The Fluctuation-Dissipation Theorem

Two definitions, one idea

The fluctuation-dissipation theorem can be stated in two ways that mean the same thing.

1. Intuitive: The microscopic processes that make a system fluctuate at rest are identical to the processes that dissipate energy when you drive it. Randomness and friction are two faces of one mechanism, so measuring one gives you the other.

2. Quantitative (Callen & Welton, 1951): For a generalized coordinate x with a linear response function χ(ω), the power spectral density of its equilibrium fluctuations is

S_x(ω) = (2k_B T / ω) · χ''(ω)   (classical, k_BT ≫ ℏω)

where χ''(ω) is the imaginary (dissipative) part of the response function — the part that describes energy absorption. The full quantum result replaces the classical 2k_BT/ω prefactor with ℏ·coth(ℏω / 2k_BT), which reduces to the classical form when thermal energy dominates and to ℏ (pure zero-point fluctuations) as T → 0.

Symbols and units:

• S_x(ω) — one-sided power spectral density of the fluctuating quantity x, units of [x]²/Hz (or [x]²·s per rad/s depending on convention).
• χ''(ω) — dissipative (imaginary) part of the linear susceptibility, units of [x]/[force].
• ω — angular frequency, rad/s.
• k_B — Boltzmann constant, 1.380649 × 10⁻²³ J/K (exact, SI 2019).
• T — absolute temperature, K.
• ℏ — reduced Planck constant, 1.0546 × 10⁻³⁴ J·s.

The single sentence to remember: noise = temperature × dissipation.

Why it matters

The FDT is remarkable because it lets you predict a system's active behavior (how it responds to and dissipates a force) from purely passive measurements (how it fluctuates when left alone). You never have to poke the system. This turns hard non-equilibrium questions into easy equilibrium ones:

• You cannot separate noise from friction. A resistor that dissipates power must generate voltage noise. A viscous fluid that drags a bead must make that bead jiggle. There is no such thing as friction without accompanying thermal noise at T > 0 — a fact that sets fundamental limits on amplifiers, sensors, and detectors.
• It is the sensitivity floor of precision instruments. LIGO's mirror coatings dissipate a minuscule amount of mechanical energy; by the FDT they therefore vibrate with Brownian "coating thermal noise," which is a dominant limit in the detector's most sensitive band (~50–200 Hz).
• It defines a primary thermometer. Because ⟨V²⟩ = 4k_BTRΔf involves only fundamental constants, measuring the noise of a known resistor gives absolute temperature directly — Johnson-noise thermometry now contributes to the redefinition of the kelvin.
• Its violation measures non-equilibrium. In glasses, active matter, and living cells, the ratio of fluctuation to response defines an effective temperature; when it exceeds the bath temperature, the excess quantifies energy injected by the system's own machinery.

How it works, step by step

Follow the logic through the canonical example, Brownian motion, described by the Langevin equation:

m (dv/dt) = -γ v + ξ(t)

Here m is the particle mass, v its velocity, γ the drag (dissipation) coefficient, and ξ(t) a random thermal force (fluctuation) delivered by molecular collisions.

1. The drag and the kicks share one origin. Both -γv and ξ(t) come from the same solvent molecules hitting the particle. The average of the collisions is the systematic drag; the deviations are the random force. So they cannot be independent.
2. Model the noise as white. Assume ⟨ξ(t)⟩ = 0 and ⟨ξ(t)ξ(t')⟩ = 2D_ξ·δ(t − t'), a delta-correlated (memoryless) random force with strength D_ξ.
3. Demand equipartition. In equilibrium, the equipartition theorem fixes the average kinetic energy: ½m⟨v²⟩ = ½k_BT, so ⟨v²⟩ = k_BT/m. This is the constraint that couples noise to friction.
4. Solve the Langevin equation. The stationary velocity variance from the equation is ⟨v²⟩ = D_ξ/(mγ). Setting this equal to k_BT/m forces D_ξ = γk_BT. The noise strength is locked to the friction times the temperature — you cannot choose one without the other. This is fluctuation-dissipation in its rawest form.
5. Read off diffusion. The position spreads as ⟨Δx²⟩ = 2Dt at long times, with diffusion coefficient D = k_BT/γ — Einstein's relation. High friction means slow diffusion; both are set by the same γ.

The generalization (Kubo, 1957) drops all model assumptions: any linear transport coefficient equals a time integral of an equilibrium correlation function of the relevant flux, computed in the unperturbed ensemble. That is the modern face of the theorem.

The three classic special cases

Result	Fluctuation	Dissipation	Relation
Einstein (1905)	Diffusion coefficient D	Drag coefficient γ	D = k_BT / γ
Stokes-Einstein	Diffusion of a sphere	Viscous drag γ = 6πηa	D = k_BT / (6πηa)
Johnson-Nyquist (1928)	Voltage noise ⟨V²⟩	Resistance R	⟨V²⟩ = 4k_BTR·Δf
Nyquist current form	Current noise ⟨I²⟩	Conductance G = 1/R	⟨I²⟩ = 4k_BTG·Δf
Kubo (1957)	Correlation ⟨J(0)J(t)⟩	Conductivity σ	σ = (1/k_BT)∫⟨J(0)J(t)⟩dt
Green-Kubo (viscosity)	Stress correlation ⟨σ_xy(0)σ_xy(t)⟩	Shear viscosity η	η = (V/k_BT)∫⟨σ_xy(0)σ_xy(t)⟩dt

Every row is the same statement: a dissipative coefficient on the right equals a fluctuation on the left, divided or multiplied by k_BT. Change the physical arena — colloids, circuits, electron gases, flowing liquids — and only the labels change.

Worked example — Johnson-Nyquist noise of a real resistor

How much thermal voltage noise does a 1 kΩ resistor produce at room temperature over a 10 kHz bandwidth? Use the Nyquist formula:

⟨V²⟩ = 4 k_B T R Δf

with k_B = 1.381 × 10⁻²³ J/K, T = 300 K, R = 1000 Ω, Δf = 10⁴ Hz:

⟨V²⟩ = 4 × (1.381e-23) × 300 × 1000 × 1e4
     = 1.657 × 10⁻¹³ V²
V_rms = √⟨V²⟩ ≈ 4.07 × 10⁻⁷ V ≈ 0.41 µV

A convenient shortcut: the noise spectral density is √(4k_BTR) ≈ 4.07 nV/√Hz for a 1 kΩ resistor at 300 K, a number every analog-circuit designer memorizes. Multiply by √Δf to get the RMS voltage in any bandwidth. Cooling the resistor to liquid-helium temperature (4.2 K) shrinks the noise by √(300/4.2) ≈ 8.5× — which is exactly why sensitive front-end amplifiers are cryocooled.

A short history

• 1905 — Einstein. In his "miracle year" paper on Brownian motion, Einstein derived D = k_BT/γ (via D = μk_BT, with mobility μ = 1/γ), the first fluctuation-dissipation relation. It let Jean Perrin (1908–1913) measure Avogadro's number and clinch the reality of atoms, earning Perrin the 1926 Nobel Prize.
• 1908 — Langevin. Paul Langevin recast Brownian motion as a stochastic differential equation, making the split between systematic drag and random force explicit.
• 1928 — Johnson & Nyquist. John B. Johnson measured the thermal voltage noise of resistors at Bell Labs; Harry Nyquist explained it the same year with a thermodynamic argument, giving ⟨V²⟩ = 4k_BTRΔf.
• 1951 — Callen & Welton. Herbert Callen and Theodore Welton proved the general theorem, valid quantum-mechanically, linking any linear dissipative response to its equilibrium fluctuation spectrum.
• 1957 — Kubo. Ryogo Kubo formulated linear response theory, expressing transport coefficients as equilibrium correlation-function integrals — the Green-Kubo relations — completing the modern framework.

Python — verifying D = k_BT/γ from a Langevin simulation

import numpy as np

k_B = 1.380649e-23   # J/K
T   = 300.0          # K
gamma = 1.0e-8       # drag coefficient (kg/s), e.g. ~1 µm bead in water
m   = 1.0e-15        # mass (kg)
dt  = 1.0e-7         # s
N   = 2_000_000

# Predicted diffusion coefficient (Einstein relation)
D_pred = k_B * T / gamma
print(f"Predicted D = k_BT/gamma = {D_pred:.3e} m^2/s")

# Noise strength locked by FDT:  = 2*gamma*k_BT*delta(t-t')
sigma = np.sqrt(2 * gamma * k_B * T / dt)  # per-step force std

v = 0.0
x = 0.0
xs = np.empty(N)
for i in range(N):
    xi = sigma * np.random.randn()
    v += (-gamma * v + xi) / m * dt        # Langevin update
    x += v * dt
    xs[i] = x

# Measure D from the long-time mean-square displacement:  = 2 D t
t = np.arange(N) * dt
D_meas = (xs[N//2:]**2).mean() / (2 * t[N//2:].mean())
print(f"Measured  D             = {D_meas:.3e} m^2/s")
# The two agree: the noise you PUT IN (fluctuation) and the drag (dissipation)
# reproduce Einstein's D — you cannot pick them independently.

Common misconceptions

• "It's just about Brownian motion." Einstein's relation is one special case. The theorem covers voltage noise in circuits, spin fluctuations in magnets, density fluctuations in fluids, radiation from a warm body, and any linear response — all unified by Callen-Welton and Kubo.
• "Fluctuations and dissipation are independent knobs." They are not. At a given temperature they are rigidly proportional. You cannot build a resistor that dissipates but produces no noise, nor a bath with friction but no jiggling — thermodynamics forbids it.
• "It works out of equilibrium too." The standard FDT strictly requires thermal equilibrium at a well-defined T. In driven, aging, or active systems it is violated, and the violation is what people measure to define an effective temperature and quantify non-equilibrium.
• "The classical formula is always right." When ℏω becomes comparable to k_BT (high frequency or low temperature), you must use the quantum coth(ℏω/2k_BT) form; zero-point fluctuations survive even at T = 0, which the classical version misses.
• "χ'' is the response you measure directly." χ''(ω) is the dissipative (out-of-phase, energy-absorbing) part of the complex susceptibility. The full response also has a reactive part χ'(ω); the two are tied by the Kramers-Kronig relations, and only χ'' enters the FDT.
• "Confusing shot noise with thermal noise." Johnson-Nyquist noise is equilibrium thermal noise (∝ T, present with no current). Shot noise arises from discrete charge crossing a barrier under an applied current and is not a fluctuation-dissipation effect — different physics.