Kolmogorov-Nagumo Mean: Definition & Applications

• Kolmogorov-Nagumo mean is a quasi-arithmetic mean defined via a continuous, strictly monotonic generator function, encapsulating various averages such as arithmetic, geometric, and harmonic means.
• It unifies classical measures under Kolmogorov’s axioms, ensuring properties like continuity, symmetry, and associativity which guarantee the existence of an inverse for the generator function.
• The method extends to weighted means and finds applications in statistical mechanics, information theory, and neural modeling, offering insights into entropy, vulnerability metrics, and equilibrium systems.

The Kolmogorov-Nagumo mean, also called the quasi-arithmetic mean or "regular mean" (Editor's term), is defined for a finite tuple $(x_1,\dots,x_n)$ as $$M_{\phi}(x_1,\dots,x_n) = \phi^{-1}( \frac{1}{n}\sum_{i=1}^n \phi(x_i) )$$, where $\phi$ is a continuous, strictly monotonic generator function with a well-defined inverse. This framework subsumes the arithmetic, geometric, harmonic, and power means as particular choices of $\phi$, and characterizes every symmetric, reflexive, monotonic, associative mean under Kolmogorov’s axioms. The Kolmogorov-Nagumo construction is central in applications ranging from statistical mechanics, information theory, and quantitative information flow to multifractal analysis and neural modeling.

1. Axiomatic Foundations and Representation

The Kolmogorov-Nagumo mean is uniquely determined by Kolmogorov’s axioms:

• Continuity and Monotonicity: The mean is continuous in each coordinate and strictly increasing.
• Symmetry: Invariance under permutation of inputs.
• Idempotence: $M(x,\dots,x) = x$.
• Replacement-invariance (Associativity): If $m = M(x_1,\dots,x_{n_0})$ for $1 \le n_0 < n$, then $$M(x_1,\dots,x_n) = M(m, ..., m, x_{n_0+1}, ..., x_n)$$.

These properties imply that any such mean must be expressible as $$M_f(x_1,\dots,x_n) = f^{-1}(\frac{1}{n}\sum_i f(x_i))$$ for a continuous, strictly monotone generator $f$ (Carvalho, 14 Jan 2026). Monotonicity guarantees $f^{-1}$ exists, ensuring the functional representation is well-defined (Singpurwalla et al., 2020).

2. Classical Examples and Structural Properties

Special cases arise by choosing $\phi$ appropriately (Singpurwalla et al., 2020, Carvalho, 14 Jan 2026):

• Arithmetic mean: $\phi(x) = x$ yields $M(x_1,\dots,x_n) = \frac{1}{n}\sum x_i$.
• Geometric mean: $\phi(x) = \ln x$ on $(0,\infty)$ gives $M(x_1,\dots,x_n) = (\prod x_i)^{1/n}$.
• Harmonic mean: $\phi(x) = 1/x$, $M(x_1,\dots,x_n) = n/(\sum x_i^{-1})$.
• Power mean: $\phi(x) = x^p$, $M(x_1,\dots,x_n) = (\frac{1}{n}\sum x_i^p)^{1/p}$.
• Exponential-log mean: $\phi(x) = e^x$, $M(x_1,\dots,x_n) = \ln(\frac{1}{n}\sum e^{x_i})$.

Means outside the Kolmogorov-Nagumo class (e.g., the median) fail the associativity property and cannot be represented in this framework (Singpurwalla et al., 2020). The regular mean class is continuously stable under small perturbations of the generator function: small changes in $f$ yield small changes in $M_f$, as quantified by uniform continuity in Lipschitz spaces (Carvalho, 14 Jan 2026).

3. Connections to Functional Equations and Superposition Theorem

Kolmogorov’s superposition theorem states that any continuous function of several variables can be written as a sum of compositions of continuous univariate and outer functions. Imposing Kolmogorov-Nagumo axioms collapses this structure into the form $$M_f(x_1,\dots,x_n) = f^{-1}(\frac{1}{n}\sum f(x_i))$$ (Carvalho, 14 Jan 2026). This illustrates a deep link between the theory of regular means and neural architectures: a two-layer feed-forward network with identical hidden activations $f$ and output activation $f^{-1}$ computes a Kolmogorov-Nagumo mean.

4. Extension to Weighted Means and Statistical Applications

The Kolmogorov-Nagumo mean generalizes to weighted averages with positive weights $w_i$, $\sum w_i = 1$: $M_f(x_1,\dots,x_n; w) = f^{-1}(\sum_i w_i f(x_i))$ (Scarfone et al., 2022, Zarrabian et al., 2024). In probability theory and statistical mechanics, weights typically correspond to probability masses $p_i$, and expectations become nonlinear under this framework.

The universal central limit theorem applies to Kolmogorov-Nagumo means: for i.i.d. $X_i$ and generator $f$ with finite variance, $\sqrt{n}(M_f(X_1, ..., X_n) - \mu_f)$ converges in distribution to a normal law $N(0, \sigma^2)$, where $\mu_f = f^{-1}( E[f(X)] )$ (Carvalho, 14 Jan 2026). Rate of convergence depends on the third cumulant of $f(X)$, with heavier tails and skewness decelerating convergence.

5. Kolmogorov-Nagumo Means in Information Theory and Quantitative Information Flow

Within quantitative information flow (QIF), Kolmogorov-Nagumo means underpin generalized notions of vulnerability and information leakage (Zarrabian et al., 2024). The $f$-mean framework defines a one-parameter family of vulnerabilities and leakages via the choice of generator $f$:

• Prior vulnerability: $$V_{f,g}(\pi) = \sup_{w \in \mathcal W} f^{-1}(\sum_x \pi_x f(g(w,x)))$$.
• Posterior vulnerability: $$V_{f,g}(\delta_y) = \sup_w f^{-1}(\sum_x \delta_x^y f(g(w,x)))$$.
• Leakage measures: Additive and multiplicative leakages are constructed based on differences or ratios of $f$-means.

Arimoto mutual information, Sibson mutual information, and Rényi divergence arise as special cases by suitable selection of $f$, $h$, and gain functions. The framework unifies various entropy and privacy measures including min-entropy, Shannon entropy, maximal $\alpha$-leakage, and local Rényi differential privacy.

6. Nonlinear Averages in Thermodynamics and Statistical Mechanics

Kolmogorov-Nagumo means generalize expectations in thermodynamic formalisms (Morales et al., 2023, Scarfone et al., 2022). For exponential generators $f(x) = e^{\gamma x}$, the mean reduces to $$\langle X \rangle_\gamma = \frac{1}{\gamma}\ln \sum_i p_i e^{\gamma x_i}$$, interpolating between arithmetic ($\gamma \to 0$) and more nonlinear averages. The corresponding entropy becomes Rényi entropy; canonical equilibrium is still governed by the Boltzmann distribution.

Kolmogorov-Nagumo-based averages maintain the Legendre transform structure of thermodynamics, preserving conjugate relations between entropy, thermodynamic variables, and Lagrange multipliers (Scarfone et al., 2022). In non-equilibrium settings, generalized second laws and H-theorems hold; γ-thermodynamic length and information-geometric quantities (Fisher-Rao metric) are determined by the Bregman divergence associated to the chosen $f$ (Morales et al., 2023).

7. Implications, Limitations, and Theoretical Significance

The Kolmogorov-Nagumo mean formalism is mathematically elegant: it uniquely characterizes the most general class of means satisfying symmetry, associativity, continuity, and monotonicity. It provides a unifying structure linking classical averages, statistical functionals, entropy measures, privacy criteria, equilibria in statistical mechanics, and learning architectures via functional equations and superposition theorems.

However, the abstraction is limited to means that satisfy the associativity/replacement-invariance axiom; statistics like the median, trimmed mean, or mode fall outside this class (Singpurwalla et al., 2020). The choice of generator has significant impact on robustness, sensitivity, and the operational interpretation of the mean in applications, as demonstrated by the stability and central limit results (Carvalho, 14 Jan 2026).

Kolmogorov-Nagumo means continue to inform contemporary research in information theory, learning, physics, and privacy, offering foundational insight into the functional-equation perspective underlying aggregation and expectation in mathematics and the physical sciences.