Physics 127b: Statistical Mechanics

Linear Response Theory

Useful references are Callen and Greene [1], and Chandler [2], chapter 16.

Task
To calculate the change in a measurement hB (t)i due to the application of a small “field” F (t) that gives
a perturbation to the Hamiltonian 1H = −F (t) A. Here both A and B arePdetermined by the phase
space coordinates denoted rEN (t) , pEN (t) (e.g. B might be the electric current N i=1 (e/m) p
Ei ). The time
N N
dependence is given by the evolution of rE (t) , pE (t) according to Hamilton’s equations. In statistical

mechanics we deal with an ensemble of systems given for example by a known distribution ρ rEN , pEN at
t = 0. The expectation value at a later time t is then
Z



hB (t)i = d rEN d pEN ρ rEN , pEN B rEN (t) ← rEN , pEN (t) ← pEN (1)

where the cumbersome notation rEN (t) ← rEN means we must evaluate B at the phase space coordinate that
evolves from the value rEN at t = 0. For shorthand this will be denoted B (t). We could equivalently follow
the time evolution of ρ through Liouville’s equation and instead evaluate
Z



hB (t)i = d rEN d pEN ρ rEN , pEN , t B rEN , pEN (2)

but the first form is more convenient.

Onsager regression

F(t)
equilibrium
under H 0 + ∆H
ρ= ρH (r N ,p N )

H = H 0 + ∆H
H=H0

First consider the special case of a force F (t) switched on to the value F in the distant past, and then
switched off at t = 0. We are interested in measurements in the system for t > 0 as it relaxes to

equilibrium.
For t ≤ 0 the distribution is the equilibrium one for the perturbed Hamiltonian H rEN , pEN = H0 + 1H ,
with H0 the unperturbed Hamiltonian, i.e. (for a canonical distribution)

e−β(H0 +1H )
ρ rEN , pEN = R (3)
d rEN d pEN e−β(H0 +1H )

1
 R
so that introducing a convenient notation for the integral over phase space T r ≡ d rEN d pEN


T re−β(H0 +1H ) B rEN , pEN
hB (0)i = . (4)
T re−β(H0 +1H )
For t ≥ 0 we let rEN (t) , pEN (t) for each member of the ensemble evolve under the Hamiltonian, now H0
from its value rEN , pEN at t = 0,so that


T re−β(H0 +1H ) B rEN (t) ← rEN , pEN (t) ← pEN
hB (t)i = . (5)
T re−β(H0 +1H )


Note that the integral is over rEN , pEN which we can denote rEN (0) , pEN (0), and 1H = 1H rEN , pEN etc.
It is now a simple matter to expand the exponentials to first order in 1H (F small!)


T re−βH0 (1 − β1H )B rEN (t) ← rEN , pEN (t) ← pEN
hB (t)i ' (6)
T re−βH0 (1 − β1H )
to give  
hB (t)i = hBi0 − β h1H B (t)i0 − hBi0 h1H i0 + O (1H )2 (7)
where <>0 denotes the average over the ensemble for a system for which no perturbation was applied i.e.
ρ0 = e−βH0 /T re−βH0 . Note in this unperturbed system the Hamiltonian remains H0 for all time, so that
one-time averages such as hB (t)i0 are in fact time independent. Finally putting in the form of 1H, writing
δB (t) = B (t) − hBi0 etc., and noticing that A(Er N , pEN ) is equivalent to A(0),



1H = −F A rEN , pEN = −FA rEN (0) , pEN (0) = −FA (0) , (8)

gives for the change in the measurement 1 hB (t)i = hB (t)i − hBi0

1 hB (t)i = βF hδA (0) δB (t)i0 , (9)

where we have written the result in terms of δA = A − hAi0 .

This result which equates the time dependence of the decay of a prepared perturbation to the time
dependence of a correlation function in the unperturbed system proves the Onsager regression hypothesis.

Kubo formula
For a general F (t) we write the linear response as
Z ∞



1 hB (t)i = χAB t, t 0 F t 0 dt 0 (10)
−∞

with χAB the susceptibility or response function with the properties

χAB t, t 0
= χAB t − t 0 stationarity of unperturbed system
0 0
χAB t − t = 0 for t < t causality

. (11)
∗
χ̃AB (−f ) = χ̃AB (f ) χAB t, t 0 real

The causality condition leads to interesting properties in the complex f plane: for the sign convention I am
using for the Fourier transforms, χ̃ (f ) must be analytic, (has no poles) in the upper half plane. We will not
need this result here.

2
 For the step function force turned off at t = 0
Z 0

1 hB (t)i = F χAB t − t 0 dt 0 (12a)
Z−∞
∞
=F χAB (τ ) dτ . (12b)
t

Differentiating then gives the classical Kubo expression


−β dtd hδA (0) δB (t)i0 t ≥ 0
χAB (t) = . (13)
0 t <0

Connection with energy absorption: fluctuation-dissipation

Consider a sinusoidal force F (t) = Re Ff e−2π if t = 21 (Ff e−2π if t + c.c.). The rate of doing work on the
system is “force × velocity” W = F Ȧ i.e.
Z ∞
d


W = F (t) χ t, t 0 F t 0 dt 0 (14)
dt −∞
writing simply χ for χAA . Substituting in the sinusoidal form for the force, and remembering that χ (t, t 0 )
only depends on the time difference τ = t − t 0
 Z 
1 −2πif t d −2π if t ∞ 2π if τ
W = (Ff e + c.c.) e χ (τ ) Ff e dτ + c.c. (15)
4 dt −∞

we recognize the integral as giving the Fourier transform of χ, so that the rate of working averaged over a
period is
2 
W̄ = − 41 2π if Ff χ̃ (f ) − χ̃ (−f ) (16a)
χ̃ 00 (f )
2
= πf Ff (16b)

where it is conventional to write χ̃ = χ̃ 0 + i χ̃ 00 with χ̃ 0 the real part Re χ̃ (not the driviative!) and χ̃ 00 the
imaginary part Im χ̃ , and terms varying as e±4π if t average to zero. Thus the imaginary part of χ̃ tells us
about the energy absorption or dissipation.
But from the definition of the Fourier transform
Z ∞
00
χ̃ (f ) = χ (t) sin (2πf t) dt (17a)
−∞
Z ∞
d
= −β hδA (0) δA (t)i0 sin (2πf t) dt (17b)
dt
0
Z ∞
= β (2πf ) hδA (0) δA (t)i0 cos (2πf t) dt (17c)
0

where we have used the Kubo expression Eq. (13) for χ = χAA and then integrated by parts. Finally we
recognize the integral as giving the spectral density of A fluctuations, so that (including necessary factors)
χ̃ 00 (f )
GA (f ) = 4kB T (18)
2πf
relating the spectral density of fluctuations to the susceptibility component giving energy absorption, i.e. the
classical fluctuation-dissipation theorem.

3
Langevin Force
If we suppose the fluctuations in A derive from a fluctuating Langevin contribution F 0 to the force F we can
write Z ∞



1 hA (t)i = χ t, t 0 F 0 t 0 dt 0 . (19)
−∞
Since the Fourier transform of a convolution is just the product of the Fourier transforms the spectral density
of hA(t)i is given by
GA (f ) = |χ̃ (f )|2 GF (f ) , (20)
and using the expression for GA derived above leads to
 
1 1
GF (f ) = 4kB T Im − . (21)
2πf χ̃ (f )

Instead of the susceptibility we can introduce the impedance Z = F / Ȧ so that

1 1
Z̃ (f ) = − . (22)
2π if χ̃ (f )

Then defining the “resistance” R̃ (f ) = Re Z̃ (f ) we can write the spectral density of the fluctuating force
as
GF (f ) = 4kB T R̃ (f ) . (23)
The analogy with the Johnson noise expression should be apparent.
Note that the derivations have been classical, so we arrive at the classical version of the fluctuation-
dissipation theorem, the force expression etc. In a quantum treatment A and B, as well as H are operators
that may not commute, so that we have to be more careful in the expansion of the exponentials. Thus Eq. (13)
will involve commutators of operators. The simplest form of the expression is to define a slightly different
susceptibility χ 00 (t) = (2i)−1 [χ (t) − χ (−t)] (the Fourier inverse of the dissipative part χ 00 (f )) which is then
related to the commutator-correlation function
00 1
χAB (t) = h[A(0), B(t)]i0 , (24)
2h̄
where A, B are Heisenberg (time dependent) operators. The change to Eq. (18) is quite simple: make the
replacement  
hf hf
kB T → coth (25)
2 2kB T
which can be interpreted as taking into account the Bose occupation factor of the modes. The quantum
approach was pioneered by Kubo [3], and the set of ideas are often called the Kubo formalism.

Applications to LIGO
The fluctuation-dissipation theorem is an important tool in understanding the noise that limits the sensitivity
of the LIGO experiment—at the operation frequency thermal noise is the dominant source. Therefore:
understand the dissipation and you understand the noise. Some references that should now be readable are:

1. A. Abramovici et al., Science 256, 325 (1992) — general review of LIGO

2. P.R. Slauson, Phys. Rev. D42, 2437 (1990) — discussion of mechanical dissipation and noise (available
online)

4
 3. A. Gillespie and F. Raab, Phys. Lett. A190, 213 (1994) — theory and measurements of suspension
losses and noise in a LIGO test bed

4. Y. Levin, Phys. Rev. D57, 659 (1998) — application of the fluctuation dissipation theorem leading to
a novel calculation of mass-mode noise (available online)

The paper by Yuri Levin, who was a student here a few years ago, is a particularly nice application of the
ideas we have been talking about, and I strongly urge you to read it.

References
[1] H.B. Callen and R.F. Greene, Phys. Rev. 86, 702 (1952)

[2] D. Chandler, Introduction to modern statistical mechanics (Oxford University Press, 1987)

[3] R. Kubo, J. Phys. Soc. Japan 12, 570 (1957)