Wiener Chaos Expansion Overview

• Wiener Chaos Expansion is an infinite series representation of square-integrable functionals of Gaussian processes using orthogonal Hermite polynomials for spectral decomposition.
• It decouples randomness from dynamic propagators in systems like SDEs and SPDEs, allowing efficient computation of deterministic chaos coefficients via methods such as Monte Carlo simulation.
• The technique supports advanced applications including optimal prediction, parameter estimation, and surrogate modeling with rigorous L² error analysis and spectral convergence guarantees.

A Wiener Chaos Expansion (WCE) is an infinite series representation of square-integrable functionals of Gaussian processes in terms of orthogonal polynomials—typically probabilists’ Hermite polynomials—of isonormal Gaussian modes. This framework provides a canonical spectral decomposition for generalized stochastic processes, forming the mathematical foundation for both theoretical and computational stochastic analysis. WCEs are central to Malliavin calculus, numerical resolution of stochastic differential and partial differential equations, non-linear prediction, and advanced data-driven surrogate modeling of stochastic systems.

1. Mathematical Foundation and Formalism

The WCE is constructed for an isonormal Gaussian process $X$ over a Hilbert space $\mathcal H$. For $F \in L^2(\Omega)$ measurable w.r.t.\ the σ-algebra generated by $X$, the expansion

$F = \sum_{n=0}^\infty I_n(f_n)$

holds, where $I_n$ denotes the $n$-fold symmetric multiple Wiener–Itô integral and $f_n \in \mathcal H^{\odot n}$ are uniquely determined symmetric kernel functions. The chaos spaces $\mathscr W^n := I_n(\mathcal H^{\odot n})$ are mutually orthogonal in $L^2(\Omega)$: $\mathcal H$0 For finite-dimensional Gaussian vectors $\mathcal H$1 with arbitrary covariance, the expansion is in terms of generalized multivariate Hermite polynomials, which remain complete and weakly orthogonal even under dependence (Rahman, 2017).

This chaos decomposition enables a Parseval-type isometry: $\mathcal H$2 guaranteeing $\mathcal H$3-convergence and uniqueness (Briand et al., 2012). In practical implementations, the expansion is truncated in both order and the number of basis functions, with precise error estimates available (Huschto et al., 2019).

2. Chaos Expansions for Gaussian-Driven Systems

For stochastic differential equations (SDEs) or stochastic partial differential equations (SPDEs) driven by Gaussian noise, WCE separates the randomness from the dynamical propagators. The solution $\mathcal H$4 to a stochastic PDE under Gaussian stimulation can be written as

$\mathcal H$5

where $\mathcal H$6 is a tensor product of Hermite polynomials in the Gaussian modes $\mathcal H$7, and $\mathcal H$8 are deterministic "chaos coefficients" or "propagators" (Nijimbere, 2019, Ji et al., 2 Aug 2025, Neufeld et al., 2024, Ferrucci et al., 2022).

The stochastic evolution equations induce a triangular coupled system for the coefficients $\mathcal H$9, which can be derived via projection and Malliavin calculus duality: $F \in L^2(\Omega)$0 with explicit combinatorial coefficients obtained through chaos product rules (i.e., via Wick products and Hermite polynomial identities) (Nijimbere, 2019, Coriasco et al., 2023). Generic $F \in L^2(\Omega)$1 error bounds for order truncation $F \in L^2(\Omega)$2 have a form

$F \in L^2(\Omega)$3

with geometric decay for analytic input data (Ji et al., 2 Aug 2025, Nijimbere, 2019).

3. Structural and Algorithmic Aspects

Basis Selection and Orthogonality: Adaptation of the basis (via Hilbert space rotation) can concentrate system variance into fewer leading chaos modes, facilitating low-dimensional surrogates. This approach is effective for random fields and partial differential equations (Tsilifis et al., 2016) and is implemented by constructing a smooth family of orthonormal matrices $F \in L^2(\Omega)$4 that rotates the coordinate system of the underlying Gaussian space.

Truncation and Monte Carlo Estimation: Finite chaos expansions are computed for a maximum order $F \in L^2(\Omega)$5 and number of temporal/spectral basis elements $F \in L^2(\Omega)$6. The computation of the coefficients is highly parallelizable: the coefficients are given by scalar $F \in L^2(\Omega)$7 inner products and can be estimated via independent Monte Carlo simulations (Lelong, 2019, Delgado-Vences et al., 2024). For path-dependent and high-dimensional cases, this bypasses the curse of dimensionality present in regression-based algorithms.

Propagation Systems and Neural Surrogates: The infinite hierarchy of deterministic ODEs for the chaos coefficients ("propagators") can be truncated and solved by standard deterministic solvers or approximated via neural networks. Neural operator architectures with FiLM layers and Wick–Hermite feature conditioning have proven effective for data-driven surrogate modeling even in singular SPDE regimes (Neufeld et al., 2024, Shi et al., 9 Mar 2026).

4. Applications and Numerical Schemes

Stochastic Evolution Problems

• Stochastic PDEs: WCE transforms SPDEs to deterministic PDE systems for the coefficients, allowing for high-accuracy computation of moments and probability distributions (Nijimbere, 2019, Ji et al., 2 Aug 2025).
• BSDEs and Variants with Jumps: Combined use of Wiener and Charlier chaos permits handling of jump diffusions. WCE-based Picard schemes provide closed-form updates for the conditional expectations and fully parallel coefficient computation, with explicit error analysis in all approximation parameters (Briand et al., 2012, Geiss et al., 2015).

Inverse and Inference Problems

• Parameter Estimation in SDEs: WCE maps the stochastic inverse problem to deterministic optimization of the finite chaos propagator system, with Tikhonov regularization enforcing correct mean and variance matching. Stochastic gradient descent using analytic gradients from the chaos system provides rapid and precise parameter recovery (Delgado-Vences et al., 27 Mar 2026).

Prediction and Filtering

• Optimal Prediction: For Gaussian processes (including stationary-increment and fBm), constructing a prediction-adapted basis in Fock space allows direct computation of the Wiener chaos expansion of conditional expectations given past data (Alpay et al., 2014).

Bridge Simulation

• Diffusion Bridges: A truncated Fourier–Hermite expansion of a guided SDE allows exact satisfaction of endpoint constraints with no rejection rate. The chaos coefficients are computed by numerically solving a deterministic triangular ODE system for each basis index (Delgado-Vences et al., 2024).

Fractional and Generalized Chaos

• Fractional Wiener Chaos: Via power-normalized parabolic cylinder functions, a fractional generalization of Hermite polynomials produces a new orthogonal basis for heavy-tailed or fractional stochastic analysis, preserving martingale and orthogonality properties (Boguslavskaya et al., 2023).
• Non-Gaussian Extensions: Variants may use alternative families (e.g., Charlier polynomials for Poissonian components) or multidimensional polynomials when extending to non-Gaussian or dependent input spaces (Rahman, 2017, Geiss et al., 2015).

5. Error Analysis and Theoretical Guarantees

WCE admits detailed a priori $F \in L^2(\Omega)$8 error bounds for all sources of approximation:

• Chaos order $F \in L^2(\Omega)$9: factorial or algebraic decay in $X$0 under regularity assumptions (Huschto et al., 2019, Nijimbere, 2019).
• Basis truncation $X$1: decay governed by the smoothness of stochastic input, and the quadrature error in spectral projection (Huschto et al., 2019).
• Monte Carlo coefficient error: variance $X$2, independent over multi-indices (Lelong, 2019, Geiss et al., 2015).
• Dimension reduction: adapted chaos bases can capture $X$3 of variance using orders of magnitude fewer terms for elliptic PDEs (Tsilifis et al., 2016).

These estimates yield spectral convergence rates for analytic data and stability as all parameters are increased. For SPDEs, error in moments is directly computable from the chaos coefficients (Nijimbere, 2019, Ji et al., 2 Aug 2025). In the context of SDE parameter estimation, matching empirical moments with chaos moments regularizes the inference problem and provides provable convexity of loss (Delgado-Vences et al., 27 Mar 2026).

6. Advances, Limitations, and Extensions

Dimension, Adaptivity, and Curse of Chaos: The principal computational challenge is the combinatorial growth in the number of coefficients as chaos order and dimensionality increase (the "curse of chaos"). Adapted or reduced chaos representations by basis rotation, and dimension reduction via variance concentration, are critical advances for high-dimensional problems (Tsilifis et al., 2016). Sparse-grid and adaptive truncation can further alleviate this bottleneck (Nijimbere, 2019).

Singular SPDEs and Neural Operators: For equations like dynamical $X$4, incorporating Wick powers through chaos-featured neural operators with FiLM conditioning (WCE–FiLM–NO) offers state-of-the-art predictions—recovering correlation functions and renormalized limits—without explicit access to renormalization counterterms (Shi et al., 9 Mar 2026).

Non-Gaussian and Jump Processes: Integrated Hermite–Charlier bases enable extension to Lévy-driven and Poissonian systems, with forward Picard schemes providing efficient parallelizable solvers (Geiss et al., 2015).

Fractional and Heavy-Tail Analysis: Fractional chaos expansions promise orthogonal series for wider classes of functionals, including power-law distributions and fractional diffusions (Boguslavskaya et al., 2023).

Outlook: WCE forms the analytic backbone for both canonical infinite-dimensional stochastic analysis and scalable computational methods for high-dimensional, path-dependent, and non-linear stochastic systems. Its adaptability—under Gaussian and non-Gaussian noise, with advanced basis construction and neural surrogates—continues to support advances across stochastic modeling, inference, and uncertainty quantification.

---

Key references:

• “On the Wiener Chaos Expansion of the Signature of a Gaussian Process” (Ferrucci et al., 2022)
• “Simulation of BSDEs by Wiener chaos expansion” (Briand et al., 2012)
• “Wiener-Chaos Approach to Optimal Prediction” (Alpay et al., 2014)
• “Reduced Wiener Chaos representation of random fields via basis adaptation and projection” (Tsilifis et al., 2016)
• “Fractional Wiener Chaos” (Boguslavskaya et al., 2023)
• “Pricing path-dependent Bermudan options using Wiener chaos expansion” (Lelong, 2019)
• “The asymptotic error of chaos expansion approximations for stochastic differential equations” (Huschto et al., 2019)
• “Solving stochastic partial differential equations using neural networks in the Wiener chaos expansion” (Neufeld et al., 2024)
• “Wiener Chaos Expansion based Neural Operator for Singular SPDEs” (Shi et al., 9 Mar 2026)