Cumulants and the Linked-Cluster Connection

Two problems that look nothing alike turn out to run on the same engine. In probability, you describe a random variable through its moments $\langle X\rangle,\langle X^2\rangle,\dots$ — but the moments mix genuine information with trivial bookkeeping, so the sharper description uses cumulants, the Taylor coefficients of $\log\langle e^{tX}\rangle$ . In statistical mechanics, you describe an interacting gas through its partition function, a sum over every configuration — but the physically clean object is $\log Z$ , which keeps only the connected clusters of particles. These are not two analogies. They are the same theorem, and this article proves the identity: a cumulant is a connected correlation, and a connected cluster is a cumulant. The bridge is the combinatorics of set partitions. A moment of order $n$ is a sum over every way to partition $n$ points into blocks, with each block contributing a cumulant; a partition function is a sum over every way to group particles into clusters, with each cluster contributing a connected integral. In both cases a generic object factorizes into independent connected pieces, and summing over how many of each piece appear rebuilds an exponential. That single fact — the exponential formula, $(\text{all pieces})=\exp(\text{connected pieces})$ — is why the logarithm extracts connectivity in one language and cumulants in the other. The payoff is structural and physical. It is why $\log Z$ is extensive (each connected cluster is local, so the sum scales with volume), why free energies are sums of connected diagrams (disconnected vacuum bubbles cancel), and why a Gaussian has no cumulants beyond the second (Wick's theorem: nothing is irreducibly connected past a pair). The Mayer cluster expansion, the virial series, the cumulant expansion of perturbation theory, and the linked-cluster theorem of many-body physics are one identity in four costumes — and this article is where they meet.

Key takeaways

The cumulant generating function $K(t)=\log\langle e^{tX}\rangle$ has Taylor coefficients $\kappa_n$ ; taking the logarithm is exactly the operation that strips away disconnected contributions.
A moment is a sum over all set partitions of products of cumulants ( $\mu_n=\sum_\pi\prod_B\kappa_{|B|}$ ); the cumulant is the single-block, connected piece — the inversion runs on the partition lattice.
The exponential formula $\sum_N Z_N/N!=\exp(\sum_\ell C_\ell/\ell!)$ turns a sum over all graphs into the exponential of a sum over connected graphs, mirroring $M(t)=e^{K(t)}$ .
Applied to the grand partition function this is the linked-cluster theorem, $\log\Xi=V\sum_\ell b_\ell z^\ell$ : the Mayer cluster integrals $b_\ell$ are the cumulants of the fugacity expansion.
Cumulants add for independent parts and vanish beyond second order for Gaussians, which is why they are the natural language for fluctuations, response functions, and diagrammatic perturbation theory.
Key idea: taking the logarithm of a generating function keeps only connected pieces. Cumulants are connected moments and connected clusters are cumulants; the exponential formula $(\text{all pieces})=\exp(\text{connected pieces})$ is the one theorem behind the cumulant expansion, the Mayer cluster expansion, and the connected-diagram rule of field theory.

The core insight

Taking the logarithm of a generating function is the same operation as keeping only connected pieces: cumulants are connected moments, and connected clusters are cumulants. The exponential formula $(\text{all pieces})=\exp(\text{connected pieces})$ unifies the moment–cumulant relation, the Mayer cluster expansion, and the connected-diagram rule of quantum field theory.

Where this lives today

The connected-piece theorem is the quiet workhorse of quantitative many-body physics. Coupled-cluster theory — the most accurate routinely used method in quantum chemistry, and the reason molecular energies can be computed to chemical accuracy — is built on the exponential ansatz $|\Psi\rangle=e^{\hat T}|\Phi_0\rangle$ , whose name and structure descend directly from Mayer's clusters: the exponential automatically resums the disconnected excitations, leaving only the connected amplitudes to be solved for. In condensed matter and field theory, the generating functional of connected Green's functions, $W=\log Z$ , is the object from which every physical correlation is built, precisely because the logarithm discards disconnected vacuum factors. Cumulants have also become a direct experimental probe. In heavy-ion collisions, event-by-event cumulants of the net-proton number — the third and fourth in particular — are measured as signatures of the QCD critical point, where higher cumulants are predicted to diverge faster than lower ones and can even change sign; the same $\log$ -of-generating-function structure that defines a virial coefficient defines these fluctuation observables. In quantum transport, "full counting statistics" characterizes a current not by its mean alone but by the cumulants of the transferred charge, whose generating function is again a logarithm. Wherever a system's fluctuations, not just its averages, carry the physics, the connected structure of this article is doing the work.

The physical picture

Picture a configuration of an interacting gas as a graph: each particle is a vertex, and a Mayer bond $f_{ij}=e^{-\beta u(r_{ij})}-1$ is drawn wherever two particles are close enough to interact. A typical configuration splits into several disconnected blobs that do not know about one another — the integral over the whole graph factorizes into a product over its connected pieces. Summing over all graphs gives the partition function $\Xi$ ; but because the blobs factorize, that sum is the exponential of the sum over a single connected blob. The logarithm then undoes the exponential and leaves only connected blobs. That is the linked-cluster theorem, and it is the reason $\log\Xi$ counts connected clusters and nothing else. The probabilistic mirror is exact. Write $\langle e^{tX}\rangle$ for a quantity built from many independent contributions. Moments mix everything, including products of independent pieces; but $\log\langle e^{tX}\rangle$ is additive over independent variables, so it can only retain genuinely joint — that is, connected — information. Cumulants of a sum of independent variables add, and cross-cumulants between independent variables vanish. "Additive under $\log$ " and "connected under graph-summation" are the same statement, and the figure below makes the dictionary explicit: cumulants, gas clusters, and Feynman diagrams are one connected structure wearing three costumes.

Figure

The one theorem, three costumes: connected pieces exponentiate. Mayer clusters (f-bonds), statistical cumulants κₙ (connected correlations), and connected Feynman diagrams (propagators) all obey all-pieces = exp(connected).

A useful slogan: the moment of a product is a sum over all pairings and partitions; the cumulant is the part that cannot be factored. For a Gaussian, every higher moment is exactly the sum over pairings — Wick's theorem — so every cumulant beyond the second vanishes: there is nothing irreducibly connected past the two-point function. The next figure shows the smallest nontrivial case, $\mu_4=\kappa_4+3\kappa_2^2$ : one connected four-point piece, plus three disconnected pairings that the logarithm removes. Reading a moment "minus its disconnected pieces" is exactly reading a cumulant.

Figure

The fourth moment splits into one connected part (κ₄) plus three disconnected pairings (3 κ₂²). Taking the logarithm removes the three pairings and leaves κ₄ — a cumulant is a moment with its disconnected pieces subtracted.

Extensivity now comes for free. A connected cluster of $\ell$ particles, after one center-of-mass integration, contributes a volume-independent number times a single factor of $V$ . So $\log\Xi\sim V$ , while $\Xi=e^{\log\Xi}$ grows like $e^{cV}$ . Extensivity, connectivity, and the logarithm are three names for one thing.

The core equations

The starting point is the pair of generating functions. The moment generating function packages every moment; its logarithm packages the cumulants, and equating Taylor coefficients turns cumulants into explicit polynomials in the moments.

M(t)=\langle e^{tX}\rangle=\sum_{n=0}^{\infty}\frac{\mu_n}{n!}\,t^n,\qquad K(t)=\log M(t)=\sum_{n=1}^{\infty}\frac{\kappa_n}{n!}\,t^n \qquad (1)

where: $M(t)$ — moment generating function; $K(t)$ — cumulant generating function, its logarithm; $\mu_n=\langle X^n\rangle$ — the $n$ -th moment (with $\mu_0=1$ ); $\kappa_n$ — the $n$ -th cumulant. The single operation " $\log$ " is what converts the moment ladder into the cumulant ladder — everything in this article is a consequence of that one step. Reading off the low-order coefficients gives the cumulants as moment polynomials. The first is the mean, the second the variance; the third is the third central moment, but the fourth is not the fourth central moment.

\kappa_1=\mu_1,\quad \kappa_2=\mu_2-\mu_1^2,\quad \kappa_3=\mu_3-3\mu_2\mu_1+2\mu_1^3,\quad \kappa_4=\mu_4-4\mu_3\mu_1-3\mu_2^2+12\mu_2\mu_1^2-6\mu_1^4 \qquad (2)

where: $\kappa_2$ — variance; $\kappa_3$ — skewness numerator; $\kappa_4$ — excess-kurtosis numerator. For a centered variable ( $\mu_1=0$ ) these reduce to $\kappa_2=\mu_2$ , $\kappa_3=\mu_3$ , and crucially $\kappa_4=\mu_4-3\mu_2^2$ — the $-3\mu_2^2$ subtracts the Gaussian (disconnected-pairing) baseline, so $\kappa_4$ measures departure from Gaussianity. The inverse relation is the combinatorial heart of the whole subject: a moment is a sum over every way to partition its indices into blocks, each block contributing one cumulant. The cumulant is the single-block ("connected") term of the moment.

\mu_n=\sum_{\pi\in\mathcal{P}(n)}\ \prod_{B\in\pi}\kappa_{|B|},\qquad \kappa_n=\sum_{\pi\in\mathcal{P}(n)}(-1)^{|\pi|-1}(|\pi|-1)!\prod_{B\in\pi}\mu_{|B|} \qquad (3)

where: $\mathcal{P}(n)$ — the set of partitions of $\{1,\dots,n\}$ (there are $B_n$ of them, the Bell numbers $B_2=2,B_3=5,B_4=15$ ); $\pi$ — a partition into blocks $B$ ; $|\pi|$ — number of blocks. The first identity is the exponential formula in disguise; the second is Möbius inversion on the partition lattice, with the alternating signs that pick out connectivity. The same algebra applied to the grand partition function is the linked-cluster theorem. Summing over particle number exponentiates the disconnected configurations, so the logarithm keeps only connected clusters — the cluster integrals $b_\ell$ play the role of the cumulants.

\Xi=\sum_{N=0}^{\infty}\frac{z^N}{N!}Q_N=\exp\!\Big(V\sum_{\ell\ge1}b_\ell z^\ell\Big),\qquad \log\Xi=V\sum_{\ell\ge1}b_\ell\,z^\ell \qquad (4)

where: $z$ — fugacity; $Q_N$ — the $N$ -particle configuration integral (owned by aq-sm-configuration-integral); $b_\ell$ — the $\ell$ -th cluster integral (owned by aq-sm-mayer-cluster-expansion), the connected $\ell$ -cluster per unit volume; $V$ — volume. Compare (4) with (1): $\ell\leftrightarrow n$ , $b_\ell\leftrightarrow\kappa_\ell$ , $\Xi\leftrightarrow M$ . The linked-cluster theorem is the moment–cumulant relation for the fugacity expansion. The identification is literal: the cluster integral is a cumulant of the interaction, computed in the ideal-gas measure. Writing $\Xi=Z_0\langle e^{-\beta U}\rangle_0$ and taking the log gives the cumulant series of $-\beta U$ .

\log Z-\log Z_0=\log\big\langle e^{-\beta U}\big\rangle_0=\sum_{n\ge1}\frac{(-\beta)^n}{n!}\,\langle U^n\rangle_{0,c},\qquad \langle U^n\rangle_{0,c}\ \Longleftrightarrow\ \text{connected }f\text{-bond graphs} \qquad (5)

where: $\langle\cdots\rangle_0$ — average over the non-interacting configuration measure; $\langle U^n\rangle_{0,c}$ — the connected (cumulant) average of the total interaction energy. Re-expressing $U=\sum_{i<j}u_{ij}$ with $e^{-\beta u}=1+f$ converts these energy cumulants into connected $f$ -bond graphs, recovering the $b_\ell$ . Cluster integrals are cumulants; the linked-cluster theorem is the cumulant theorem in disguise.

Derivation AThe moment is a sum over partitions and the cumulant is its connected part

The goal is to establish the master combinatorial identity, $\mu_n=\sum_{\pi\in\mathcal P(n)}\prod_{B\in\pi}\kappa_{|B|}$ , and to check it explicitly for $n=2,3,4$ so that "the cumulant is the connected piece of the moment" becomes a precise, verifiable statement rather than a slogan. The strategy is to take the definition $M(t)=e^{K(t)}$ , expand the exponential of the cumulant sum, and organize the resulting terms by which cumulants appear at each order — the organizing principle turns out to be exactly the set partitions of the index set. The subtlety to watch is the counting of identical blocks: the $1/m!$ symmetry factors of the exponential are precisely what makes the raw expansion collapse into a clean sum over distinct partitions, with each partition counted once.

Step 1. Exponentiate the cumulant sum. From $M=e^{K}$ with $K=\sum_{m\ge1}\kappa_m t^m/m!$ , write $M=\prod_{m\ge1}\exp(\kappa_m t^m/m!)=\prod_{m\ge1}\sum_{p_m\ge0}\frac{1}{p_m!}\big(\tfrac{\kappa_m}{m!}\big)^{p_m}t^{m p_m}$ . Why: the exponential of a sum is the product of exponentials, and each factor is a Taylor series in the number $p_m$ of size- $m$ blocks that will appear.

Step 2. Collect the coefficient of $t^n/n!$ . A term has total degree $n=\sum_m m\,p_m$ ; its coefficient of $t^n/n!$ is $\dfrac{n!}{\prod_m (m!)^{p_m}p_m!}\prod_m \kappa_m^{p_m}$ . That multinomial factor is exactly the number of set partitions of $\{1,\dots,n\}$ with $p_m$ blocks of size $m$ . Why: choosing which of the $n$ labelled points fall into each block, then dividing by the internal orderings $(m!)^{p_m}$ and the interchange of identical blocks $p_m!$ , is the standard count of partitions of a given block-type.

Step 3. Sum over block types. Summing over all $\{p_m\}$ with $\sum m p_m=n$ therefore reproduces $\mu_n=\sum_{\pi\in\mathcal P(n)}\prod_{B\in\pi}\kappa_{|B|}$ , one term per partition $\pi$ . Why: every partition of the index set arises exactly once, and its weight is the product of one cumulant per block — the exponential formula for exponential generating functions.

Step 4. Read the low orders. For $n=2$ : partitions $\{12\}$ and $\{1\}\{2\}$ give $\mu_2=\kappa_2+\kappa_1^2$ . For $n=3$ : $\mu_3=\kappa_3+3\kappa_2\kappa_1+\kappa_1^3$ (one 3-block, three ways to split $3=2+1$ , one all-singletons). For $n=4$ : $\mu_4=\kappa_4+4\kappa_3\kappa_1+3\kappa_2^2+6\kappa_2\kappa_1^2+\kappa_1^4$ , the $B_4=15$ partitions grouped by type. Why: inverting these gives (2); in particular the " $3$ " in $\mu_4=\kappa_4+3\kappa_2^2+\cdots$ counts the three pairings of four points — the disconnected pieces the logarithm removes. *Verified: the partition-sum formulas for $\mu_2,\mu_3,\mu_4$ are checked against the direct cumulant inversion by enumerating all set partitions, and the Bell counts $2,5,15$ confirmed, in verify_cumulants-linked-cluster.py.*

Go deeper · Derivation B · 2 minThe exponential formula: all pieces = exp of connected pieces, and its Mayer face

Derivation BThe exponential formula: all pieces = exp of connected pieces, and its Mayer face

The goal here is to prove the exponential (linked-cluster) formula, $\sum_N Z_N/N!=\exp(\sum_\ell C_\ell/\ell!)$ , and to show it is the same identity as $M=e^K$ , so that applying it to the grand partition function yields $\log\Xi=V\sum_\ell b_\ell z^\ell$ with the cluster integrals in the role of cumulants. The route is to observe that any labelled graph is a disjoint union of connected components, count the ways to assemble a graph from a given inventory of connected pieces, and recognize the resulting sum as an exponential. The point to watch is that "connected" here means what it does for cumulants — a single block — so the Mayer $b_\ell$ , the joint cumulants, and the connected diagrams are literally the same objects in different weights.

Step 1. Decompose a graph into connected components. A graph on $N$ labelled vertices is fixed by a partition of the vertices into blocks and a connected graph on each block; let $C_\ell$ denote the sum (integral) over connected graphs on $\ell$ labelled vertices. Why: connectivity partitions the vertex set uniquely, so summing over graphs = summing over partitions and, within each block, over connected structures.

Step 2. Count assemblies with symmetry factors. A configuration with $m_\ell$ connected pieces of size $\ell$ (so $\sum_\ell \ell m_\ell=N$ ) contributes $\prod_\ell (C_\ell/\ell!)^{m_\ell}/m_\ell!$ after dividing by the interchange of identical pieces. Why: $\ell!$ handles the labels inside each connected block and $m_\ell!$ avoids double-counting identical blocks — exactly the symmetry factors of Step 2 in the previous card.

Step 3. Resum into an exponential. Summing freely over all $\{m_\ell\}$ , $\sum_N Z_N/N!=\sum_{\{m_\ell\}}\prod_\ell\frac{1}{m_\ell!}\big(\tfrac{C_\ell}{\ell!}\big)^{m_\ell}=\prod_\ell\exp\!\big(\tfrac{C_\ell}{\ell!}\big)=\exp\!\big(\sum_\ell \tfrac{C_\ell}{\ell!}\big)$ . Why: each factor is the Taylor series of an exponential; the product over sizes collapses to a single exponential of the connected sum — the exponential formula, identical in form to $M=e^K$ .

Step 4. Take the log and read the Mayer face. Therefore $\log\big(\sum_N Z_N/N!\big)=\sum_\ell C_\ell/\ell!$ — connected pieces only. Applied to $\Xi=\sum_N z^N Q_N/N!$ with the fugacity weight, this is $\log\Xi=V\sum_\ell b_\ell z^\ell$ , so $b_\ell=C_\ell/(\ell!\,V)$ is the connected $\ell$ -cluster: the $\ell$ -th cumulant of the fugacity expansion. Why: the logarithm inverts the exponential of connected pieces, discarding every disconnected product; the same log that turns moments into cumulants turns the grand partition function into the connected-cluster series (owned in full by aq-sm-mayer-cluster-expansion). *Verified: the exp/log inverse $\log(\exp(\text{connected}))=\text{connected}$ , the reproduction of the disconnected $z^2$ term as $b_2+\tfrac12 b_1^2$ (equivalently $\mu_2=\kappa_2+\kappa_1^2$ ), and the Mayer form $\log\Xi=V\sum b_\ell z^\ell$ are checked symbolically in verify_cumulants-linked-cluster.py.*

Limiting cases

The correspondence specializes cleanly, and each limit sharpens the meaning of "connected."

Case	Result	Meaning
Gaussian / ideal	$\kappa_n=0\ (n\ge3)$ ; $f_{ij}\to0$	Only the two-point cumulant survives (Wick's theorem); with no bonds, no clusters, so $\log\Xi=zV$ (ideal gas).
Independent parts	$\kappa_n(X+Y)=\kappa_n(X)+\kappa_n(Y)$	Cumulants add; equivalently $\log\Xi$ of non-interacting subsystems is a sum — no connected cluster spans independent parts.
Poisson	$\kappa_n=\lambda$ for all $n$	Every cumulant equals the rate; a "maximally connected" ladder — the diagrammatic opposite of the Gaussian.
Low density	$\log\Xi/V\approx z+b_2 z^2+\cdots$	The first nontrivial cumulant of the Mayer bond gives $b_2$ , hence $B_2=-b_2$ , the leading virial coefficient.
Large deviations	$\tfrac1N\log\langle e^{tS_N}\rangle\to\lambda(t)$	The scaled cumulant generating function is the Legendre dual of the rate function — the same $\log$ -of-MGF structure governs rare events.

Common trapConnected is not the same as irreducible

A connected cluster is one whose graph cannot be split into disjoint pieces — this is what the logarithm keeps, and it defines the cumulant. An irreducible (biconnected, "star") cluster is a stronger condition: it stays connected even after deleting any single vertex. The linked-cluster theorem for $\log\Xi$ keeps all connected clusters $b_\ell$ ; but the virial coefficients $B_k$ in the density expansion keep only the irreducible ones (the open 3-chain is connected yet reducible, and cancels from $B_3$ ). So there are two successive topological reductions: $\log$ keeps connected, and the fugacity-to-density inversion further keeps irreducible. Reading " $\kappa_n$ = connected" as " $\kappa_n$ = irreducible" is the classic error — the irreducible reduction is a second step, owned by aq-sm-irreducible-cluster-integrals.

Worked examples

1 · The fourth moment of an exponential variable (confidence builder)

We verify $\mu_4=\kappa_4+3\kappa_2^2+\cdots$ concretely for $X\sim\mathrm{Exp}(1)$ (mean $1$ ), where the cumulants are known exactly, to see the connected and disconnected pieces separate.

Step 1. For $\mathrm{Exp}(1)$ the cumulant generating function is $K(t)=-\log(1-t)=\sum_{n\ge1}t^n/n$ , so $\kappa_n=(n-1)!$ . Thus $\kappa_1=1$ , $\kappa_2=1$ , $\kappa_3=2$ , and $\kappa_4=6$ .

Step 2. Assemble the fourth moment from the partition sum: $\mu_4=\kappa_4+4\kappa_3\kappa_1+3\kappa_2^2+6\kappa_2\kappa_1^2+\kappa_1^4=6+4(2)(1)+3(1)^2+6(1)(1)^2+(1)^4$ .

Step 3. Add the pieces: $6+8+3+6+1=\mathbf{24}$ . Directly, $\mu_4=\langle X^4\rangle=4!=24$ for $\mathrm{Exp}(1)$ — they agree. The connected part is $\kappa_4=6$ ; the term $3\kappa_2^2=3$ is the three disconnected pairings that the logarithm removes.

Step 4. The centered fourth cumulant is $\kappa_4=\mu_4^{c}-3(\mu_2^{c})^2$ . Here $\mu_2^c=\mathrm{Var}=1$ and $\mu_4^c=9$ (fourth central moment of $\mathrm{Exp}(1)$ ), giving $\kappa_4=9-3=\mathbf{6}$ — matching Step 1. The $-3\sigma^4$ subtraction is exactly the disconnected-pairing removal, in central-moment language.

2 · Cumulant additivity versus moment non-additivity (exam level)

We show that the fourth cumulant of a sum of independent variables is additive while the fourth central moment is not — the operational signature that cumulants are connected.

Step 1. Let $S=X_1+X_2+X_3$ with independent $X_i\sim\mathrm{Exp}(1)$ . By additivity of cumulants, $\kappa_4(S)=3\kappa_4(X)=3\times6=\mathbf{18}$ , and this scales linearly with the number of terms.

Step 2. The fourth central moment does not add. Using $\mu_4^c(S)=N\mu_4^c(X)+3N(N-1)(\sigma^2)^2$ with $N=3$ , $\mu_4^c(X)=9$ , $\sigma^2=1$ : $\mu_4^c(S)=3(9)+3\cdot3\cdot2\cdot1=27+18=\mathbf{45}$ .

Step 3. Check consistency: $\kappa_4(S)=\mu_4^c(S)-3(\sigma_S^2)^2$ with $\sigma_S^2=N\sigma^2=3$ , so $\kappa_4(S)=45-3(3)^2=45-27=\mathbf{18}$ — matching Step 1. The $3(N-1)$ extra central-moment pieces are precisely the disconnected pairings across distinct variables, which cancel in the cumulant.

Step 4. The physical reading: $\kappa_4(S)\propto N$ is why intensive cumulants (per particle) are finite and extensive quantities have well-behaved fluctuations, while raw central moments grow faster and mix connected and disconnected information. A Monte-Carlo check of $10^6$ samples reproduces $\kappa_4(S)\approx18$ and $\mu_4^c(S)\approx45$ , confirming the split numerically.

Origins · 1 mincumulant generating function and how this was discovered

cumulant generating function

Thiele 1889 (half-invariants); the log-of-a-generating-function that produces the cumulants — and the same object as the free energy

-\beta F=\log Z

· the Taylor coefficients of

K(t)=\log\langle e^{tX}\rangle

are exactly the connected correlators

The cumulant generating function is the logarithm of the moment generating function, $K(t)=\log M(t)=\log\langle e^{tX}\rangle$ , and its power-series coefficients $K(t)=\sum_{n\ge1}\kappa_n\,t^n/n!$ are the cumulants that this article's whole apparatus is built on. Taking the logarithm is precisely the operation that strips the disconnected (product) structure out of the moments and leaves the connected pieces, which is why the linked-cluster theorem, Mayer's $\log\Xi$ , and the connected-diagram rule of field theory are all the same identity viewed through this one function. Its probabilistic seed is Thiele's 1889 "half-invariants" — what we now call cumulants — and its statistical-mechanical incarnation is the free energy itself: $-\beta F=\log Z$ is a cumulant generating function in the coupling, so the mean energy is $\langle H\rangle$ (first cumulant), the heat capacity is the energy variance (second cumulant), and higher response functions are higher cumulants. The additivity $\kappa_n(X+Y)=\kappa_n(X)+\kappa_n(Y)$ for independent variables is nothing but extensivity of $\log Z$ for independent subsystems — the same reason free energies add while partition functions multiply. In every guise the content is one identity: the log of a generating function keeps only the connected content, so cumulants, Mayer clusters, and connected Feynman diagrams are the same mathematics wearing different physical costumes.

History

1889

— Thiele introduces the "half-invariants" (cumulants) as the natural descriptors of a distribution.

1927

— Ursell defines cluster (Ursell) functions, the connected building blocks of the gas expansion.

1937

— Mayer proves the linked-cluster theorem for the classical gas: log of the grand partition function keeps only connected clusters.

1955

— Brueckner formulates the linked-cluster expansion for nuclear matter, removing unlinked-diagram divergences.

1957

— Goldstone gives the diagrammatic proof that the ground-state energy is a sum of connected (linked) diagrams.

Going further

Texts. Kardar, Statistical Physics of Particles, Ch. 2 (cumulants) and Ch. 5–6 (cluster expansion). Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 9–10 (cluster expansion, linked clusters). Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 6 and the cumulant exercises. Landau & Lifshitz, Statistical Physics Part 1, §§74–76. Negele & Orland, Quantum Many-Particle Systems (the linked-cluster theorem and connected diagrams).

Provenance. Base article: AllQFT track `sm`, "Cumulants and the Linked-Cluster Connection" (STAT.INTER.CUMLC.1260); source Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 9–10, and Kardar Ch. 2, 6. All symbolic and numeric claims re-verified here in verify_cumulants-linked-cluster.py (the low-order cumulants from $\log M$ ; the moment = sum-over-set-partitions identity for $n=2,3,4$ with Bell counts $2,5,15$ ; the exp-of-connected inverse and the $b+\tfrac12 a^2$ disconnected term; Gaussian $\kappa_{n\ge3}=0$ and Poisson $\kappa_n=\lambda$ ; cumulant additivity of a sum with a Monte-Carlo cross-check; the Mayer $\log\Xi=V\sum b_\ell z^\ell$ form); no base physics error found.

Prerequisites. Configuration Integral and the Interacting Partition Function · The Mayer f-Function · Mayer Cluster Expansion · Cluster Integrals and Their Combinatorics.

Leads to. Linked-Cluster Theorem · Irreducible Cluster Integrals and the Virial Series · Virial Expansion of the Equation of State · Second Virial Coefficient.

Cumulant, cluster, connected diagram — are these literally the same object, or just analogous?

Literally the same, once you fix the weights. In all three cases you have a generating function that sums over "everything," and its logarithm keeps only the connected part. For a random variable, "everything" is the moments and the connected part is the cumulants. For a gas, "everything" is all Mayer graphs and the connected part is the cluster integrals $b_\ell$ . In field theory, "everything" is all Feynman diagrams and the connected part is the connected diagrams. The combinatorics — sum over set partitions, with the exponential formula relating the full sum to the connected sum — is identical. Only the objects living on the vertices and edges differ (numbers, $f$ -bonds, propagators). That is why the coupled-cluster method in chemistry carries "cluster" in its name: it is Mayer's theorem applied to electronic wavefunctions.

Why does taking a logarithm remove disconnected pieces? It still feels like a trick.

It is a consequence of one fact: independent things multiply, and the logarithm turns products into sums. A disconnected configuration is a product of its independent connected pieces. When you sum over how many copies of each connected piece can appear, you are summing a Taylor series that reassembles an exponential — so the full generating function is $\exp(\text{connected})$ . The logarithm simply inverts that exponential and returns the single connected sum. Nothing is thrown away by hand; the disconnected terms were never independent data, they were just repeated products of the connected ones, and the exponential is the bookkeeping device that packages such repetition.

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

No, and this trips up many people. The third cumulant does equal the third central moment, $\kappa_3=\mu_3^c$ . But the fourth cumulant is $\kappa_4=\mu_4^c-3(\mu_2^c)^2$ — the fourth central moment minus three times the variance squared. The subtraction removes the "Gaussian baseline": a Gaussian has $\mu_4^c=3\sigma^4$ exactly, so its $\kappa_4=0$ . The fourth cumulant therefore measures excess kurtosis, how far the tails depart from Gaussian, whereas the fourth central moment does not vanish even for a Gaussian. The general rule: cumulants equal central moments only up to third order.

How does this connect to the extensivity of the free energy?

Directly. The free energy is $F=-k_BT\log Z$ , and $\log Z$ is a sum over connected clusters by the linked-cluster theorem. Each connected cluster is a local object: after one center-of-mass integration it contributes a volume-independent number times a single factor of $V$ . So $\log Z\propto V$ , hence $F\propto V$ — extensivity. If $\log Z$ had kept disconnected clusters, you would get terms scaling like $V^2$ and worse, and the free energy per particle would not have a thermodynamic limit. Extensivity is not an assumption here; it is a theorem, and its proof is exactly "only connected clusters contribute to $\log Z$ ."

Where do the alternating signs and factorials in the cumulant-from-moments formula come from?

They are the Möbius function of the partition lattice. When you invert " $\mu_n=$ sum over partitions of products of cumulants" to solve for $\kappa_n$ , you are performing Möbius inversion on the poset of set partitions ordered by refinement. The Möbius function of that lattice is $\mu(\hat0,\hat1)=(-1)^{|\pi|-1}(|\pi|-1)!$ for a partition into $|\pi|$ blocks — precisely the coefficient in $\kappa_n=\sum_\pi(-1)^{|\pi|-1}(|\pi|-1)!\prod_B\mu_{|B|}$ . So the signs and factorials are not arbitrary; they are the inverse, on the partition lattice, of the plain sum-over-partitions that builds the moments.

Does the connected-cluster structure survive for quantum systems, or is it classical only?

It survives, and in fact becomes even more central. For quantum gases the configuration integrals are replaced by traces over symmetrized/antisymmetrized states, and the same exponential formula gives the quantum cluster (Kahn–Uhlenbeck) expansion, with connected clusters now carrying exchange as well as interaction. In quantum field theory the theorem is foundational: the generating functional $W[J]=\log Z[J]$ produces connected correlation functions, and the effective action (its Legendre transform) produces one-particle-irreducible ones — two successive reductions exactly parallel to "connected" then "irreducible" in the classical virial story. The logarithm-keeps-connected principle is one of the few statements that reads identically from a dilute classical gas to interacting quantum fields.