08.12.02 · stat-mech / fluctuations

Equilibrium fluctuations of thermodynamic quantities

shipped3 tiersLean: none

Anchor (Master): Einstein *Ann. Phys.* 33 (1910) on critical opalescence; Smoluchowski *Ann. Phys.* 25 (1908) on density fluctuations; Landau-Lifshitz Vol. 5 §§111-114 and §116 (Ornstein-Zernike form); Ornstein-Zernike *Proc. Akad. Sci. Amsterdam* 17 (1914); Cramér *Actualités Sci. Indust.* 736 (1938) (large-deviation rate functions); Onsager-Machlup *Phys. Rev.* 91 (1953) (action for fluctuation trajectories); Touchette *Phys. Rep.* 478 (2009)

Intuition Beginner

A bottle of gas at room temperature, sitting on a shelf, looks at rest. Pressure, temperature, volume, and number of molecules are macroscopic numbers that an experimentalist quotes without flinching. Look at the bottle through a powerful enough microscope and you would see molecules darting in every direction, bumping the wall, and trading energy with one another. The macroscopic numbers are averages over a population so large that the average looks stable. Step back from the microscope and the noise vanishes. Step closer and the noise is everything that happens.

Equilibrium fluctuations are the small departures of those macroscopic numbers from their average values. The total energy of the gas wobbles by a tiny amount; the volume of a balloon held at fixed pressure ripples by a tiny amount; the number of molecules inside an imagined sub-volume of the bottle drifts up and down as molecules wander in and out. None of these wobbles violates conservation of energy or any other law. They reflect the fact that a macroscopic average is built from many microscopic events, and the average shifts a little whenever the events do not perfectly balance.

The size of each wobble is set by how stiff the macroscopic state is. A rigid object resists deformation: push on it and it barely budges. A floppy object yields easily: push on it and it gives. The stiffness of a thermodynamic state against energy, volume, or particle-number changes is encoded in the response functions called specific heat, compressibility, and susceptibility. The reason this matters is that fluctuations and response functions turn out to measure the same thing.

Visual Beginner

A schematic shows a sealed box of gas with a probe region drawn in the middle. The number of molecules inside the probe region drifts up and down with time, tracing a noisy line that oscillates around an average value $\overset{ˉ}{N}$ . To one side, a thermometer reads a single temperature; to another side, a meter reads a single pressure. The picture suggests that one fixed temperature and pressure do not pin down the number of molecules in the probe region: the molecule count fluctuates, and the spread of the fluctuation is exactly what this unit computes.

The picture captures the essential idea: macroscopic variables fluctuate even when the system is in equilibrium, and the size of the fluctuation is set by how stiffly the system responds to external pushes on the same variable.

Worked example Beginner

Consider a sealed bottle of ideal gas with volume $V$ at temperature $T = 300$ K. The total number of molecules is $N$ . Imagine drawing a small sub-volume inside the bottle and counting the molecules inside it. The number of molecules in the sub-volume is not a fixed constant; it changes from moment to moment as molecules drift in and out.

Step 1. Pick the sub-volume to contain on average $N = 1 0^{20}$ molecules. The mean-square fluctuation in particle number for an ideal gas equals $N$ itself: this is the result that the variance of a Poisson count is the mean. The standard deviation is therefore $N = 1 0^{20} = 1 0^{10}$ molecules.

Step 2. The relative fluctuation is the ratio of the standard deviation to the mean: $N / N = 1/ N = 1 0^{- 10}$ . The particle count varies by one part in ten billion. No ordinary measurement detects this jitter; it is invisible at the macroscopic scale.

Step 3. Repeat with a much smaller sub-volume containing on average $N = 100$ molecules. Now $N / N = 1/ 100 = 0.1$ . The particle count varies by ten percent: a noticeable jitter.

Step 4. Shrink the sub-volume further to contain on average $N = 10$ molecules. Then $N / N = 1/ 10 \approx 0.32$ . The count varies by about thirty percent. Even short snapshots see the count drift between roughly $7$ and $13$ .

Step 5. The pattern is universal: the relative fluctuation falls as $1/ N$ . Large bottles look quiet; tiny droplets are noisy. The transition is gradual, set entirely by the square root of how many particles you have. Near a critical point this scaling breaks down because the compressibility blows up; the unit returns to that case in the intermediate sections.

What this tells us: equilibrium fluctuations are not a defect of measurement. They are real, and their relative size is fixed by the number of degrees of freedom in the sub-system. Macroscopic measurements average over so many molecules that fluctuations look like zero; mesoscopic measurements see the noise directly.

Check your understanding Beginner

Formal definition Intermediate+

Let $S$ be a macroscopic system in thermal equilibrium at temperature $T$ , volume $V$ , and particle number $N$ . The microscopic state is governed by a Hamiltonian $H$ , and the canonical ensemble assigns probabilities $p_{i} = Z^{- 1} e^{- H_{i} / k_{B} T}$ to microstates $i$ , with partition function $Z = Z (T, V, N) = \sum_{i} e^{- H_{i} / k_{B} T}$ . Macroscopic observables $X_{1}, X_{2}, \dots, X_{m}$ — energy, volume, particle number, magnetisation, etc. — have ensemble means $\overset{ˉ}{X}_{a} = ⟨ X_{a} ⟩$ and fluctuations $Δ X_{a} = X_{a} - \overset{ˉ}{X}_{a}$ . The variance and the covariance of the fluctuations are

⟨(Δ X_{a})^{2} ⟩ = ⟨ X_{a}^{2} ⟩ - ⟨ X_{a} ⟩^{2}, ⟨ Δ X_{a} Δ X_{b} ⟩ = ⟨ X_{a} X_{b} ⟩ - ⟨ X_{a} ⟩ ⟨ X_{b} ⟩ .

The basic identity is the Einstein fluctuation formula: the probability of a configuration in which the macroscopic variables take values $X_{1}, \dots, X_{m}$ different from their equilibrium values is

W (X_{1}, \dots, X_{m}) \propto exp (\frac{Δ S _{tot}}{k _{B}}),

where $Δ S_{tot}$ is the entropy change of the system plus the surrounding heat reservoir compared to the equilibrium configuration. Maximum-entropy principles fix $Δ S_{tot}$ to be non-positive (vanishing at equilibrium), and the Gaussian quadratic expansion around equilibrium gives

Δ S_{tot} \approx - \frac{1}{2} a, b \sum β_{ab} Δ X_{a} Δ X_{b}, β_{ab} > 0,

where the matrix $β_{ab}$ is the second derivative of $- S_{tot} / k_{B}$ at equilibrium. The probability $W$ is Gaussian with covariance matrix $β^{- 1}$ :

⟨ Δ X_{a} Δ X_{b} ⟩ = (β^{- 1})_{ab} .

For each extensive variable $X_{a}$ there is a conjugate intensive variable $F_{a}$ defined so that the equilibrium thermodynamic identity reads $\sum_{a} F_{a} d X_{a} = 0$ on the constraint surface. Inversion of the Gaussian covariance gives the moment-derivative identity

⟨ Δ X_{a} Δ X_{b} ⟩ = k_{B} \frac{\partial X _{a}}{\partial F _{b}}_{other F},

which converts fluctuation variances into response functions. The conjugate pairs at use here are $(E, 1/ T)$ , $(V, - P / T)$ , $(N, μ / T)$ , and for a magnetic system $(M, H / T)$ .

A sign convention is adopted throughout: extensive variables $X$ carry the same sign as the standard textbook ( $E$ , $V$ , $N$ , $M$ positive when present); conjugate intensive variables are temperature $T$ , pressure $P$ , chemical potential $μ$ , and magnetic field $H$ . Boltzmann's constant $k_{B}$ has its 2019 SI fixed value $k_{B} = 1.380 649 \times 1 0^{- 23}$ J/K.

Counterexamples to common slips

The canonical ensemble fixes the temperature, not the energy. A system at fixed $T$ has energy variance $⟨(Δ E)^{2} ⟩ = k_{B} T^{2} C_{V} > 0$ ; only the microcanonical ensemble at fixed $E$ has $⟨(Δ E)^{2} ⟩ = 0$ . Choosing the right ensemble is choosing which variables are pinned and which fluctuate.
The grand-canonical ensemble adds a fluctuating particle number. The same system in the canonical ensemble has $⟨(Δ N)^{2} ⟩ = 0$ identically (the particle number is fixed). The cross-ensemble equivalence of free energies (Gibbs identity) does not mean that the variance of $N$ is the same in both descriptions; it changes from zero to $k_{B} T (\partial N / \partial μ)_{T, V}$ .
The Gaussian expansion is the second-order Taylor expansion of $Δ S_{tot}$ around the equilibrium maximum. Higher-order terms generate non-Gaussian corrections suppressed by inverse powers of $N$ in a system of size $N$ . Near a critical point the Gaussian expansion breaks down because the second-derivative matrix $β_{ab}$ acquires a vanishing eigenvalue, and non-Gaussian large-deviation theory takes over.

Key derivation Intermediate+

Theorem (Gaussian fluctuation theorem; Landau-Lifshitz §111-112). Let $S$ be a macroscopic system in equilibrium at temperature $T$ , with extensive variables $X_{1}, \dots, X_{m}$ and conjugate intensive variables $F_{1}, \dots, F_{m}$ . To leading order in $1/ N$ where $N$ is the size of the system, the fluctuation covariance matrix equals $$ \langle \Delta X_a, \Delta X_b\rangle = k_B \left(\frac{\partial X_a}{\partial F_b}\right)_{\text{other }F\text{ held fixed}}. $$ The same identity, written diagonally, gives the worked variances $$ \langle (\Delta E)^2\rangle = k_B T^2 C_V, \quad \langle (\Delta V)^2\rangle = k_B T V \kappa_T, \quad \langle (\Delta N)^2\rangle = \frac{N^2 k_B T \kappa_T}{V}, $$ and $$ \langle (\Delta T)^2\rangle = \frac{k_B T^2}{C_V}, \quad \langle (\Delta P)^2\rangle = \frac{k_B T}{V \kappa_S}, \quad \langle \Delta T, \Delta V\rangle = 0, $$ where $C_{V}$ is the heat capacity at constant volume, $κ_{T} = - V^{- 1} (\partial V / \partial P)_{T}$ is the isothermal compressibility, and $κ_{S} = - V^{- 1} (\partial V / \partial P)_{S}$ is the adiabatic compressibility.

Proof. Begin with the Einstein fluctuation formula $W \propto e^{Δ S_{tot} / k_{B}}$ , where $Δ S_{tot}$ is the entropy change of the system plus the bath relative to equilibrium. For a system in contact with a heat-and-volume reservoir, the first law of thermodynamics applied to the bath gives $$ \Delta S_{\text{bath}} = -\frac{\Delta E + P, \Delta V - \mu, \Delta N}{T}, $$ where $T$ , $P$ , and $μ$ are the reservoir temperature, pressure, and chemical potential (equal at equilibrium to those of the system). The total entropy change of system plus bath is therefore $$ \Delta S_{\text{tot}} = \Delta S_{\text{sys}} - \frac{\Delta E + P, \Delta V - \mu, \Delta N}{T}. $$ Expand $Δ S_{sys}$ in a Taylor series in the deviations $Δ E$ , $Δ V$ , $Δ N$ around equilibrium. The first-order terms cancel by the equilibrium conditions $(\partial S / \partial E)_{V, N} = 1/ T$ , $(\partial S / \partial V)_{E, N} = P / T$ , $(\partial S / \partial N)_{E, V} = - μ / T$ . The second-order remainder is $$ \Delta S_{\text{tot}} = \frac{1}{2}!\left[ \frac{\partial^2 S}{\partial X_a \partial X_b} \right]{\text{eq}} \Delta X_a, \Delta X_b = -\frac{1}{2} \sum{a,b} \beta_{ab}, \Delta X_a, \Delta X_b, $$ with summation over the extensive variables and $β_{ab}$ a positive-definite quadratic form (positive definiteness is equivalent to thermodynamic stability of the equilibrium). The probability of a fluctuation is then the multivariate Gaussian $$ W(\Delta X) \propto \exp!\left( -\frac{1}{2 k_B} \sum_{a,b} \beta_{ab}, \Delta X_a, \Delta X_b \right), $$ and the covariance matrix is $⟨ Δ X_{a} Δ X_{b} ⟩ = k_{B} (β^{- 1})_{ab}$ .

To convert the inverse matrix into a thermodynamic derivative, write the second-derivative matrix as $β_{ab} = - (1/ k_{B}) \partial^{2} S / \partial X_{a} \partial X_{b} = (\partial F_{a} / \partial X_{b})_{X \neq = b}$ using the Maxwell-type identity $F_{a} = - (1/ k_{B}) (\partial S / \partial X_{a})$ up to the conventional sign for conjugate variables. The inverse matrix is therefore $(β^{- 1})_{ab} = (\partial X_{a} / \partial F_{b})_{F \neq = b}$ , giving the general identity $$ \langle \Delta X_a, \Delta X_b\rangle = k_B \left(\frac{\partial X_a}{\partial F_b}\right)_{F \neq b}. $$

The diagonal worked entries follow by specialising. For energy, take the conjugate pair $(E, 1/ T)$ and use the canonical-ensemble identity $⟨ E ⟩ = - \partial lo g Z / \partial β$ where $β = 1/ (k_{B} T)$ . Differentiating once more gives $⟨(Δ E)^{2} ⟩ = \partial^{2} lo g Z / \partial β^{2} = - \partial ⟨ E ⟩ / \partial β = k_{B} T^{2} (\partial ⟨ E ⟩ / \partial T)_{V} = k_{B} T^{2} C_{V}$ . The same identity emerges from the moment-derivative formula via $(\partial E / \partial (1/ T))_{V} = - T^{2} (\partial E / \partial T)_{V} = - T^{2} C_{V}$ and a sign-tracking inversion.

For volume, the conjugate pair is $(V, - P / T)$ . The moment-derivative identity reads $⟨(Δ V)^{2} ⟩ = - k_{B} T (\partial V / \partial P)_{T} = k_{B} T V κ_{T}$ , using the definition of isothermal compressibility.

For particle number, the conjugate pair is $(N, μ / T)$ in the grand-canonical ensemble. The identity gives $⟨(Δ N)^{2} ⟩ = k_{B} T (\partial N / \partial μ)_{T, V}$ . To express this in terms of macroscopic response functions, use the Gibbs-Duhem relation: at fixed $T$ , a change in $μ$ relates to changes in $P$ and the density $n = N / V$ . The result is $⟨(Δ N)^{2} ⟩ = N^{2} k_{B} T κ_{T} / V$ , the standard Landau-Lifshitz §114 form.

The decorrelation $⟨ Δ T Δ V ⟩ = 0$ follows from a change of variables in the Gaussian. Trade the pair $(E, V)$ for the pair $(T, V)$ via the Jacobian. The second-derivative matrix becomes block-diagonal in $(T, V)$ at fixed $N$ , so the off-diagonal covariance vanishes. The diagonal entries are $⟨(Δ T)^{2} ⟩ = k_{B} T^{2} / C_{V}$ (the inverse of the energy variance, since $T$ and $E$ are conjugate response variables) and $⟨(Δ V)^{2} ⟩$ as computed above.

For pressure, use the pair $(S, V)$ as independent fluctuating variables instead of $(E, V)$ . At fixed entropy the compressibility is adiabatic $κ_{S}$ , and $⟨(Δ P)^{2} ⟩ = k_{B} T / (V κ_{S})$ . The choice of variable pair determines whether the response function is isothermal or adiabatic.

For magnetic systems, the conjugate pair is $(M, H / T)$ with $M$ the total magnetisation and $H$ the external field. The moment-derivative identity gives $⟨(Δ M)^{2} ⟩ = k_{B} T (\partial M / \partial H)_{T} = k_{B} T χ_{T}$ , the isothermal magnetic susceptibility. The same logic yields the static structure factor in spatially extended systems as the Fourier transform of the density-density correlation function. $□$

Bridge. This builds toward the central content of the fluctuation-dissipation theorem and appears again in 08.12.01 (the Landau-Callen-Welton FDT), where the static covariance identities of the present unit are the $ω \to 0$ limit of a more general frequency-dependent relation between equilibrium correlations and dynamic susceptibilities. The foundational reason these formulas hold is exactly the Gaussian quadratic expansion of $Δ S_{tot}$ around the equilibrium maximum: the matrix of second derivatives of entropy is the inverse of the covariance matrix, and the Maxwell-type relations of equilibrium thermodynamics identify the matrix entries with the standard response functions. This is exactly the moment-derivative identity, and putting these together produces every classical fluctuation formula in equilibrium statistical mechanics in a single line. The bridge is the recognition that the same Gaussian expansion generalises to a Gaussian field theory of fluctuating densities — covered in 08.06.01 — whose two-point function is the Ornstein-Zernike correlation function near the critical point, and whose breakdown signals the failure of Gaussian theory at criticality. The central insight is that response functions and equilibrium variances are the same object; one set of measurements determines the other.

Exercises Intermediate+

Exercise 5 (medium, short-answer).

Explain why $⟨ Δ T Δ V ⟩ = 0$ in the canonical ensemble, but $⟨ Δ E Δ V ⟩$ need not vanish.

Hint

The Gaussian quadratic form in the variables $(Δ T, Δ V)$ is block-diagonal at fixed $N$ . In the variables $(Δ E, Δ V)$ it is not.

Answer

In the variables $(Δ T, Δ V)$ , the second-derivative matrix of $- Δ S_{tot} / k_{B}$ decomposes into a $T$ -only block ( $C_{V} / T^{2}$ ) and a $V$ -only block ( $1/ (V κ_{T} T)$ ), with no cross-coupling. The covariance matrix is therefore diagonal in these variables and $⟨ Δ T Δ V ⟩ = 0$ . In the variables $(Δ E, Δ V)$ at fixed $N$ , the Jacobian connecting $(T, V)$ to $(E, V)$ introduces $\partial E / \partial V ∣_{T} = T (\partial P / \partial T)_{V} - P$ , which is generically non-zero. So $⟨ Δ E Δ V ⟩ = T ⟨ Δ T Δ V ⟩ \cdot \partial E / \partial T + \partial E / \partial V \cdot ⟨(Δ V)^{2} ⟩ = (T (\partial P / \partial T)_{V} - P) k_{B} T V κ_{T}$ , which vanishes only at points where the equation of state has $T (\partial P / \partial T)_{V} = P$ (true for the ideal gas, where $P V = N k_{B} T$ ). Rubric: full credit for the block-diagonal argument plus the Jacobian computation.

Exercise 7 (hard, short-answer).

State and derive the Landau-Lifshitz result $⟨(Δ N)^{2} ⟩ = N^{2} k_{B} T κ_{T} / V$ in the grand-canonical ensemble, using the Gibbs-Duhem relation $V d P = N d μ + S d T$ at fixed $T$ .

Hint

Start from $⟨(Δ N)^{2} ⟩ = k_{B} T (\partial N / \partial μ)_{T, V}$ , then use $(\partial μ / \partial P)_{T} = V / N$ from the Gibbs-Duhem relation at fixed $T$ , and finally the definition of $κ_{T}$ .

Answer

The grand-canonical moment-derivative identity gives $⟨(Δ N)^{2} ⟩ = k_{B} T (\partial N / \partial μ)_{T, V}$ . At fixed $T$ the Gibbs-Duhem relation reads $V d P = N d μ$ , so $(\partial μ / \partial P)_{T} = V / N$ , or equivalently $(\partial N / \partial μ)_{T, V} = (\partial N / \partial P)_{T} \cdot (\partial P / \partial μ)_{T} = (\partial N / \partial P)_{T, V} \cdot N / V$ . To compute $(\partial N / \partial P)_{T}$ at fixed $V$ , note that $N = nV$ at fixed $V$ , so $(\partial N / \partial P)_{T, V} = V (\partial n / \partial P)_{T} = V \cdot n / V \cdot (- 1/ V) (\partial V / \partial P)_{T} ∣_{N} = n κ_{T} V / V = N κ_{T} / V \cdot V$ ... cleaning up: $N$ depends on $P$ at fixed $V, T$ through the relation $μ = μ (P, T)$ and $N (μ, V, T)$ , with $(\partial N / \partial μ)_{V, T} \cdot V / N = V κ_{T} \cdot N / V$ , giving $(\partial N / \partial μ)_{V, T} = N^{2} κ_{T} / V$ . Hence $⟨(Δ N)^{2} ⟩ = k_{B} T \cdot N^{2} κ_{T} / V = N^{2} k_{B} T κ_{T} / V$ . The relative fluctuation is $⟨(Δ N)^{2} ⟩ / N = k_{B} T κ_{T} / V$ , which scales as $1/ N$ at fixed density. Rubric: full credit for the Gibbs-Duhem manipulation plus the final formula.

Exercise 8 (hard, short-answer).

Near a critical point of a fluid the isothermal compressibility diverges as $κ_{T} \sim ∣ T - T_{c} ∣^{- γ}$ with critical exponent $γ \approx 1.24$ in three dimensions (Ising universality class). Explain why this implies giant density fluctuations and macroscopic visibility (critical opalescence).

Hint

Use $⟨(Δ N)^{2} ⟩ / N = k_{B} T n κ_{T}$ at fixed density $n = N / V$ . Compute the scaling as $T \to T_{c}$ . Then compare with the correlation length $ξ$ to understand the macroscopic effect.

Answer

Approaching $T_{c}$ , the relative number variance $⟨(Δ N)^{2} ⟩ / N = k_{B} T n κ_{T} \sim ∣ T - T_{c} ∣^{- γ}$ diverges. Density fluctuations of all sizes are amplified, no longer suppressed by the central-limit $1/ N$ . The correlation length $ξ \sim ∣ T - T_{c} ∣^{- ν}$ with $ν \approx 0.63$ in three dimensions diverges as well, so fluctuations on length scales up to the wavelength of visible light become substantial. Density fluctuations scatter light strongly because the dielectric constant is a function of density and a spatially varying density refracts light. Near $T_{c}$ the fluid turns milky to the eye, an observable effect Einstein 1910 explained quantitatively from $W \propto e^{Δ S / k_{B}}$ . The breakdown of Gaussian fluctuation theory at $T_{c}$ is the signal of a critical point and the cue for renormalisation-group analysis; this is the content of 08.05.01 (critical exponents) and 08.05.02 (correlation functions). Rubric: full credit for the divergence argument plus the connection to light scattering.

Advanced results Master

Theorem (Ornstein-Zernike correlation function; Ornstein-Zernike 1914; Landau-Lifshitz §116). Outside the critical point, the equilibrium density-density correlation function of a fluid in $d$ spatial dimensions has the long-distance asymptotic form $$ G(r) = \langle \Delta n(0) \Delta n(\mathbf{r})\rangle \sim \frac{A}{r^{(d-1)/2} \xi^{(d-3)/2}}, e^{-r/\xi}, \qquad r \gg \xi, $$ where $ξ (T)$ is the correlation length and $A$ is an amplitude. The integral $χ = \int d^{d} r G (r)$ equals $n^{2} k_{B} T κ_{T}$ , recovering the total compressibility-fluctuation identity of the present unit. Approaching the critical point, $ξ \to \infty$ and the exponential decay flattens into a power law $G (r) \sim r^{- (d - 2 + η)}$ with the anomalous-dimension exponent $η$ .

The derivation in the Ornstein-Zernike 1914 paper introduces the direct correlation function $C (r)$ via the integral equation $G (r) = C (r) + n \int d^{d} r^{'} C (r - r^{'}) G (r^{'})$ , which inverts the random-phase approximation. The Fourier-space form $\hat{G} (k) = \hat{C} (k) / (1 - n \hat{C} (k))$ has a Lorentzian shape near $k = 0$ with width $1/ ξ$ , recovering the exponential decay in position space.

Theorem (Cramér's large-deviation principle; Cramér 1938). Let $X_{1}, X_{2}, \dots, X_{n}$ be i.i.d. real random variables with finite moment generating function $M (λ) = E [e^{λ X_{1}}]$ . The empirical mean $\overset{ˉ}{X}_{n} = n^{- 1} \sum_{i = 1}^{n} X_{i}$ satisfies the large-deviation principle $$ \lim_{n \to \infty} \frac{1}{n} \log P(\bar{X}n \in A) = -\inf{x \in A} I(x), \qquad I(x) = \sup_{\lambda}, [\lambda x - \log M(\lambda)], $$ with rate function $I (x)$ the Legendre transform of the log moment-generating function. The function $I$ is convex, non-negative, and zero at $x = E [X_{1}]$ .

For sums of identically distributed thermodynamic variables (energies of independent subsystems, particle counts in independent cells), Cramér's theorem replaces the Gaussian formula at any fixed deviation $∣ x - E [X_{1}] ∣ > 0$ . The Gaussian fluctuation theorem of the present unit is the local quadratic approximation to $I$ at its minimum: $I (x) \approx \frac{1}{2 σ ^{2}} (x - E [X_{1}])^{2}$ for small deviations, with $σ^{2}$ the variance. Large deviations of order one are exponentially rare with the rate set by $I$ , not the variance.

Theorem (Sanov's theorem; Sanov 1957). Let $X_{1}, \dots, X_{n}$ be i.i.d. random variables with distribution $μ$ on a Polish space. The empirical measure $\overset{μ}{^}_{n} = n^{- 1} \sum_{i = 1}^{n} δ_{X_{i}}$ satisfies the large-deviation principle on the space of probability measures, with rate function the relative entropy $$ H(\nu | \mu) = \int \log!\left(\frac{d\nu}{d\mu}\right) d\nu $$ if $ν ≪ μ$ , and $+ \infty$ otherwise. For every measurable set $B$ of probability measures, $$ \lim_{n \to \infty} \frac{1}{n} \log P(\hat{\mu}n \in B) = -\inf{\nu \in B} H(\nu | \mu). $$

Sanov's theorem is the level-2 large-deviation statement: it gives rate functions for the entire empirical distribution, not just its mean. It is the rigorous origin of relative entropy as a Lyapunov functional in stochastic thermodynamics. The Gärtner-Ellis theorem extends Cramér and Sanov to non-i.i.d. sequences when a limit of normalised log-moment-generating functions exists.

Theorem (Onsager-Machlup action; Onsager-Machlup 1953). Let $X (t)$ be a stationary Gaussian Markov process satisfying the Langevin equation $\dot{X} = - γ X + 2 D η (t)$ with $⟨ η (t) η (t^{'})⟩ = δ (t - t^{'})$ and equilibrium variance $D / γ$ . The probability density on the space of trajectories $x : [0, T] \to R$ is $$ P[x] \propto \exp!\left( -\frac{1}{4D} \int_0^T (\dot{x} + \gamma x)^2, dt \right). $$ The action functional $S [x] = (4 D)^{- 1} \int (\overset{x}{˙} + γ x)^{2} d t$ generalises the static Gaussian formula $- Δ S_{tot} / (2 k_{B})$ to entire trajectories. Most-probable paths between fluctuation states are stationary points of $S$ .

The Onsager-Machlup action provides the path-integral foundation of fluctuation theory: probabilities of trajectories of $X$ are assigned by the Boltzmann-like weight $e^{- S [x]}$ , with $S [x]$ replacing the static $Δ S_{tot}$ . Time-reversal symmetry of $S [x]$ for reversible Langevin dynamics is the origin of detailed-balance constraints on the equilibrium ensemble. Non-equilibrium extensions break this symmetry and give a precise quantitative measure of irreversibility, the entropy-production rate.

Theorem (sample-to-sample fluctuations in spin glasses; Parisi 1979-1983). For an Edwards-Anderson spin glass on $N$ sites with quenched random couplings ${J_{ij}}$ drawn from a Gaussian distribution, the free energy density $f_{N} = - lo g Z_{N} / (β N)$ has sample-to-sample fluctuations $f_{N} ({J}) - ⟨ f_{N} ⟩$ that do not vanish as $N \to \infty$ in the replica-symmetry-broken low-temperature phase. The order parameter is the Parisi function $q (x)$ for $x \in [0, 1]$ , a non-decreasing function encoding overlaps between distinct equilibrium states accessible to a typical sample.

The standard $1/ N$ self-averaging argument fails because the relevant ensemble has infinitely many pure states organised by a hierarchical (ultrametric) overlap structure; the Parisi function captures this structure. The replica method, formalised by Mézard-Parisi-Virasoro 1987 and rigorously confirmed by Talagrand and Panchenko in the 2000s for the Sherrington-Kirkpatrick model, treats $\overline{lo g Z}$ via the analytic continuation $n \to 0$ of $\overline{Z^{n}}$ for integer $n$ , yielding the saddle-point equations for the overlap matrix and its Parisi-replica-symmetry-broken solution.

Theorem (Einstein-Smoluchowski critical opalescence). Near the gas-liquid critical point of a one-component fluid, density fluctuations of length scale $ξ$ diverging with $∣ T - T_{c} ∣^{- ν}$ scatter light of wavelength $λ$ with intensity proportional to the static structure factor $S (q) = \int d^{3} r G (r) e^{- i q \cdot r}$ at $∣ q ∣ = 4 π sin (θ /2) / λ$ . The scattered intensity diverges as $∣ T - T_{c} ∣^{- γ}$ with the susceptibility exponent $γ$ .

This is the macroscopic visibility of critical fluctuations. Einstein 1910 derived the leading-order density-fluctuation formula from $W \propto e^{Δ S / k_{B}}$ together with the variance $⟨(Δ N)^{2} ⟩ = N^{2} k_{B} T κ_{T} / V$ ; the divergence of $κ_{T}$ at $T_{c}$ explains the macroscopic milky appearance of a fluid at its critical density. Modern scaling theory (Kadanoff 1966, Wilson 1971) and renormalisation-group calculations (Wilson-Fisher 1972, 08.04.02) sharpen the prediction to $κ_{T} \sim ∣ T - T_{c} ∣^{- γ}$ with $γ \approx 1.24$ in three dimensions.

Synthesis. The foundational reason every equilibrium variance has a thermodynamic-derivative expression is the Einstein principle $W \propto e^{Δ S_{tot} / k_{B}}$ combined with the Gaussian quadratic expansion of $Δ S_{tot}$ around the equilibrium maximum. The central insight is that the matrix of second derivatives of the entropy is, up to a factor of $k_{B}$ , the inverse of the covariance matrix of fluctuations; cap the inversion with the Maxwell identities of equilibrium thermodynamics and one obtains the moment-derivative formula $⟨ Δ X_{a} Δ X_{b} ⟩ = k_{B} (\partial X_{a} / \partial F_{b})$ . Putting these together, every classical equilibrium-fluctuation formula — the energy variance $k_{B} T^{2} C_{V}$ , the volume variance $k_{B} T V κ_{T}$ , the particle-number variance $N^{2} k_{B} T κ_{T} / V$ , the magnetisation variance $k_{B} T χ_{T}$ — collapses into a single derivation, and the structure generalises directly to time-dependent and spatially extended observables. This is exactly the structure that builds toward the full fluctuation-dissipation theorem of 08.12.01, in which the same identity becomes the $ω \to 0$ limit of a relation between dynamic correlations and dissipative response.

The Gaussian formula is the local quadratic approximation to a larger non-Gaussian structure. The Cramér and Sanov theorems give rigorous rate functions for finite-deviation events; the Gaussian variance is the curvature of the rate function at its minimum, exactly as the quadratic expansion of $Δ S_{tot}$ is the curvature of the entropy at equilibrium. The bridge is that large-deviation theory generalises equilibrium fluctuation theory to non-Gaussian, non-equilibrium, and disordered settings, while Gaussian theory remains the leading-order asymptotic in $1/ N$ . Near a critical point the Gaussian expansion fails — the second-derivative matrix of $- Δ S_{tot} / k_{B}$ acquires a vanishing eigenvalue along the order-parameter direction, and the resulting divergence of variances is the signal of a phase transition. The Ornstein-Zernike asymptotic form $G (r) \sim e^{- r / ξ} / r^{(d - 1) /2}$ identifies $ξ$ with the inverse correlation length, and the divergence $ξ \to \infty$ at $T_{c}$ marks the regime where renormalisation-group analysis takes over. The Parisi solution of the Sherrington-Kirkpatrick spin glass exhibits a third type of breakdown: sample-to-sample fluctuations of the free energy do not vanish, the rate function carries genuine structure, and self-averaging fails — the central feature that distinguishes ordered, critical, and disordered phases at the level of fluctuation theory.

Full proof set Master

Proposition (Gaussian fluctuation theorem), proof. Given in the Intermediate-tier derivation: expand $Δ S_{tot}$ to second order in the deviations from equilibrium using the first law applied to a system in contact with a reservoir; identify the matrix of second derivatives with the inverse covariance via the Maxwell-type identity; specialise to extract the diagonal variances $⟨(Δ E)^{2} ⟩ = k_{B} T^{2} C_{V}$ , $⟨(Δ V)^{2} ⟩ = k_{B} T V κ_{T}$ , $⟨(Δ N)^{2} ⟩ = N^{2} k_{B} T κ_{T} / V$ , and the decorrelation $⟨ Δ T Δ V ⟩ = 0$ in the variables $(T, V)$ . $□$

Proposition (Cramér's theorem in $R$ ), proof. Let $Λ (λ) = lo g M (λ) = lo g E [e^{λ X_{1}}]$ be the log moment-generating function, assumed finite on a neighbourhood of zero. The Legendre transform $I (x) = sup_{λ} [λ x - Λ (λ)]$ is convex and lower-semi-continuous. The upper bound is Chebyshev: for $λ > 0$ and $x > E [X_{1}]$ , $$ P(\bar{X}n \geq x) = P!\left(\sum{i=1}^n X_i \geq nx\right) \leq e^{-\lambda n x} \mathbb{E}!\left[e^{\lambda \sum X_i}\right] = e^{-n[\lambda x - \Lambda(\lambda)]}. $$ Optimising over $λ > 0$ gives $P (\overset{ˉ}{X}_{n} \geq x) \leq e^{- n I (x)}$ , so $lim sup_{n} n^{- 1} lo g P (\overset{ˉ}{X}_{n} \geq x) \leq - I (x)$ .

For the matching lower bound, introduce the tilted distribution $d P_{λ} = e^{λ X - Λ (λ)} d P$ for each $λ$ . Under $P_{λ}$ , $X_{1}$ has mean $Λ^{'} (λ)$ . For $x > E [X_{1}]$ choose $λ^{*} = λ^{*} (x)$ such that $Λ^{'} (λ^{*}) = x$ (existence by convexity and continuity). Restricting $\overset{ˉ}{X}_{n}$ to a small neighbourhood of $x$ and changing measure via the Cameron-Martin / Esscher tilt formula gives $P (\overset{ˉ}{X}_{n} \in [x, x + δ]) = e^{- n [λ^{*} x - Λ (λ^{*})] + O (n δ)} \cdot P_{λ^{*}} (\overset{ˉ}{X}_{n} \in [x, x + δ])$ ; the second factor tends to $1/2$ by the central limit theorem applied under $P_{λ^{*}}$ . So $lim inf_{n} n^{- 1} lo g P (\overset{ˉ}{X}_{n} \geq x) \geq - I (x)$ . Combining the bounds gives the limit. The case $x < E [X_{1}]$ is symmetric via $X \to - X$ . $□$

Proposition (Ornstein-Zernike asymptotic form), proof sketch. In Fourier space the OZ integral equation reads $\hat{G} (k) = \hat{C} (k) + n \hat{C} (k) \hat{G} (k)$ , equivalent to $\hat{G} (k) = \hat{C} (k) / (1 - n \hat{C} (k))$ . Expand $\hat{C} (k) = \hat{C} (0) - a k^{2} + O (k^{4})$ near $k = 0$ , with $a > 0$ . Define $ξ^{- 2} = (1 - n \hat{C} (0)) / (na)$ ; then $\hat{G} (k) \approx \hat{C} (0) / [1 - n \hat{C} (0) + na k^{2}] = ξ^{2} / [1 + (k ξ)^{2}] \cdot \hat{C} (0) / (na)$ — a Lorentzian (Yukawa) form in momentum space. The inverse Fourier transform in $d$ dimensions is the modified Bessel function $K_{(d - 2) /2} (r / ξ) \cdot r^{- (d - 2) /2} / ξ^{(d - 2) /2}$ , with the asymptotic form $K_{ν} (z) \sim π / (2 z) e^{- z}$ for $z ≫ 1$ giving the announced result. The integral $\int d^{d} r G (r) = \hat{G} (0) = \hat{C} (0) / (1 - n \hat{C} (0))$ recovers the static structure factor $S (0) = n k_{B} T κ_{T}$ on identifying with the moment-derivative identity. $□$

Proposition (Onsager-Machlup action), proof. Discretise the trajectory $x : [0, T] \to R$ on a grid $0 = t_{0} < t_{1} < \dots < t_{N} = T$ with spacing $δ t$ . The Euler-Maruyama discretisation of $\dot{X} = - γ X + 2 D η$ is $x_{n + 1} - x_{n} = - γ x_{n} δ t + 2 D δ t ξ_{n}$ with $ξ_{n}$ i.i.d. standard Gaussian. The probability density of $ξ_{n} = (x_{n + 1} - x_{n} + γ x_{n} δ t) / 2 D δ t$ is $(2 π)^{- 1/2} e^{- ξ_{n}^{2} /2}$ . The joint density of $(x_{1}, \dots, x_{N})$ is $$ P[x] = \prod_{n=0}^{N-1} (4\pi D \delta t)^{-1/2} \exp!\left( -\frac{(x_{n+1} - x_n + \gamma x_n \delta t)^2}{4 D \delta t} \right). $$ Taking the continuum limit $δ t \to 0$ , the exponent becomes the Riemann sum for $(4 D)^{- 1} \int_{0}^{T} (\overset{x}{˙} + γ x)^{2} d t$ , identifying $S [x] = (4 D)^{- 1} \int (\overset{x}{˙} + γ x)^{2} d t$ as the Onsager-Machlup action. The prefactor diverges as $(4 π D δ t)^{- N /2}$ and is absorbed into the normalisation of the path-space measure (Wiener measure for the underlying Brownian motion). $□$

Proposition (vanishing of $⟨ Δ T Δ V ⟩$ ), proof. Choose $(T, V)$ as the independent fluctuating variables at fixed $N$ . The free energy $F (T, V, N) = - k_{B} T lo g Z$ has $S = - (\partial F / \partial T)_{V}$ and $P = - (\partial F / \partial V)_{T}$ . The entropy as a function of $E$ and $V$ at fixed $N$ has second derivatives at equilibrium that, when re-expressed in terms of $T$ and $V$ , decouple: $(\partial^{2} S / \partial T^{2})_{V} = - C_{V} / T^{2}$ and $(\partial^{2} S / \partial V^{2})_{T} = - 1/ (V κ_{T} T)$ at fixed $N$ , with the cross-derivative vanishing because $T$ and $V$ are independent variables of the free energy. The Gaussian quadratic form in $(Δ T, Δ V)$ is block-diagonal, so the off-diagonal entry of the covariance matrix vanishes: $⟨ Δ T Δ V ⟩ = 0$ . $□$

Connections Master

Partition function 08.01.01. The canonical partition function $Z (T, V, N) = \sum_{i} e^{- E_{i} / k_{B} T}$ generates equilibrium fluctuation variances through its logarithm: $⟨(Δ E)^{2} ⟩ = \partial^{2} lo g Z / \partial β^{2}$ , $⟨(Δ N)^{2} ⟩ = (k_{B} T)^{2} \partial^{2} lo g Ξ/ \partial μ^{2}$ in the grand-canonical case. The cumulant generating function $lo g Z$ is the master object whose Taylor coefficients are exactly the fluctuation moments of the present unit.
Boltzmann distribution 08.01.03. The Boltzmann weights $e^{- E_{i} / k_{B} T} / Z$ are the underlying probability assignment whose moments are the fluctuations computed here. The Einstein fluctuation formula $W \propto e^{Δ S_{tot} / k_{B}}$ is the macroscopic-variable expression of the same Boltzmann principle, with $Δ S_{tot}$ standing in for the log of the multiplicity weight.
Free energy 08.01.04. Response functions $C_{V}$ , $κ_{T}$ , $κ_{S}$ , $χ_{T}$ are second derivatives of the free energy with respect to the conjugate intensive variables. The moment-derivative identity $⟨ Δ X_{a} Δ X_{b} ⟩ = k_{B} (\partial X_{a} / \partial F_{b})$ is the cumulant interpretation of these derivatives. Free-energy convexity in the extensive variables is equivalent to positive-definiteness of the fluctuation covariance, the macroscopic statement of thermodynamic stability.
Correlation functions 08.05.02. The spatially extended generalisation of the present unit replaces global variances with two-point functions $G (r) = ⟨ Δ n (0) Δ n (r)⟩$ . The integral $\int d^{d} r G (r) = n k_{B} T κ_{T}$ recovers the bulk compressibility-fluctuation identity. The Ornstein-Zernike asymptotic form is the natural extension and the bridge to critical-point scaling.
Fluctuation-dissipation theorem (Landau-Callen-Welton) 08.12.01. The static identities of the present unit are the $ω \to 0$ limit of the frequency-dependent FDT, which identifies the spectral density of equilibrium fluctuations with the imaginary part of the dynamic response function. Equilibrium thermodynamic-derivative formulas, dynamic susceptibility, and noise spectra are three faces of one identity.
Critical exponents 08.05.01. Approaching a critical point, the response functions of the present unit acquire power-law divergences $κ_{T} \sim ∣ T - T_{c} ∣^{- γ}$ , $χ_{T} \sim ∣ T - T_{c} ∣^{- γ}$ , $C_{V} \sim ∣ T - T_{c} ∣^{- α}$ with universal critical exponents. The Gaussian fluctuation theorem fails at the critical point because the second-derivative matrix of $- Δ S_{tot} / k_{B}$ acquires a vanishing eigenvalue along the order-parameter direction, and the variance of the order parameter diverges.

Historical & philosophical context Master

The systematic theory of equilibrium fluctuations begins with Marian Smoluchowski's 1908 paper on the kinetic theory of opalescence (Ann. Phys. (Leipzig) 25, 205) ^{[Smoluchowski 1908]} and Albert Einstein's 1910 paper on critical opalescence (Ann. Phys. (Leipzig) 338, 1275) ^{[Einstein 1910]}. Smoluchowski computed density fluctuations in a fluid by treating each cubic region as an open subsystem of a larger reservoir and applying the kinetic-theory machinery already in use for Brownian motion. Einstein generalised to the critical region and made the connection to optical observability quantitative: the milky appearance of carbon dioxide near its critical point at room temperature was the empirical phenomenon waiting for a theoretical explanation, and Einstein's formula $W \propto e^{Δ S / k_{B}}$ supplied one. The unified framework now bears the name Einstein-Smoluchowski.

Leonhard Ornstein and Frits Zernike extended the analysis in 1914 (Proc. Akad. Sci. Amsterdam 17, 793) ^{[Ornstein Zernike 1914]} by introducing the direct correlation function $C (r)$ and the integral equation that now bears their names. The Ornstein-Zernike asymptotic form $G (r) \sim r^{- (d - 1) /2} e^{- r / ξ}$ became the standard description of equilibrium correlations outside the critical region and the canonical introduction to correlation-length physics. Landau and Lifshitz codified the Gaussian fluctuation theory in Volume 5 of their Course of Theoretical Physics (§§111-114 of Statistical Physics, Part 1) ^{[Landau Lifshitz 1980]}, including the systematic derivation of all the worked variances of the present unit and the statistical independence of $T$ and $V$ in the canonical ensemble. Their presentation remains the standard reference for the subject.

The non-Gaussian extension came from probability theory. Harald Cramér's 1938 paper (Actualités Sci. Indust. 736, 5) ^{[Cramér 1938]} introduced what is now called Cramér's theorem and the rate function $I (x)$ as the Legendre transform of the log moment-generating function. Ivan Sanov in 1957 (Mat. Sbornik 42, 11) ^{[Sanov 1957]} generalised to the level of empirical measures, identifying the relative entropy as the rate function. Lars Onsager and Stefan Machlup in 1953 (Phys. Rev. 91, 1505) ^{[Onsager Machlup 1953]} supplied the path-integral language for whole fluctuation trajectories, the precursor to modern stochastic-thermodynamics formulations of entropy production. Hugo Touchette's 2009 review (Phys. Rep. 478, 1) ^{[Touchette 2009]} consolidates the large-deviation viewpoint as the natural rigorous successor to Gaussian theory in equilibrium and non-equilibrium statistical mechanics.

The disordered case has its own history. Sam Edwards and Philip Anderson in 1975 proposed the Edwards-Anderson model of a spin glass; Giorgio Parisi between 1979 and 1983 (Phys. Rev. Lett. 43, 1754; J. Phys. A 13, L115; Phys. Rev. Lett. 50, 1946) solved the mean-field Sherrington-Kirkpatrick spin glass with the replica-symmetry-breaking ansatz, identifying the Parisi function $q (x)$ as the order parameter. Michel Talagrand established a rigorous version in 2006 (Ann. of Math. 163, 221) and Dmitry Panchenko completed the picture in 2013 (Ann. Probab. 41, 1315). The sample-to-sample fluctuations of the free energy in the spin-glass phase remain the cleanest known example of a thermodynamic ensemble in which the central-limit $1/ N$ self-averaging fails.

Bibliography Master

@article{Smoluchowski1908,
  author  = {Smoluchowski, Marian},
  title   = {Molekular-kinetische Theorie der Opaleszenz von Gasen im kritischen Zustande, sowie einiger verwandter Erscheinungen},
  journal = {Annalen der Physik},
  volume  = {330},
  number  = {2},
  year    = {1908},
  pages   = {205--226}
}

@article{Einstein1910,
  author  = {Einstein, Albert},
  title   = {Theorie der Opaleszenz von homogenen Fl\"ussigkeiten und Fl\"ussigkeitsgemischen in der N\"ahe des kritischen Zustandes},
  journal = {Annalen der Physik},
  volume  = {338},
  number  = {16},
  year    = {1910},
  pages   = {1275--1298}
}

@article{OrnsteinZernike1914,
  author  = {Ornstein, Leonard S. and Zernike, Frits},
  title   = {Accidental deviations of density and opalescence at the critical point of a single substance},
  journal = {Proceedings of the Royal Netherlands Academy of Arts and Sciences},
  volume  = {17},
  year    = {1914},
  pages   = {793--806}
}

@book{LandauLifshitzVol5,
  author    = {Landau, L. D. and Lifshitz, E. M.},
  title     = {Statistical Physics, Part 1},
  edition   = {3},
  publisher = {Pergamon Press},
  series    = {Course of Theoretical Physics},
  volume    = {5},
  year      = {1980}
}

@book{Callen1985,
  author    = {Callen, Herbert B.},
  title     = {Thermodynamics and an Introduction to Thermostatistics},
  edition   = {2},
  publisher = {John Wiley \& Sons},
  year      = {1985}
}

@book{PathriaBeale2011,
  author    = {Pathria, R. K. and Beale, Paul D.},
  title     = {Statistical Mechanics},
  edition   = {3},
  publisher = {Academic Press},
  year      = {2011}
}

@article{Cramer1938,
  author  = {Cram{\'e}r, Harald},
  title   = {Sur un nouveau th{\'e}or{\`e}me-limite de la th{\'e}orie des probabilit{\'e}s},
  journal = {Actualit{\'e}s Scientifiques et Industrielles},
  volume  = {736},
  year    = {1938},
  pages   = {5--23}
}

@article{Sanov1957,
  author  = {Sanov, Ivan N.},
  title   = {On the probability of large deviations of random variables},
  journal = {Matematicheskii Sbornik},
  volume  = {42},
  number  = {84},
  year    = {1957},
  pages   = {11--44}
}

@article{OnsagerMachlup1953,
  author  = {Onsager, Lars and Machlup, Stefan},
  title   = {Fluctuations and irreversible processes},
  journal = {Physical Review},
  volume  = {91},
  number  = {6},
  year    = {1953},
  pages   = {1505--1512}
}

@article{Touchette2009,
  author  = {Touchette, Hugo},
  title   = {The large deviation approach to statistical mechanics},
  journal = {Physics Reports},
  volume  = {478},
  number  = {1--3},
  year    = {2009},
  pages   = {1--69}
}

@article{Parisi1979,
  author  = {Parisi, Giorgio},
  title   = {Infinite number of order parameters for spin-glasses},
  journal = {Physical Review Letters},
  volume  = {43},
  year    = {1979},
  pages   = {1754--1756}
}

@book{MezardParisiVirasoro1987,
  author    = {M{\'e}zard, Marc and Parisi, Giorgio and Virasoro, Miguel A.},
  title     = {Spin Glass Theory and Beyond},
  publisher = {World Scientific},
  series    = {World Scientific Lecture Notes in Physics},
  volume    = {9},
  year      = {1987}
}

@article{Talagrand2006,
  author  = {Talagrand, Michel},
  title   = {The Parisi formula},
  journal = {Annals of Mathematics},
  volume  = {163},
  number  = {1},
  year    = {2006},
  pages   = {221--263}
}

@article{Kubo1966,
  author  = {Kubo, Ryogo},
  title   = {The fluctuation-dissipation theorem},
  journal = {Reports on Progress in Physics},
  volume  = {29},
  number  = {1},
  year    = {1966},
  pages   = {255--284}
}

Prerequisites

08.01.01
08.01.03
08.01.04
08.05.02

Tier anchors

beginner: Landau-Lifshitz Vol. 5 *Statistical Physics, Part 1*, 3rd ed. (Pergamon, 1980), §111 informal opening; Reif *Fundamentals of Statistical and Thermal Physics* §1.1 and §6.6; Kittel *Thermal Physics* Ch. 12 informal pages on fluctuations
intermediate: Landau-Lifshitz Vol. 5 §§111-114 (Gaussian quadratic theory, computed variances $\langle (\Delta E)^2\rangle$, $\langle (\Delta V)^2\rangle$, $\langle (\Delta N)^2\rangle$, decorrelation of $T$ and $V$); Callen *Thermodynamics and an Introduction to Thermostatistics*, 2nd ed. (Wiley, 1985), Chs. 19-20; Pathria-Beale *Statistical Mechanics*, 3rd ed., Ch. 13
master: Einstein *Ann. Phys.* 33 (1910) on critical opalescence; Smoluchowski *Ann. Phys.* 25 (1908) on density fluctuations; Landau-Lifshitz Vol. 5 §§111-114 and §116 (Ornstein-Zernike form); Ornstein-Zernike *Proc. Akad. Sci. Amsterdam* 17 (1914); Cramér *Actualités Sci. Indust.* 736 (1938) (large-deviation rate functions); Onsager-Machlup *Phys. Rev.* 91 (1953) (action for fluctuation trajectories); Touchette *Phys. Rep.* 478 (2009)

References

Einstein, A. — Theorie der Opaleszenz von homogenen Flüssigkeiten und Flüssigkeitsgemischen in der Nähe des kritischen Zustandes · Ann. Phys. (Leipzig) 338, 1275 (1910) — the originator paper on critical opalescence, deriving density fluctuations from the entropy-deficit formula $W \propto e^{\Delta S/k_B}$ near the critical point of a fluid
Smoluchowski, M. — Molekular-kinetische Theorie der Opaleszenz von Gasen im kritischen Zustande, sowie einiger verwandter Erscheinungen · Ann. Phys. (Leipzig) 25, 205 (1908) — kinetic theory of opalescence and the first systematic computation of equilibrium density fluctuations in a fluid
Ornstein, L. S. & Zernike, F. — Accidental deviations of density and opalescence at the critical point of a single substance · Proc. Akad. Sci. Amsterdam 17, 793 (1914) — the integral equation for the direct correlation function and the asymptotic form $G(r) \sim r^{-(d-1)/2} e^{-r/\xi}/\xi^{(d-3)/2}$ for the equilibrium density-density correlation outside the critical point
Landau, L. D. & Lifshitz, E. M. — Statistical Physics, Part 1, 3rd ed. · Pergamon, 1980 — Vol. 5, §111 (fluctuation probability and the Gaussian quadratic form), §112 (general formula $\langle \Delta X_i \Delta X_j\rangle = k_B (\partial X_i/\partial F_j)$), §113 (worked variances of $E$, $V$, $N$, $T$, $P$), §114 (statistical independence of $\Delta T$ and $\Delta V$), §116 (Ornstein-Zernike correlation function)
Callen, H. B. — Thermodynamics and an Introduction to Thermostatistics, 2nd ed. · Wiley, 1985 — Chs. 19-20: fluctuation moments in terms of thermodynamic response functions, the moment-matrix inversion identifying canonical-conjugate variable pairs, microcanonical and canonical comparisons
Pathria, R. K. & Beale, P. D. — Statistical Mechanics, 3rd ed. · Academic Press, 2011 — Ch. 13 fluctuations and the canonical-grand-canonical fluctuation relations, Einstein-Smoluchowski density fluctuations, magnetisation fluctuations and susceptibility
Cramér, H. — Sur un nouveau théorème-limite de la théorie des probabilités · Actualités Scientifiques et Industrielles 736, 5 (1938) — Cramér's theorem: empirical means $\bar{X}_n$ of i.i.d. random variables satisfy $P(\bar{X}_n \in A) \asymp e^{-n I(A)}$ with rate function $I(x) = \sup_\lambda (\lambda x - \log M(\lambda))$, the Legendre transform of the log moment-generating function
Sanov, I. N. — On the probability of large deviations of random variables · Mat. Sbornik N.S. 42, 11 (1957) — large-deviation principle for empirical measures: $P(\hat{\mu}_n \in B) \asymp e^{-n H(B \| \mu)}$ with rate function the relative entropy / Kullback-Leibler divergence
Onsager, L. & Machlup, S. — Fluctuations and irreversible processes · Phys. Rev. 91, 1505 (1953) — the action functional $S[x] = \frac{1}{4} \int (\dot{x} + \gamma x)^2/D\, dt$ assigning probabilities to whole fluctuation trajectories of a Gaussian Markov process, the origin of path-integral large-deviation theory in non-equilibrium thermodynamics
Touchette, H. — The large deviation approach to statistical mechanics · Phys. Rep. 478, 1 (2009) — comprehensive modern review of large-deviation rate functions in equilibrium and non-equilibrium statistical mechanics, Gärtner-Ellis theorem, contraction principle, application to the Gibbs free energy as a Legendre transform of the cumulant generating function
Kubo, R. — The fluctuation-dissipation theorem · Rep. Prog. Phys. 29, 255 (1966) — the $\omega \to 0$ limit of the fluctuation-dissipation theorem recovering the equilibrium variance formulas of the present unit, identifying static susceptibilities with mean-square fluctuations
Mézard, M., Parisi, G. & Virasoro, M. A. — Spin Glass Theory and Beyond · World Scientific, 1987 — replica theory and sample-to-sample fluctuations in disordered systems; the order parameter is the Parisi function $q(x)$ encoding overlaps between replicas, the prototype non-self-averaging situation where individual realisations differ from the ensemble mean by amounts not suppressed by $1/\sqrt{N}$

Estimated time

beginner: 22m
intermediate: 50m
master: 90m