Cumulant-Generating Function (CGF)

• CGF is defined as the logarithm of the moment-generating function, providing a complete description of distribution cumulants and facilitating statistical inference.
• The CGF exhibits convexity and smoothness, enabling derivation of moments, large-deviation rate functions, and tight exponential tail bounds.
• CGF plays a pivotal role across fields—from statistical mechanics to quantum many-body physics and coding theory—bridging theoretical insights with practical applications.

A cumulant-generating function (CGF) is a central object in probability theory, statistics, information theory, statistical mechanics, quantum many-body physics, signal processing, and machine learning. Defined as the logarithm of the moment-generating function (MGF), the CGF provides a complete characterization of a distribution’s cumulants, unifies moment-based and large-deviation asymptotics, and underpins statistical inference, coding theory, and many-body field theory. Beyond classical applications, CGFs govern the operational behavior of estimators, encode fluctuation theorems in nonequilibrium physics, and provide a bridge to convex duality in information-theoretic bounds.

1. Mathematical Definition and Core Properties

Let $X$ be a real-valued random variable. The moment-generating function is $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$ whenever the expectation exists in a neighborhood of $\lambda=0$. The cumulant-generating function is

$$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$

with domain $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$, always containing $0$. The CGF inherits several fundamental properties:

• Convexity: $\Lambda_X(\lambda)$ is convex by Hölder's inequality; $$\frac{\partial^2}{\partial \lambda^2} \Lambda_X(\lambda) = \operatorname{Var}[e^{\lambda X} X]/\mathbb{E}[e^{\lambda X}] \geq 0$$ (Hellström et al., 2023).
• Smoothness: On the interior of its domain, $\Lambda_X$ is $C^\infty$; the $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$0-th derivative at $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$1 yields the $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$2-th cumulant $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$3:

$M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$4

In particular, - $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$5, - $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$6, - $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$7 and higher cumulants encode skewness, kurtosis, etc. (Guruacharya et al., 2016, Guruacharya et al., 2015, Rodríguez et al., 2014).

• Relationship with MGF and Moments: The MGF $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$8 and the CGF $M_X(\lambda) = \mathbb{E}[e^{\lambda X}]$9 encode the same information, but the latter has additivity under independence and direct access to cumulants through its Taylor expansion.

In the multivariate case, for $\lambda=0$0 and $\lambda=0$1,

$\lambda=0$2

and mixed partial derivatives at $\lambda=0$3 yield joint cumulants (Rodríguez et al., 2014, Moor et al., 2024).

2. CGF in Statistical Theory and Large Deviations

The CGF underpins exponential inequalities, tail bounds, and large deviations by its connection with the Legendre–Fenchel (convex conjugate) transform. The supremum

$\lambda=0$4

is the Cramér rate function or convex conjugate, governing large-deviation probabilities of empirical means via the Gärtner–Ellis theorem (Hellström et al., 2023). For example, PAC-Bayesian and information-theoretic generalization bounds express discrepancies between empirical and population quantities in terms of the CGF and its conjugate, recovering classical inequalities (Hoeffding, Bernstein, Catoni, Maurer–Langford–Seeger) and yielding optimally tight bounds for various exponential family tail behaviors:

• Bernoulli: $\lambda=0$5, conjugate is binary KL-divergence.
• Sub-Gaussian: $\lambda=0$6.
• Sub-Poisson, sub-gamma, sub-Laplacian, and other NEFs are directly captured (Hellström et al., 2023).

The CGF framework unifies the derivation, optimality, and logarithmic correction structure in high-probability bounds and generalization error theory, with the tightest comparator always being the convex conjugate of a bounding distribution’s CGF.

3. Applications in Coding Theory, Information Theory, and Data Compression

In universal coding and lossy compression, CGFs are essential for analyzing codeword length fluctuations and excess distortion events. For a code $\lambda=0$7 and source random variable $\lambda=0$8, the normalized CGF of codeword length is

$\lambda=0$9

with $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$0 recovering mean length and $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$1 the maximum length (Saito et al., 2018). In variable-length lossy source coding allowing positive excess distortion probability, exact achievability and converse bounds relate the CGF directly to Rényi entropy of order $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$2: $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$3 with $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$4 the minimal $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$5-Rényi entropy under admissible reproduction distributions (Saito et al., 2018).

In the individual sequence (finite-state encoder) regime, empirical CGFs analyze worst-case and fluctuation-sensitive code length complexity, relating to empirical Rényi entropy in the fixed-to-variable setting, and to Lempel-Ziv finite-state compressibility in the variable-to-variable regime (Merhav, 2016).

4. CGF in Quantum Many-Body Systems

In the analysis of retarded Green’s functions, particularly within quantum field theory and ab initio GW approximations, the CGF appears as the exponent in diagrammatic expansions. For a retarded Green function $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$6,

$$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$7

with $$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$8 the cumulant generating function, explicitly given by

$$\Lambda_X(\lambda) = \log \mathbb{E}[e^{\lambda X}]$$9

This representation resums classes of diagrams to infinite order and relates the cumulant to the improper Dyson self-energy via convolution identities (Mayers et al., 2016). The CGF approach preserves critical spectral moments, yields refined quasiparticle properties, and underlies the treatment of satellite structure and correlation energies in electron gases.

In the theory of geometric polarization and quantum phase transitions, generating functions such as generalized Bargmann invariants act as quantum CGFs. Their cumulants yield moments and higher fluctuations, enabling the construction of geometric Binder cumulants highly sensitive to gap closings and criticality in the system (Hetényi, 7 Apr 2026).

5. CGF-Based Inference and Saddle-Point Techniques

CGFs form the foundation for saddle-point approximations in statistical inference, offering computational and theoretical advantages in tail and likelihood computations. The saddle-point equation

$$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$0

locates the point of maximum likelihood for an extreme event. The Lugannani–Rice formula expresses tail probabilities

$$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$1

where $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$2, $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$3, and $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$4, $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$5 are the standard normal CDF and PDF. This method applies to outage probability calculations in wireless communications (Nakagami-m, Rician, Hoyt fading) and is validated against direct integration methods while offering superior computational efficiency and sharpness for moderate-to-rare events (Guruacharya et al., 2015, Guruacharya et al., 2016).

In frequency-domain analysis of time series, the Legendre transform of the CGF of the Whittle likelihood score underpins density approximations and hypothesis testing, extending traditional empirical likelihood and bootstrap methodologies while maintaining higher-order accuracy in small samples (Moor et al., 2024).

6. Multivariate and High-Dimensional Generalizations

The CGF’s multivariate framework enables joint cumulant extraction and low-dimensional modeling of complex dependence structures in high-dimensional networks. Mixed partial derivatives of $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$6 at $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$7 directly yield joint cumulants, while parametric models such as nested expansions in quadratic forms (generalizing Gaussians) allow parsimonious but flexible representation and inference of higher-order interactions (Rodríguez et al., 2014). Model fitting via “cumulant matching” aligns theoretical cumulants with empirical “interaction manifestations” (e.g., quantiles or entropy of sums), extending the method-of-moments rationale and anchoring statistical mechanics, network modeling, and statistical learning.

In the context of the Erdős–Rényi random graph, CGFs (known as pressure functions) characterize the large-deviation behavior of subgraph counts (e.g., triangles). The infinite-volume limit reduces to a variational problem over graphon space, with symmetry-breaking phenomena (replica phase transitions) mapped directly via the CGF functional (Giardinà et al., 2020).

7. CGF in Nonparametric Bayesian Analysis and Beyond

In Bayesian nonparametrics, the CGF for Dirichlet processes has recently been bounded non-asymptotically using superadditivity arguments, Fekete’s lemma, and large-deviation theory. The finite-$$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$8 CGF is bounded by the convex conjugate of the scaled reversed KL divergence: $$\operatorname{Dom} \Lambda_X = \{\lambda \in \mathbb{R}: \mathbb{E}[e^{\lambda X}] < \infty\}$$9 bridging pointwise concentration with large-deviation asymptotics and enabling sharp confidence regions for functionals of sums of independent Dirichlet processes (Perrault et al., 2024). Practical computation of these bounds often reduces to convex optimization, even in high-dimensional or feature-rich settings, with direct consequences for uncertainty quantification and regret bounds in nonparametric Thompson sampling and related algorithms.

---

Key References by Application Domain

Domain	Paper Title & arXiv ID	Primary CGF Role
Generalization Bounds	"Comparing Comparators..." (Hellström et al., 2023)	PAC-Bayes, convex conjugate bounds
Quantum Green's Functions	"Description of quasiparticle..." (Mayers et al., 2016)	Cumulant ansatz in GW/MBPT
Coding Theory	"Cumulant Generating Function..." (Saito et al., 2018); "On empirical cumulant..." (Merhav, 2016)	Lossy coding, compressibility
Statistical Inference	"On the use of the cumulant generating function..." (Moor et al., 2024)	Saddlepoint, density approximation
Stochastic Geometry	"Analysis of SINR..." (Guruacharya et al., 2016); "Saddle Point Approximation..." (Guruacharya et al., 2015)	Interference, outage probability
Multivariate Statistics	"Multivariate interactions modeling..." (Rodríguez et al., 2014)	Beyond-pairwise dependence
Nonparametric Bayesian	"A New Bound on the Cumulant Generating Function..." (Perrault et al., 2024)	DP mean functional, superadditivity

The cumulant-generating function is a mathematically rigorous and versatile tool, providing direct access to cumulants, controlling exponential moments, and underpinning advanced methodology across disciplines. Its operational role in convex analysis, duality, large deviations, and risk-sensitive analysis makes it indispensable for precision modeling, inference, and uncertainty quantification in modern theoretical and applied research.