General Polynomial Chaos Expansion

• General Polynomial Chaos Expansion (gPCE) is a spectral framework that represents square-integrable random variables using orthonormal multivariate polynomials consistent with their probability laws.
• It employs methods like Gram–Schmidt orthogonalization, whitening transformations, and adaptive quadrature to accurately construct basis functions and estimate expansion coefficients.
• gPCE efficiently addresses high-dimensional challenges through decomposition and sparsity-promoting techniques, facilitating robust uncertainty propagation, statistical analysis, and sensitivity computation.

General Polynomial Chaos Expansion (gPCE) is a rigorous spectral framework for representing square-integrable random variables and processes in terms of orthonormal multivariate polynomials that are constructed to be consistent with the input probability law, including arbitrary, possibly dependent, random vectors. It generalizes classical polynomial chaos (Wiener–Hermite) beyond the product measures and enables efficient uncertainty propagation, statistical analysis, and sensitivity computation for complex stochastic models in computational science and engineering.

1. Mathematical Foundation and Basis Construction

At its core, gPCE represents a random variable or model response $Y(\xi)$, where $\xi=(\xi_1,\ldots,\xi_N)$ is a random vector with joint distribution $\mu$ (density $f(\xi)$ when it exists), as a (possibly infinite) expansion: $$Y(\xi) = \sum_{\alpha \in \mathcal{A}} c_\alpha \Psi_\alpha(\xi)$$ where:

• $\mathcal{A} \subseteq \mathbb{N}_0^N$ is a (finite or infinite) multi-index set, often chosen as $\{|\alpha| \leq p\}$ for total degree $p$,
• $\Psi_\alpha$ are multivariate polynomials orthonormal in $L^2(\mu)$:

$\xi=(\xi_1,\ldots,\xi_N)$0

• $\xi=(\xi_1,\ldots,\xi_N)$1 are the expansion coefficients.

Basis construction methods depend crucially on the law of $\xi=(\xi_1,\ldots,\xi_N)$2:

• For independent inputs with known marginals (e.g., Gaussians, uniforms), $\xi=(\xi_1,\ldots,\xi_N)$3 can be built as tensor products of univariate orthonormal polynomials (Hermite, Legendre, etc.) (Mühlpfordt et al., 2020, Fagiano et al., 2012).
• For dependent or non-product laws, explicit tensorization fails; one constructs measure-consistent orthogonal polynomials via:
  • Gram–Schmidt orthogonalization of the monomials with respect to $\xi=(\xi_1,\ldots,\xi_N)$4 (Navarro et al., 2014),
  • Whitening transformations based on the moment Gram matrix and Cholesky factorization (Lee et al., 2022, Choi et al., 26 Oct 2025),
  • Recurrences for the multivariate Hermite family in dependent Gaussians (Rahman, 2017).

This ensures completeness: under mild conditions, the multivariate orthogonal polynomials form a complete basis in $\xi=(\xi_1,\ldots,\xi_N)$5, guaranteeing mean-square convergence of the truncated expansion (Rahman, 2017, Breden, 2022).

2. Numerical Construction and Quadrature

The practical calculation of expansion coefficients and basis functions requires quadrature or sampling schemes adapted to the input distribution:

• Stieltjes or three-term recurrences for building univariate bases from arbitrary densities (Mühlpfordt et al., 2020, Zhang et al., 2014).
• Tensorized or sparse-grid quadrature for high-dimensional integration, leveraging recurrence-generated nodes and weights (Mühlpfordt et al., 2020).
• For arbitrary or data-driven inputs, kernel density estimation (KDE) and sample-based moment estimation provide plug-in surrogates for Gram matrices and basis construction (Choi et al., 26 Oct 2025).
• Christoffel Sparse Approximation (CSA) leverages equilibrium measures and Christoffel function preconditioning in sampling with $\xi=(\xi_1,\ldots,\xi_N)$6-minimization to recover sparse expansions robustly with fewer samples (Jakeman et al., 2016).

A summary of core computational approaches is provided in the table below:

Law of $\xi=(\xi_1,\ldots,\xi_N)$7	Basis Construction	Coefficient Estimation
Product	Tensor-product Askey	Quadrature, least-squares, Galerkin
Arbitrary	Gram–Schmidt, Whitening	Projection, regression, convex/CSA
Data-driven	KDE + Whitening	Regression, projection

3. Expansion Properties, Moments, and Statistics

For a given truncated expansion $\xi=(\xi_1,\ldots,\xi_N)$8: $\xi=(\xi_1,\ldots,\xi_N)$9 the key statistical quantities follow directly from orthonormality:

• Mean: $\mu$0,
• Variance: $\mu$1 (in fully orthonormalized bases),
• For Gaussian and other coupled bases (with nontrivial off-diagonal Gram blocks), $\mu$2 (Rahman, 2017).

Higher moments (skewness, kurtosis, etc.) are given by summing the appropriate products of expansion coefficients and higher-order moment (linearization) coefficients; analytic and quadrature-free formulas are available for classical continuous orthogonal polynomials (Savin et al., 2016).

Sensitivity indices generalize to the correlated-input setting using the PCE coefficients and joint moments, allowing for decomposition of total variance into orthogonal, covariate, and conditional contributions (Navarro et al., 2014).

4. Computation in High Dimensions and Sparsity Enhancements

The curse of dimensionality in gPCE arises from the exponential growth of multi-index sets. Mitigation strategies include:

• Dimensionally Decomposed gPCE (DD-GPCE): restricts the expansion to interaction orders $\mu$3 (i.e., terms involving at most $\mu$4 variables), dramatically reducing the basis size (Choi et al., 26 Oct 2025, Lee et al., 2022).
• Active subspace and basis adaptation: constructs rotated variables concentrating most dependence into a low-dimensional subspace, followed by univariate or low-dimensional PC expansion (Tsilifis, 2016).
• Sparsity-promoting regularization: $\mu$5-norm penalty or weighted regression in coefficient estimation exploits compressibility of typical physical systems (Fagiano et al., 2012, Jakeman et al., 2016, Yang et al., 2017).
• Reduced basis methods: further accelerate non-intrusive gPCE by replacing repeated high-fidelity solves with surrogates, especially for PDEs with low Kolmogorov width solution manifolds (Jiang et al., 2016).

Dimension-independent best $\mu$6-term convergence rates are obtained for $\mu$7-holomorphic maps with $\mu$8-summable coefficient decay: error $\mu$9 in $f(\xi)$0 for $f(\xi)$1, independent of nominal input dimension (Aftab et al., 2 Jun 2025).

5. Extensions: Dependent Inputs, Data-Driven Bases, and Applications

gPCE is fully defined for dependent, possibly strongly correlated inputs:

• Multivariate Hermite (Wiener–Hermite) expansions for general normal laws require block-diagonalization at each total order due to weak orthogonality and involve blockwise linear system solution for the coefficients (Rahman, 2017).
• Measure-consistent polynomial bases for arbitrary joint laws entail Gram–Schmidt on monomials or whitening transforms of the sample moment matrix; these approaches enable sensitivity analysis and risk estimation for correlated, non-Gaussian settings (Lee et al., 2022, Choi et al., 26 Oct 2025).

gPCE is also utilized in:

• Propagation of uncertainty in dynamical systems, including stiff and multi-scale ODEs and Hamiltonian/chaotic regimes; for Hamiltonian systems, the expansion coefficients obey an induced Hamiltonian ODE system (Pasini et al., 2013).
• Chance-constrained optimization and risk analysis in power systems, where gPCE is embedded into the nonlinear constraints and enables fast, non-sample-based, enforcement of operational chance constraints (Mohy-ud-din et al., 26 Sep 2025).
• A posteriori validation and branch continuation: fully computer-assisted proof of parameter-dependent and stochastic-invariant sets in differential equations, using validated Newton–Kantorovich-type radii polynomial bounds in the gPCE coefficient Banach algebra (Breden, 2022).

6. Algorithmic and Software Ecosystem

Modern algorithmic workflows span:

• Symbolic and numerical Gram matrix computation (e.g., via quadrature or sample moments),
• Monomial vector assembly and Cholesky/Lanczos whitening,
• Coefficient estimation via regression, matched quadrature, or convex optimization embedding constraints for mean, variance, and physical bounds (Fagiano et al., 2012, Yang et al., 2017, Choi et al., 26 Oct 2025),
• Sparse-grid or preconditioned compressed-sensing sampling for underdetermined settings (Jakeman et al., 2016),
• Rigorous error bounds and a posteriori validation for solutions to parametric PDEs/ODEs in the gPCE framework (Breden, 2022).

Robust open-source implementations for gPCE are available in scientific computing languages, including PolyChaos.jl in Julia, which provides modular types for measures, basis polynomials, quadrature, tensorization, and surrogate construction (Mühlpfordt et al., 2020).

7. Limitations, Open Problems, and Best Practices

• Truncation error: Spectral convergence is guaranteed for analytic responses, but truncation in degree, index set, or interaction order can produce bias in capturing nonlinearity or rare-event statistics (e.g., for highly chaotic or stiff systems) (Pasini et al., 2013).
• Non-orthogonality in transformations: Rotated variable-based adaptivity or basis adaptation may destroy explicit orthogonality except in Gaussian cases; numerical loss of orthogonality requires careful management for stable regression (Yang et al., 2017).
• Sampling considerations: The choice of quadrature or sampling measure, regularization parameters, and equilibrium densities is critical, especially in underdetermined or high-order expansions (Jakeman et al., 2016, Fagiano et al., 2012).
• Empirical basis construction for unknown distributions: Data-driven whitening and KDE-based basis formation are effective when sufficient samples are available, but may be unstable or biased in very high-dimensional settings with limited data (Choi et al., 26 Oct 2025).
• Interpretation of sensitivity: Sobol' indices and other variance-based decompositions require extended definitions in the correlated case and must be interpreted with care (Navarro et al., 2014).

Best practices thus include adaptive determination of degree and index sets, exploitation of compressibility, dimensional decomposition, careful sampling or quadrature, and validation via a posteriori error bounds or surrogate Monte Carlo as needed.

---

For comprehensive expositions and rigorous algorithmic formulations, see (Navarro et al., 2014, Rahman, 2017, Lee et al., 2022, Choi et al., 26 Oct 2025, Aftab et al., 2 Jun 2025, Mühlpfordt et al., 2020, Jakeman et al., 2016).