Professional Documents
Culture Documents
Entropy and H Theorem: The Mathematical Legacy of Ludwig Boltzmann
Entropy and H Theorem: The Mathematical Legacy of Ludwig Boltzmann
C
edric Villani
University of Lyon
& Institut Henri Poincare (Paris)
R3v
3
X
i=1
S2
f
vi
= Q(f, f ) =
xi
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
R3v
3
X
i=1
S2
f
= Q(f, f ) =
vi
xi
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
R3v
3
X
i=1
S2
f
vi
= Q(f, f ) =
xi
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
R3v
3
X
i=1
S2
f
vi
= Q(f, f ) =
xi
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
v + v |v v |
v =
2
2
v + v |v v |
v =
2
2
(v, v ) (v , v )
R3v
3
X
i=1
S2
f
vi
= Q(f, f ) =
xi
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
R3v
3
X
i=1
S2
f
vi
= Q(f, f ) =
xi
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
R3v
3
X
i=1
S2
f
vi
= Q(f, f ) =
xi
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
Conceptually tricky!
Molecular chaos: No correlations between particles
...
However correlations form due to collisions!
f X f
vi
= Q(f, f ) =
+
t
xi
i=1
R3v
S2
h
i
B(vv , ) f (t, x, v )f (t, x, v )f (t, x, v)f (t, x, v ) dv d
Conceptually tricky!
One-sided chaos: No correlations on pre-collisional
configurations, but correlations on post-collisional
configurations (so which functional space!?),
and this should be preserved in time,
although correlations continuously form!
... Valid approximation? 100 years of controversy
Nontrajectorial proof??
Qualitative properties?
Fundamental properties of Newtonian mechanics:
preservation of volume in configuration space
What can be said about the volume in this
infinite-dimensional limit??
Boltzmanns H functional
Z
S(f ) = H(f ) :=
f (x, v) log f (x, v) dv dx
x R3v
Information theory
H () =
log d;
= .
fj = 1
fj = 1
N (f ) = 1
fj = 1
8!
8 (f ) =
6! 2!
fj = 1
N!
N (f ) =
N1 ! . . . Nk !
fj = 1
#N (f1 , . . . , fk ) e
fj log fj
X
1
log #N (f1 , . . . , fk )
fj log fj
N
Sanovs theorem
A mathematical translation of Boltzmanns intuition
x1 , x2 , . . . (microscopic variables) independent, law ;
N
X
1
bN :=
xi (random measure, empirical)
N i=1
Sanovs theorem
A mathematical translation of Boltzmanns intuition
x1 , x2 , . . . (microscopic variables) independent, law ;
N
X
1
xi (random measure, empirical)
bN :=
N i=1
Fuzzy writing: P [b
N ] eN H ()
Sanovs theorem
A mathematical translation of Boltzmanns intuition
x1 , x2 , . . . (microscopic variables) independent, law ;
N
X
1
N
b :=
xi (random measure, empirical)
N i=1
Fuzzy writing: P [b
N ] eN H ()
Rigorous writing: H () = lim lim sup lim sup
k
0
N
h
1
j (x1 ) + . . . + j (xN )
j k,
log PN
N
N
j d <
Universality of entropy
Lax entropy condition
Entropy is used to select physically relevant shocks in
compressible fluid dynamics
(Entropy should increase,
i.e. information be lost, not created!!)
Universality of entropy
Voiculescus classification of II1 factors
States
Type II1 factor := infinite-dimensional VN algebra with
trivial center and a tracial state = linear form s.t.
(A A) 0
(I) = 1
(positivity)
(unit mass)
(AB) = (BA)
(A, ): noncommutative probability space
States
Type II1 factor := infinite-dimensional VN algebra with
trivial center and a tracial state = linear form s.t.
(A A) 0
(I) = 1
(positivity)
(unit mass)
(AB) = (BA)
(A, ): noncommutative probability space
Noncommutative probability spaces as macroscopic limits
(N )
(N )
X1 , . . . , Xn
random N N matrices
1
(N )
(P (A1 , . . . , An )) := lim
E tr P (X1 , . . . , Xn(N ) )
N N
may define a noncommutative probability space.
Voiculescus entropy
Think of (A1 , . . . , An ) in (A, ) as the observable limit of
a family of microscopic systems = large matrices
(N, , k) :=
(X1 , . . . , Xn ), N N Hermitian; P polynomial of degree k,
1
tr P (X1 , . . . , Xn ) (P (A1 , . . . , An )) <
N
( ) := lim lim sup lim sup
k
n
1
log vol((N, , k)) log N
2
N
2
Universality of entropy
BallBartheNaors quantitative central limit theorem
X1 , X2 , . . . , Xn , . . . identically distributed, independent
real random variables;
EXj2 < ,
Then
EXj = 0
X1 + . . . + XN
X1 + . . . + XN
increases with N
f (v)f (v )
f (v)f (v ) f (v )f (v ) log
B d dv dv dx
f (v )f (v )
Obviously 0
|vu(x)|2
2T (x)
e
Moreover, = 0 if and only if f (x, v) = (x)
(2 T (x))3/2
= Hydrodynamic state
Loschmidts paradox
At time t, reverse all velocities (entropy is preserved),
start again (entropy increases) but go back to initial
configuration (and initial entropy!)
Loschmidts paradox
At time t, reverse all velocities (entropy is preserved),
start again (entropy increases) but go back to initial
configuration (and initial entropy!)
Boltzmann: Go on, reverse them!
Loschmidts paradox
At time t, reverse all velocities (entropy is preserved),
start again (entropy increases) but go back to initial
configuration (and initial entropy!)
Boltzmann: Go on, reverse them!
... Subtleties of one-sided chaos
... Importance of the prepared starting state:
Loschmidts paradox
At time t, reverse all velocities (entropy is preserved),
start again (entropy increases) but go back to initial
configuration (and initial entropy!)
Boltzmann: Go on, reverse them!
... Subtleties of one-sided chaos
... Importance of the prepared starting state:
low macroscopic entropy (unlikely observed state),
high microscopic entropy (no initial correlations)
Large-time behavior
The state of maximum entropy given the conservation of
total mass and energy is a Gaussian distribution
... and this Gaussian distribution is the only one which
prevents entropy from growing further
0.01
0.001
0.0001
1e-05
1e-06
1e-07
1e-08
0
0.2
0.4
0.6
0.8
1.2
0.1
1.4
1.6
10
12
14
16
...
...
Ricci curvature
One of the three most popular notions of curvature. At
each x, Ric is a quadratic form on Tx M .
(Ric)ij = (Riem)kkij
It measures the rate of separation of geodesics in a given
direction, in the sense of volume (Jacobian)
Distortion
geodesics are distorted
by curvature effects
111
000
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
000
111
location of
the observer
[ Ric 0 ]
t=1
t=0
t = 1/2
S=
log
t=0
t=1
force F [f ](t, x) =
ZZ
W (x y) f (t, y, w) dy dw