Heat Capacity and the Fluctuation-Response Relation

Poke a system with a change in temperature and it absorbs energy; how much it absorbs per degree is the heat capacity $C_V$ . Leave that same system utterly alone in a heat bath and its energy still trembles — trading quanta with the bath, wandering above and below its mean. These look like two unrelated facts, one about response to a deliberate push, one about spontaneous jitter at rest. The central surprise of this article is that they are the same number: the heat capacity is exactly the variance of the equilibrium energy fluctuations, divided by $k_BT^2$ . This identity, $C_V = \langle(\Delta E)^2\rangle/k_BT^2$ , is the simplest fluctuation–response (or fluctuation–dissipation) relation in physics — the prototype of a family that also links magnetic susceptibility to spin fluctuations and compressibility to density fluctuations. Its companion article, Energy Fluctuations in the Canonical Ensemble, derives the identity from $\ln Z$ ; here we read it the other way. Instead of asking "how big are the fluctuations?" we ask "what does the fluctuation tell us?" — and the answer is a response coefficient you can measure with a calorimeter, or extract from a simulation without ever touching the temperature. Two consequences fall out at once. Because a variance can never be negative, $C_V \ge 0$ is guaranteed — thermodynamic stability is baked into the statistics, not assumed. And because energies of independent parts add, the relative energy fluctuation shrinks as $1/\sqrt N$ , which is precisely why the fluctuating canonical energy behaves like the fixed microcanonical one for macroscopic matter. The article's owned result makes the "response tracks fluctuation" slogan quantitative on the two-level system, and every number here is python-verified.

⚡ Key Takeaways

The heat capacity $C_V=(\partial\langle E\rangle/\partial T)_V$ is a response: energy absorbed per degree of warming.
In the canonical ensemble it equals a fluctuation: $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ — the simplest fluctuation–dissipation relation.
Because a variance is non-negative, this forces $C_V\ge0$ — thermodynamic stability guaranteed by statistics, not postulated.
Relative energy fluctuations scale as $1/\sqrt N$ , so canonical and microcanonical ensembles agree for macroscopic systems (ensemble equivalence).
A diverging $C_V$ at a continuous phase transition is a diverging fluctuation — the gateway to critical phenomena.
Key idea: the heat capacity you measure by warming a system is identically the size of the energy jitter the system undergoes at rest — response equals fluctuation, $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ .

💡 The Core Insight

The constant-volume heat capacity, a response to changing temperature, is identically the fluctuation of the equilibrium energy: $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ . Since a variance is non-negative, $C_V\ge0$ automatically (stability), and since $C_V,\langle E\rangle\propto N$ the relative fluctuation vanishes as $1/\sqrt N$ (ensemble equivalence). This is the prototype fluctuation–dissipation relation.

The physical picture

Picture a small copper block clamped in thermal contact with a large water bath. Its energy is not a fixed number — it drifts up and down as it swaps thermal quanta with the bath. Now ask two questions about that block. Response: if I nudge the bath temperature up by one kelvin, how much energy does the block soak up? That is $C_V$ . Fluctuation: if I just watch the block sit at fixed temperature, how widely does its energy wander around the mean? That is $\langle(\Delta E)^2\rangle$ . The claim of this article is that these are one measurement in disguise. The reason is that both are controlled by the same feature of the energy spectrum. A block whose accessible states are densely spread around the current thermal energy can readily shuffle energy around: it fluctuates a lot, and it also absorbs a lot when warmed. A block with a big gap up to its next states is stiff both ways — it barely fluctuates, and it barely responds. Softness and jitter travel together. Formally, the Boltzmann weight $e^{-\beta E}$ is a moment-generating engine in the variable $-\beta$ : one derivative of $\ln Z$ gives the mean energy, a second gives its variance, and since temperature enters only through $\beta$ , the temperature-derivative of the mean (the heat capacity) is forced to be the variance up to the fixed factor $k_BT^2$ .

Figure

Two questions, one answer. Warm the system (response, left): it absorbs C_V per degree. Watch it at rest (fluctuation, right): its energy jitters with variance ⟨ΔE²⟩. The two are identical, C_V = ⟨ΔE²⟩ / k_BT².

The most striking payoff is the two-level (Schottky) system: $N$ ions each with a ground state and one excited state a gap $\varepsilon$ above. As you warm it from absolute zero, the heat capacity rises, reaches a peak, then falls again — a distinctive bump. Why a peak and not a monotone rise? The energy variance per ion actually climbs monotonically with $T$ and saturates at $\varepsilon^2/4$ (both levels equally full). It is the $1/T^2$ factor in $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ that turns that monotone fluctuation into a peaked response — the fluctuation is largest at high $T$ , but the response is weighted toward lower $T$ . The peak sits at $k_BT\approx0.42\,\varepsilon$ (that is $\beta\varepsilon\approx2.40$ ), and experimentalists read the level gap $\varepsilon$ straight off the peak temperature. This is how a two-level structure — nuclear spins, crystal-field doublets — is detected in a real material.

Figure

The owned result. The per-ion energy variance ⟨ΔE²⟩(T) (amber) climbs monotonically and saturates at ε²/4. Multiply by 1/k_BT² and it becomes the peaked Schottky heat capacity C_V(T) (blue), peaking at βε ≈ 2.40. Response is fluctuation, reweighted by 1/T².

Finally, the $1/\sqrt N$ smallness is why thermodynamics is possible at all. A kilogram of copper has $\sim10^{25}$ atoms, so its relative energy fluctuation is about one part in $10^{12}$ — completely invisible on any macroscopic scale. Yet those microscopic, invisible fluctuations are exactly what fix the macroscopic $C_V$ you read off a calorimeter. The tiny jitter you can never see determines the response you always measure.

The core equations

The heat capacity at constant volume is the thermodynamic definition we start from — the response of the mean energy to a change in the bath temperature at fixed $V,N$ :

C_V=\left(\frac{\partial\langle E\rangle}{\partial T}\right)_{V,N} \qquad (1)

where: $C_V$ — constant-volume heat capacity (a response coefficient); $\langle E\rangle$ — mean internal energy; $T$ — absolute temperature; $V,N$ held fixed. This is what a calorimeter measures: divide the heat supplied by the temperature rise. The mean energy and its variance are the first two $\beta$ -derivatives of $\ln Z$ (derived in full in Energy Fluctuations in the Canonical Ensemble). We recall them because they are the bridge from the spectrum to the response:

\langle E\rangle=-\frac{\partial\ln Z}{\partial\beta},\qquad \langle(\Delta E)^2\rangle\equiv\langle E^2\rangle-\langle E\rangle^2=\frac{\partial^2\ln Z}{\partial\beta^2},\qquad \beta=\frac1{k_BT} \qquad (2)

where: $Z=\sum_i e^{-\beta E_i}$ — canonical partition function; $\langle(\Delta E)^2\rangle$ — variance of the energy in the canonical distribution; $k_B$ — Boltzmann constant. That $\ln Z$ generates the mean and the variance is the cumulant-generating structure that makes the whole identity inevitable. The central result of the article — the fluctuation–response relation — states that the response $C_V$ is the fluctuation $\langle(\Delta E)^2\rangle$ , scaled by $k_BT^2$ :

C_V=\frac{\langle(\Delta E)^2\rangle}{k_BT^2} \qquad (3)

where: the numerator is a spontaneous equilibrium fluctuation, the left side is a driven thermal response, and $k_BT^2$ is the exact conversion factor. Read left-to-right it is the fluctuation–dissipation theorem for energy; read right-to-left it is a recipe for computing $C_V$ from a variance you can sample in a simulation without differentiating anything. Because the right-hand side is a variance over $k_BT^2$ , and both are non-negative, stability is immediate — a statistical guarantee, not a thermodynamic postulate:

C_V=\frac{\langle(\Delta E)^2\rangle}{k_BT^2}\ge0 \qquad (4)

where: $\ge0$ follows because $\langle(\Delta E)^2\rangle\ge0$ (a variance) and $k_BT^2>0$ . A material with negative $C_V$ at fixed $V$ would need a negative variance — impossible. (Self-gravitating systems appear to violate this, but they are not in the canonical ensemble; see the warning below.) The same content appears as the concavity of the free energy in temperature, the route most useful for thermodynamic manipulations:

C_V=-T\left(\frac{\partial^2 F}{\partial T^2}\right)_V=T\left(\frac{\partial S}{\partial T}\right)_V,\qquad F=-k_BT\ln Z \qquad (5)

where: $F$ — Helmholtz free energy; $S=-(\partial F/\partial T)_V$ — entropy. $C_V\ge0$ is equivalent to $F(T)$ being concave — the temperature-stability condition of thermodynamics.

Derivation AHeat capacity equals the scaled energy variance

This is the load-bearing identity of the article: we prove that the thermodynamic response $C_V$ is exactly the statistical fluctuation $\langle(\Delta E)^2\rangle$ divided by $k_BT^2$ , starting only from the definition of $C_V$ and the partition-function expression for $\langle E\rangle$ . The strategy is to differentiate the mean energy with respect to $T$ , convert the $T$ -derivative into a $\beta$ -derivative via the chain rule, recognize the second log-derivative of $Z$ as the variance, and watch two minus signs conspire to make the result manifestly positive.

Step 1. Write the response as a $\beta$ -derivative. By definition $C_V=\partial\langle E\rangle/\partial T$ . Since $\beta=1/k_BT$ , the chain rule gives $\partial/\partial T=(\partial\beta/\partial T)\,\partial/\partial\beta$ with $\partial\beta/\partial T=-1/k_BT^2=-k_B\beta^2$ , so $C_V=-\dfrac{1}{k_BT^2}\dfrac{\partial\langle E\rangle}{\partial\beta}$ . Why: temperature enters the canonical distribution only through $\beta$ , so every temperature-response is really a $\beta$ -response, carried across by the Jacobian $-1/k_BT^2$ .

Step 2. Differentiate the mean energy. With $\langle E\rangle=-\partial_\beta\ln Z$ , we have $\partial_\beta\langle E\rangle=-\partial_\beta^2\ln Z$ . From the generating-function structure, $\partial_\beta^2\ln Z=\langle E^2\rangle-\langle E\rangle^2=\langle(\Delta E)^2\rangle$ , so $\partial_\beta\langle E\rangle=-\langle(\Delta E)^2\rangle$ . Why: the second derivative of $\ln Z$ is the variance — the quotient rule turns $\partial_\beta^2 Z/Z$ minus $(\partial_\beta Z/Z)^2$ into $\langle E^2\rangle-\langle E\rangle^2$ .

Step 3. Combine the two minus signs. Substituting into Step 1, $C_V=-\dfrac{1}{k_BT^2}\cdot\big(-\langle(\Delta E)^2\rangle\big)=\dfrac{\langle(\Delta E)^2\rangle}{k_BT^2}$ . Why: the Jacobian carries one minus sign and $\partial_\beta\langle E\rangle$ carries another; they cancel, leaving a manifestly positive ratio of a variance to a positive number.

Step 4. Read off stability. Because $\langle(\Delta E)^2\rangle\ge0$ (it is a mean square) and $k_BT^2>0$ , we conclude $C_V\ge0$ with no further assumption. Equivalently, $C_V=-T(\partial^2 F/\partial T^2)_V\ge0$ means $F(T)$ is concave — the thermodynamic stability condition emerges as a corollary of the positivity of a variance. Why: statistical mechanics does not need to postulate $C_V\ge0$ ; it inherits it from the unassailable fact that a variance cannot be negative. *Verified: the three routes $C_V=\partial\langle E\rangle/\partial T$ , $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ , and $C_V=k_B\beta^2\,\partial_\beta^2\ln Z$ are shown to be symbolically identical on the two-level system in verify_aq-sm-heat-capacity-fluctuation-response.py.*

Derivation BThe Schottky bump: why the response peaks though the fluctuation does not

Here is the article's owned structural result, made quantitative on the two-level system. The goal is to explain the distinctive peak in the Schottky heat capacity: we compute the per-ion energy variance and the heat capacity as functions of temperature, show that the variance is monotone (it only grows and saturates), and prove that the peak in $C_V$ is entirely manufactured by the $1/T^2$ weighting in $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ . This is "response tracks fluctuation" made exact — the two curves are the same object, reweighted.

Step 1. Set up the spectrum. Each ion has levels $0$ and $\varepsilon$ , so $Z_1=1+e^{-\beta\varepsilon}$ and $\langle E\rangle_1=\varepsilon/(e^{\beta\varepsilon}+1)$ . The per-ion energy variance is $\langle(\Delta E)^2\rangle_1=\varepsilon^2\,\dfrac{e^{\beta\varepsilon}}{(e^{\beta\varepsilon}+1)^2}$ . Why: differentiating $\langle E\rangle_1$ once more with respect to $-\beta$ , or directly forming $\langle E^2\rangle_1-\langle E\rangle_1^2$ with $\langle E^2\rangle_1=\varepsilon^2/(e^{\beta\varepsilon}+1)$ , gives the standard two-level variance.

Step 2. Show the variance is monotone in $T$ . Writing $x=\beta\varepsilon=\varepsilon/k_BT$ , the per-ion variance is $\varepsilon^2\,e^{x}/(e^{x}+1)^2$ . As a function of $x$ this decreases monotonically for $x>0$ ; as a function of temperature it therefore increases monotonically, rising from $0$ at $T=0$ and saturating at $\varepsilon^2/4$ as $T\to\infty$ . There is no interior peak in the fluctuation. Why: at high $T$ both levels are equally populated ( $p_0=p_1=\tfrac12$ ), giving the maximal spread $\varepsilon^2/4$ ; the fluctuation just grows toward that ceiling.

Step 3. Form the heat capacity and find its peak. Now $C_V^{(1)}=\langle(\Delta E)^2\rangle_1/k_BT^2=k_B x^2\,\dfrac{e^{x}}{(e^{x}+1)^2}$ . The extra $x^2=(\varepsilon/k_BT)^2$ suppresses the high- $T$ region and creates an interior maximum. Setting $dC_V^{(1)}/dx=0$ gives a transcendental equation whose root is $x=\beta\varepsilon\approx2.3994$ , i.e. the peak sits at $k_BT_{\text{peak}}\approx0.417\,\varepsilon$ , with peak value $C_V^{(1)}\approx0.4392\,k_B$ . Why: multiplying a monotone-decreasing $e^{x}/(e^{x}+1)^2$ by a rising $x^2$ produces a single interior peak — the Schottky anomaly — located where the two tendencies balance.

Step 4. State the connection honestly. The fluctuation $\langle(\Delta E)^2\rangle(T)$ and the response $C_V(T)$ are not peaked at the same temperature — the fluctuation has no finite- $T$ peak at all. What is exact is the pointwise identity $C_V(T)=\langle(\Delta E)^2\rangle(T)/k_BT^2$ at every temperature: the response is the fluctuation, seen through the lens of $1/T^2$ . The famous Schottky peak is the fingerprint of that reweighting, and its location $\beta\varepsilon\approx2.40$ delivers the level gap. Why: "response tracks fluctuation" is precise as an identity between functions, not as a coincidence of peak positions; the $1/T^2$ factor is a physical, not cosmetic, part of the relation. *Verified: $C_V^{(1)}=k_B x^2 e^{x}/(e^{x}+1)^2$ , the variance saturation at $\varepsilon^2/4$ , the $C_V$ peak at $x=2.3994$ (peak value $0.4392\,k_B$ ), and the pointwise identity $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ at several temperatures are all confirmed in the verify script.*

Limiting cases

The fluctuation–response relation $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ across regimes:

Case	Result	Meaning
Thermodynamic limit	$\dfrac{\sqrt{\langle(\Delta E)^2\rangle}}{\langle E\rangle}=\dfrac{\sqrt{k_BT^2C_V}}{\langle E\rangle}\sim\dfrac1{\sqrt N}\to0$	$C_V\propto N$ and $\langle E\rangle\propto N$ , so relative fluctuations vanish; canonical $\equiv$ microcanonical.
Classical ideal gas	$C_V=\tfrac32Nk_B,\ \ \langle(\Delta E)^2\rangle=\tfrac32N(k_BT)^2$	Relative fluctuation $=\sqrt{2/3N}$ , explicitly $O(N^{-1/2})$ ; $=8.2\times10^{-13}$ at $N=10^{24}$ .
Two-level (Schottky)	$C_V=k_B(\beta\varepsilon)^2\dfrac{e^{\beta\varepsilon}}{(e^{\beta\varepsilon}+1)^2}$	Peaks at $\beta\varepsilon\approx2.40$ ( $k_BT\approx0.42\varepsilon$ ); the level gap is read off the peak.
Low $T$ (quantum solid)	$C_V\to0$	Energy frozen into the ground state; fluctuations suppressed, $\langle(\Delta E)^2\rangle\to0$ — third law.
Critical point	$C_V\sim\lvert T-T_c\rvert^{-\alpha}$	Diverging long-ranged energy fluctuations drive $C_V\to\infty$ ; the $1/\sqrt N$ picture breaks.

⚠️ “A variance is always small, so C_V can never blow up, and it can never be negative anyway”

Both halves need care. The relation $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ does force $C_V\ge0$ — in the canonical ensemble. But it does not cap $C_V$ : at a continuous phase transition the correlation length diverges, energy fluctuations become long-ranged instead of the sum of $N$ independent small pieces, the $1/\sqrt N$ central-limit scaling fails, and $C_V\to\infty$ . A finite variance is not guaranteed. The other trap is the ensemble. This identity is specifically canonical — it needs a system that genuinely exchanges energy with a reservoir so that $E$ fluctuates. In the microcanonical ensemble energy is fixed, $\langle(\Delta E)^2\rangle=0$ identically, yet $C_V\ne0$ : there $C_V$ is read from $\partial T/\partial E$ , not from a variance. Finally, systems with long-range interactions (self-gravitating stars, some nuclear matter) can show negative microcanonical heat capacity — no contradiction, because they are not described by a canonical distribution with a normalizable $Z$ , so the positivity argument simply does not apply. Match the ensemble to the physics before quoting the relation.

Worked examples

1 · Reading a level gap off a Schottky peak (a real crystal-field measurement)

A dilute paramagnetic salt has a doublet of ions with a crystal-field gap $\varepsilon$ such that $\varepsilon/k_B=20\ \mathrm{K}$ . Where (in temperature) does the molar heat capacity peak, and how large is the peak? One mole of ions.

Step 1. Locate the peak in $x=\varepsilon/k_BT$ . The two-level heat capacity $C_V=R\,x^2 e^{x}/(e^{x}+1)^2$ (per mole, $Nk_B\to R$ ) peaks where $dC_V/dx=0$ , whose numerical root is $x=\beta\varepsilon\approx2.3994$ .

Step 2. Convert to a temperature. Since $x=\varepsilon/k_BT=(\varepsilon/k_B)/T=20\,\mathrm{K}/T$ , the peak is at $T_{\text{peak}}=20\,\mathrm{K}/2.3994=\mathbf{8.34\ K}$ .

Step 3. Compute the peak height. At $x=2.3994$ , the factor $x^2 e^{x}/(e^{x}+1)^2=0.4392$ , so $C_V^{\text{peak}}=0.4392\,R=0.4392\times8.314\ \mathrm{J\,mol^{-1}K^{-1}}=\mathbf{3.65\ J\,mol^{-1}K^{-1}}$ .

Step 4. The physics. The heat capacity bump at $8.34\ \mathrm{K}$ is a direct fingerprint of the $20\ \mathrm{K}$ level gap — the very fact that a two-level structure exists is what produces a peaked (rather than monotone) $C_V$ , and its location fixes $\varepsilon$ . Below the peak the gap freezes out excitations ( $C_V\to0$ ); above it both levels saturate ( $C_V\to0$ ); the bump in between is where the response, weighted by $1/T^2$ , is maximal.

2 · How invisible are the energy fluctuations of a mole of gas? (exam level)

Take one mole of monatomic ideal gas at $T=300\ \mathrm{K}$ . Compute the RMS energy fluctuation $\sqrt{\langle(\Delta E)^2\rangle}$ and express it as a fraction of the internal energy $\langle E\rangle$ .

Step 1. Internal energy. $\langle E\rangle=\tfrac32 N_A k_BT=\tfrac32(6.022\times10^{23})(1.381\times10^{-23}\,\mathrm{J/K})(300\,\mathrm{K})=\mathbf{3742\ J}$ .

Step 2. Variance from the fluctuation–response relation. With $C_V=\tfrac32 N_A k_B$ , $\langle(\Delta E)^2\rangle=k_BT^2 C_V=\tfrac32 N_A(k_BT)^2$ . Numerically $\tfrac32(6.022\times10^{23})(4.14\times10^{-21}\,\mathrm{J})^2=1.55\times10^{-17}\ \mathrm{J^2}$ , so $\sqrt{\langle(\Delta E)^2\rangle}=\mathbf{3.94\times10^{-9}\ J}$ .

Step 3. The fraction. $\dfrac{\sqrt{\langle(\Delta E)^2\rangle}}{\langle E\rangle}=\dfrac{3.94\times10^{-9}}{3742}=\mathbf{1.05\times10^{-12}}$ , which equals $\sqrt{2/3N_A}$ exactly.

Step 4. The lesson. The energy of a mole of gas is defined to about one part in $10^{12}$ — a spread of a few nanojoules on top of kilojoules of energy — so the fluctuating canonical energy is, for all macroscopic purposes, as sharp as the fixed microcanonical energy. This $1/\sqrt N$ smallness, read straight off the heat capacity through the fluctuation relation, is the quantitative content of ensemble equivalence: it is why we may use the mathematically convenient canonical ensemble to compute the properties of systems that are, physically, at fixed energy.

fluctuation-dissipation theorem

response theory · Callen–Welton 1951 · the fluctuation–response principle

The fluctuation–dissipation theorem is the general statement that the way a system responds to a small external perturbation is completely determined by the spontaneous fluctuations it undergoes in thermal equilibrium — no driving required. The heat-capacity relation $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ is its simplest and oldest instance: Einstein wrote the energy-fluctuation formula in 1904, and the same structure appeared in Johnson and Nyquist's 1928 analysis of thermal ("Johnson–Nyquist") noise in resistors, where the voltage fluctuations across a resistor at temperature $T$ fix its dissipative resistance. The general theorem was proved by Herbert Callen and Theodore Welton in 1951, and cast in its modern linear-response form by Ryogo Kubo in 1957, linking any response function to an equilibrium time-correlation of the conjugate fluctuation. The principle is one of the deepest unifying ideas in statistical physics: susceptibility is spin fluctuation, compressibility is density fluctuation, mobility is velocity fluctuation (the Einstein relation), and heat capacity is energy fluctuation. It is why we can extract macroscopic transport and response coefficients from simulations that merely watch a system fluctuate at equilibrium, and why the same $k_BT$ that measures thermal jitter also measures the strength of dissipation.

⏳ History

1904

— Einstein derives the mean-square energy fluctuation of a system in a heat bath, ⟨ΔE²⟩ = k_BT²C_V.

1910

— Einstein applies fluctuation theory to critical opalescence, where density fluctuations diverge.

1928

— Johnson measures, and Nyquist explains, thermal voltage noise in resistors — an early fluctuation-dissipation relation.

1951

— Callen and Welton prove the general fluctuation-dissipation theorem in the Physical Review.

1957

— Kubo formulates linear response, expressing response functions as equilibrium correlation functions.

Where this lives today

The fluctuation–response relation is the daily workhorse of computational statistical mechanics. In Monte Carlo and molecular-dynamics simulations, one almost never obtains the heat capacity by numerically differentiating $\langle E\rangle(T)$ — that amplifies statistical noise ruinously. Instead $C_V$ is read directly from the measured energy variance in a single run at fixed $T$ : $C_V=(\langle E^2\rangle-\langle E\rangle^2)/k_BT^2$ . The same trick gives the magnetic susceptibility from magnetization variance and the compressibility from particle-number variance. Locating a phase transition then becomes a matter of watching for a variance peak: the specific-heat divergence of the 2D Ising model, the $\lambda$ -anomaly of liquid helium, and the melting of model solids are all found this way, and finite-size scaling of the variance peak yields the critical exponent $\alpha$ and the transition temperature $T_c$ . Beyond equilibrium thermodynamics, the same principle powers modern measurement. Passive microrheology infers the viscoelastic response of a cell or polymer network purely from the thermal jiggling of an embedded bead — the fluctuation spectrum, via the fluctuation–dissipation theorem, is the response spectrum. In electronics, Johnson–Nyquist noise sets the temperature-dependent noise floor of every resistor, and is used in reverse as a primary thermometer (Johnson-noise thermometry) traceable to $k_B$ itself: fluctuations and response are two faces of the single number $k_BT$ .

Check yourself

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., §3.6 · Kardar, Statistical Physics of Particles, Ch. 4 · Reif, Fundamentals of Statistical and Thermal Physics, §6.5 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 6 · Landau & Lifshitz, Statistical Physics Part 1, §112 · Callen & Welton, Phys. Rev. 83, 34 (1951) · Kubo, J. Phys. Soc. Japan 12, 570 (1957).

Provenance. The relation $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ follows from $\langle E\rangle=-\partial_\beta\ln Z$ and $\langle(\Delta E)^2\rangle=\partial_\beta^2\ln Z$ (Pathria §3.6; Kardar Ch. 4; Reif §6.5). A verifier should reproduce: the three equivalent routes to $C_V$ (definition, variance, $k_B\beta^2\partial_\beta^2\ln Z$ ); the two-level Schottky $C_V=k_B(\beta\varepsilon)^2 e^{\beta\varepsilon}/(e^{\beta\varepsilon}+1)^2$ with peak at $\beta\varepsilon\approx2.3994$ (peak value $0.4392\,k_B$ ), and the monotone variance saturating at $\varepsilon^2/4$ ; the ideal-gas $C_V=\tfrac32Nk_B$ , variance $\tfrac32N(k_BT)^2$ , and relative fluctuation $\sqrt{2/3N}$ . Numerical instances — the crystal-field peak at $T=8.34\ \mathrm{K}$ and $C_V^{\text{peak}}=3.65\ \mathrm{J\,mol^{-1}K^{-1}}$ ; the mole-of-gas RMS fluctuation $3.94\times10^{-9}\ \mathrm{J}$ and fraction $1.05\times10^{-12}$ — are recomputed in verify_aq-sm-heat-capacity-fluctuation-response.py. Callen–Welton 1951, Kubo 1957, Johnson–Nyquist 1928 dates web-verified. $k_B$ , $N_A$ , $R$ (CODATA/SI).

Builds on: Canonical Ensemble · Canonical Partition Function · Boltzmann Factor · Thermodynamic Quantities from the Partition Function · Energy Fluctuations in the Canonical Ensemble. Leads to: Grand Canonical Ensemble · Grand Potential · Helmholtz Free Energy from the Partition Function · Particle-Number Fluctuations.

If C_V is just a fluctuation, why can I measure it with a calorimeter without ever seeing any fluctuations?

Because the two are the same number, and the calorimeter is measuring the aggregate effect of the microscopic fluctuations. When you supply heat $\delta Q$ and watch the temperature rise $\delta T$ , you are probing exactly how easily the system's energy distribution shifts — and how easily it shifts is set by how wide it already is. A wide (large-variance) energy distribution moves a lot per unit temperature change (large $C_V$ ); a narrow one barely moves (small $C_V$ ). You never see the individual fluctuations because their relative size is $\sim1/\sqrt N\sim10^{-12}$ , far below any calorimeter's resolution. But their collective statistical width is precisely the quantity $\langle(\Delta E)^2\rangle$ , and $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ tells you that measuring the response and measuring the fluctuation are two experiments returning one answer.

The energy variance of the two-level system only grows with temperature, yet C_V peaks. How can the response peak if the fluctuation doesn't?

This is the subtle and important point. The fluctuation and the response are linked by $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ — an identity at every temperature, not a statement that they peak together. The per-ion variance $\varepsilon^2 e^{x}/(e^{x}+1)^2$ (with $x=\varepsilon/k_BT$ ) rises monotonically with $T$ and saturates at $\varepsilon^2/4$ once both levels are equally populated. What makes $C_V$ peaked is the explicit $1/T^2=(x/\varepsilon)^2 k_B^2$ factor: it suppresses the high-temperature region where the variance is largest, and the product $x^2\times$ (variance-shape) has a single interior maximum at $x\approx2.40$ . So the response is the fluctuation reweighted toward lower temperature. The Schottky peak is the fingerprint of that reweighting — and it is physically real, because the $1/T^2$ is a genuine part of the fluctuation–response relation, not an artifact.

Does C_V ≥ 0 really follow with no assumptions, and does it always hold?

In the canonical ensemble it follows with essentially no assumptions: $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ is a variance divided by a positive number, and a variance is a mean of a square, hence non-negative. So any system describable by a canonical distribution with a well-defined partition function has $C_V\ge0$ — thermodynamic stability is a theorem, not a postulate. Where it "fails" is where the premise fails. Systems with long-range interactions — self-gravitating star clusters, some nuclear and cluster systems — can exhibit negative microcanonical heat capacity, but they are not in a canonical state (their $Z$ does not converge / the ensembles are inequivalent), so the argument simply does not apply to them. Within its proper domain — short-range interactions, a normalizable $Z$ — $C_V\ge0$ is airtight.

How is this the same relation as magnetic susceptibility being spin fluctuations?

It is the same structure with a different conjugate pair. In every case you differentiate the partition function twice with respect to a control parameter, and the second derivative of $\ln Z$ produces a variance. For energy, the control is $-\beta$ (temperature) and the variance is energy fluctuations, giving $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ . For magnetization, the control is the magnetic field $h$ : $\partial\langle M\rangle/\partial h=\chi=\langle(\Delta M)^2\rangle/k_BT$ — susceptibility equals magnetization fluctuations. For particle number, the control is the chemical potential $\mu$ : the isothermal compressibility is set by $\langle(\Delta N)^2\rangle$ . All are instances of the fluctuation–dissipation theorem: a linear response coefficient equals an equilibrium fluctuation of the conjugate observable, divided by $k_BT$ (times $T$ for the energy case). The recurring theme is that the same Boltzmann weight $e^{-\beta E}$ that fluctuates is the one you differentiate to get a response.

Why is C_V a good detector of phase transitions?

Because $C_V\propto\langle(\Delta E)^2\rangle$ , so a peak or divergence in the heat capacity is a direct readout of anomalously large energy fluctuations. Away from a transition the energy is a sum of many nearly-independent local pieces, its variance grows only as $N$ , and $C_V$ is smooth and finite. At a continuous transition the correlation length diverges: distant regions become correlated, energy fluctuations turn long-ranged and super-extensive, the $1/\sqrt N$ scaling breaks down, and $\langle(\Delta E)^2\rangle$ — hence $C_V$ — diverges as $\lvert T-T_c\rvert^{-\alpha}$ . Watching $C_V(T)$ is watching the fluctuation amplitude, which is why the specific heat is among the first observables computed when hunting a transition — experimentally as a $\lambda$ -shaped anomaly, in simulation as a variance peak whose finite-size scaling gives $\alpha$ .

Is the microcanonical heat capacity the same as the canonical one, given they compute it so differently?

In the thermodynamic limit, yes — this is ensemble equivalence, and the fluctuation relation is exactly why. Canonically, energy fluctuates and $C_V=\langle(\Delta E)^2\rangle/k_BT^2$ ; microcanonically, energy is fixed and $C_V$ comes instead from $1/C_V\propto\partial^2 S/\partial E^2$ . Genuinely different objects, yet for a large short-range system they give the same $C_V$ , because the canonical energy distribution is a sharp Gaussian of relative width $1/\sqrt N$ centered on the microcanonical energy. The two disagree only where fluctuations are not small — near critical points or for long-range interactions — exactly where the ensembles become inequivalent.