every point of recurs at some time t . The above statement of the Poincare recurrence theorem means, returning to our example of particles that are initially conned to one half of a box, that eventually the particles will all return to that half of the box. How can this be reconciled with the increase of entropy in the Boltzmann equation? Until the 1950s or so, the answer given was simply that the Poincare recurrence time is astronomically long (which is true), and that the Boltzmann equation is simply an approximate description valid for short times. But why does it work so well? To understand why, consider the image of an initial rectangular region of phase space. For the harmonic oscillator, the shape of this image never becomes very complicated, and for an integral number of periods of the oscillator, the image takes exactly its original shape. The harmonic oscillator is, however, atypical in that it is integrable. A consequence of integrability is that nearby trajectories do not separate rapidly with increasing times. However, many problems are now known to be chaotic: nearby trajectories diverge exponentially rapidly with increasing times. As a result, after a long time the image of the initial region must be a very complicated shape. The specic area of dynamical systems dealing with area-preserving maps is often referred to as Hamiltonian chaos because of its importance for Hamiltonian evolution of classical physical systems. The validity of the predictions of the Boltzmann equation is related to the idea of coarsegraining in chaotic systems: as time goes by, the size over which features in phase space average out becomes smaller and smaller. Put another way, if we have only some nite precision in our resolution of phase space, eventually statistical averages such as those contained in the Boltzmann equation will become valid (in practice quite rapidly). Showing that this really happens in a given problem is quite dicult. An example of a chaotic system is the Sinai billiard: a billiard ball moves in the plane in a shape made by cutting a quarter circle o from one corner of a rectangle. A distribution of noninteracting billiard balls starting at nearly the same point in phase space will appear to diuse as a result, even though Liouvilles theorem still holds. Liouvilles theorem will be the basis for the BBGKY hierarchy to be derived below, which can be thought of as justication and improvement of Boltzmanns equation. Suppose we have a problem
H =
i=1
Ui = U (ri ),
(19)
Fi
N j=1,j=i
Kij
pi g
pi m
ri g
(20)
Here g is the distribution function in N -particle phase space. Liouvilles theorem can be restated as + hN (r1 , p1 , . . . , rN , pN ) g = 0, t where the dierential operator hN is dened as
N
(21)
hN (r1 , p1 , . . . , rN , pN ) =
i=1
pi m
ri
+ Fi
pi
pi
1 N + Kij ( 2 i,j=1
pj ).
(22)
Now dene the single-particle distribution function as, with z = (r, p) and g normalized to 1,
N
f1 (t, p, r)
i=1
(p pi )(r ri ) = N
(23)
The factor of N appears naturally since g is symmetric under permutations of the zi . To obtain an equation for the evolution of f1 , we integrate out all but one coordinate in the Liouville equation (21). In general, to nd the evolution equation of fn it is useful to write
n N
hN = hn + hN n +
i=1 j=n+1
Kij (
pi
pj )
(24)
where we have suppressed the coordinate dependence. In the next lecture we will show that the combinatorial factors cancel appropriately and yield
n + hn fn (z1 , . . . , zn ) = t i=1
dzs+1 Ki,s+1
(25)
The set of these equations for all values of n is referred to as the BBGKY hierarchy (BogoliubovBorn-Green-Kirkwood-Yvon).