--- title: Thermodynamic Fluctuations and Response Functions module: Fluctuations and Response moduleNumber: 12 lessonNumber: 1 order: 1201 summary: > Thermodynamic variables are sharp only on average; a macroscopic system in equilibrium fluctuates about its mean values. Einstein inverted Boltzmann's $S=k_B\ln\Omega$ into a Gaussian probability for a fluctuation, $w\propto e^{\Delta S/k_B}$, and the second moments it predicts reproduce the response functions: $\langle\Delta E^2\rangle=k_BT^2C_V$, $\langle\Delta V^2\rangle=k_BTV\kappa_T$, $\langle\Delta M^2\rangle=k_BT\chi_T$. The variances diverge where the responses diverge, at a critical point, producing critical opalescence and the breakdown of the thermodynamic description. topics: [Fluctuations and Response] sources: - book: Landau & Lifshitz ref: "Statistical Physics, Part 1 (Vol. 5); §110–114 Fluctuations" - book: Reif ref: "Ch. 10 — Classical Theory of the Specific Heats; §10.6–10.8, Ch. 15 §15.1–15.3" - book: Kardar (Statistical Physics of Particles) ref: "Ch. 4 — Classical Statistical Mechanics; §4.7, and Ch. 3" - book: Pathria & Beale ref: "Ch. 13 — Fluctuations and Nonequilibrium Statistical Mechanics; §13.1–13.2" draft: false --- Equilibrium thermodynamics assigns a single number to each state variable. The statistical description behind it assigns a distribution: the energy of a canonical system, the volume of a subsystem at fixed pressure, the magnetization of a paramagnet all fluctuate about their means from instant to instant. The mean is what thermodynamics reports; the variance is a separate, measurable quantity, and it is fixed by the same equilibrium data. The [energy variance in the canonical ensemble](/statistical-mechanics/canonical/energy-fluctuations-and-ensemble-equivalence) already appeared as $\langle\Delta E^2\rangle=k_BT^2C_V$, a static fluctuation equal to a thermal response. That identity is one case of a general theorem. This lesson derives the general result from Einstein's inversion of the Boltzmann entropy formula, tabulates the fluctuation–response identities, and follows them to the critical point, where the responses diverge and the fluctuations grow until they invalidate the thermodynamic description that produced them. ## Einstein's inversion of the entropy formula The Boltzmann relation $S=k_B\ln\Omega$ counts the microstates $\Omega$ consistent with a macrostate. Read forward it computes entropy from a count. Einstein read it backward: the number of microstates consistent with a value $x$ of some internal variable is $\Omega(x)=e^{S(x)/k_B}$, and since every microstate of an isolated system is equally probable, the probability of observing the value $x$ is proportional to that count.[^ll-einstein] > **Definition (Einstein fluctuation probability).** For an isolated system whose > total entropy is $S(x)$ when an internal variable takes the value $x$, the > equilibrium probability density of $x$ is > $$w(x)\;\propto\;\exp\!\left[\frac{S(x)}{k_B}\right].$$ The entropy is maximal at the equilibrium value $\bar x$, so $S'(\bar x)=0$ and $S''(\bar x)<0$. Expanding to second order in the fluctuation $\Delta x=x-\bar x$, $$ S(x)=S(\bar x)-\tfrac12\,\lvert S''(\bar x)\rvert\,(\Delta x)^2+\cdots, $$ turns the probability into a Gaussian, $$ w(\Delta x)=\frac{1}{\sqrt{2\pi\sigma^2}}\exp\!\left[-\frac{(\Delta x)^2}{2\sigma^2}\right], \qquad \sigma^2=\langle\Delta x^2\rangle=\frac{k_B}{\lvert S''(\bar x)\rvert} =-\frac{k_B}{(\partial^2 S/\partial x^2)_{\bar x}}. $$ The variance of a fluctuation is $k_B$ divided by the curvature of the entropy at its maximum. A stiff entropy maximum (large curvature) suppresses fluctuations; a flat one permits large ones. The sign works out because stability requires the maximum, $S''<0$, so $\langle\Delta x^2\rangle>0$. The whole of Gaussian fluctuation theory is the systematic evaluation of that curvature for each choice of $x$. $$ % caption: The total entropy $S(x)$ is a concave maximum at the equilibrium value $\bar x$; exponentiating $S/k_B$ turns the parabolic top into a Gaussian probability $w(x)$ whose variance is $k_B$ over the curvature $|S''|$. \begin{tikzpicture}[>=Latex,font=\footnotesize,scale=1.0] \definecolor{acc}{HTML}{4A6FA5} \draw[->,black] (0,0)--(6.4,0) node[right,black]{$x$}; \draw[->,black] (0,0)--(0,4.2) node[above,black]{$S(x)$}; \draw[very thick] plot[domain=0.5:5.9,samples=80] (\x,{3.6-0.42*(\x-3.2)*(\x-3.2)}); \node at (1.5,3.35) {entropy}; \draw[black,dashed] (3.2,0)--(3.2,3.6); \node[black!70,font=\scriptsize,below] at (3.2,0) {equilibrium value}; \begin{scope}[yshift=-0.1cm] \draw[->,black] (7.2,0)--(13.6,0) node[right,black]{$x$}; \draw[->,black] (7.2,0)--(7.2,4.2) node[above,black]{$w(x)$}; \draw[very thick] plot[domain=7.7:13.1,samples=90] (\x,{3.8*exp(-0.9*(\x-10.4)*(\x-10.4))}); \node at (12.0,3.0) {probability}; \draw[<->,acc] (10.4-1.05,1.2)--(10.4+1.05,1.2) node[midway,below,acc,font=\scriptsize]{width set by curvature}; \draw[black,dashed] (10.4,0)--(10.4,3.8); \end{scope} \end{tikzpicture} $$ ## Fluctuations of a subsystem at fixed temperature and pressure Most systems of interest are not isolated but held at a fixed temperature $T_0$ and pressure $P_0$ by a large medium. The isolated-system formula still applies if $x$ is a variable of the small subsystem, provided $S(x)$ is the total entropy of subsystem plus medium. The change in total entropy when the subsystem fluctuates away from equilibrium is $-R_{\min}/T_0$, where $R_{\min}$ is the minimum work that an external agent would have to supply to produce that fluctuation reversibly. For a fluctuation at fixed $T_0,P_0$,[^ll-rmin] $$ R_{\min}=\Delta E-T_0\,\Delta S+P_0\,\Delta V, $$ with $\Delta E,\Delta S,\Delta V$ the changes in the subsystem's own energy, entropy, and volume. Expanding to second order in the subsystem's independent variations, the first-order terms cancel against the equilibrium conditions $T=T_0$, $P=P_0$, leaving a quadratic form. The compact result is $$ R_{\min}=\tfrac12\big(\Delta T\,\Delta S-\Delta P\,\Delta V\big), $$ so the fluctuation probability is $$ w\;\propto\;\exp\!\left[-\frac{\Delta T\,\Delta S-\Delta P\,\Delta V}{2k_BT}\right]. $$ Here $\Delta T,\Delta S,\Delta P,\Delta V$ are the deviations of the subsystem's temperature, entropy, pressure, and volume from their equilibrium values; only two of the four are independent. The quadratic form is diagonalized by choosing the right pair. > **Result.** With temperature and volume as the independent variables, expand > $\Delta S=(C_V/T)\,\Delta T+(\partial P/\partial T)_V\,\Delta V$ and > $\Delta P=(\partial P/\partial T)_V\,\Delta T+(\partial P/\partial V)_T\,\Delta V$. > The cross terms cancel, and > $$\Delta T\,\Delta S-\Delta P\,\Delta V=\frac{C_V}{T}(\Delta T)^2-\left(\frac{\partial P}{\partial V}\right)_T(\Delta V)^2.$$ > The probability factorizes into independent Gaussians in $\Delta T$ and > $\Delta V$. Because the quadratic form is diagonal, temperature and volume fluctuations are statistically independent: $\langle\Delta T\,\Delta V\rangle=0$. Reading off the two variances, $$ \langle(\Delta T)^2\rangle=\frac{k_BT^2}{C_V}, \qquad \langle(\Delta V)^2\rangle=-k_BT\left(\frac{\partial V}{\partial P}\right)_T =k_BT\,V\kappa_T, $$ where $\kappa_T=-V^{-1}(\partial V/\partial P)_T$ is the isothermal compressibility. The temperature fluctuation is inversely proportional to the heat capacity, the volume fluctuation directly proportional to the compressibility. Both variances are manifestly positive because stability forces $C_V>0$ and $\kappa_T>0$; the same inequalities that guarantee a stable equilibrium guarantee real fluctuations about it. $$ % caption: In the $(\Delta T,\Delta V)$ plane the fluctuation probability is a Gaussian whose contours are axis-aligned ellipses; the vanishing cross-correlation $\langle\Delta T\,\Delta V\rangle=0$ makes the principal axes lie along the coordinate directions, with half-widths $k_BT^2/C_V$ and $k_BTV\kappa_T$. \begin{tikzpicture}[>=Latex,font=\footnotesize,scale=1.0] \definecolor{acc}{HTML}{4A6FA5} \draw[->,black] (-3.2,0)--(3.4,0) node[right,black]{$T$ deviation}; \draw[->,black] (0,-1.7)--(0,2.8) node[above,black]{$V$ deviation}; \draw[acc,thick] (0,0) ellipse (2.6 and 1.4); \draw[acc] (0,0) ellipse (1.7 and 0.92); \draw[acc!70] (0,0) ellipse (0.85 and 0.46); \fill[black!75] (0,0) circle (1.6pt); \node[black!75,above right,font=\scriptsize] at (0,0) {equilibrium}; \draw[<->,black] (-2.6,-1.95)--(2.6,-1.95); \node[black!70,font=\scriptsize,below] at (0,-2.15) {temperature spread}; \draw[<->,black] (-3.15,-1.4)--(-3.15,1.4); \node[black!70,font=\scriptsize,above] at (-3.15,1.4) {volume spread}; \end{tikzpicture} $$ The conjugate choice, $(S,P)$ as the independent pair, diagonalizes the same form the other way and gives $$ \langle(\Delta S)^2\rangle=k_B C_P, \qquad \langle(\Delta P)^2\rangle=-k_B\left(\frac{\partial P}{\partial V}\right)_S, \qquad \langle\Delta S\,\Delta P\rangle=0. $$ The four variances of $(\Delta T,\Delta V,\Delta S,\Delta P)$ are set by the four response functions $C_V$, $\kappa_T$, $C_P$, $\kappa_S$. A cross-correlation that does not vanish is $\langle\Delta T\,\Delta S\rangle=k_BT$, obtained directly from the diagonal form; it restates $\langle\Delta E^2\rangle=k_BT^2C_V$ once $\Delta E$ is expressed through $\Delta S$ and $\Delta V$. ## The fluctuation–response identities Each extensive variable pairs with an intensive conjugate field through the energy, and the variance of the extensive variable equals $k_BT$ times its response to that field. The pattern is uniform across the ensembles. - **Energy and temperature.** In the canonical ensemble $\langle\Delta E^2\rangle=k_BT^2C_V$, derived from $\partial^2\ln Z/\partial\beta^2$. The spread of the energy is the heat capacity. - **Volume/number and pressure.** For a subsystem at fixed pressure $\langle\Delta V^2\rangle=k_BTV\kappa_T$; equivalently, in the [grand ensemble](/statistical-mechanics/grand-canonical/chemical-potential-fugacity-and-number-fluctuations) the particle number obeys $\langle\Delta N^2\rangle=k_BT\,(\partial N/\partial\mu)_{T,V}=k_BT\,N^2\kappa_T/V$. Density fluctuations are the compressibility. - **Magnetization and field.** For a magnet in a field $H$, $\langle\Delta M^2\rangle=k_BT\,(\partial M/\partial H)_T=k_BT\,\chi_T$. The spread of the magnetization is the isothermal susceptibility. > **Theorem (Fluctuation–response identity).** For an extensive variable $A$ > conjugate to an intensive field $f$ through the free energy, the equilibrium > variance of $A$ equals $k_BT$ times the static response of $A$ to $f$: > $$\langle\Delta A^2\rangle=k_BT\left(\frac{\partial\langle A\rangle}{\partial f}\right)_T.$$ The proof is one differentiation. Adding a field term $-fA$ to the Hamiltonian shifts every Boltzmann weight by $e^{\beta fA}$, so the partition function generates the moments of $A$: $\langle A\rangle=\beta^{-1}\partial\ln Z/\partial f$ and $\partial\langle A\rangle/\partial f=\beta(\langle A^2\rangle-\langle A\rangle^2)$. The identity is the statement that the first derivative of the mean (the response) and the second cumulant (the variance) come from consecutive derivatives of the same $\ln Z$. The response function is a susceptibility measured by applying a field; the variance is a fluctuation measured with no field applied. That they are equal means an equilibrium fluctuation predicts a nonequilibrium response, the seed of the [fluctuation–dissipation theorem](/statistical-mechanics/fluctuations/linear-response-and-the-fluctuation-dissipation-theorem) in its full dynamical form. $$ % caption: Each extensive variable and its conjugate field form a pair; the equilibrium variance of the extensive variable (measured with no field) equals $k_BT$ times its response function (measured by applying the field), so a fluctuation and a susceptibility carry the same information. \begin{tikzpicture}[>=Latex,font=\footnotesize,scale=1.0] \definecolor{acc}{HTML}{4A6FA5} \node[draw,thick,fill=black!6,align=center,minimum width=2.6cm,minimum height=1.0cm] (v1) at (0,3) {energy variance}; \node[draw,thick,fill=black!6,align=center,minimum width=2.6cm,minimum height=1.0cm] (v2) at (0,1.5) {volume variance}; \node[draw,thick,fill=black!6,align=center,minimum width=2.6cm,minimum height=1.0cm] (v3) at (0,0) {magnetization variance}; \node[draw,black,thick,fill=black!6,align=center,minimum width=2.6cm,minimum height=1.0cm] (r1) at (6,3) {heat capacity}; \node[draw,black,thick,fill=black!6,align=center,minimum width=2.6cm,minimum height=1.0cm] (r2) at (6,1.5) {compressibility}; \node[draw,black,thick,fill=black!6,align=center,minimum width=2.6cm,minimum height=1.0cm] (r3) at (6,0) {susceptibility}; \draw[<->,acc,very thick] (v1)--(r1) node[midway,above,black!70,font=\scriptsize]{$k_BT^2$}; \draw[<->,acc,very thick] (v2)--(r2) node[midway,above,black!70,font=\scriptsize]{$k_BTV$}; \draw[<->,acc,very thick] (v3)--(r3) node[midway,above,black!70,font=\scriptsize]{$k_BT$}; \node[black,font=\scriptsize] at (0,3.9) {no drive applied}; \node[black,font=\scriptsize] at (6,3.9) {drive applied}; \end{tikzpicture} $$ > **Worked example.** A cube of water of side $\ell=1\ \mathrm{\mu m}$ at > $T=300\ \mathrm{K}$ has isothermal compressibility > $\kappa_T\approx4.5\times10^{-10}\ \mathrm{Pa^{-1}}$ and volume > $V=10^{-18}\ \mathrm{m^3}$. Its volume fluctuation is > $$ > \langle\Delta V^2\rangle=k_BTV\kappa_T > =(1.38\times10^{-23})(300)(10^{-18})(4.5\times10^{-10})\ \mathrm{m^6} > \approx1.9\times10^{-48}\ \mathrm{m^6}, > $$ > so the root-mean-square relative fluctuation is > $\sqrt{\langle\Delta V^2\rangle}/V\approx1.4\times10^{-6}$. A micron-scale water > droplet holds its volume to about one part in a million; the same formula > applied to a mole gives one part in $10^{12}$, matching the $1/\sqrt{N}$ estimate > below. ## Relative fluctuations and the thermodynamic limit Every extensive variance is itself extensive: $\langle\Delta A^2\rangle\sim N$ because both $\langle A\rangle$ and the response are proportional to $N$. The root-mean-square fluctuation therefore grows as $\sqrt{N}$, while the mean grows as $N$, so the relative fluctuation shrinks: $$ \frac{\sqrt{\langle\Delta A^2\rangle}}{\langle A\rangle}\;\propto\;\frac{1}{\sqrt{N}}. $$ For a macroscopic sample with $N\sim10^{23}$ this is of order $10^{-12}$, and the distinction between a fluctuating variable and a sharp one is undetectable. The sharpness of thermodynamic quantities is the $1/\sqrt{N}$ law, and it is what lets equilibrium thermodynamics report single numbers rather than distributions. The law rests on two assumptions: that the response functions are finite, and that correlations extend over a microscopic range so that $N$ independent regions each fluctuate on their own. Both fail at a critical point. ## Critical divergence and critical opalescence The volume and density fluctuations are proportional to the isothermal compressibility $\kappa_T$. Along a liquid–gas coexistence curve $\kappa_T$ grows as the critical point is approached, and at the critical point the isotherm has a horizontal inflection, $(\partial P/\partial V)_T=0$, so $\kappa_T\to\infty$. The density variance diverges with it: $$ \langle\Delta N^2\rangle=\frac{k_BT\,N^2\kappa_T}{V}\;\xrightarrow[\;T\to T_c\;]{}\;\infty. $$ A diverging variance means the Gaussian theory itself breaks down — the quadratic expansion of the entropy loses its leading term — but the approach to the divergence is physical. As $\kappa_T$ grows, density fluctuations become correlated over a length $\xi$, the correlation length, which diverges as $\xi\sim\lvert T-T_c\rvert^{-\nu}$. When $\xi$ reaches the wavelength of visible light, the fluid scatters light strongly and turns milky. This is **critical opalescence**: a transparent fluid at its critical point becomes cloudy because density fluctuations on the scale of hundreds of nanometres refract light in every direction.[^ll-opal] The scattering intensity is proportional to the structure factor, which by the fluctuation identity is proportional to $\kappa_T$, so the opalescence is a direct optical readout of the diverging compressibility. $$ % caption: The isothermal compressibility and the density variance $\langle\Delta N^2\rangle=k_BTN^2\kappa_T/V$ both diverge as the temperature approaches $T_c$; the correlation length $\xi$ grows until density fluctuations scatter visible light, the critical opalescence. \begin{tikzpicture}[>=Latex,font=\footnotesize,scale=1.0] \definecolor{acc}{HTML}{4A6FA5} \draw[->,black] (0,0)--(7.4,0) node[right,black]{$T$}; \draw[->,black] (0,0)--(0,4.3) node[above,black]{compressibility}; \draw[acc,very thick] plot[domain=0.35:2.45,samples=110] (\x,{0.55+0.9/((3.0-\x)*(3.0-\x))}); \draw[acc,very thick] plot[domain=3.55:6.9,samples=110] (\x,{0.55+0.9/((\x-3.0)*(\x-3.0))}); \draw[black,dashed,thick] (3.0,0)--(3.0,4.1); \node[black!70,below] at (3.0,0) {$T_c$}; \node[black!70,font=\scriptsize,align=center] at (5.5,2.7) {variance diverges\\at criticality}; \end{tikzpicture} $$ ## The limits of the thermodynamic description Away from criticality the relative fluctuation of an extensive variable is $O(N^{-1/2})$, and thermodynamics is exact in the limit $N\to\infty$. Near a continuous transition the argument fails in a specific way. The independent fluctuating regions have volume $\xi^d$ in $d$ dimensions, so the effective number of independent regions in a sample of volume $V$ is $V/\xi^d$, not $N$. As $\xi\to\infty$ this number falls toward one, and the relative fluctuation within a correlation volume approaches unity. The order parameter is then not sharp, and the mean-field (Gaussian) treatment of its fluctuations becomes self-inconsistent. The condition that fluctuations remain small compared with the mean order parameter is the **Ginzburg criterion**, and where it is violated the critical exponents are no longer those of Gaussian theory. This is the quantitative reason the mean-field exponents of [Landau theory](/statistical-mechanics/phase-transitions/critical-exponents-and-landau-theory) disagree with experiment below the upper critical dimension: fluctuations, small everywhere else, dominate the critical region. ## Summary - **Einstein's inversion** of $S=k_B\ln\Omega$ gives the fluctuation probability $w\propto e^{S/k_B}$; expanding the entropy about its maximum yields a Gaussian with $\langle\Delta x^2\rangle=-k_B/(\partial^2 S/\partial x^2)$ — the variance is $k_B$ over the entropy curvature. - For a subsystem at fixed $T,P$ the probability is $w\propto\exp[-(\Delta T\,\Delta S-\Delta P\,\Delta V)/2k_BT]$; choosing $(T,V)$ as independent variables diagonalizes it, giving $\langle(\Delta T)^2\rangle=k_BT^2/C_V$ and $\langle(\Delta V)^2\rangle=k_BTV\kappa_T$ with $\langle\Delta T\,\Delta V\rangle=0$. - The **fluctuation–response identities** $\langle\Delta E^2\rangle=k_BT^2C_V$, $\langle\Delta V^2\rangle=k_BTV\kappa_T$, $\langle\Delta N^2\rangle=k_BTN^2\kappa_T/V$, $\langle\Delta M^2\rangle=k_BT\chi_T$ all follow from $\langle\Delta A^2\rangle=k_BT\,(\partial\langle A\rangle/\partial f)_T$: an equilibrium variance equals $k_BT$ times a static susceptibility. - Relative fluctuations scale as $1/\sqrt{N}$ and are negligible for macroscopic samples, which is why thermodynamic variables are effectively sharp. - At a critical point the responses diverge, so the variances diverge; the correlation length grows to the scale of visible light (critical opalescence), and the Ginzburg criterion marks where fluctuations invalidate the mean-field description. [^ll-einstein]: **Landau & Lifshitz**, _Statistical Physics, Part 1_ (Vol. 5), §110–112 — the probability of a fluctuation as $w\propto e^{\Delta S/k_B}$ and its reduction to a Gaussian through the entropy curvature. [^ll-rmin]: **Landau & Lifshitz**, _Statistical Physics, Part 1_ (Vol. 5), §112 — the minimum work $R_{\min}=\Delta E-T_0\Delta S+P_0\Delta V$ and the quadratic form $R_{\min}=\tfrac12(\Delta T\,\Delta S-\Delta P\,\Delta V)$; **Reif**, _Fundamentals of Statistical and Thermal Physics_, §15.2 gives the same result through the availability. [^ll-opal]: **Landau & Lifshitz**, _Statistical Physics, Part 1_ (Vol. 5), §116, and **Pathria & Beale**, _Statistical Mechanics_ (4th ed.), §13.2 — the divergence of the density fluctuation with the compressibility, the Ornstein–Zernike correlation function, and critical opalescence.