In many physical and mathematical systems, we encounter processes that are neither purely deterministic nor simple white noise. When a non-linear system is driven by random fluctuations—such as a wave in a noisy plasma or a financial market with "long memory"—the resulting behavior becomes difficult to track. Standard tools like the Central Limit Theorem often fail because the data points are not independent; they exhibit long-range dependence where the past continues to influence the present across vast scales.
To manage this complexity, researchers use Wiener chaos expansion (and its generalizations). This technique acts like a Fourier series for random variables, decomposing a complex stochastic process into a sum of simpler, orthogonal components. By breaking a random signal into these "building blocks," we can isolate how non-linearity and noise interact to produce macroscopic patterns.
The Intuition of Orthogonal Decomposition
The core idea is to represent a random variable or process as a series of orthogonal polynomials. For systems driven by Gaussian noise, these are typically Hermite polynomials. As Grinberg et al. demonstrate in the context of the stochastic Korteweg-de Vries equation, this allows us to decompose stochastic input into a sum of deterministic functions multiplied by random variables.
This separation is powerful because of orthogonality. In a Hermite expansion, each term represents a different "order" of randomness. Because these terms are orthogonal with respect to the underlying Gaussian measure, their expected values are zero for any term of order one or higher. Khandelwal et al. leverage this property to solve a common problem in non-linear modeling: simulation bias. By transforming non-linear models into a Hermite basis, the authors can extract an unbiased deterministic response simply by isolating the zero-order term, effectively filtering out the stochastic noise that usually skews results.
Identifying the Hermite Rank
When dealing with non-linear transformations of Gaussian processes, the "shape" of the transformation determines the system's long-term behavior. This is characterized by the Hermite rank—the index of the first non-vanishing term in the chaos expansion.
The rank acts as a structural parameter that dictates the eventual distribution of the system. Clausel et al. show that the Hermite rank determines whether a non-linear time series will eventually settle into a standard Gaussian distribution or converge to more "exotic" non-Gaussian limits, such as the Rosenblatt process. As Anh et al. argue, when variables exhibit long-range dependence, the standard bell curve of the Central Limit Theorem is replaced by these Hermite-type distributions. The chaos expansion allows researchers to track exactly how these functionals stabilize as the observation area increases.
Bridging Microscopic Noise and Macroscopic Limits
Chaos expansions provide a rigorous bridge between microscopic fluctuations and macroscopic equations. In the study of heat-like parabolic equations with highly oscillatory potentials, Bal uses orthogonal expansions to show how long-range correlations in the potential lead to the emergence of multiplicative noise in the macroscopic limit.
Similarly, for singular stochastic partial differential equations (SPDEs) like the 2D Parabolic Anderson Model, Huang uses Wiener chaos to represent the solution as an infinite sum of orthogonal stochastic integrals. This provides a "one-shot" way to understand the statistical properties of the system, identifying specific kernels for each order of the expansion that correspond to physical paths through a random environment.
Recent advances have even brought this framework into deep learning. Shi et al. developed a Neural Operator that solves SPDEs by disentangling random noise from deterministic dynamics using Wick-Hermite polynomials. This allows the model to learn the underlying "propagator" of the system, achieving high accuracy in fluid dynamics and image generation by treating the stochastic forcing as a structured expansion rather than an opaque source of error.
Beyond the Gaussian Assumption
While Hermite polynomials are the standard for Gaussian noise, the principle of chaotic decomposition extends to other measures. Nourdin et al. explore Walsh chaos expansion for Rademacher functionals, bridging Malliavin calculus with Stein’s method to calculate error bounds for non-linear functionals.
However, one must be careful when moving away from Gaussianity. Djah et al. point out that "graphical Wick ordering"—the systematic removal of self-contraction loops in Feynman diagrams—is only equivalent to orthogonal polynomial decomposition in the Gaussian case. For non-Gaussian measures, these two approaches diverge, suggesting that the geometric intuition of "removing loops" does not always align with the algebraic intuition of "orthogonal components."
Key Foundational Work
Feynman graph representation of the perturbation series for general functional measures (2004) — contributed 1 concept elements to this direction.
Numerical studies of the stochastic Korteweg-de Vries equation (2006) — contributed 1 concept elements to this direction.