Theory of Random Functions Overview

• Theory of random functions is a framework analyzing probability measures on infinite-dimensional spaces, emphasizing Gaussian processes and invariance properties.
• It applies to machine learning regression, random analytic function zero-statistics, and fractal singular functions through robust spectral and statistical methods.
• The study leverages spectral analysis, algorithmic randomness, and Choquet integration to uncover universal laws and computational complexity relationships.

The theory of random functions investigates probability measures on spaces of functions, typically infinite-dimensional, and analyzes their structural, geometric, analytic, and statistical properties. This field provides foundational frameworks for understanding Gaussian processes, random analytic functions, algorithmic randomness on function spaces, approximation theory under capacity, spectral identities, and the emergence of universal laws from random inputs. It interfaces closely with stochastic process theory, functional analysis, mathematical statistics, ergodic theory, random matrix theory, and effective/integrable computability.

1. Infinite-Dimensional Function Spaces and Gaussian Random Functions

A canonical setting is the space of real-valued continuous functions on $\mathbb{R}^n$, $F = C(\mathbb{R}^n)$. A random function is formalized as a probability measure $P$ on $(F, \sigma\text{-algebra})$, frequently assumed to be a (centered) Gaussian measure. For any finite collection $\{x_1,\ldots,x_k\} \subset \mathbb{R}^n$, the random vector $(f(x_1),\ldots,f(x_k))$ is multivariate Gaussian, determined by a covariance kernel $K(x,x') = \mathbb{E}_P[f(x)f(x')]$.

Central to this framework are symmetry postulates: translation, rotation, and scale invariance, combined with the requirement of Gaussianity. These constraints uniquely specify the structure of admissible random functions. For instance, translation and rotation invariance render $K(x,x') = k_f(x-x')$ with $k_f$ radial, while scale invariance requires $k_f(cu) = c^\alpha k_f(u)$ for a specific scaling exponent $\alpha$.

Spectral analysis translates these symmetries into the spectral domain: only power-law spectral densities $S(\|\omega\|) = A\|\omega\|^{-(n+\alpha)}$ are permitted, with $\alpha=2$ yielding the unique generalized polyharmonic spline (thin-plate spline) kernel $K(x,x') = \|x-x'\|^2\ln\|x-x'\|$. This construction is canonical and emerges solely from the indifference principle—no other kernel can satisfy the required invariances and Gaussianity (Bakhvalov, 14 Dec 2025).

2. Random Analytic Functions and Universality

Random analytic functions, including random entire and random polynomial models, instantiate probability measures on spaces of holomorphic functions, typically via random coefficients in power or orthogonal expansions. For example, $$f_\omega(z) = \sum_{j=0}^\infty \chi_j(\omega)a_jz^j$$ with i.i.d. complex-valued random variables $\chi_j$ (including Gaussian, Rademacher, or Steinhaus types), yields families exhibiting highly regular statistical properties (Li et al., 2020, Starr, 2011).

Universality phenomena are prominent: local statistics of zeros, critical values, and maximum modulus display asymptotics independent of the precise law of the coefficients, governed instead by symmetry, covariance structure, and regularity of the function basis. For Gaussian analytic functions (GAFs), zeros form determinantal point processes, and all non-Gaussian models with matching covariance exhibit identical fine-scale zero correlations near boundaries (Starr, 2011).

A general framework (Nguyen et al., 2017) provides universality theorems relating the root statistics of diverse families—random trigonometric polynomials, Kac/Weyl/elliptic polynomials, Weyl series, random Taylor series—to those of the corresponding Gaussian models via Green's representation and Lindeberg replacement. This theory yields precise results for mean zero density, local statistics, and scaling limits.

3. Algorithmic Randomness, Effective Capacities, and Computable Random Functions

Random functions also appear in algorithmic randomness and computable analysis through probabilistic measures on spaces of (partial) continuous functions, particularly $2^\omega \to 2^\omega$, induced by Bernoulli measures on labeled trees or codes (Cenzer et al., 2015). The notion of "online random functions" (bitwise, prefix-dependent output) provides a paradigm for algorithmic randomness in function space with important structural consequences: such functions are neither surjective nor injective; their ranges are disjoint from computable reals and their fibers coincide with random closed sets.

Generalizations extend to "online partial random functions" with delay/partiality, parameterized by computable effective capacities (hit probabilities for clopen sets), admitting fine-grained control over the resulting structure of random closed sets. The close relation to Martin-Löf randomness and Kolmogorov complexity yields complexity bounds on elements of the range, tightly linked to the capacity structure (Cenzer et al., 2015).

4. Machine Learning, Variational Principles, and Optimal Estimation

The theory underpins variational and Bayesian methods in machine learning and regression. For data $(x_i, y_i)$ modeled as $y_i = f(x_i) + \varepsilon_i$, with $f$ drawn from a Gaussian process prior determined by optimal symmetry, the maximum a posteriori estimate reduces to minimizing a penalized least-squares functional $$\frac{1}{\sigma^2}\sum_{i}(y_i - f(x_i))^2 + \|f\|^2_{\mathcal{H}_K}$$ over the relevant RKHS. The noise parameter $\sigma^2$ acts as the natural Tikhonov regularizer and arises intrinsically from the probabilistic model.

The resulting interpolant or smoother is the unique minimizer that adherently fits data consistent with the noninformative, symmetry-respecting Gaussian prior. The solution kernel (thin-plate spline) is not a heuristic or empirical choice but a rigorous consequence of the maximum entropy principle under the full group of invariances (Bakhvalov, 14 Dec 2025).

5. Spectral Equivalence, Operator Methods, and Laws of Quadratic Forms

Spectral theory supplies deep equivalence relations for Gaussian random functions defined via covariance operators in Hilbert spaces (Nazarov et al., 2020). Spectral equivalence—equality of the nonzero eigenvalues of the respective covariance operators—guarantees equality in law for quadratic functionals (e.g., $L^2$-norms, integrated squares, V-statistics for goodness-of-fit tests). The operator calculus (composition, tensor products, centering, integration) enables the derivation of broad classes of identities in distribution, including pinned vs. centered path processes, multidimensional Brownian sheets, and integrated bridges.

This methodology also provides a unified lens for understanding null distributional laws in nonparametric statistics and for tracing algebraic (operator-theoretic) sources of equivalence underlying apparently disparate models (Nazarov et al., 2020).

6. Random Iteration, Self-Similarity, and Singular Distributions

Random iteration of deterministic or parametrized function families, under random symbolic sequences, is central to constructing self-similar (often singular) limit functions. For families of strictly increasing maps and nondegenerate Bernoulli measures on symbolic codes, random iteration almost surely leads to divergence to $+\infty$ or $-\infty$ for each initial point; the distribution of the exit direction satisfies functional equations encoding self-similarity (Mitrea et al., 22 Aug 2025). Classical fractal singular functions—the Cantor function, Lebesgue singular functions, Minkowski's question-mark—arise as solutions to these equations under suitable tuples of maps and weights.

This random-iteration viewpoint generalizes to multiple maps, vector-valued iterations, and non-affine (possibly nonlinear) dynamics, providing a probabilistic and ergodic-theoretic mechanism for constructing and analyzing singular measures and distributions.

7. Approximation, Capacity, and Nonlinear Integration of Random Functions

Random functions defined on measurable spaces with capacities (not necessarily additive measures) generalize classical probabilistic integration and approximation. The Choquet integral (relative to a submodular capacity) is the central apparatus; it continues to satisfy many desirable properties with appropriate modifications. Multivariate random Bernstein polynomial approximations provide quantitative rates of uniform convergence both in Choquet-mean and in capacity norm, determined by modulus of continuity and optimal constants (Gal et al., 2020). The theory bridges classical deterministic approximation theorems (e.g., Bernstein’s theorem) and modern uncertainty quantification where probability may be non-additive.

Applications range from stochastic approximation and robust statistics to random field modeling under ambiguity, where standard probabilistic assumptions are insufficient or unavailable.

8. Arithmetic Characterizations: Random Functions in Complexity Theory

In computational complexity, random (probabilistic) functions are characterized axiomatically via bounded arithmetic theories augmented with probabilistic constructs. The theory RS$^1_2$ provides a language and a set of axioms such that the $\Sigma^b_1$-representable functions are exactly those computed by polynomial-time probabilistic Turing machines (Antonelli et al., 2023). The central algebra of random functions, POR, abstracts all poly-time random function constructions (including random bit queries, composition, and bounded recursion), mirroring the role played by Cobham's algebra in deterministic PTIME.

This arithmetic foundation enables translation between probabilistic machine models, algebraic function definitions, and logical proof-theoretic statements concerning the class of random functions, and provides a basis for further study of randomness in computational logic.

---

References

• (Bakhvalov, 14 Dec 2025) Solving a Machine Learning Regression Problem Based on the Theory of Random Functions
• (Li et al., 2020) Inequalities Concerning Maximum Modulus and Zeros of Random Entire Functions
• (Starr, 2011) Universality of Correlations for Random Analytic Functions
• (Nguyen et al., 2017) Roots of random functions: A framework for local universality
• (Cenzer et al., 2015) Algorithmically random functions and effective capacities
• (Mitrea et al., 22 Aug 2025) Singular functions obtained via random function iteration
• (Nazarov et al., 2020) Spectral equivalence of Gaussian random functions: operator approach
• (Gal et al., 2020) Approximation of Random Functions by Random Polynomials in the Framework of Choquet's Theory of Integration
• (Antonelli et al., 2023) An Arithmetic Theory for the Poly-Time Random Functions