Cumulant Expansion

Take the logarithm of an average of an exponential — that single, innocuous-looking operation is one of the most consequential moves in statistical mechanics. Whenever a partition function factorizes into a solvable reference times a thermal average of an exponentiated perturbation, $Z = Z_0\,\langle e^{-\beta V}\rangle_0$ , the free energy is $F = F_0 - k_BT\ln\langle e^{-\beta V}\rangle_0$ , and everything hinges on expanding that logarithm. Expand the exponential naively and you get the moments $\langle V^n\rangle_0$ ; but moments grow with the wrong power of system size — $\langle V^2\rangle_0 \sim N^2$ — and their sum stubbornly refuses to be extensive. The cumulant expansion is the reorganization that fixes this: it rewrites $\ln\langle e^{tX}\rangle$ as a sum of cumulants $\kappa_n$ , the genuinely connected pieces of the moments, each individually extensive. The cumulants are a ladder. The first is the mean, the second the variance, the third measures asymmetry, the fourth tailedness. For a Gaussian the ladder stops after two rungs — every cumulant beyond the variance vanishes — which is exactly why harmonic reference systems are so tractable. Above all, cumulants are additive: for independent parts, $\kappa_n(X+Y) = \kappa_n(X)+\kappa_n(Y)$ , whereas moments carry cross terms and do not add. That additivity is the algebraic guarantee of extensivity: $\ln Z$ for weakly coupled subsystems is the sum of the parts, and interactions are precisely the higher connected cumulants that couple them. This article owns the cumulant machinery: the moment-to-cumulant combinatorics, the additivity that makes cumulants the natural variables of any extensive quantity, and the identity $-\beta F = \ln Z$ that makes the free energy a cumulant generator. Applied to $X = -\beta V$ , this one identity is the engine behind first-order perturbation theory, the Gibbs–Bogoliubov variational bound, the linked-cluster theorem, and the virial series. The abstract combinatorics of the next few pages is the hinge that turns generating-function algebra into concrete thermodynamics.

Key takeaways

A cumulant expansion rewrites $\ln\langle e^{tX}\rangle = \sum_n \kappa_n\, t^n/n!$ ; the coefficients $\kappa_n$ are the cumulants, the connected parts of the moments.
The first four are the mean $\kappa_1=\langle X\rangle$ , the variance $\kappa_2=\langle X^2\rangle-\langle X\rangle^2$ , the skewness numerator $\kappa_3$ , and the excess-kurtosis numerator $\kappa_4=\langle X^4\rangle-3\langle X^2\rangle^2$ (for a centered variable).
For a Gaussian variable every cumulant beyond the second vanishes; this exact truncation is why harmonic/Gaussian reference systems are solvable.
Cumulants ADD over independent subsystems, $\kappa_n(X+Y)=\kappa_n(X)+\kappa_n(Y)$ , while moments do not — this additivity is the algebraic form of extensivity of $\ln Z$ .
Applied to $Z=Z_0\langle e^{-\beta V}\rangle_0$ , the logarithm gives the free-energy cumulant series $F=F_0+\langle V\rangle_0-\tfrac{\beta}{2}(\langle V^2\rangle_0-\langle V\rangle_0^2)+\cdots$ , whose truncation is first-order perturbation theory and the Gibbs–Bogoliubov bound.
Key idea: the logarithm turns products into sums — $\ln\langle e^{X}\rangle$ keeps only the connected (cumulant) parts of the moments, so every term is extensive and disconnected pieces vanish. Perturbation theory, the linked-cluster theorem, and the virial series are all this one identity applied to $X=-\beta V$ .

The core insight

The logarithm turns products into sums: $\ln\langle e^{X}\rangle$ keeps only the connected (cumulant) parts of the moments, so every term is extensive and disconnected pieces never appear. Perturbation theory, the linked-cluster theorem, and the virial series are all this one identity applied to $X=-\beta V$ .

Where this lives today

The cumulant expansion is quietly everywhere. In modern statistical field theory the connected correlation functions — the direct descendants of cumulants — are the fundamental objects: the generating functional $W[J]=\ln Z[J]$ produces only connected diagrams, and its Legendre transform builds the effective action from which every observable follows. The identity "the log keeps only connected pieces" that we proved here for a four-point moment is, verbatim, the linked-cluster theorem of many-body theory and the connected-diagram rule of quantum field theory (developed for gases in aq-sm-cumulants-linked-cluster). Beyond equilibrium physics, cumulants are the language of full counting statistics — the study of the entire distribution of charge or particles transferred through a mesoscopic conductor, where the higher cumulants ( $\kappa_3$ skewness, $\kappa_4$ kurtosis of the transferred charge) are measured directly and reveal correlations invisible to the average current and its noise. In quantum information, Kubo's second-cumulant (Gaussian) approximation governs how a qubit dephases under environmental noise, exact when the noise is Gaussian and quantified by non-Gaussianity ( $\kappa_{n\ge3}\ne0$ ) otherwise. And in quantitative finance and the physics of complex systems, the failure of the central limit theorem — heavy tails, nonzero higher cumulants — is precisely a large-cumulant phenomenon. In every case the strategy is the same one this article makes precise: keep the Gaussian part exactly, and treat the higher connected cumulants as controlled corrections.

The physical picture

Picture a fluctuating quantity $X$ — the energy of a subsystem, the interaction of a gas, a noisy signal. The moments $\langle X\rangle, \langle X^2\rangle, \langle X^3\rangle,\dots$ each mix together information you would rather keep separate: the second moment already contains the square of the mean, the fourth contains pieces built from the variance. The cumulants peel these apart. The mean tells you where the distribution sits; the variance how wide it is; the third cumulant its asymmetry; the fourth how heavy its tails are relative to a Gaussian. Each cumulant is a fresh, irreducible piece of information, and taking the logarithm of $\langle e^{tX}\rangle$ is precisely the operation that isolates them. The deepest reason cumulants matter in statistical mechanics is extensivity. The quantity $\ln Z$ must scale linearly with the number of particles $N$ — free energies are additive. But the raw moments of an interaction $V=\sum_{i<j}u_{ij}$ are not extensive: $\langle V^2\rangle_0 \sim N^2$ because it counts all pairs of pairs, including pairs that share no particle at all. The logarithm performs an exact cancellation. The disconnected $O(N^2)$ contributions in $\langle V^2\rangle_0$ are precisely matched and removed by $\langle V\rangle_0^2$ , leaving only the connected $O(N)$ piece. So the second cumulant — variance — is extensive even though the second moment is not. This is the same cancellation that, drawn as diagrams, keeps only connected diagrams in $\ln Z$ . The first figure shows the ladder itself: how each moment decomposes into cumulants, and how the logarithm strips away the disconnected pairings. The headline is the fourth-moment identity $m_4 = \kappa_4 + 3\kappa_2^2$ : the fourth moment is the genuinely connected four-point cumulant $\kappa_4$ plus three disconnected pairings of the variance (the Wick pairings). For a Gaussian only those three pairings survive and $\kappa_4=0$ ; any nonzero $\kappa_4$ is a direct, quantitative measure of non-Gaussianity.

Figure

Moments carry disconnected pieces; cumulants are the connected parts. The fourth moment m₄ = κ₄ + 3κ₂² splits into one connected four-point cumulant plus three disconnected pairings of the variance — for a Gaussian only the pairings survive

For a gas the same picture becomes literal. The interaction $V$ is a sum over pairs, so a cumulant of $V$ is a sum over ways the pair terms can touch through shared particles. Two pair terms that share a particle form a connected three-body cluster; two pair terms with no particle in common contribute zero after the connected subtraction. So the cumulants of $V$ are nothing but connected clusters of interacting particles — the exact statement of the Mayer cluster expansion and the linked-cluster theorem, developed for gases in aq-sm-mayer-cluster-expansion and unified with cumulants in the companion article aq-sm-cumulants-linked-cluster. The second figure makes the Gaussian truncation vivid: a Gaussian variable has only two nonzero cumulants, while a skewed or heavy-tailed distribution lights up the higher rungs of the ladder. This is why the cumulant expansion is a controlled non-Gaussianity expansion — you keep the Gaussian part exactly and treat everything else as corrections.

Figure

The cumulant ladder as a fingerprint of a distribution: a Gaussian has only κ₁ (mean) and κ₂ (variance); every higher cumulant is exactly zero. A skewed distribution lights up κ₃; a heavy-tailed one lights up κ₄

Finally, because cumulants of independent variables simply add, the cumulant expansion is the natural language for any extensive thermodynamic quantity. Non-interacting subsystems contribute additively to $\ln Z$ ; interactions are exactly the corrections that couple them, and those corrections are the higher connected cumulants. This is the thread we now make quantitative.

The core equations

The whole subject is generated by one object: the logarithm of the moment generating function. Start with a random variable $X$ , form the average $\langle e^{tX}\rangle$ of an exponential (a formal power series in the source $t$ ), and take its logarithm. The Taylor coefficients of that logarithm are the cumulants.

K(t) = \ln\langle e^{tX}\rangle = \sum_{n=1}^{\infty}\frac{t^n}{n!}\,\kappa_n \qquad (1)

where: $X$ — the random variable (in perturbation theory $X=-\beta V$ ); $t$ — a formal source / expansion parameter; $\kappa_n \equiv \langle X^n\rangle_c$ — the $n$ -th cumulant (the connected average); $\langle\cdot\rangle$ — the ensemble average over the reference distribution. The generating function $M(t)=\langle e^{tX}\rangle=\sum_n m_n t^n/n!$ (with moments $m_n=\langle X^n\rangle$ ) is the exponential of $K(t)$ : moments exponentiate cumulants, cumulants are the log of moments. Reading off the first two coefficients of $K(t)=\ln M(t)$ gives the two everyone already knows: the mean and the variance. The mean is just the first moment; the variance is the first place the logarithm does real work, subtracting off the square of the mean.

\kappa_1 = \langle X\rangle, \qquad \kappa_2 = \langle X^2\rangle - \langle X\rangle^2 \qquad (2)

where: $\kappa_1$ — the mean; $\kappa_2$ — the variance, the second cumulant. Note $\kappa_2\ge 0$ always: it is a variance, and this non-negativity has a thermodynamic consequence below (it is why fluctuations lower the free energy). The third and fourth cumulants are where "connected" stops being trivial. Expanding $\ln M(t)$ to fourth order and collecting powers of $t$ gives the standard combinatorial formulae; for a centered variable ( $\langle X\rangle=0$ ) they collapse to the third central moment and the excess-kurtosis numerator.

\kappa_3 = \langle X^3\rangle - 3\langle X^2\rangle\langle X\rangle + 2\langle X\rangle^3, \qquad \kappa_4 = \langle X^4\rangle - 4\langle X^3\rangle\langle X\rangle - 3\langle X^2\rangle^2 + 12\langle X^2\rangle\langle X\rangle^2 - 6\langle X\rangle^4 \qquad (3)

where: $\kappa_3$ — the skewness numerator (asymmetry); $\kappa_4$ — the excess-kurtosis numerator. For a centered variable these reduce to $\kappa_3=\langle X^3\rangle$ and $\kappa_4=\langle X^4\rangle-3\langle X^2\rangle^2$ ; the subtracted $3\langle X^2\rangle^2$ is exactly the three Gaussian (Wick) pairings, so $\kappa_4$ measures departure from Gaussianity. The inverse relation — moments in terms of cumulants — is the combinatorial heart, and it is beautiful: each moment is a sum over all set partitions, with each block replaced by a cumulant. This is the "moments = all groupings of connected pieces" statement.

m_n = \langle X^n\rangle = \sum_{\pi\,\in\,P(n)}\ \prod_{B\in\pi}\kappa_{|B|} \qquad (4)

where: $P(n)$ — the set of all partitions of $\{1,\dots,n\}$ into non-empty blocks; $B$ — a block of the partition $\pi$ ; $|B|$ — the size of block $B$ ; $\kappa_{|B|}$ — the cumulant of that size. For $n=4$ this reads $m_4=\kappa_4+4\kappa_3\kappa_1+3\kappa_2^2+6\kappa_2\kappa_1^2+\kappa_1^4$ , which for a centered variable is $\kappa_4+3\kappa_2^2$ — the connected part plus the three pairings. Now the physics. Apply all of this to a partition function that splits into a reference times an average of an exponentiated perturbation, $Z=Z_0\langle e^{-\beta V}\rangle_0$ . Since $-\beta F=\ln Z$ is itself a cumulant generator with $X=-\beta V$ , the free energy comes out as a cumulant series in the interaction.

F = F_0 - k_BT\,\ln\langle e^{-\beta V}\rangle_0 = F_0 + \langle V\rangle_0 - \frac{\beta}{2}\big(\langle V^2\rangle_0 - \langle V\rangle_0^2\big) + \cdots \qquad (5)

where: $F_0$ — the reference free energy; $V$ — the perturbation ( $V=H-H_0$ ); $\beta=1/k_BT$ ; $\langle\cdot\rangle_0$ — the average in the reference ensemble. The first cumulant $\langle V\rangle_0$ is first-order perturbation theory (the mean-field correction); the second cumulant $\langle V^2\rangle_0-\langle V\rangle_0^2\ge 0$ is a fluctuation correction that always lowers $F$ — the seed of the Gibbs–Bogoliubov variational bound $F\le F_0+\langle V\rangle_0$ that grounds every mean-field theory (aq-sm-mean-field-theory).

The core equations — reading them together

A useful "which relation when" summary of the four structural identities above:

relation	statement	use it for
$K=\ln M$	cumulants are the log of moments	defining and extracting $\kappa_n$
moment → cumulant (eq. 3)	$\kappa_n$ from central moments	computing skewness, kurtosis, corrections
cumulant → moment (eq. 4)	sum over set partitions	Wick's theorem, Gaussian truncation
additivity	$\kappa_n(X+Y)=\kappa_n(X)+\kappa_n(Y)$	extensivity, central-limit scaling

Derivation AThe first four cumulants from the logarithm of the moment generating function

The goal here is to derive the explicit moment-to-cumulant formulae of equation (3) directly from the defining identity $K(t)=\ln M(t)$ , by expanding both sides as power series in the source $t$ and matching coefficients order by order. The strategy is purely algebraic: write $M(t)=1+m_1t+m_2t^2/2+\cdots$ , substitute into $\ln(1+u)=u-u^2/2+u^3/3-\cdots$ , and collect. The one thing to watch is bookkeeping of factorials — the cumulant $\kappa_n$ is $n!$ times the coefficient of $t^n$ in $K(t)$ , not the coefficient itself — because both series are exponential generating functions. Getting that factor right is the entire content of the " $3$ " in $\kappa_4$ 's $-3\langle X^2\rangle^2$ .

Step 1. Write both generating functions as exponential power series. $M(t)=\sum_{n\ge0} m_n t^n/n!$ with $m_0=1$ , and $K(t)=\sum_{n\ge1}\kappa_n t^n/n!$ . The defining relation is $M(t)=e^{K(t)}$ , equivalently $K(t)=\ln M(t)$ . Why: the exponential form is what makes cumulants additive later — the log of a product of MGFs is a sum of CGFs.

Step 2. Substitute $u=M-1$ into the log series. With $u=m_1t+m_2t^2/2+m_3t^3/6+m_4t^4/24+\cdots$ , expand $\ln(1+u)=u-\tfrac12u^2+\tfrac13u^3-\tfrac14u^4+\cdots$ . Each power of $u$ contributes to several orders in $t$ ; the $-\tfrac12u^2$ term is what subtracts $\langle X\rangle^2$ from the second moment. Why: the subtractions in equations (2)–(3) come entirely from the higher powers of $u$ in the log — this is where "connected" is enforced algebraically.

Step 3. Collect the coefficient of $t^1$ and $t^2$ . At order $t^1$ : only $u$ contributes, giving $\kappa_1=m_1$ . At order $t^2$ : the coefficient of $t^2/2$ is $m_2-m_1^2$ , so $\kappa_2=m_2-m_1^2$ , the variance. Why: the mean is untouched (nothing to subtract at first order), but at second order the log removes the disconnected product $m_1^2$ — the origin of extensivity.

Step 4. Collect $t^3$ and $t^4$ . The same procedure yields $\kappa_3=m_3-3m_1m_2+2m_1^3$ and $\kappa_4=m_4-4m_1m_3-3m_2^2+12m_1^2m_2-6m_1^4$ . For a centered variable ( $m_1=0$ ) these are $\kappa_3=m_3$ and $\kappa_4=m_4-3m_2^2$ — the third central moment and the excess kurtosis. Why: the $-3m_2^2$ is precisely the three ways to pair four points (Wick), so a Gaussian, whose fourth moment is exactly $3m_2^2$ , has $\kappa_4=0$ . *Verified: the four moment-to-cumulant relations of equation (3), the centered forms, and the decomposition $m_4=\kappa_4+3\kappa_2^2$ are all confirmed symbolically in verify_cumulant-expansion.py by expanding $\ln M(t)$ to $O(t^4)$ and reading $\kappa_n=n!\,[t^n]K(t)$ .*

Go deeper · Derivation B · 2 minAdditivity, the Gaussian truncation, and the free-energy cumulant series

Derivation BAdditivity, the Gaussian truncation, and the free-energy cumulant series

The purpose of this card is to establish the two properties that make cumulants indispensable in statistical mechanics — additivity over independent variables and exact truncation for Gaussians — and then to assemble the free-energy series of equation (5). The route is to use the exponential structure $M=e^K$ : for independent variables the MGFs multiply, so the CGFs add, so the cumulants add termwise; and a Gaussian has a CGF that is a polynomial of degree two, so its cumulants beyond the second are identically zero. What to watch is the direction of the free-energy correction: the second cumulant is a variance and enters $F$ with a minus sign, so fluctuations always lower the free energy — a fact with a rigorous name (Gibbs–Bogoliubov).

Step 1. Additivity for independent variables. If $X$ and $Y$ are independent, $\langle e^{t(X+Y)}\rangle=\langle e^{tX}\rangle\langle e^{tY}\rangle$ , so $K_{X+Y}(t)=K_X(t)+K_Y(t)$ and therefore $\kappa_n(X+Y)=\kappa_n(X)+\kappa_n(Y)$ for every $n$ . Why: moments do NOT add — the second moment of $X+Y$ carries a cross term $2\langle X\rangle\langle Y\rangle$ — but the logarithm converts the product of MGFs into a sum, killing the cross terms; this is exactly why $\ln Z$ is extensive.

Step 2. Gaussian truncation. A Gaussian with mean $\mu$ and variance $\sigma^2$ has the exact CGF $K(t)=\mu t+\tfrac12\sigma^2 t^2$ , a polynomial of degree two. Hence $\kappa_1=\mu$ , $\kappa_2=\sigma^2$ , and $\kappa_n=0$ for all $n\ge3$ . Why: the entire cumulant ladder collapses to two rungs, which is why a Gaussian (harmonic) reference truncates the expansion exactly at second order — the basis of every Gaussian variational method.

Step 3. Free energy as a cumulant generator. From $Z=Z_0\langle e^{-\beta V}\rangle_0$ and $F=-k_BT\ln Z$ , one gets $-\beta F=-\beta F_0+\ln\langle e^{-\beta V}\rangle_0$ . The second term is the CGF of $V$ evaluated at source $t=-\beta$ : $\ln\langle e^{-\beta V}\rangle_0=\sum_n \tfrac{(-\beta)^n}{n!}\kappa_n^{(V)}$ , with the cumulants taken in the reference ensemble. Why: $-\beta F=\ln Z$ literally IS a cumulant generating function, so the free energy inherits the whole cumulant structure of the interaction.

Step 4. Read the first two orders. The $n=1$ term is $-\beta\langle V\rangle_0$ and the $n=2$ term is $+\tfrac12\beta^2(\langle V^2\rangle_0-\langle V\rangle_0^2)$ . Dividing by $-\beta$ gives $F=F_0+\langle V\rangle_0-\tfrac{\beta}{2}(\langle V^2\rangle_0-\langle V\rangle_0^2)+\cdots$ . Since the variance is non-negative, truncating at first order overestimates $F$ : $F\le F_0+\langle V\rangle_0$ , the Gibbs–Bogoliubov bound. Why: keeping only $\kappa_1$ is first-order perturbation theory and the variational mean-field free energy; the neglected second cumulant, being a variance, guarantees the truncation is an upper bound. *Verified: the Gaussian cumulants ( $\kappa_{1}=\mu,\kappa_2=\sigma^2,\kappa_{n\ge3}=0$ ), the additivity $\kappa_2(X+Y)=\kappa_2(X)+\kappa_2(Y)$ (with the non-adding moment cross term $2m_{x1}m_{y1}$ exhibited), and the free-energy series with its sign are all confirmed in verify_cumulant-expansion.py.*

Limiting cases

The cumulant expansion specializes in the regimes that matter most in statistical mechanics.

Case	Result	Meaning
Gaussian reference	$\kappa_n=0$ for $n\ge3$	Only mean and variance survive; harmonic/Gaussian reference systems truncate the expansion exactly at second order.
First order (mean field)	$F\approx F_0+\langle V\rangle_0$	Keeping only $\kappa_1$ is first-order perturbation theory; the neglected $\kappa_2\ge0$ makes this a rigorous upper bound (Gibbs–Bogoliubov).
Independent subsystems	$\kappa_n(X_A{+}X_B)=\kappa_n(X_A)+\kappa_n(X_B)$	Cumulants add over independent parts, guaranteeing extensivity of $\ln Z$ .
Sum of $N$ i.i.d. terms	$\kappa_n(S_N)=N\kappa_n(X)$	The cumulant of a sum scales as $N$ ; normalizing by $\sqrt{N}$ sends $\kappa_{n\ge3}\to0$ — the central limit theorem.
Dilute classical gas	$\ln Z=\sum_{\ell}\tfrac{V}{\lambda^3}b_\ell z^\ell$	Cumulants of $V=\sum_{i<j}u_{ij}$ regroup into Mayer cluster integrals $b_\ell$ , giving the virial series.

Common trapCumulants are not central moments beyond third order

It is tempting to think " $\kappa_2$ is the variance, $\kappa_3$ is the third central moment, so $\kappa_4$ must be the fourth central moment." It is not. The fourth cumulant is $\kappa_4=\mu_4-3\mu_2^2$ , the fourth central moment minus three times the squared variance — the excess kurtosis, not the kurtosis. The subtracted $3\mu_2^2$ is exactly the Gaussian (Wick) part, which is why $\kappa_4$ vanishes for a Gaussian while $\mu_4=3\sigma^4$ does not. The identity $\kappa_n=\mu_n$ (central moment) holds only for $n\le3$ ; from fourth order on, cumulants and central moments diverge, and conflating them is the classic error. A second trap: the naive power series of $\langle e^{-\beta V}\rangle_0$ gives moments, which are not extensive and double-count disconnected configurations — you must take the logarithm (or equivalently subtract the disconnected pieces) to recover the extensive cumulants.

Worked examples

1 · All cumulants of a Poisson distribution equal its mean (confidence builder)

We compute the first four cumulants of a Poisson distribution with rate $\mu$ and discover a remarkable fact: every one of them equals $\mu$ .

Step 1. The moment generating function of Poisson $(\mu)$ is $M(t)=\langle e^{tX}\rangle=e^{\mu(e^t-1)}$ . Its logarithm is immediate: $K(t)=\ln M(t)=\mu(e^t-1)$ .

Step 2. Expand $e^t-1=t+\tfrac{t^2}{2}+\tfrac{t^3}{6}+\tfrac{t^4}{24}+\cdots$ , so $K(t)=\mu\big(t+\tfrac{t^2}{2}+\tfrac{t^3}{6}+\tfrac{t^4}{24}+\cdots\big)$ . The coefficient of $t^n/n!$ in $K(t)$ is $\mu$ for every $n$ .

Step 3. Therefore $\kappa_n=\mu$ for all $n$ : mean $\kappa_1=\mu$ , variance $\kappa_2=\mu$ , third cumulant $\kappa_3=\mu$ , fourth $\kappa_4=\mu$ . The famous "variance equals mean" of the Poisson is just the $n=2$ case of a much stronger statement. Every cumulant of a Poisson equals its rate $\mu$ .

Step 4. Cross-check numerically for $\mu=3$ using the moments. The Poisson moments are $m_1=3$ , $m_2=12$ , $m_3=57$ , $m_4=309$ (from $m_n=\sum_k S(n,k)\mu^k$ with Stirling numbers of the second kind). Then $\kappa_2=m_2-m_1^2=12-9=\mathbf{3}$ , $\kappa_3=m_3-3m_1m_2+2m_1^3=57-108+54=\mathbf{3}$ , and $\kappa_4=m_4-4m_1m_3-3m_2^2+12m_1^2m_2-6m_1^4=309-684-432+1296-486=\mathbf{3}$ . All four cumulants equal $3$ , confirming the general result and providing a clean stress test of the moment-to-cumulant formulae.

2 · Second-order free-energy correction for a two-level perturbation (exam level)

We apply the free-energy cumulant series to a concrete perturbation and compute the first- and second-order corrections, exhibiting the Gibbs–Bogoliubov sign.

Step 1. Take a reference system whose perturbation $V$ takes two values in the reference ensemble: $V=+\varepsilon$ with probability $p$ and $V=-\varepsilon$ with probability $1-p$ . Compute the reference averages: $\langle V\rangle_0=\varepsilon(2p-1)$ and $\langle V^2\rangle_0=\varepsilon^2$ .

Step 2. The first cumulant (first-order perturbation theory) is $\langle V\rangle_0=\varepsilon(2p-1)$ . For the symmetric case $p=\tfrac12$ this vanishes: to first order the perturbation does not shift the free energy at all.

Step 3. The second cumulant is $\kappa_2^{(V)}=\langle V^2\rangle_0-\langle V\rangle_0^2=\varepsilon^2-\varepsilon^2(2p-1)^2=\varepsilon^2\cdot 4p(1-p)$ . For $p=\tfrac12$ this is $\varepsilon^2$ , the maximum possible variance.

Step 4. The free energy to second order is $F=F_0+\varepsilon(2p-1)-\tfrac{\beta}{2}\varepsilon^2\cdot 4p(1-p)+\cdots$ . At $p=\tfrac12$ this is $F=F_0-\tfrac{\beta\varepsilon^2}{2}+\cdots=F_0-\dfrac{\varepsilon^2}{2k_BT}+\cdots$ . The correction is strictly negative and grows as the temperature drops: fluctuations lower the free energy, exactly as the Gibbs–Bogoliubov bound $F\le F_0+\langle V\rangle_0$ demands. The second cumulant, being a variance, can only push $F$ down.

Origins · 1 mincumulant expansion and how this was discovered

cumulant expansion

introduced by T. N. Thiele (1889) as "half-invariants"; named "cumulants" by R. A. Fisher and J. Wishart (1932); central to statistical mechanics via the free-energy and linked-cluster connections · κ_n are the connected correlations that add for independent variables

The cumulant expansion is the systematic reorganization of the logarithm of a moment generating function into its connected, additive parts — the cumulants. The Danish astronomer and statistician Thorvald N. Thiele introduced them in 1889 under the name "half-invariants," recognizing that these quantities, unlike the moments, are additive over independent variables and vanish above second order for the normal law. Ronald Fisher and John Wishart gave them the modern name "cumulants" in 1932, and Fisher exploited their additivity throughout his work on statistical estimation. In physics the same combinatorics arrived through the theory of imperfect gases: Ursell's 1927 cluster functions and Mayer's 1937 cluster expansion are the cumulant expansion applied to the interaction of a classical gas, where the logarithm of the grand partition function keeps only connected clusters. Ryogo Kubo's 1962 "generalized cumulant expansion" extended the method to time-ordered exponentials of fluctuating quantum Hamiltonians, making it the standard tool for NMR line shapes and, today, for the dephasing of qubits by Gaussian noise. Across all these settings the content is a single identity — the log of an average of an exponential is the sum of the connected averages — which is why cumulants, connected correlation functions, connected Feynman diagrams, and linked clusters are all the same object wearing different clothes.

History

1889

— Thiele introduces cumulants as "half-invariants," noting their additivity and Gaussian truncation.

1927

— Ursell introduces cluster functions, the connected pieces of the classical gas partition function.

1932

— Fisher and Wishart give cumulants their modern name and exploit their additivity.

1937

— Mayer's cluster expansion realizes the cumulant expansion for the interacting classical gas.

1962

— Kubo formulates the generalized (quantum, time-ordered) cumulant expansion for fluctuating Hamiltonians.

Going further

Texts. Kardar, Statistical Physics of Particles, §5–6 (cumulant and cluster expansions). Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 10 (cluster expansion). Reif, Fundamentals of Statistical and Thermal Physics, Ch. 10. Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity. Landau & Lifshitz, Statistical Physics Part 1, §74–76. R. Kubo, "Generalized Cumulant Expansion Method," J. Phys. Soc. Jpn. 17, 1100 (1962).

Provenance. Base article: AllQFT track `sm`, "Cumulant Expansion" (STAT.INTER.CUMUL.1259); source Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 10. All symbolic and numeric claims re-verified here in verify_cumulant-expansion.py: the four moment-to-cumulant relations by expanding $\ln M(t)$ to $O(t^4)$ ; the centered forms and $m_4=\kappa_4+3\kappa_2^2$ ; the Gaussian truncation $\kappa_{n\ge3}=0$ ; additivity for independent variables (with the non-adding moment cross term shown); the free-energy series and its sign; and the Poisson result $\kappa_n=\mu$ (all cumulants equal, numeric check at $\mu=3$ ). No base physics error found; the base's equations (moment–cumulant partition formula, free-energy series, Gibbs–Bogoliubov bound, hard-sphere $B_2=\tfrac{2\pi}{3}\sigma^3$ ) are all standard and correct.

Prerequisites. Configuration Integral and the Interacting Partition Function · The Mayer f-Function · Partition function and free energy · Moments and probability distributions.

Leads to. Linked-Cluster Theorem · Mayer Cluster Expansion · Virial Expansion of the Equation of State · Van der Waals from the Virial Expansion.

Why does taking a logarithm turn products into connected sums? It still feels like magic.

It is the exponential structure, not magic. Moments exponentiate cumulants: $M(t)=e^{K(t)}$ , so a moment of order $n$ is built from all ways of grouping $n$ points into cumulant blocks — connected pieces multiplied together (the partition sum of equation 4). The logarithm inverts the exponential and therefore returns only the single connected block, stripping away every product of smaller pieces. The same identity — "the log of a sum over all groupings is the sum over connected groupings" — is why the log of the grand partition function keeps only connected clusters (Mayer) and why $\ln Z$ in field theory keeps only connected Feynman diagrams. One theorem, many costumes.

Is the fourth cumulant the same as the fourth central moment?

No, and this is the most common slip. The fourth cumulant is $\kappa_4=\mu_4-3\mu_2^2$ — the fourth central moment minus three times the squared variance. The subtracted term is the three Gaussian pairings (Wick's theorem), so $\kappa_4$ is the excess kurtosis: it measures how much heavier or lighter the tails are than a Gaussian with the same variance. A Gaussian has $\mu_4=3\sigma^4$ but $\kappa_4=0$ . The identification $\kappa_n=\mu_n$ (central moment) is true only through third order; from fourth order onward the disconnected pairings make them differ.

When is truncating the free energy at first cumulant actually a good approximation?

Whenever the fluctuations of the perturbation in the reference ensemble are small, i.e. $\beta^2\langle V^2\rangle_{0,c}\ll |\beta\langle V\rangle_0|$ or, more usefully, whenever a well-chosen reference makes the relative second cumulant small. The magic of the Gibbs–Bogoliubov bound is that even when the fluctuations are not small, the first-order result is still a rigorous upper bound on $F$ — so you can variationally minimize $F_0+\langle V\rangle_0$ over a family of reference Hamiltonians and get the best mean-field estimate. Mean-field theory is exactly this variational first-cumulant truncation; the neglected second cumulant is the fluctuation correction it ignores.

How does the cumulant expansion connect to the central limit theorem?

Directly and cleanly. For a sum $S_N=X_1+\cdots+X_N$ of i.i.d. variables, additivity gives $\kappa_n(S_N)=N\kappa_n(X)$ . Now standardize: $Z_N=(S_N-N\mu)/\sqrt{N}\sigma$ . Under this rescaling the $n$ -th cumulant of $Z_N$ scales as $N^{1-n/2}$ , so $\kappa_1(Z_N)=0$ , $\kappa_2(Z_N)=1$ , and every higher cumulant $\kappa_{n\ge3}(Z_N)\to0$ as $N\to\infty$ . A distribution whose cumulants beyond the second vanish is a Gaussian — that is the central limit theorem, read straight off the cumulant ladder.

What is the physical meaning of the second cumulant of the interaction being a variance?

Its non-negativity is a statement about fluctuations and free energy. In $F=F_0+\langle V\rangle_0-\tfrac{\beta}{2}\langle V^2\rangle_{0,c}+\cdots$ , the second cumulant $\langle V^2\rangle_{0,c}=\langle V^2\rangle_0-\langle V\rangle_0^2\ge0$ enters with a minus sign, so the fluctuations of the interaction in the reference ensemble always lower the free energy relative to the first-order (mean-field) estimate. Physically, a system with an interaction that fluctuates can lower its free energy by correlating with those fluctuations — it does better than the static mean-field value. That improvement is bounded by $-\tfrac{\beta}{2}\langle V^2\rangle_{0,c}$ at leading order, and its non-negativity is the microscopic reason the Gibbs–Bogoliubov inequality points the way it does.

Why are cumulants the natural variables for thermodynamics rather than moments?

Because thermodynamic quantities are extensive — they scale with system size — and cumulants are the objects that respect extensivity automatically. $\ln Z$ , and hence the free energy, must be proportional to $N$ ; the moments of an interaction are not (they scale as higher powers of $N$ because they count disconnected configurations), but the cumulants are, because they keep only connected pieces that scale linearly with $N$ at fixed density. Additivity over independent subsystems is the same statement. So expressing $\ln Z$ as a cumulant series is not a trick — it is writing the free energy in the only variables that can be extensive term by term.

Does the cumulant expansion always converge?

Not necessarily — like any perturbation series it can be asymptotic rather than convergent. It is genuinely controlled when the perturbation is small or the reference is well chosen (small relative second cumulant), and for the classical gas the cluster form has a finite radius of convergence in density (Penrose, Ruelle) at low density and short range. But near a phase transition — where fluctuations of the order parameter become large and correlations long-ranged — the higher cumulants stop being small and the truncated expansion fails. That failure is not a defect of the method; it is the physics telling you that mean-field/perturbative thinking has broken down and you need the renormalization group. The cumulant expansion is exact as a reorganization but only useful as a truncation when the fluctuations it neglects are controlled.