large deviations in stochastic...

36
Large deviations in stochastic thermodynamics Andreas Engel University of Oldenburg http://www.statphys.uni-oldenburg.de Thermodynamics is a funny subject. The first time you go through it, you don't understand it at all. The second time you go through it, you think you understand it, except for one or two small points. The third time you go through it, you know you don't understand it, but by that time you are so used to it, so it doesn't bother you any more.“ (Arnold Sommerfeld)

Upload: others

Post on 19-Aug-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Large deviations in stochastic thermodynamicsAndreas Engel

University of Oldenburghttp://www.statphys.uni-oldenburg.de

„Thermodynamics is a funny subject. The first time you go through it, you don't understand it at all. The second time you go through it, you think you understand it, except for one or two small points. The third time you go through it, you know you don't understand it, but by that time you are so used to it, so it doesn't bother you any more.“ (Arnold Sommerfeld)

Page 2: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Traditional thermodynamics

Page 3: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

The balances of energy and entropy

S ! 0, @S/@X ! 0 for T ! 0

dU = dW + dQ

dS = diS + deS, diS � 0

Page 4: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Small Systems

W ' �F ' kBT, �S ' kB

Page 5: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Features of small system thermodynamics

N ⇠ 1023 N = 7

−1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

W

P

−1 0 1 2 3 4 5 60

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

0.5

W

P

W � �F hW i � �F

1. Fluctuating thermodynamic quantities,in particular in non-equilibrium processes.

2. Strong coupling to the reservoir(s).

3. Information acquired in measurements becomes thermodynamically relevant.

W ' �F ' kBT, �S ' kB

Page 6: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

System with Reservoirs

SHeat

Inf

Work

Heat ! ! !

!

T1 T2

(T = 0)

(T = 1)

E

SS

S

E E

Page 7: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Engine

SHeat

Inf

Work

Heat

!

T1 T2

(T = 0)

(T = 1)

E

SS

S

E E�!�!

�!

Page 8: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Refrigerator

SHeat

Inf

Work

Heat

!

T1 T2

(T = 0)

(T = 1)

E

SS

S

E E�!

�!

�!

Page 9: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Maxwell’s Demon

SHeat

Inf

Work

Heat !T1 T2

(T = 0)

(T = 1)

E

SS

S

E E

�!�!

�!

Page 10: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Eraser

SHeat

Inf

Work

Heat !T1 T2

(T = 0)

(T = 1)

E

SS

S

E E

�!

�!

�!

Landauer limit: �S � kB ln 2

Page 11: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Information driven heat pump

SHeat

Inf

Work

HeatT1 T2

(T = 0)

(T = 1)

E

SS

S

E E�!

�!

�!

Page 12: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Quantitative analysis: Langevin dynamics

• select relevant degrees of freedom• subsume the rest into a heat bath• model the interaction with the bath by friction and noise (FDT)

works nicely if timescales separate, Here: overdamped version

x = �V

0(x,�) +p

2/� ⇠(t) h⇠(t)⇠(t0)i = �(t � t

0)

Page 13: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Continuous stochastic processes

PT [x(·)] = NT [x(·)] exp ��

4

Z T

0dt⇣˙

x� f(x, t)⌘2!

Stochastic differential equation:

Fokker-Planck equation:

Path measure in function space:

x = f(x, t) +

r2

�⇠(t) h⇠i(t)⇠j(t0)i = �ij�(t� t0)

@tP (x, t) = �r✓f(x, t)P (x, t)� 1

�rP (x, t)

0 0.2 0.4 0.6 0.8 1t

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

x

Page 14: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Stochastic thermodynamics

First Law of thermodynamics for a single fluctuating trajectory.(Sekimoto, 1994)

Work and heat become stochastic variables .

What are their distributions?

W [x(·)], Q[x(·)]

Let with some protocol (driven system).�(t)f(x, t) = �rV (x,�(t))

dU = dV =@V

@�

d� +@V

@x

dx = dW + dQ

Change of energy of the system:

Page 15: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Transformation of probability

PT [x(·)] = NT [x(·)] exp ��

4

Z T

0dt⇣˙

x+rV (x,�)⌘2!

What to do with it?

P (W ) :=

Z (xT ,T )

(x0,0)Dx(·)PT [x(·)] �(W �W [x(·)])

P (Q) :=

Z (xT ,T )

(x0,0)Dx(·)PT [x(·)] �(Q�Q[x(·)])

W [x(·)] =Z T

0dt

@V

@�(x(t),�(t)) �(t)

Q[x(·)] =Z T

0dt rV (x(t),�(t)) · x(t)

Page 16: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Time inversion

−1.5 −1 −0.5 0 0.5 1 1.50

0.5

1

1.5

x

V0,V

1

Forward process

−1.5 −1 −0.5 0 0.5 1 1.50

0.5

1

1.5

x

V0,V

1

Backward process

−1.5 −1 −0.5 0 0.5 1 1.50

0.5

1

1.5

x

V0,V

1

Forward process

−1.5 −1 −0.5 0 0.5 1 1.50

0.5

1

1.5

x

V0,V

1

Backward process

Reverse process: , mirror trajectory: x(t) := x(T � t)�(t) := �(T � t)

Page 17: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

The detailed fluctuation theorem

PT [x(·)]¯PT [¯x(·)]

=

NT [x(·)] exp✓��

4

R T0 dt

⇣˙

x+rV (x,�)⌘2◆

NT [¯x(·)] exp✓��

4

R T0 dt

⇣˙

¯

x+rV (

¯

x, ¯�)⌘2◆

=

NT [x(·)] exp✓��

4

R T0 dt

⇣˙

x+rV (x,�)⌘2◆

NT [x(·)] exp✓��

4

R T0 dt

⇣� ˙

x+rV (x,�)⌘2◆

= exp

��

Z T

0dt ˙x ·rV (x,�)

!= e���Q

Exact for arbitrarily large deviations from equilibrium!

Page 18: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Boundary terms

Include distribution of initial and final states.

Start in equilibrium: p[x(·)] = p0(x0)PT [x(·)] =1

Z0e��V0 PT [x(·)]

p[x(·)]p[¯x(·)] = exp[�(F0 � V0)� �(F

T

� VT

)� ��Q] = exp[�(�V ��Q��F )] = e�Wdiss[x(·)]

Start in any distribution (Seifert,2005):

�Sm[x(·)] :=�Q[x(·)]

T�s[x(·)] := ln

p0(x0)

pT (xT )�S[x(·)] := �Sm[x(·)] +�s[x(·)]

p[x(·)]p[x(·)] = e�S[x(·)]/kB

Page 19: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Consequences

p[x(·)]p[x(·)] = e�S[x(·)]/kB

0 X

f

X X

f(<X>)

<X>

<f(X)>

1 2

• The „emergence of irreversibility“:

• The integral fluctuation theorem:

• The Second Law as equality:

• Three faces of the Second Law:

�S[x(·)] ⇡ kB ! �S[x(·)]� kB

h�Si � 0

he��S[x(·)]/kB i = 1

S = Sa + Sna �! he��Sa[x(·)]/kB i = 1 , he��Sna[x(·)]/kB i = 1

Page 20: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

More consequences

• Jarzynski equality (1997): equilibrium information from non-equilibrium processes.

• efficiency of molecular motors: running reliably forward.

• chemical thermodynamics at the molecular level.

• statistical mechanics very far from equilibrium.

• non-equilibrium steady states: house-keeping heat, linear response, Onsager reciprocity.

he��W i = e���F

j 6= 0

Page 21: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Microcanonical perspective(Cleuren et a., Phys. Rev. Lett. 96, 050601 (2006))

W

−W

E E+W

E+WE

P(W)=

P(−W)=

Page 22: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

(Cleuren et a., Phys. Rev. Lett. 96, 050601 (2006))

P(−W)= = =P(W) e ∆ S/ kB

W

−W

E E+W

E+WE

P(W)=

P(−W)=

Microcanonical perspective

Page 23: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

More consequences

• Jarzynski equality (1997): equilibrium information from non-equilibrium processes.

• efficiency of molecular motors: running reliably forward.

• chemical thermodynamics at the molecular level.

• statistical mechanics very far from equilibrium.

• non-equilibrium steady states: house-keeping heat, linear response, Onsager reciprocity.

he��W i = e���F

j 6= 0

Page 24: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Why not Gibbs? - A tale of tails

Large deviations become important in statistical mechanics.

Prob(�S/kB ��) e

��

−2 0 2 4 6 8 100

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Stot /N

Prob

N ≅ 1N ≅ 100

dominant trajectories in are atypical for .P [x(·)]he��S/kB i

Page 25: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Large deviations

standard statistical physics: small fluctuations O(1/pN) �! hf(W )i ' f(hW i)

he��W i ' e��hW i �! hW i ' �F WRONG!

Why? Rare events contribute substantially to averages!

Mathematical framework: Large deviation theory

0 0.5 1 1.5 2w

0

0.5

1

1.5

2

2.5

3

3.5

4

I

P (W ) = e�NI(w)+o(N)

I(hwi) = 0, I(w) � 0

I(w) ⇡ I 00

2(w � hwi)2

Page 26: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Transformation of rare probabilities

P

x

(x) ⇠ e

�NI

x

(x)

x, P

x

(x); y = f(x) �! P

y

(y) =?P

y

(y) =

ZdxP

x

(x) �(y � f(x))

�! P

y

(y) =

Zdx e

�NI

x

(x)�(y � f(x))

saddle-point approximation:

Py(y) ⇠ e�NIy(y)

contraction principle

I

y

(y) = minx:y=f(x)

I

x

(x)

Page 27: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

The gist of it

Page 28: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Asymptotics of work distributions

−4 −2 0 2 4 6 8 10 120

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

W

hist P(W)P(W)e−β W P(W)

D. Nickelsen, A.E., Eur. Phys. J. B, 82,207 (2011)

Combine analytical information on the tail of P(W) with the histogram.

• is not known exactly.• Asymptotics for small is crucial.• This region is badly sampled.

P (W )

W

Page 29: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

The idea

−15 −10 −5 0

10−6

10−4

10−2

100

W

P(W

)

atmpar: t0 = 0 ; t1 = 100 ; r = 0 ; β = 1 ; M = 5e7 ; RelTol = 1e−07 % ; bins = 100 ; jakob = 3 ; N = 1 ; n = 5e7 ; S = 100 ; W* = −10..−2

SimulationAsymptotik

1

−7 −6 −5 −4 −3 −2 −1 00

0.1

0.2

0.3

0.4

0.5

0.6

W

P(W

)

atmpar: t0 = 0 ; t1 = 100 ; r = 0 ; β = 1 ; M = 5e7 ; RelTol = 1e−07 % ; bins = 100 ; jakob = 3 ; N = 1 ; n = 5e7 ; S = 100 ; W* = −10..−2

SimulationAsymptotik

1

Crucial: existence of an overlap

Get analytical information about the tail of .Combine it with the histogram in integrals.

P (W )

Page 30: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Method of optimal fluctuation

Contraction principle: The probability of an unlikely event is dominated by the probability of its most probable cause.

Here: Tails of are dominated by maximizing under the constraint .

Formally: and saddle-point approximation for functional integral.

P (W ) x(·)W [x(·)] = W

� !1

Includes contributions from the optimal trajectory and its neighbourhood.

P (W ) =Z

dx0 p0(x0)Z

dx

f

(xf ,tf )Z

(x0,t0)

Dx(·) P [x(·)] �(W �W [x(·)])

P (W ) =e��S[x(·)]

Z0

pdetM/�

(1 +O(1/�))

P [x(·)]

Page 31: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

The optimal trajectory

�S

�x

= 0 �!

@S

@q= 0 �!

P (W ) = NZ

dx0

Z0

Zdx

T

Zdq

4⇡/�

x(T )=xTZ

x(0)=x0

Dx(·) e

��S[x(·),q]

0

x

ttf

S[x(·), q] = V0(x0) +TZ

0

dt

h14(x + V

0)2 +iq

2V

i� iq

2W

One solution of ELE for each value of the work combining unlikely initial conditions with strange realization of the noise.

¨x + (1� iq) ˙

V

0 � V

0V

00 = 0˙x0 � V

00 = 0,

˙xT + V

0T = 0

W =Z T

0dt ˙V

Page 32: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

The pre-exponential factor

• Contributions from quadratic neighbourhood of the optimal trajectory.• Includes neighbourhood of initial and final points.• Constraint suppresses fluctuations orthogonal to it.

A := � d2

dt2+ (V 00)2 + V 0V 000 � (1� iq) ˙V 00

V 000 'n(0)� 'n(0) = 0, V 00

T 'n(T ) + 'n(T ) = 0

dn :=Z T

0dt 'n(t) ˙V 0(t)

P (W ) =Np2Z0

e��S

qdetA h ˙V 0|A�1| ˙V 0i

�1 +O(1/�)

X

n

d2n

�n= h ˙V 0|A�1| ˙V 0i

Determine eigenvalues and eigenfunctions of the Hessian, as well as the projections of on the gradient of the constraint:

�n, 'n'ndn

Page 33: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

The breathing parabola

−20 −15 −10 −5 0 5 10 15 200

5

10

15

20

25

30

35

40

x

V

V0

V1•

• Experimentally accessible.

• is not Gaussian.

• is not known analytically.

P (W )

V (x, t) =k(t)2

x

2

P (W )

P (W ) =NZ0

px0 xT

|W | e

��iq2 |W | �1 +O(1/�)

�⇠ C1

s�

|W | e

�� C2 |W |

Exact results for the asymptotics (for all protocols):

Page 34: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

−7 −6 −5 −4 −3 −2 −1 00

0.1

0.2

0.3

0.4

0.5

0.6

W

P(W

)atmpar: t0 = 0 ; t1 = 100 ; r = 0 ; β = 1 ; M = 5e7 ; RelTol = 1e−07 % ; bins = 100 ; jakob = 3 ; N = 1 ; n = 5e7 ; S = 100 ; W* = −10..−2

SimulationAsymptotik

1

−15 −10 −5 0

10−6

10−4

10−2

100

W

P(W

)

atmpar: t0 = 0 ; t1 = 100 ; r = 0 ; β = 1 ; M = 5e7 ; RelTol = 1e−07 % ; bins = 100 ; jakob = 3 ; N = 1 ; n = 5e7 ; S = 100 ; W* = −10..−2

SimulationAsymptotik

1

−20 −15 −10 −5 00

0.5

1

1.5

2

2.5

3

W

P(W

) ⋅e−

W

atmpar: t0 = 0 ; t1 = 100 ; r = 0 ; β = 1 ; M = 5e7 ; RelTol = 1e−07 % ; bins = 100 ; jakob = 3 ; N = 1 ; n = 5e7 ; S = 100 ; W* = −10..−2

SimulationAsymptotik

1

−25 −20 −15 −10 −5 010−3

10−2

10−1

100

101

102

103

W

P(W

) ⋅e−

W

atmpar: t0 = 0 ; t1 = 100 ; r = 0 ; β = 1 ; M = 5e7 ; RelTol = 1e−07 % ; bins = 100 ; jakob = 3 ; N = 1 ; n = 5e7 ; S = 100 ; W* = −10..−2

SimulationAsymptotik

1

Comparison with simulations

Work distribution in quasi-static processes always Gaussian?

Page 35: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Entropic saddle-points

1.0

U(x)/(f2 c/4)

x

f∗ = 1/2

(a)

0f∗ = 1

0.2

0.4

f∗ = 2

0.6

0.8h(σ)

1

-1

ARW

0 1

h∗(σ)

2 3

h(σ)

σ

(b)

-4

-2

0

2

0.0 0.2 0.4 0.6 0.8

f∗ = 0

10time τ/τ0

0

1

2

0 1 2 3potential

0

1

2

-20 -10 0

positionx

Driving a colloidal particle around a periodic potential. Large deviation function for the entropy production .�

For a single optimal trajectory dominates.For an ensemble of nearly optimal trajectories dominates.

|�| � 1|�| ⌧ 1

T. Speck, A.E., U. Seifert, JSTAT, P12001 (2012)

Page 36: Large deviations in stochastic thermodynamicspperso.th.u-psud.fr/.../files/hh12/Andrea_talk_1.pdf · 2017. 6. 12. · Continuous stochastic processes P T [x(·)] = N T [x(·)] exp

Thank you!

C. Jarzinsky, Ann. Rev. Condens. Matter Physics 2, 329 (2011)U. Seifert, Rep. Prog. Phys. 75, 126001 (2012)J. M. Parrondo et al., Nature Physics, 11, 131 (2015)

Recent reviews: