The Thermodynamic Limit

Boil a kettle and the water changes phase at a sharp temperature — a kink in the graph of energy against heat, a jump you can point to. Yet the partition function of any finite collection of molecules is a finite sum of smooth exponentials, and a finite sum of smooth functions is itself perfectly smooth: it has no kinks, no jumps, no phase transition anywhere. The paradox is real, and its resolution is a limit. Sharp thermodynamics — extensive entropy, size-independent temperature, genuine phase transitions — is not a property of any finite system. It is what emerges when the particle number $N$ and the volume $V$ grow without bound together, holding the density $N/V$ fixed. This is the thermodynamic limit, and it is the precise procedure that lets microscopic mechanics become macroscopic thermodynamics. The limit does two things at once. First, it sorts every quantity into extensive (grows in proportion to size: energy, entropy, free energy) or intensive (converges to a size-independent value: temperature, pressure, density). That dichotomy, which the whole edifice of thermodynamics takes for granted, is only meaningful because the limit exists. Second, it makes fluctuations vanish: relative fluctuations of extensive quantities fall as $1/\sqrt{N}$ , so a macroscopic observable becomes effectively deterministic — a theory with no fluctuations emerges from one built entirely on them. But the limit's existence is a theorem, not a definition, and it can fail. Interactions that are too attractive at short range let matter collapse ( $E\sim -N^2$ , entropy non-extensive); interactions that are too long-ranged — unscreened gravity is the archetype — make energy super-extensive and break ensemble equivalence, allowing negative specific heats. The mathematical conditions that guarantee a well-behaved limit are stability (energy bounded below by $-BN$ ) and temperedness (the potential decays faster than $r^{-d}$ ). Under them, Van Hove, Ruelle, Fisher, and Lebowitz–Lieb proved that the free energy per particle converges to a definite convex function, independent of the shape of the growing box and of its boundary conditions. Every time we write $S = k_B\ln\Omega$ and call it extensive, or invoke the equivalence of ensembles, or speak of a sharp critical temperature, we are silently standing inside this limit.

Key takeaways

The thermodynamic limit is the joint limit $N\to\infty$ , $V\to\infty$ at fixed density $n=N/V$ (and fixed intensive parameters), in which thermodynamics emerges from statistical mechanics.
Extensive quantities ( $E$ , $S$ , $F$ ) scale linearly with size; intensive quantities ( $T$ , $P$ , $\mu$ , densities) converge to size-independent values.
Relative fluctuations of extensive quantities fall as $1/\sqrt{N}\to0$ , so macroscopic observables become sharp and the ensembles become equivalent.
A finite partition function is analytic, so sharp phase transitions and spontaneous symmetry breaking are strictly possible only in the limit.
Existence requires stable ( $E\ge -BN$ ) and tempered (potential decays faster than $r^{-d}$ ) interactions; otherwise energy is super-extensive and the limit fails.
Mathematically, $f = -k_BT\lim_{N\to\infty}N^{-1}\ln Z_N$ exists and is convex (Van Hove, Ruelle, Fisher, Lebowitz–Lieb), independent of boundary conditions and box shape.
Key idea: Take $N,V\to\infty$ with $N/V$ fixed — extensive quantities then grow linearly and intensive ones converge, so $f = -k_BT\lim_{N\to\infty}N^{-1}\ln Z_N$ exists, is convex, and is ensemble- and boundary-independent. This is where thermodynamics and phase transitions become exact.

The core insight

Take $N,V\to\infty$ with $N/V$ fixed: extensive quantities grow linearly in size and intensive ones converge, so the free energy per particle $f=-k_BT\lim_{N\to\infty}N^{-1}\ln Z_N$ exists, is convex, and is independent of ensemble and boundary conditions — this is where thermodynamics and phase transitions become exact.

Where this lives today

The thermodynamic limit is the silent assumption behind every phase diagram, but its breakdown is where the modern action is. Nanoscience deliberately builds systems in the crossover region — clusters, quantum dots, single molecules — where the surface fraction is neither negligible nor total, and where "temperature" and "phase" require careful redefinition because the sharp limit has not set in. Finite-size scaling, the workhorse of computational statistical mechanics, turns the very corrections that vanish in the limit into a tool: by watching how a quantity depends on system size $L$ , one extracts critical exponents and locates transitions from simulations of a few thousand particles. At the other extreme, long-range and gravitational systems are studied precisely because the ordinary limit fails there. Self-gravitating star clusters exhibit negative specific heat and gravothermal collapse; trapped-ion and cavity-QED experiments engineer tunable long-range interactions to realize ensemble inequivalence in the laboratory. Understanding exactly which interactions admit a thermodynamic limit — and what replaces thermodynamics when they do not — remains an active frontier, from black-hole thermodynamics to the statistical mechanics of neural networks with all-to-all coupling.

The physical picture

Tile space with identical boxes, each holding fluid at the same density and temperature. The total energy is the sum of the box energies plus whatever leaks across the interfaces between them. If the forces are short-ranged, the interfaces scale like surface area — $\sim N^{(d-1)/d}$ — while the bulk scales like volume — $\sim N$ . Their ratio, the surface fraction $\sim N^{-1/d}$ , goes to zero: for a mole of gas it is about $10^{-8}$ . Energy and entropy therefore become exactly additive, and the answer stops caring about the shape of the container or what happens at its walls. This "bulk drowns out surface" picture is the physical heart of extensivity, and the reason a chemistry table can list one entropy for argon rather than one per flask. Fluctuations tell the same story from the probabilistic side. A macroscopic observable is a sum of $\sim N$ weakly correlated microscopic contributions; its mean grows as $N$ but its standard deviation only as $\sqrt{N}$ , so the relative spread falls as $1/\sqrt{N}$ . For $N\sim10^{23}$ that is $\sim10^{-12}$ — utterly invisible. The distribution of the energy density collapses onto a single value, and the sharp, fluctuationless relations of classical thermodynamics appear.

Figure

Surface drowns in bulk — the surface fraction falls as N to the minus one-third, and relative fluctuations as one over root N

Now the phase-transition puzzle. A finite partition function $Z_N=\sum_i e^{-\beta E_i}$ is a finite sum of analytic exponentials — analytic, positive, nowhere zero for real temperature. No kink, no jump can hide in it. A sharp phase transition is a non-analyticity, so it can only appear as the limit of an infinite sequence of ever-larger analytic functions. Yang and Lee gave the mechanism a face: the zeros of $Z_N$ sit off the real axis in the complex-fugacity plane, and only as $N\to\infty$ do they migrate inward and pinch the real axis. Where they pinch, the pressure develops its kink and a phase transition is born.

Figure

Yang–Lee zeros pinch the real axis only as N grows — a finite system stays analytic

The limit can also fail spectacularly. For self-gravitating matter the potential energy scales as $N^2/R$ , not $N$ ; entropy is non-extensive, ensembles are inequivalent, and specific heats can go negative. This is not a defect of nature but a precise warning: "thermodynamics" as usually formulated presupposes stable, tempered, short-ranged interactions.

The core equations

The definition is a joint limit — both $N$ and $V$ diverge, locked together so that the density stays put. Taking one without the other gives a dilute gas or a collapse, not thermodynamics.

n=\frac{N}{V}=\text{const},\qquad N\to\infty,\;V\to\infty \qquad (1)

where: $N$ — number of particles; $V$ — volume; $n$ — number density, held fixed as the system grows. Every intensive control parameter ( $T$ , and $n$ itself) is frozen while the extensive size runs to infinity. The central object is the free energy per particle. Its existence as a finite limit is the whole game: if it exists, the extensive $F = Nf$ and all of thermodynamics follows.

f(n,T)=-k_BT\,\lim_{N\to\infty}\frac{1}{N}\ln Z_N(V,T),\quad V=N/n \qquad (2)

where: $f$ — Helmholtz free energy per particle; $Z_N$ — canonical partition function of $N$ particles; $\beta=1/k_BT$ . The limit converges because stability and temperedness make $\ln Z_N$ super-additive, and a super-additive sequence bounded above (per particle) has a limit — this is Fekete's lemma applied to statistical mechanics. Entropy is extensive in exactly the same sense: the per-particle entropy converges, and the total scales linearly.

S(E,N,V)=k_B\ln\Omega(E,N,V),\qquad s(\varepsilon,n)=\lim_{N\to\infty}\frac{S}{N} \qquad (3)

where: $\Omega$ — number (or phase-space volume) of microstates in the energy shell; $s$ — entropy per particle; $\varepsilon=E/N$ — energy per particle. The Boltzmann entropy is only called extensive because this limit exists. Fluctuations are what make the limit deliver sharp thermodynamics. The canonical energy variance is set by the heat capacity, and both grow only linearly in $N$ .

\frac{\langle\Delta E^2\rangle}{\langle E\rangle^2}=\frac{k_BT^2 C_V}{\langle E\rangle^2}\propto\frac{1}{N} \qquad (4)

where: $C_V$ — heat capacity, extensive ( $\propto N$ ); $\langle\Delta E^2\rangle$ — canonical energy variance. Because the numerator is $\propto N$ and the denominator $\propto N^2$ , the relative fluctuation falls as $1/N$ (its square root, $\sigma_E/\langle E\rangle$ , as $1/\sqrt N$ ). Different ensembles agree in the limit because the canonical free energy is the Legendre transform of the microcanonical entropy — a saddle-point statement that becomes exact as $N\to\infty$ .

-\beta f(n,T)=\max_{\varepsilon}\!\left[\frac{s(\varepsilon,n)}{k_B}-\beta\varepsilon\right] \qquad (5)

where: $s(\varepsilon,n)$ — microcanonical entropy density; $f(n,T)$ — canonical free energy density; the maximum over $\varepsilon$ is the Legendre transform. Where $s$ is concave the transform is invertible and every ensemble gives the same thermodynamics; where $s$ has a non-concave region, ensembles can become inequivalent. Yang and Lee expressed the transition as zeros of the grand partition function in the complex-fugacity plane, which reach the real axis only in the limit.

Z_N(z)=\sum_{k=0}^{N}\frac{z^k}{k!}\,Q_k=\prod_{i}\left(1-\frac{z}{z_i}\right) \qquad (6)

where: $z=e^{\beta\mu}$ — fugacity; $z_i$ — complex zeros; $Q_k$ — configurational integral for $k$ particles. For finite $N$ the zeros avoid the positive real axis (so $Z_N>0$ and everything is analytic); a transition is a zero reaching the axis as $N\to\infty$ . The existence of the limit rests on two inequalities. Stability forbids collapse; temperedness forbids super-extensive energy.

E\ge -B\,N\ \ (\text{stability}),\qquad \phi(r)\le A\,r^{-d-\epsilon}\ (r\to\infty)\ \ (\text{temperedness}) \qquad (7)

where: $\phi(r)$ — pair potential; $B,A,\epsilon>0$ ; $d$ — spatial dimension. Violate stability and matter implodes ( $E\sim -N^2$ ); violate temperedness (as gravity does) and energy grows faster than $N$ , breaking extensivity and ensemble equivalence. Finally, spontaneous symmetry breaking is encoded in the order of two limits that do not commute.

m=\lim_{h\to 0^+}\;\lim_{N\to\infty}\;\frac{1}{N}\frac{\partial \ln Z_N(h)}{\partial(\beta h)} \qquad (8)

where: $m$ — order parameter (e.g. spontaneous magnetization); $h$ — symmetry-breaking field. Take $N\to\infty$ first and the free energy develops a cusp at $h=0$ with a nonzero one-sided slope; take $h\to0$ first at finite $N$ and symmetry forces $m=0$ . The limit remembers which order you chose.

Derivation AExtensivity and the 1/√N suppression of fluctuations

The claim that thermodynamics — a theory with no fluctuations — emerges from a probabilistic microtheory rests on one estimate: relative fluctuations of extensive quantities vanish in the limit. We prove it for energy in the canonical ensemble, where the variance is a second derivative of $\ln Z$ and evaluates to the heat capacity. The key is that both the mean energy and the heat capacity are extensive, so their ratio carries a clean $1/\sqrt N$ .

Step 1. Mean energy as a derivative. From $Z=\sum_i e^{-\beta E_i}$ , the average energy is $\langle E\rangle=-\partial\ln Z/\partial\beta$ . Why: differentiating $\ln Z$ pulls down a factor $-E_i$ inside the Boltzmann-weighted average — the standard generating-function trick.

Step 2. Variance as a second derivative. Differentiating once more, $\langle\Delta E^2\rangle=\langle E^2\rangle-\langle E\rangle^2=\partial^2\ln Z/\partial\beta^2$ . Why: the second derivative of a log-generating function is always the variance of the conjugate variable — here, the energy.

Step 3. Relate the variance to the heat capacity. Using $\partial/\partial\beta=-k_BT^2\,\partial/\partial T$ ,

\langle\Delta E^2\rangle=-\frac{\partial\langle E\rangle}{\partial\beta}=k_BT^2\,C_V.

Why: this is the fluctuation–response relation — spontaneous energy fluctuations are tied to how the system responds to a temperature change, i.e. its heat capacity.

Step 4. Insert extensivity. Both energy and heat capacity are extensive: $\langle E\rangle=N\varepsilon$ and $C_V=Nc_V$ . Why: away from a critical point $\varepsilon$ and $c_V$ are intensive, finite per-particle values; the entire size dependence sits in the explicit $N$ .

Step 5. Form the relative fluctuation.

\frac{\sigma_E}{\langle E\rangle}=\frac{\sqrt{k_BT^2\,Nc_V}}{N\varepsilon}=\frac{1}{\sqrt N}\,\frac{\sqrt{k_BT^2\,c_V}}{\varepsilon}\;\xrightarrow{N\to\infty}\;0.

Why: the size dependence is purely $1/\sqrt N$ , with an intensive, finite prefactor away from criticality — so the energy density becomes sharp, the macrostate becomes effectively deterministic, and thermodynamics emerges. The one exception is a critical point, where $c_V$ diverges and the argument breaks — which is exactly why critical fluctuations are macroscopically visible (critical opalescence). *Verified: reproduces Pathria & Beale §3.6 ( $\langle\Delta E^2\rangle=k_BT^2C_V$ , relative fluctuations $\sim N^{-1/2}$ ); the $N^{-1/2}$ scaling and its breakdown at diverging $c_V$ are checked numerically in verify_aq-sm-thermodynamic-limit.py.*

Go deeper · Derivation B · 1 minConvexity of the free energy density and thermodynamic stability

Derivation BConvexity of the free energy density and thermodynamic stability

A structural payoff of the limit is that the limiting free energy is convex in its intensive variables — which is exactly the mathematical content of thermodynamic stability (positive compressibility, positive susceptibility, positive heat capacity). We show this by coupling an external field $h$ to an extensive observable $M$ and recognizing the second derivative of $\ln Z$ as a manifestly non-negative variance. Convexity then survives the limit because a pointwise limit of convex functions is convex.

Step 1. Field-dependent partition function. Couple a field $h$ to the extensive conjugate $M$ (say magnetization): $Z_N(h)=\sum_i e^{-\beta(E_i-hM_i)}$ . Why: this is the standard way to probe response — the field biases the ensemble toward larger $M$ .

Step 2. Per-particle generating function. Define $g(h)=N^{-1}\ln Z_N(h)$ . Why: dividing by $N$ gives the quantity with a finite limit; its derivatives are the intensive response functions.

Step 3. Curvature is a variance.

\frac{\partial^2 g}{\partial h^2}=\frac{\beta^2}{N}\,\langle\Delta M^2\rangle\ge 0.

Why: the second derivative of a log-partition function is the variance of the conjugate variable $M$ , which cannot be negative — so $g$ is convex in $h$ , at every finite $N$ .

Step 4. Convexity survives the limit. Since $f=-k_BT\,g$ , we get $\partial_h^2 f\le 0$ , and because a pointwise limit of convex functions is convex, the limiting $f(h)$ inherits the property. Why: no smoothing is needed — convexity is preserved under pointwise limits, so the infinite-system free energy is convex even though it may now have a cusp.

Step 5. Convexity is stability. The susceptibility $\chi=\partial_h m=-\partial_h^2 f/(k_BT)\ge0$ . Why: a stable phase must respond to a field by aligning with it (positive susceptibility) and must resist compression (positive compressibility); both are the statement that $f$ is convex. The limit does not merely allow stable thermodynamics — it enforces it. *Verified: matches the convexity theorems for free energies in Ruelle, Statistical Mechanics: Rigorous Results, and the stability inequalities ( $C_V\ge0$ , $\kappa_T\ge0$ ) in Landau–Lifshitz; the variance-equals-curvature identity is checked symbolically and the 1D-Ising free energy's convexity numerically in verify_aq-sm-thermodynamic-limit.py.*

Limiting cases

How quantities behave as the size $\lambda$ grows (number $N=\lambda N_0$ , volume $V=\lambda V_0$ at fixed density):

Case	Result	Meaning
Extensive quantities	$X(\lambda N,\lambda V)=\lambda\,X(N,V)$	Energy $E$ , entropy $S$ , free energies $F,\Omega$ , particle number scale linearly with size.
Intensive quantities	$y(\lambda N,\lambda V)=y(N,V)$	Temperature $T$ , pressure $P$ , chemical potential $\mu$ , density $n$ are size-independent.
Relative fluctuations	$\sigma_E/\langle E\rangle\sim N^{-1/2}\to 0$	Macroscopic observables become sharp; ensembles become equivalent.
Surface vs. bulk	$E_\text{surf}/E_\text{bulk}\sim N^{-1/d}\to 0$	Boundary corrections are subextensive for short-ranged forces; results independent of boundary conditions.
Long-range / gravity	$E_\text{grav}\sim -N^2/R\sim -N^{5/3}$	Non-tempered interactions make energy super-extensive; the limit and ensemble equivalence break down (negative specific heat possible).

Common trapThe thermodynamic limit is just "let N get large"

Two traps. First, it is the joint limit $N\to\infty$ AND $V\to\infty$ at fixed density $n=N/V$ — not $V\to\infty$ at fixed $N$ (which gives an infinitely dilute gas, density $\to0$ ) nor $N\to\infty$ at fixed $V$ (which drives density $\to\infty$ and collapse). Second, the order of limits matters: to obtain spontaneous symmetry breaking you must send $N\to\infty$ first and only then take the symmetry-breaking field $h\to0^+$ . Reversing the order restores the unbroken, symmetric state ( $m=0$ ) because any finite system has zero order parameter at $h=0$ by symmetry. The two orders give genuinely different answers below $T_c$ — that non-commutativity is symmetry breaking.

Worked examples

1 · How fast does the surface stop mattering? (a real convergence)

A cubic sample of gas contains $N$ particles arranged as an $L\times L\times L$ block, $N=L^3$ . Estimate the surface-to-bulk ratio for $N=8$ , $N=1000$ , and $N=6\times10^{23}$ , and read off how quickly extensivity sets in.

Step 1. Count bulk and surface. Bulk scales as the volume, $\sim N=L^3$ ; surface scales as the area, $\sim 6L^2 = 6N^{2/3}$ . The surface fraction is $\sim 6N^{2/3}/N = 6N^{-1/3}$ .

Step 2. Small block, $N=8$ ( $L=2$ ). $6N^{-1/3}=6/2 = 3.0$ — the "surface" correction is as big as the bulk; a handful of particles is all surface, and extensivity is meaningless.

Step 3. Modest block, $N=1000$ ( $L=10$ ). $6N^{-1/3}=6/10 = 0.60$ — still a 60% correction; a nanocluster is not yet a thermodynamic system.

Step 4. Macroscopic, $N=6\times10^{23}$ . $N^{-1/3}=(6\times10^{23})^{-1/3}=1.18\times10^{-8}$ , so the surface fraction is $\approx 7.1\times10^{-8}$ . The boundary contributes about one part in $14$ million — utterly negligible, and this is why a mole of argon has a single, container-independent entropy. The lesson is quantitative: extensivity is not switched on at some threshold but approached as $N^{-1/3}$ , painfully slowly for small systems and overwhelmingly for macroscopic ones. Nanoscience lives precisely in the crossover where that correction is neither 1 nor 0.

2 · Why a finite Ising magnet cannot truly break symmetry

Consider $N$ Ising spins in a field $h$ at temperature $T$ . Show that the magnetization per spin obeys $m_N(0)=0$ for every finite $N$ , yet a nonzero spontaneous magnetization appears in the limit below $T_c$ . Illustrate with the mean-field free energy.

Step 1. Finite- $N$ symmetry. At $h=0$ the energy is invariant under flipping every spin ( $s_i\to-s_i$ ), so states $+M$ and $-M$ have equal weight and $\langle M\rangle=0$ exactly. Hence $m_N(0)=\langle M\rangle/N=0$ for every finite $N$ .

Step 2. Mean-field free energy. In the limit the free energy per spin is $f(m)=\tfrac12 a\,m^2 + \tfrac14 b\,m^4 - hm$ , with $a\propto(T-T_c)$ and $b>0$ . Minimize over $m$ .

Step 3. Above $T_c$ ( $a>0$ ). The only minimum of $f(m)$ at $h=0$ is $m=0$ : a single symmetric well. No spontaneous magnetization.

Step 4. Below $T_c$ ( $a<0$ ). The potential becomes a double well with minima at $m=\pm\sqrt{-a/b}$ . Taking $N\to\infty$ first, then $h\to0^+$ , selects the $+$ well: $\boxed{m_0=\sqrt{-a/b}>0}$ .

Step 5. The order of limits. Reverse it — $h\to0$ first at finite $N$ — and Step 1 forces $m=0$ . The two answers, $\sqrt{-a/b}$ and $0$ , are both correct for their order; the physical, symmetry-broken magnet is the $N\to\infty$ -first answer. A finite magnet only appears to have a fixed direction because the tunnelling time between the two wells is astronomically long ( $\sim e^{+cN}$ ), not because $\langle M\rangle\ne0$ .

Origins · 1 minthermodynamic limit and how this was discovered

thermodynamic limit

foundational program, 1949–1969 · Utrecht, Geneva, Princeton

The existence of the thermodynamic limit was turned from a physicist's assumption into a mathematical theorem over two decades. Léon Van Hove (1949) first proved, for a classical gas with a hard core and short-range attraction, that the free energy per particle converges to a well-defined function independent of the container shape. Chen-Ning Yang and Tsung-Dao Lee (1952) showed how phase transitions arise from the zeros of the grand partition function reaching the real axis only in the limit. David Ruelle and Michael Fisher (1960s) then established the general conditions — stability and temperedness — under which the limit exists and the free energy is convex and boundary-independent, and Joel Lebowitz and Elliott Lieb extended the results to charged (Coulomb) matter once neutrality and quantum mechanics are included. Their collective achievement is the rigorous foundation on which every casual use of "extensive," "phase transition," and "equivalence of ensembles" quietly rests.

History

1902

— Gibbs frames statistical mechanics in terms of ensembles, implicitly requiring a large-system limit for thermodynamics to emerge.

1949

— Van Hove proves the free energy per particle converges for a classical gas with hard core and short-range forces.

1952

— Yang and Lee locate phase transitions in the zeros of the grand partition function, which reach the real axis only as N grows.

1963

— Fisher and Ruelle establish general stability and temperedness conditions guaranteeing the limit and a convex free energy.

1969

— Lebowitz and Lieb prove the thermodynamic limit exists for real (Coulomb) matter, given charge neutrality and quantum mechanics.

Going further

Texts. Pathria & Beale, Statistical Mechanics, 4th ed., §1.1, §3.1 and the thermodynamic-limit appendix · Kardar, Statistical Physics of Particles, Ch. 4 & 7 · Reif, Fundamentals of Statistical and Thermal Physics, Ch. 7 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 6 · Ruelle, Statistical Mechanics: Rigorous Results (1969), Ch. 2–3 · Yang & Lee, Phys. Rev. 87, 404 & 410 (1952).

Provenance. Definitions and scaling from Pathria & Beale §1.1, §3.1; the fluctuation derivation reproduces $\langle\Delta E^2\rangle=k_BT^2C_V$ and the $N^{-1/2}$ scaling (Pathria Ch. 3). Existence theorems (stability, temperedness, subadditivity) follow Ruelle Ch. 2–3 and Fisher's lemma; Yang–Lee zero structure from Yang & Lee, Phys. Rev. 87 (1952); the Lebowitz–Lieb Coulomb result from J. Math. Phys. 9 (1968)/Adv. Math. (1972). Convexity and stability inequalities cross-checked against Landau–Lifshitz. A verifier should recompute: $\langle\Delta E^2\rangle=\partial^2\ln Z/\partial\beta^2=k_BT^2C_V$ and the $N^{-1/2}$ scaling; that $\partial_h^2\ln Z=\beta^2\langle\Delta M^2\rangle\ge0$ implies convexity; the surface fraction $6N^{-1/3}$ for cubic samples ( $3.0$ , $0.60$ , $7.1\times10^{-8}$ at $N=8,10^3,6\times10^{23}$ ); and the mean-field double-well minima $m=\pm\sqrt{-a/b}$ . No fabricated numerical constants are used.

Prerequisites. Microstates and Macrostates · Statistical Ensembles · Accessible Phase Space and Energy Shells · Boltzmann Entropy. Leads to. Gibbs Entropy · Phase transitions and critical phenomena · Equivalence of ensembles.

If real systems are finite, why do we care about an idealized N → ∞ limit at all?

Because the limit is not a description of any one system but of the regularities that finite systems approach — and approach fast. For $N\sim10^{23}$ the surface correction is $\sim10^{-8}$ and the relative fluctuation $\sim10^{-12}$ , so a macroscopic sample is indistinguishable from the ideal limit to any measurement. The limit is what makes "the entropy of argon" a single number rather than a container-dependent quantity, and what makes a sharp boiling point exist rather than a smeared crossover. It is the mathematical statement of why thermodynamics — a fluctuation-free, size-independent theory — is even possible.

How can a phase transition be "impossible" for finite N when I can clearly boil a small pot of water?

A small pot still contains $\sim10^{25}$ molecules, so it is already deep in the effective thermodynamic limit — the transition looks perfectly sharp. The statement "no transition at finite $N$ " is about mathematical non-analyticity: for genuinely small $N$ (say a hundred molecules) the boiling would be a smooth, rounded crossover spread over a range of temperatures, not a knife-edge. As $N$ grows the rounding narrows as a power of $1/N$ until, only in the limit, it becomes a true singularity. Your pot is sharp because $10^{25}$ is, for all practical purposes, infinity.

Why does the free energy per particle converge — what stops it from drifting as N grows?

Sub-additivity plus a lower bound. Stability ( $E\ge -BN$ ) guarantees $\ln Z_N/N$ is bounded below, and temperedness (fast-decaying interactions) makes the interface energy between two joined subsystems grow slower than the volume, so $\ln Z_{N_1+N_2}\ge\ln Z_{N_1}+\ln Z_{N_2}-o(N)$ . A super-additive sequence bounded per particle has a limit (Fekete's lemma). Physically: because surface drowns in bulk, doubling the system almost exactly doubles $\ln Z$ , so the per-particle value settles down instead of drifting.

What exactly goes wrong for gravity, and why is negative specific heat allowed there?

Gravity is not tempered: the potential $\sim -1/r$ decays too slowly, so the total potential energy scales as $N^2/R\sim N^{5/3}$ rather than $N$ . Energy is then super-extensive, the free energy per particle does not converge, and the microcanonical entropy $s(\varepsilon)$ develops a non-concave region. Where $s$ is non-concave the microcanonical and canonical ensembles genuinely disagree, and a self-gravitating system at fixed energy can have $\mathrm{d}T/\mathrm{d}E<0$ — add energy and it cools. A star cluster does exactly this (it heats up as it loses energy to escaping stars), which is why stellar dynamics needs its own statistical mechanics.

Why must N → ∞ come before h → 0 for symmetry breaking, and not the other way?

At any finite $N$ the two symmetry-related states ( $+m$ and $-m$ ) are connected by a finite free-energy barrier, so the system tunnels between them and $\langle m\rangle=0$ at $h=0$ by symmetry. Taking $N\to\infty$ first makes that barrier (and the tunnelling time $\sim e^{+cN}$ ) infinite, trapping the system in one well; then sending $h\to0^+$ leaves it there with $m=\sqrt{-a/b}\ne0$ . Reverse the order — $h\to0$ first at finite $N$ — and the system still explores both wells, averaging to zero. The non-commuting limits are the precise mathematical encoding of "an infinite system chooses a phase."

Is the equivalence of ensembles always guaranteed in the limit?

No — only where the microcanonical entropy density $s(\varepsilon)$ is concave. There the canonical free energy is its Legendre transform and every ensemble gives the same equation of state; at finite $N$ they differ merely by $O(\ln N/N)$ surface terms. But if $s$ has a non-concave (convex) region — which happens at first-order transitions and in long-range systems — the Legendre transform "skips over" it, and the microcanonical ensemble can access states (with negative specific heat, say) that the canonical ensemble cannot represent. Ensemble inequivalence is thus a precise diagnostic that you have left the comfortable short-range world.

The partition function has complex zeros (Yang–Lee) — but temperature and fugacity are real. Why do complex zeros matter?

Because the real-axis behaviour of an analytic function is controlled by where its zeros sit in the complex plane, even if you never evaluate it there. For finite $N$ the zeros of $Z_N(z)$ avoid the positive real $z$ -axis, so $\ln Z_N$ is analytic and smooth for all real fugacities — no transition. As $N\to\infty$ the zeros migrate and their accumulation set can pinch the real axis at one point $z_c$ ; there $\ln Z$ loses analyticity and the pressure develops a kink. So the complex zeros are the bookkeeping that tells you whether and where a real-variable singularity — a phase transition — will appear.