Paris, December 9 th 2011. fsu-logo Local collisions dynamics Introduced by Prosen-Campbell T L T L...
Transcript of Paris, December 9 th 2011. fsu-logo Local collisions dynamics Introduced by Prosen-Campbell T L T L...
fsu-logo
Large deviations of the current in collisional dynamics
Raphael Lefevere1
1Laboratoire de Probabilites et modeles aleatoires. Universite Paris Diderot (Paris 7).
Paris, December 9 th 2011.
fsu-logo
Collaborations
T. Gilbert, R. Lefevere, Heat conductivity from molecular chaos hypothesis inlocally confined billiard systems. Physical Review Letters (2008) 101, 200601
R.Lefevere, L.Zambotti Hot scatterers and tracers for the transfer of heat incollisional dynamics. Journal of Statistical Physics (2010) 139, 686-713
R. Lefevere, M. Mariani and L. Zambotti, Macroscopic fluctuations theory ofaerogel dynamics. Journal of Statistical Mechanics (2010) L12004
R. Lefevere, M. Mariani and L. Zambotti, Large deviations of the current instochastic collisional dynamics. Journal of Mathematical Physics (2011) 52,033302
fsu-logo
Aerogels
fsu-logo
Local collision dynamics.
fsu-logo
Local collisions dynamics
Introduced by Prosen-Campbell
TL
TL
TR
TR
2b
TL 6= TR, b > 0, b < a < 2b.
(qi, pi)1≤i≤N , qi ∈ [−b, b] et |qi − qi+1| ≤ a.
Ballistic motion+ reflections on the interval’s boundaries.
Interaction if |qi − qi+1| = a (pi = v, pi+1 = v′)→ (pi = v′, pi+1 = v)
fsu-logo
Local collisions dynamics
2b
2b− a
B = {(q1, q2) : q1 ∈ [−b, b], q2 ∈ [−b, b], |q1 − q2| ≤ a}
fsu-logo
Local collision dynamics.
Energy of particle n at time t :
E(n, t)− E(n, 0) = J(n− 1, [0, t])− J(n, ([0, t])
Time-integrated current:
J(n, [0, t]) = − 1
2
Cn(t)Xk=1
[p2n+1(τkn)− p2
n(τkn)]
(τkn)k is the sequence of collision times between particles n and n+ 1. Cn(t)number of collisions up to time t.
fsu-logo
fsu-logo
Fourier law
Define TN : [0, 1]× R+ → R+ and JN : [0, 1]× R+ → R.
TN (x, t) =1
εN
Xi∈Bε(x)
E(i, N2t), Bε(x) = {i : |i
N− x| ≤ ε}
JN (x, t) = N ·1
N2∆t· J([Nx], [N2t,N2(t+ ∆t)]), ∆t arbitrary
Want to show :
Fourier law
When N →∞, ∆t→ 0, ε→ 0, (TN ,JN ) converge in L2 to the unique solution
(T , J ) of
∂tT (x, t) = −∂xJ (x, t)
J (x, t) = −κ(T (x, t))∂xT (x, t)
with κ(T ) ∼ (T )12 and suitable b.c.
Much too hard!
fsu-logo
Stochastic collisions dynamics
Evolution of energy :
E(n, s+ ∆s)− E(n, s) = J(n, [s, s+ ∆s])− J(n− 1, [s, s+ ∆s])
Replace by stochastic model :
1 Thermal baths having temperatures T (n, s), piecewise constant in time
2 Tracer particles carrying energy current J(n, [s, s+ ∆s]) and interactingrandomly with the heat baths
T1 T2 T3 T4 T5 T6
fsu-logo
Stochastic collisions dynamics
TL TR
Update of the particle’s speed with the law (β = T−1):
ϕ(v) = βve−βv22 .
T T
Markov process (q(s), p(s)) with invariant measure :
γ(dq, dp) =
rβ
2π1[0,1](q)e
−β p22
fsu-logo
Renewal process
Waiting times distributed with (β± = T−1± )
ψ−(τ) :=β−
τ3exp(−
β−
2τ2) and ψ+(τ) :=
β+
τ3exp(−
β+
2τ2).
Total time elapsed at the k + 1-st collision at side ± :
S±k := S±0 +kXi=1
(τ±i + τ∓i ), τ±i ∼ ψ±
Renewal processes :N±t = sup{k : S±k ≤ t}
with waiting times distributed with ψ+ ∗ ψ−.
fsu-logo
Convergence results.
Let
Sk =kXn=1
τn
andNt = sup{k ≥ 1 : Sk ≤ t}
Strong law of large numbers and renewal theorem
limt→∞
Nt
t=
1
E(τ), a.s.
limt→∞
ENtt
=1
E(τ)
fsu-logo
Average current
J [0, t] =
N−tXk=1
(v−k )2
2−N+
tXk=1
(v+k )2
2
where v±k ∼ ϕ±(v) = β±ve
−β± v22 .
Proposition
limt→∞
J [0, t]
t=
T− − T+
( π2T−
)12 + ( π
2T+)
12
a.s.
where TL and TR are the left and right temperatures.
Proof :
limt→∞
N±tt
=1
E(τ+ + τ−)=
1
( π2TL
)12 + ( π
2TR)
12
:= κ, a.s.
E((v±1 )2
2) = T±
fsu-logo
Fluctuations of the current
Study LDF of the current I(j, τ, T ):
Pτ,T„J [0, t]
t= j
«∼ e−tI(j,τ,T ), t→∞.
Pτ,T stochastic dynamics with a fixed temperature difference τ = TL − TR,
average temperature T = TL+TR2
.
Theorem
If τ 6= 0 then,
limε↓0
ε−2I(εj, ετ, T ) = G(j, τ, T ) =
8>>>>><>>>>>:
(j−κτ)2
4κT2 if jτ > κτ2
0 if jτ ∈ [0, κτ2]
− jτ2T2 if jτ ∈ [−κτ2, 0]
j2+κ2τ2
4κT2 if jτ < −κτ2,
where κ = ( T2π
)12 .
Note : ε will be N in the diffusive scaling limit below.
fsu-logo
-3 -2 -1 0 1 2 3
0.5
1.0
1.5
2.0
Figure: Plot of G as a function of j for κτ = κT 2 = 1
Gallavotti-Cohen symmetry:
G(j, τ, T )− G(−j, τ, T ) =jτ
2T 2.
fsu-logo
Flat part
Origin:
1 Current :
J [0, t] =
N−tXk=1
(v−k )2
2−N+
tXk=1
(v+k )2
2
2 Large deviations of Nt for a renewal process whose renewal times (τk)k∈N aredistributed with density :
ψ(τ) ∼1
τγ, as τ →∞, γ > 2.
fsu-logo
Flat part
Take a i.i.d sequence (τi)i≥1 distributed with density ψ(τ) and compute
ϕ(λ) = limt→∞
1
tlog E(expλNt)
E(expλNt) =∞Xn=0
P(Nt = n) exp(λn)
Note first that
E(expλNt) ≥ P(Nt = 0) = P(τ1 > t) ∼1
tγ−1, as t→∞
so that
ϕ(λ) ≥ lim inft→∞
1
tlog E(expλNt) ≥ lim
t→∞−
1
tlog tγ−1 = 0
But for λ ≤ 0, one has ϕ(λ) ≤ 0. For λ > 0, ϕ(λ) is the unique solution of theimplicit equation for x ∈ R :
expλ
Z ∞0
e−xτψ(τ)dτ = 1
which is strictly positive.
fsu-logo
Dynamics of heat baths and diffusive scaling
Fix ∆t > 0, tk = k∆t, sk = N2tk, k ∈ N,
1
T (n, s) = T (n, sk), sk ≤ s < sk+1, k ∈ N2 Total energy exchanged (stochastic) between the wall n and n+ 1 at fixed
temperatures over the time interval [sk, sk+1]
J(n, [sk, sk+1]) = . . .
J(n, [sk, sk+1]) is computed with {T (n, sk), T (n+ 1, sk)}.3 Evolution of temperatures
T (n, sk+1)− T (n, sk) = J(n, [sk, sk+1])− J(n− 1, [sk, sk+1]), k ∈ N
fsu-logo
Diffusive scaling
Define TN : [0, 1]× R+ → R+ and JN : [0, 1]× R+ → R.
TN (x, t) = T ([Nx], N2t)
JN (x, t) = N ·1
N2∆t· J(n, [N2tk, N
2(tk + ∆t)]), tk ≤ t < tk + ∆t
We can show :
Proposition
When N →∞, ∆t→ 0, (TN ,JN ) converge in L2 to the unique solution (T , J ) of
∂tT (x, t) = −∂xJ (x, t)
J (x, t) = −κ(T (x, t))∂xT (x, t)
with κ(T ) = ( T2π
)12 and suitable b.c.
fsu-logo
Large deviations : thermodynamic potentials in equilibrium
Take the Ising model : spin variables σi = ±1, i ∈ Zd distributed according toBoltzmann-Gibbs at temperature β−1
µ(σΛ) =e−βHΛ(σΛ)
ZΛ(β, h)
with Hamiltonian:
HΛ(σ) = −X
<i,j>∈Λ
σiσj + hXi∈Λ
σi, +b.c.
Large deviations of 1Nd
Pi∈Λ σi
P(1
Nd
Xi∈Λ
σi ∈ dm) ∼ exp(−NdI(m,β))
I(m,β) is the Legendre transform of the Helmholtz free energy :
F (h, β) = − limN→∞
1
NdlogZΛ(h, β)
Compute the cumulants (correlation functions)
See phase transitions (lack of strict convexity of I(m,β))
fsu-logo
Non-equilibrium and large deviations : diffusive description.
Macroscopic fluctuations theoryOnsager-Machlup 1953Bodineau,DerridaBertini, De Sole, Gabrielli, Jona-Lasinio, Landim
For diffusive systems (i.e TN and JN are related to microscopic energy andcurrent by a diffusive scaling):
P ({TN ' T,JN ' j} on[0, 1]× [0, 1]) ∼ exp[−N I(j, T )]
where I(j, T ) is given by
I(j, T ) =
Z 1
0dt
Z 1
0dx G(j(x, t), ∂xT (x, t), T (x, t))
if j and T satisfy ∂sT (x, s) = −∂xj(x, s), and I = +∞ otherwise.“Almost” all known examples :
G(j, τ, T ) =(j − κτ)2
4κT 2.
Local billard dynamics might be different!
fsu-logo
Fluctuations at finite N
Want to look at :
P“{TN ' T ,JN ' j} on[0, 1]× [0, 1]
”∼ exp[−N I(j, T )]
At finite N , and for each (x, t) ∈ [0, 1]× [0, 1], JN (x, t) is a random variable (andso is TN (x, t)).“Independence” over small space-time windows :
P“{TN ' T ,JN ' j} on[0, 1]× [0, 1]
”=Yk,l
P[J(l, [sk, sk+1])
N2∆t=j(xl, tk)
N]
Current computed with temperatures {T (xl, tk), T (xl+1, tk)} and “compatibilityconditions” for temperatures.Compute :
log P“{TN ' T ,JN ' j} on[0, 1]× [0, 1]
”= ∆t ·N2
Xk,l
1
N2∆tlog P[
J(l, [sk, sk+1])
N2∆t=j(xl, tk)
N]
.Remember : the theorem allows to compute:
limε→0
lims→∞
1
ε21
slog P[
J(l, [sk, sk + s]
s= εj(xl, tk)]
= G(j(xl, tk), ∂xT (xl, tk), T (xl, tk))
if T (xl, tk)− T (xl+1, tk) = ε∂xT (xl, tk).
fsu-logo
Macroscopic fluctuation theory
Putting everything together :
P“{TN ' T , jN ' j}
”∼ exp[−N I(j, T )]
where I(j, T ) is given by
I(j, T ) =
Z 1
0dt
Z 1
0dx G(j(x, t), ∂xT (x, t), T (x, t))
if j and T satisfy ∂sT (x, s) = −∂xj(x, s), and I = +∞ otherwise.
G(j, τ, T ) =
8>>>>><>>>>>:
(j−κτ)2
4κT2 if jτ > κτ2
0 if jτ ∈ [0, κτ2]
− jτ2T2 if jτ ∈ [−κτ2, 0]
j2+κ2τ2
4κT2 if jτ < −κτ2,
G(j, τ, T ) 6=(j − κτ)2
4κT 2!
fsu-logo
Conclusions.
Argument based on the fact that under (local) equilibrium distributions thereare slow particles with sufficiently large probability.
Apply to “tracer models” introduced by Larralde, Mejia-Monasterio, Leyvraz
Draw experimental consequences and observe them in numerical simulations.
Applications to continuous time random walk