Equivalence of Ensembles

There are three ways to describe a system in equilibrium, and they look completely different. The microcanonical ensemble fixes the energy exactly and counts microstates on an energy shell. The canonical ensemble lets energy flow to a heat bath and weights each microstate by $e^{-\beta E}$ . The grand canonical ensemble opens the walls to particles too and weights by $e^{-\beta(E-\mu N)}$ . One holds $E$ rigid; another lets it fluctuate; the third lets both $E$ and $N$ fluctuate. Yet — and this is one of the quiet miracles of statistical mechanics — for a macroscopic system they all predict the same thermodynamics: the same equation of state, the same entropy, the same heat capacity. Why should fixing the energy and merely fixing its average give identical answers? The resolution is the sharpness of large systems. In the canonical ensemble the energy is not truly free to roam: its probability distribution is a razor-thin spike around the mean, with a relative width that shrinks as $1/\sqrt{N}$ . To exponential accuracy the canonical system is at its mean energy, which is exactly the microcanonical energy. The integral $Z=\int dE\,\Omega(E)e^{-\beta E}$ is dominated by a single sharp maximum — a saddle point — and evaluating it recovers the microcanonical entropy. Mathematically the connection is a Legendre transform, the same operation that links internal energy to free energy in thermodynamics. The canonical free energy is the Legendre transform of the microcanonical entropy, with the saddle-point (Laplace) approximation providing the bridge. Ensemble equivalence is therefore not a coincidence but a theorem — one that holds whenever the entropy is a strictly concave function of energy, and that fails, instructively, when it is not.

Key takeaways

The microcanonical, canonical, and grand canonical ensembles fix $E$ / $\langle E\rangle$ / $\langle E\rangle$ and $\langle N\rangle$ respectively, yet give identical thermodynamics for macroscopic systems.
The reason: the canonical energy distribution is a sharp Gaussian of relative width $\sigma_E/\langle E\rangle\sim1/\sqrt{N}$ , so the system sits essentially at its mean energy — the microcanonical value.
Formally $Z=\int dE\,\Omega(E)e^{-\beta E}$ is evaluated by a saddle point at $E^*$ where $\partial S/\partial E=1/T$ ; this recovers $-\beta F = S(E^*)/k_B - \beta E^*$ .
That relation is a Legendre transform: the canonical free energy is the Legendre transform of the microcanonical entropy.
Equivalence holds when $S(E)$ is strictly concave; it can break near first-order transitions or for long-range forces.
Key idea: fixing the average energy is as good as fixing the energy itself, because for large $N$ the energy fluctuations are negligibly small and the partition-function integral is dominated by one sharp saddle at the microcanonical energy.

The core insight

Different ensembles constrain different variables — exact energy versus mean energy versus mean energy and particle number — but they agree on all thermodynamics for macroscopic systems, because the fluctuating quantities have relative widths $\sim1/\sqrt{N}$ and the partition function is dominated by a single saddle point at the microcanonical energy. The canonical free energy is the Legendre transform of the microcanonical entropy.

Where this lives today

Ensemble equivalence is the daily working assumption behind almost every equilibrium simulation: molecular dynamics naturally samples (nearly) microcanonical or canonical distributions, Monte Carlo samples canonical or grand canonical, and results are compared freely because equivalence guarantees the bulk thermodynamics matches. The systematic study of inequivalence has become a field of its own — long-range-interacting systems (self-gravitating stars, two-dimensional turbulence, plasmas, some spin models) show microcanonical negative heat capacity and ensemble-dependent phase diagrams, which is why astrophysics and plasma physics take the ensemble choice seriously. Large-deviation theory now provides the rigorous language: the microcanonical entropy is a rate function, the canonical free energy its Legendre transform, and equivalence is exactly the condition that this transform be invertible.

The physical picture

Imagine three laboratories studying the same block of copper. Lab A seals it in a perfect thermos and measures its fixed energy (microcanonical). Lab B clamps it to a temperature-controlled bath and lets heat flow (canonical). Lab C additionally connects it to a particle reservoir (grand canonical). Each lab writes down a different probability rule for the microstates. But when they compare the pressure, the entropy per atom, the heat capacity — every extensive quantity per particle, every intensive quantity — the three notebooks agree to many decimals. The differences live only in the fluctuations, which are invisible at macroscopic scale. The mechanism is concentration. In Lab B the energy is allowed to fluctuate, but the number of microstates $\Omega(E)$ rises so steeply with $E$ while $e^{-\beta E}$ falls so steeply that their product is a needle-sharp peak. The width of that peak, relative to its position, is $\sim1/\sqrt{N}$ . For $N\sim10^{23}$ that is $\sim10^{-12}$ : the energy is fixed to a part in a trillion even though nothing forces it to be.

Figure

Ensemble equivalence by concentration: the rising density of states Ω(E) times the falling Boltzmann factor e^{-βE} gives a razor-sharp product peaked at the saddle E*, whose relative width shrinks as 1/√N.

The second figure is the punchline: the relative energy width plotted against system size. It slides down a $1/\sqrt{N}$ line, so that by laboratory scales the three ensembles are indistinguishable — the width has fallen below any measurable resolution.

Figure

Punchline: the relative energy fluctuation σ_E/⟨E⟩ falls as 1/√N (straight line on log axes), reaching ~10⁻¹² at N = N_A — this vanishing width is why the three ensembles collapse onto one thermodynamics.

The core equations

The bridge between the microcanonical entropy $S(E)$ and the canonical partition function is a single integral over energy, weighting each shell of $\Omega(E)$ microstates by its Boltzmann factor. Writing the density of states as $\Omega(E)=e^{S(E)/k_B}$ turns the integrand into a competition between two exponentials.

Z(\beta)=\int dE\,\Omega(E)\,e^{-\beta E}=\int dE\,\exp\!\Big[\frac{S(E)}{k_B}-\beta E\Big] \qquad (1)

where: $S(E)=k_B\ln\Omega(E)$ — the microcanonical entropy; $\beta=1/k_BT$ . The exponent is extensive ( $\propto N$ ), so the integral is completely dominated by the energy $E^*$ that maximizes it — the saddle point. Maximizing the exponent sets its derivative to zero, which is precisely the thermodynamic definition of temperature. The saddle energy is where the microcanonical temperature equals the bath temperature.

\frac{\partial}{\partial E}\Big[\frac{S(E)}{k_B}-\beta E\Big]_{E=E^*}=0 \;\Longrightarrow\; \frac{\partial S}{\partial E}\bigg|_{E^*}=\frac{1}{T} \qquad (2)

where: $E^*$ — the most probable (and, to $O(1/N)$ , the mean) energy. Reading it physically: the canonical system settles at the energy whose microcanonical temperature matches the reservoir. Evaluating the integral at the saddle gives the free energy as a Legendre transform of the entropy. The Gaussian fluctuations around the saddle contribute only a $\ln N$ correction, negligible beside the extensive leading term.

-\beta F(\beta)=\frac{S(E^*)}{k_B}-\beta E^* + \tfrac{1}{2}\ln\!\Big(\frac{2\pi}{|S''|/k_B}\Big)+\dots \;\approx\; \max_E\Big[\frac{S(E)}{k_B}-\beta E\Big] \qquad (3)

where: the leading term is the Legendre transform; the $\tfrac12\ln$ term is the saddle-width correction, $O(\ln N)\ll N$ . This is the exact statement of equivalence: canonical free energy = Legendre transform of microcanonical entropy. The width of the saddle is set by the second derivative of the entropy, which is why concavity matters and why the relative fluctuation shrinks with size.

\frac{\sigma_E}{\langle E\rangle}=\frac{\sqrt{k_BT^2C_V}}{U}\sim\frac{1}{\sqrt{N}} \qquad (4)

where: $\sigma_E=\sqrt{\langle\Delta E^2\rangle}$ ; $C_V$ and $U$ are both extensive, so their combination scales as $N/\!\sqrt{N}\!\cdot\!1/N = 1/\sqrt{N}$ . Vanishing relative width is the engine of equivalence.

Derivation AThe canonical partition function as a saddle-point integral

The goal is to show that the canonical partition function, an integral over all energies, is dominated by a single sharp peak, and that evaluating it there reproduces the microcanonical entropy — proving the two ensembles agree to leading order in $N$ . The strategy is Laplace's method: when an integrand is the exponential of an extensive function, the integral equals its maximum value times a small Gaussian correction, and the maximum condition is exactly the equality of temperatures.

Step 1. Write $\Omega(E)=e^{S(E)/k_B}$ so that $Z=\int dE\,e^{\phi(E)}$ with $\phi(E)=S(E)/k_B-\beta E$ . Since $S\propto N$ and $E\propto N$ , $\phi$ is extensive.

Step 2. Expand $\phi$ about its maximum $E^*$ : $\phi(E)=\phi(E^*)+\tfrac12\phi''(E^*)(E-E^*)^2+\dots$ , with $\phi'(E^*)=0$ giving $\partial S/\partial E|_{E^*}=1/T$ (equation 2) and $\phi''(E^*)=S''(E^*)/k_B<0$ for a genuine maximum (concave $S$ ).

Step 3. The Gaussian integral gives $Z\approx e^{\phi(E^*)}\sqrt{2\pi/|\phi''(E^*)|}$ . Taking the logarithm, $\ln Z = \phi(E^*)+\tfrac12\ln(2\pi/|\phi''|)$ , i.e. $-\beta F=S(E^*)/k_B-\beta E^*+O(\ln N)$ — equation (3). Why the correction is negligible. The leading term $\phi(E^*)$ is $O(N)$ ; the Gaussian correction is $\tfrac12\ln(\text{something}\propto N)=O(\ln N)$ . Their ratio $\to0$ , so the free energy per particle is identical in the two ensembles.

Step 4. Test on the ideal gas, $\Omega(E)\propto E^{3N/2}$ so $S/k_B=\tfrac{3N}{2}\ln E$ . The saddle condition $\tfrac{3N}{2E^*}=\beta$ gives $E^*=\tfrac{3}{2}N/\beta=\tfrac32 Nk_BT$ — the equipartition energy, and exactly the canonical mean $\langle E\rangle=-\partial\ln Z/\partial\beta$ . Verified: symbolic saddle of $\tfrac{3N}{2}\ln E-\beta E$ gives $E^*=\tfrac32 N/\beta$ , equal to $-\partial_\beta\ln Z$ for $Z\propto\beta^{-3N/2}$ ; the correction/main ratio $\to0$ as $\ln N/N$ (numeric $N=10^6,10^{12},10^{23}$ ) — verify_aq-sm-equivalence-of-ensembles.py.

Go deeper · Derivation B · 1 minEquivalence as a Legendre transform of the entropy

Derivation BEquivalence as a Legendre transform of the entropy

The goal is to recast the saddle-point result as a clean thermodynamic statement: the canonical free energy is the Legendre transform of the microcanonical entropy, the same operation that generates all thermodynamic potentials. This reframing shows that ensemble equivalence is the statistical-mechanical shadow of the Legendre duality between energy and temperature representations, and it makes the concavity requirement transparent.

Step 1. From equation (3), $-\beta F(\beta)=\max_E[S(E)/k_B-\beta E]$ . The maximization over $E$ at fixed $\beta$ is exactly the Legendre transform of the concave function $s(E)\equiv S(E)/k_B$ with conjugate variable $\beta$ .

Step 2. The transform pairs the extensive energy $E$ with the intensive inverse temperature $\beta$ . Its stationarity condition $s'(E^*)=\beta$ is equation (2); the transformed function is $-\beta F$ .

Step 3. Because a Legendre transform of a strictly concave function is single-valued and invertible, the map $E^*\leftrightarrow\beta$ is one-to-one, and the inverse transform recovers $s(E)$ exactly. No thermodynamic information is lost passing between ensembles. Why concavity is the hinge. Legendre duality is an involution ( $f^{**}=f$ ) only on convex/concave functions. If $S(E)$ has a non-concave dip, the forward transform loses that region (it returns the concave hull), and the inverse cannot rebuild it — this is precisely where ensembles become inequivalent.

Step 4. Consistency: differentiating $-\beta F$ with respect to $\beta$ returns $-E^*=-\langle E\rangle$ , the conjugate variable, confirming the transform is set up correctly and that $\langle E\rangle_{\text{canonical}}=E^*_{\text{microcanonical}}$ . Verified: the ideal-gas $s(E)=\tfrac{3N}{2}\ln E$ has $s'=\tfrac{3N}{2E}>0$ and $s''=-\tfrac{3N}{2E^2}<0$ (strictly concave); the Legendre value equals $\tfrac{3N}{2}\ln E^*-\tfrac{3N}{2}$ and $\partial(-\beta F)/\partial\beta=-E^*$ — verify_aq-sm-equivalence-of-ensembles.py.

Limiting cases

Equivalence is generic but not universal; the same saddle-point analysis tells you exactly when it holds and when it fails.

Case	Result	Meaning
Large $N$ , short-range forces	ensembles agree to $O(\ln N/N)$	The standard case; relative fluctuations $\sim1/\sqrt{N}\to0$ .
Ideal gas, any $N$	$E^*=\langle E\rangle=\tfrac32 Nk_BT$ exactly	Saddle sits at the equipartition energy; $S(E)$ strictly concave.
Finite $N$	$O(\ln N)$ corrections	Small-system thermodynamics differs slightly between ensembles.
First-order transition	canonical $E$ jumps; MC interpolates	$S(E)$ has a linear (non-strictly-concave) segment; ensembles differ in the coexistence gap.
Long-range forces (gravity)	inequivalent	$S(E)$ non-concave; microcanonical negative $C_V$ has no canonical counterpart.

Common trapCommon pitfall

Ensemble equivalence is about thermodynamics (free energies, equations of state, entropy per particle), not about every observable. Fluctuations themselves are ensemble-dependent by construction: the microcanonical energy variance is zero, the canonical one is $k_BT^2C_V$ , and the grand canonical adds particle-number fluctuations. It is wrong to say "the ensembles are identical" — they are identical in their thermodynamic limits, i.e. in the intensive quantities and the extensive quantities per particle. Choosing the ensemble that fixes the quantity you actually control (energy for an isolated system, temperature for one in a bath) is still physically correct and often far more convenient; equivalence guarantees the bulk answers won't depend on that choice.

Worked examples

Example 1 — The saddle energy of an ideal gas (enumerate, apply, boundary number). Take $N$ argon atoms with density of states $\Omega(E)\propto E^{3N/2}$ in contact with a bath at temperature $T$ .

Step 1. The exponent to maximize is $\phi(E)=\tfrac{3N}{2}\ln E-\beta E$ . Set $\phi'(E)=\tfrac{3N}{2E}-\beta=0$ .

Step 2. Solve: $E^*=\tfrac{3N}{2\beta}=\tfrac32 Nk_BT$ . This is the microcanonical energy at temperature $T$ .

Step 3. Compare to canonical: with $Z\propto\beta^{-3N/2}$ , $\langle E\rangle=-\partial_\beta\ln Z=\tfrac{3N}{2\beta}=\tfrac32 Nk_BT$ — identical to $E^*$ . The two ensembles agree on the internal energy exactly.

Step 4. Boundary number: the saddle width is $\sigma_E=\sqrt{-1/\phi''}=\sqrt{2E^{*2}/(3N)}=E^*\sqrt{2/(3N)}$ . For $N=N_A=6.02\times10^{23}$ , $\sigma_E/E^*=\sqrt{2/(3\times6.02\times10^{23})}=\mathbf{1.05\times10^{-12}}$ — the energy is pinned to a part in $10^{12}$ , so fixing $\langle E\rangle$ is as good as fixing $E$ .

Example 2 — When does the small-system correction matter? Estimate the fractional difference between the canonical and microcanonical free energy per particle for a nanocluster of $N=100$ atoms.

Step 1. The relative width of the canonical energy is $\sigma_E/\langle E\rangle=\sqrt{2/(3N)}=\sqrt{2/300}=\mathbf{0.082}$ , i.e. about $8\%$ .

Step 2. The free-energy-per-particle discrepancy is set by the $O(\ln N)$ saddle correction relative to the extensive term, $\sim\ln N/N=\ln100/100=\mathbf{0.046}$ — a few percent.

Step 3. Interpret: for a 100-atom cluster the ensemble choice shifts intensive quantities at the percent level, so simulations of small systems must report which ensemble they used. For $N=10^{23}$ the same estimate is $\ln(10^{23})/10^{23}\approx5\times10^{-22}$ — utterly negligible, and equivalence is exact for all practical purposes.

Origins · 1 minequivalence of ensembles and how this was discovered

equivalence of ensembles

statistical-mechanics concept · Gibbs 1902 · Ruelle/Lanford rigor 1960s–70s

The equivalence of ensembles is the principle that the microcanonical, canonical, and grand canonical descriptions yield the same thermodynamics in the thermodynamic limit. Gibbs asserted it heuristically in his 1902 Elementary Principles in Statistical Mechanics; making it a theorem took the rigorous statistical mechanics of the 1960s–70s (Ruelle, Lanford, Fisher), who proved that for systems with short-range interactions the free-energy densities of the three ensembles coincide and are convex functions of their intensive parameters. The modern, sharp form is a statement about Legendre-transform duality and large-deviation rate functions: equivalence holds exactly when the microcanonical entropy density is concave, and its failure — for long-range or first-order-transition systems — is now a well-mapped part of the theory.

History

1902

— Gibbs's Elementary Principles introduces all three ensembles and argues they agree for large systems.

1937

— Darwin and Fowler develop the steepest-descent (saddle-point) method for partition functions.

1949

— van Hove proves the existence of the thermodynamic limit for a class of short-range interactions.

1969

— Ruelle's Statistical Mechanics: Rigorous Results establishes ensemble equivalence rigorously.

1973

— Lanford's large-deviation formulation recasts equivalence as a property of entropy concavity.

Going further

References. Pathria & Beale, Statistical Mechanics, 4th ed., Ch. 3–4 (the ensembles and their equivalence) · Reif, Fundamentals of Statistical and Thermal Physics, §6.1–6.10 · Kardar, Statistical Physics of Particles, Ch. 4 · Ruelle, Statistical Mechanics: Rigorous Results (1969) · Touchette, "The large deviation approach to statistical mechanics," Phys. Rep. 478 (2009).

Provenance. The saddle-point relation $-\beta F=\max_E[S(E)/k_B-\beta E]$ and the $1/\sqrt{N}$ fluctuation argument are standard (Pathria Ch. 3–4, Reif §6.5). A verifier should confirm: (i) the saddle of $\tfrac{3N}{2}\ln E-\beta E$ is $E^*=\tfrac32 N/\beta$ , equal to $-\partial_\beta\ln Z$ for $Z\propto\beta^{-3N/2}$ ; (ii) $\partial^2\phi/\partial E^2<0$ at the saddle (genuine maximum), width $\sigma_E/E^*=\sqrt{2/3N}$ ; (iii) the ratio of the $\tfrac12\ln(2\pi\sigma^2)$ correction to the extensive term $\to0$ as $\ln N/N$ ; (iv) $s(E)=\tfrac{3N}{2}\ln E$ is strictly concave ( $s''<0$ ). All in verify_aq-sm-equivalence-of-ensembles.py.

Builds on: Canonical Ensemble · Derivation of the Canonical Distribution from the Microcanonical Ensemble · Grand Canonical Ensemble · Canonical Partition Function · Energy Fluctuations in the Canonical Ensemble Leads to: Helmholtz Free Energy from the Partition Function · Thermodynamic Quantities from the Partition Function · Heat Capacity and Fluctuation-Response Relation · Grand Partition Function

If the ensembles are equivalent, why do we need three of them?

Because they are matched to different physical situations and different calculations. The microcanonical ensemble is the fundamental one (equal a priori probabilities on an energy shell) but is awkward to compute with. The canonical ensemble, fixing temperature, is by far the easiest — the Boltzmann factor factorizes over independent degrees of freedom. The grand canonical ensemble, fixing chemical potential, is essential for quantum gases and for any open or reacting system. Equivalence is what lets you pick the convenient ensemble and trust that the bulk thermodynamics is the same as the correct one.

Does equivalence mean the energy in the canonical ensemble is really constant?

No — the energy genuinely fluctuates in the canonical ensemble, with variance $k_BT^2C_V$ . What is true is that the relative fluctuation $\sigma_E/\langle E\rangle\sim1/\sqrt{N}$ is astronomically small for macroscopic $N$ . So the energy is not fixed, but it is pinned so tightly around its mean that the system behaves, thermodynamically, as if it were at the microcanonical energy. The fluctuations are real and physically important (they carry the heat capacity), just negligible on the scale of $\langle E\rangle$ .

What is the saddle point, physically?

It is the single most probable energy $E^*$ at which the product $\Omega(E)e^{-\beta E}$ is maximal. The rising density of states favors high energy; the Boltzmann factor favors low energy; their competition selects an intermediate $E^*$ where $\partial S/\partial E=1/T$ . Because the exponent is extensive, this maximum is overwhelmingly sharp — the integral is dominated by an infinitesimal neighborhood of $E^*$ . "Saddle point" is the name from complex analysis (steepest descent); here it is simply the peak of a sharply-peaked integrand.

How is this connected to Legendre transforms in thermodynamics?

Directly. In thermodynamics the Helmholtz free energy is the Legendre transform of the internal energy with respect to entropy, $F(T)=\min_S[U-TS]$ . In statistical mechanics the same duality appears as $-\beta F=\max_E[S(E)/k_B-\beta E]$ : the canonical free energy is the Legendre transform of the microcanonical entropy. Ensemble equivalence is the statement that this transform is invertible — that you can go from $S(E)$ to $F(T)$ and back without losing information. Both statements require the relevant function to be convex/concave.

Why does concavity of the entropy matter so much?

Because a Legendre transform is a faithful, invertible operation only on a convex (or concave) function. If $S(E)$ is strictly concave, the map between energy and temperature is one-to-one and the ensembles are equivalent. If $S(E)$ has a non-concave "dip" — as it does across a first-order phase transition or for long-range forces — the forward Legendre transform silently replaces the dip by its straight-line concave hull, losing the information in that energy window. The canonical ensemble then cannot see the coexistence region that the microcanonical ensemble resolves, and the two disagree.

Do quantum systems obey ensemble equivalence too?

Yes, with the same logic. For quantum gases the grand canonical ensemble is the natural one, and it is equivalent to the canonical and microcanonical descriptions in the thermodynamic limit for short-range systems. The one famous subtlety is Bose–Einstein condensation: the fluctuations of the ground-state occupation are anomalously large in the grand canonical ensemble, so that particular observable is ensemble-dependent even though the thermodynamics (energy, pressure, condensate fraction) is not. It is a reminder that equivalence is a statement about bulk thermodynamics, not about every fluctuation.