Stochastic Polynomial Chaos Expansions

• Stochastic Polynomial Chaos Expansions are spectral representations that extend classical polynomial chaos to cases where outputs are inherently stochastic.
• They offer both intrusive and non-intrusive formulations, applicable to quantum dynamics, reliability analysis, fragility modeling, and stochastic optimization.
• Adaptive sparse techniques and orthogonal polynomial bases are employed to mitigate the curse of dimensionality and enhance convergence efficiency.

Searching arXiv for recent and foundational papers on stochastic polynomial chaos expansions. {"^^^^^^^^4query4 polynomial chaos expansions\"^^^^^^^^4 OR ti:\4^^^^^^^^"stochastic polynomial chaos\"^^^^^^^^4 OR abs:\4^^^^^^^^"stochastic polynomial chaos\"","max_results":^^^^4all:\4query4^^^^,"sort_by":"submittedDate","sort_order":"descending"} Relevant arXiv records include recent work on active learning for stochastic reliability (&&&^^^^4query4^^^^&&&), stochastic-simulator emulation (&&&^^^^4all:\4^^^^&&&), seismic fragility analysis (&&&^^^^4 OR ti:\4^^^^&&&), stochastic optimization (&&&^^^^4 OR abs:\4^^^^&&&), and earlier intrusive Galerkin and quantum-dynamics formulations (Young et al., 2012, Slim et al., 2017). Stochastic Polynomial Chaos Expansions (SPCE) are spectral representations that extend polynomial chaos methodology to settings in which the quantity of interest is itself stochastic, either because a dynamical system is driven by random inputs or because a simulator returns a random output for fixed physical parameters. In intrusive formulations, the random state is expanded in orthogonal polynomials of underlying random variables and projected to a deterministic hierarchy for chaos coefficients. In non-intrusive distributional formulations, the target is the conditional law PRESERVED_PLACEHOLDER_^^^^4query4^^^^, represented on an augmented stochastic space that may include an artificial latent variable and an additive Gaussian regularization term. Under these interpretations, SPCE has been developed for stochastic quantum dynamics, stochastic evolution equations, stochastic simulators, reliability analysis, fragility modeling, stochastic optimization, and stochastic control (Young et al., 2012, &&&^^^^4all:\4^^^^&&&, &&&^^^^4 OR abs:\4^^^^&&&).

^^^^4all:\4^^^^. Conceptual foundations and measure-adapted bases

The starting point is the polynomial chaos expansion of a square-integrable random quantity in a basis orthonormal with respect to the input probability measure. Representative formulations in the literature include

PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^

with tensor-product structure for independent inputs,

PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^

The basis is chosen from the underlying measure: Gaussian variables are paired with Hermite polynomials, uniform variables with Legendre polynomials, and the Askey-scheme correspondences also include Laguerre and Jacobi families for gamma and beta laws. In optimization and intrusive stochastic Galerkin settings alike, orthogonality gives direct access to low-order moments and underlies the reduction of stochastic problems to deterministic algebraic, ODE, or PDE systems (&&&^^^^4 OR abs:\4^^^^&&&, &&&^^^^4all:\4query4^^^^&&&).

A central technical distinction is between independent and dependent inputs. For dependent random variables, tensor-product bases are no longer orthogonal to the true joint law. Three approaches are explicitly compared in the literature: mapping methods such as Rosenblatt and Nataf transformations, domination methods based on a simpler tensor-product measure whose support dominates the true support, and numerical Gram–Schmidt orthogonalization (GSO) on the original dependent measure. The reported conclusion is that mapping can distort the target function and degrade approximation quality, domination methods can be suboptimal because they orthogonalize with respect to the wrong measure, and GSO combined with weighted Leja sequences yields interpolants that are often orders of magnitude more accurate in the correct PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4^^^^ sense (&&&^^^^4all:\4all:\4^^^^&&&).

SPCE also generalizes beyond Euclidean random variables. For circular quantities on the unit circle, orthogonal polynomials on the unit circle are required. Rogers–Szegő polynomials provide the wrapped-normal basis, while more general circular densities such as the von Mises law can be handled by numerically generating orthogonal polynomials from the characteristic function via Toeplitz determinants and Szegő recurrences. This unit-circle formulation preserves periodicity and supports direct estimation of circular moments, mean direction, and circular standard deviation from expansion coefficients (&&&^^^^4all:\4 OR ti:\4^^^^&&&).

^^^^4 OR ti:\4^^^^. Intrusive stochastic Galerkin formulations

In intrusive SPCE, the random solution itself is expanded in a chaos basis and substituted into the governing equations. A prototypical example is the simulation of quantum systems driven by classical Gaussian noise. There the stochastic process is first compressed by a Karhunen–Loève expansion,

$$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$

and the density matrix is expanded as

$$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$

Galerkin projection onto a multivariate Hermite basis yields a sparsely coupled hierarchy of deterministic linear ODEs. The stochastic dimension $S$ is selected not merely by retaining the largest KLE eigenvalues, but by a physics-based ranking through cumulative transition rates $\Gamma_n$ derived from time-dependent perturbation theory. For truncation order $P$, the number of coupled equations is

$N=\frac{(S+P)!}{S!\,P!},$

which makes the curse of dimensionality explicit. In the one-qubit dephasing example PRESERVED_PLACEHOLDER_^^^^4all:\4query4^^^^ with Ornstein–Uhlenbeck noise and PRESERVED_PLACEHOLDER_^^^^4all:\4all:\4^^^^, Monte Carlo required approximately PRESERVED_PLACEHOLDER_^^^^4all:\4 OR ti:\4^^^^ iterations to converge, whereas the most accurate SPCE result shown, with PRESERVED_PLACEHOLDER_^^^^4all:\4 OR abs:\4^^^^ and PRESERVED_PLACEHOLDER_^^^^4all:\44^^^^, required solving only PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^5 coupled equations and ran about PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^6 times faster (Young et al., 2012).

Closely related intrusive constructions appear in stochastic evolution equations. For parabolic problems with uniformly distributed inputs, Wiener–Legendre chaos is combined with deterministic operator splitting. The stochastic solution is expanded as

PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^7

then truncated on

PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^8

with PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^9 retained modes. Each chaos coefficient satisfies a deterministic evolution equation, which is advanced by resolvent Lie splitting or trapezoidal resolvent splitting. The total error separates into a chaos truncation term and a time-discretization term, with first-order convergence in time for Lie splitting and second-order convergence for trapezoidal splitting (&&&^^^^4all:\44^^^^&&&).

Particle, spin, and interface dynamics lead to more structured intrusive systems. In RF Wien-filter beam and spin simulations, Galerkin projection introduces sparse third- and fourth-order tensors

PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4query4^^^^

which govern the deterministic coefficient system for positions, velocities, spins, and field components. In random level-set dynamics, generalized polynomial chaos is applied to the hyperbolic formulation of the level-set equation. The conservative Galerkin system is not automatically hyperbolic, whereas the modified capacity form is proven strongly hyperbolic and leads to a finite-volume scheme tailored to uncertain velocities. This distinction is substantive rather than cosmetic: the conservative form can exhibit complex characteristic speeds, while the capacity formulation retains a real spectrum under the stated positivity assumptions (Slim et al., 2017, &&&^^^^4all:\4^^^^6&&&).

^^^^4 OR abs:\4^^^^. Distributional SPCE for stochastic simulators

A second major interpretation of SPCE targets stochastic simulators, namely models for which the output remains random even when the physical inputs are fixed. In this setting the object of approximation is not a deterministic map PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4all:\4^^^^, but the full conditional distribution PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4 OR ti:\4^^^^. The central construction introduces an artificial latent variable PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4 OR abs:\4^^^^ and additive Gaussian noise PRESERVED_PLACEHOLDER_^^^^4 OR ti:\44^^^^, yielding

PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^5

The latent-variable term provides stochasticity at fixed PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^6; the Gaussian term regularizes the likelihood because the density induced by the latent polynomial alone may be singular and unbounded. Coefficients are estimated by maximum likelihood, with the integral over the latent variable approximated by Gaussian quadrature. The noise level PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^7 is selected by cross-validation rather than joint optimization, and the implementation uses a warm-start continuation from larger to smaller PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^8 values. In the reported case studies, this SPCE formulation accurately represents non-Gaussian and even bimodal output distributions, and generally outperforms both the generalized lambda model and kernel conditional density estimation (&&&^^^^4all:\4^^^^&&&).

In fragility analysis under stochastic ground-motion models, the same distributional idea is specialized to positive structural response quantities by applying SPCE to the logarithm of the engineering demand parameter: PRESERVED_PLACEHOLDER_^^^^4 OR ti:\4^^^^9 The fragility function is then

PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4query4^^^^

The model is calibrated by likelihood-based estimation with one-dimensional quadrature over PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4all:\4^^^^, and cross-validation selects both PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4 OR ti:\4^^^^ and the truncation set. An explicit bridge to classical cloud analysis is given: if only linear terms are retained, SPCE reduces to the classical linear cloud model. In the reported Bouc–Wen and OpenSees case studies, SPCE provides the best overall conditional-distribution and fragility estimates once the training set is moderately large, while linear models can remain competitive only in small-data regimes and probit models deteriorate when failures are rare (&&&^^^^4 OR ti:\4^^^^&&&).

A further extension handles stochastic models with both parametric uncertainty and intrinsic noise, including random fields. There the intrinsic randomness is represented through a Rosenblatt transformation to i.i.d. variables PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4 OR abs:\4^^^^, producing a stochastic PCE at each parameter setting,

PRESERVED_PLACEHOLDER_^^^^4 OR abs:\44^^^^

followed by a second PCE of the coefficient maps PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4^^^^5 over the parametric variables. For random fields, the output is first compressed by a Karhunen–Loève expansion and then each KLE coefficient is given a joint PCE, yielding a KLPC surrogate that remains generative and admits closed-form Sobol indices. In the catalysis example, this construction quantifies the contribution of intrinsic noise to overall variance and shows that parametric uncertainty dominates the variance in the reported setting (&&&^^^^4all:\4^^^^9&&&).

4. Sparse, adaptive, and high-dimensional construction

The dominant algorithmic challenge in SPCE is basis growth. Combinatorial counts such as PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4^^^^6 or PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4^^^^7 appear repeatedly across intrusive and non-intrusive settings, and motivate sparsity-promoting constructions. One route is weighted PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4^^^^8-minimization for approximately sparse polynomial chaos coefficients. Given a measurement matrix PRESERVED_PLACEHOLDER_^^^^4 OR abs:\4^^^^9 from non-adapted random samples, the coefficient vector is recovered by

$$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$^^^^4query4^^^^

with smaller weights for coefficients expected a priori to be large and larger weights for coefficients expected to be small. The theoretical guarantees are formulated through restricted isometry and null-space conditions, including the sufficient condition $$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$^^^^4all:\4^^^^. Numerical results for elliptic and cavity-flow problems show that weighted $$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$^^^^4 OR ti:\4^^^^ is especially advantageous when the sample budget is small and useful coefficient-decay information is available (&&&^^^^4 OR ti:\4query4^^^^&&&).

A complementary route is adaptive least-squares construction with sequential experimental design. In this framework the polynomial index set is kept downward-closed, admissible neighbors are appended greedily according to coefficient energy, and the data set is enriched only when the least-squares problem becomes insufficiently stable. Stability is monitored through criteria derived from the normalized information matrix $$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$^^^^4 OR abs:\4^^^^, notably $$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$4, $$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$5, and $$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$6. Relaxed criteria can improve average accuracy, while strict criteria reduce variability across experimental designs; in the reported high-frequency electromagnetic examples, the thresholds

$$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$7

provide a favorable accuracy–robustness trade-off (&&&^^^^4 OR ti:\4all:\4^^^^&&&).

Interpolation-based adaptivity yields a different sparse paradigm. Leja sequences furnish nested collocation nodes, one polynomial term per collocation point, and exact interpolation on the selected nodes. A weighted Leja rule

$$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$8

extends the construction to arbitrary continuous input densities. Because each Leja node is associated with a unique hierarchical Newton-like polynomial and hence with a unique orthogonal polynomial degree, the interpolating PCE can be computed directly and adaptively. In the reported experiments, these interpolating PCEs are consistently more accurate than pseudo-spectral projection and least-squares regression at equal model-evaluation budgets, while retaining strong performance for moment and Sobol-index estimation (&&&^^^^4 OR ti:\4 OR ti:\4^^^^&&&).

5. Hybrid, reduced-order, and decision-oriented extensions

SPCE is frequently embedded in larger surrogate architectures. One common pattern is reduced-order modeling. In POD-PCE, a stochastic field is first written as

$$\Omega(t)=\sum_{n=1}^\infty \sqrt{\lambda_n} g_n(t)\xi_n,$$9

where POD compresses the physical dimension, and then each stochastic amplitude $$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$^^^^4query4^^^^ is approximated by a PCE. The combined surrogate

$$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$^^^^4all:\4^^^^

delivers mean and covariance formulas directly from the coefficients while avoiding a full PCE at every spatial degree of freedom. In the reported examples, it preserves statistical accuracy comparable to full PCE while substantially reducing storage and computational requirements (&&&^^^^4 OR ti:\4 OR abs:\4^^^^&&&).

For stochastic dynamical systems, regular time-frozen PCEs can degrade quickly in time because they do not encode causality or memory. PC-NARX addresses this by modeling the dynamics with a NARX structure and then expanding the stochastic NARX coefficients by sparse PCE: $$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$^^^^4 OR ti:\4^^^^ Least angle regression is used both to select the NARX regressors and to build the sparse PCEs for their coefficients. In quarter-car, Duffing, and Bouc–Wen examples, this hybridization markedly outperforms regular time-frozen PCEs, especially for long-time trajectory prediction and maxima (&&&^^^^4 OR ti:\44^^^^&&&).

Frequency-domain dynamics motivate yet another hybridization. For frequency response functions, direct PCE struggles with resonance sharpness and frequency shifts across realizations. A stochastic frequency transformation first aligns resonant and antiresonant frequencies by piecewise-linear scaling, then sparse adaptive PCE is applied to the transformed FRFs and to the selected frequencies themselves, with PCA reducing the large output dimension. The reported ^^^^4 OR ti:\4^^^^-DOF and 6-DOF studies show accurate recovery of individual FRFs and faster convergence of FRF moments than Monte Carlo, although very low damping exposes sensitivity to frequency resolution and interpolation (&&&^^^^4 OR ti:\4^^^^5&&&).

Decision-oriented variants use SPCE to reformulate optimization and control. In stochastic optimization, the optimizer itself is expanded,

$$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$^^^^4 OR abs:\4^^^^

so that the stochastic program reduces to deterministic optimization over chaos coefficients, with the mean given by the zeroth coefficient and the standard deviation by the Euclidean norm of the nonconstant coefficients. Convexity of the reduced objective is preserved when the original objective is convex in $$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$4. In stochastic nonlinear MPC, Hermite polynomial chaos is embedded as the GP mean function in a GPPCE surrogate, allowing closed-form mean and variance estimates for nonlinear objectives and constraints and a chance-constrained formulation via Chebyshev-based tightening; in the reported batch-reactor study, the GPPCE-SNMPC controller showed no constraint violations in $$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$5 Monte Carlo closed-loop simulations (&&&^^^^4 OR abs:\4^^^^&&&, &&&^^^^4 OR ti:\4^^^^7&&&).

6. Reliability analysis, representative applications, and limitations

Recent work has pushed SPCE into reliability analysis for nondeterministic models. The stochastic limit-state function is written as $$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$6, and failure is expressed through the conditional failure probability

$$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$7

SPCE supplies a conditional density and hence $$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$8; active learning then enriches the design by maximizing

$$\rho(t;\vec\xi)=\sum_{n=0}^\infty \phi_n(t)\Phi_n(\vec\xi).$$9

The variance term is approximated by sampling SPCE coefficients from the asymptotic Gaussian law induced by the Fisher information of the maximum-likelihood estimator. In the stochastic $S$^^^^4query4^^^^-SS^^^^4all:\4^^^^ function, AL-SPCE reaches accurate estimates with as few as $S$^^^^4 OR ti:\4^^^^–SS^^^^4 OR abs:\4^^^^ points; at $S$4, the median estimate is $S$5 with CoV $S$6, close to the analytical value $S$7. In the wind-turbine dataset, the method achieves a median $S$8 at $S$9, close to the reference $\Gamma_n$^^^^4query4^^^^, and outperforms both static-design SPCE and direct Monte Carlo in efficiency (&&&^^^^4query4^^^^&&&).

The application range is correspondingly broad. SPCE has been used for one-qubit dephasing under long-correlated classical noise, stochastic parabolic evolution equations, beam and spin transport in an RF Wien filter, stochastic high-frequency electromagnetics, stochastic simulators in finance and epidemiology, PBEE fragility analysis, seismic and epidemic reliability, orbit-state uncertainty with angular variables on the unit circle, catalytic random fields, and universal stochastic kriging with sparse PCE trends (Young et al., 2012, &&&^^^^4all:\44^^^^&&&, Slim et al., 2017, &&&^^^^4all:\4^^^^&&&, &&&^^^^4 OR ti:\4^^^^&&&, &&&^^^^4all:\4 OR ti:\4^^^^&&&, &&&^^^^4 OR abs:\4^^^^5&&&).

Several limitations recur across this literature. First, SPCE is not a single algorithm but a family of intrusive, non-intrusive, latent-variable, and hybrid constructions, so comparisons are meaningful only within a fixed problem class. Second, the curse of dimensionality remains fundamental: even sparse or adaptive variants rely on favorable coefficient decay, low effective interaction order, or low stochastic dimension. Third, intrusive truncations are typically uncontrolled and require systematic convergence checks in the retained stochastic modes and polynomial order; in the quantum setting, positivity of the density matrix is not mathematically guaranteed under truncation, even though no violations were observed in the reported experiments. Fourth, dependence handling is delicate: transforming dependent variables to independent ones can degrade accuracy, rather than improve it. Fifth, direct PCE can fail for long-time dynamics or misaligned FRFs, which explains the emergence of PC-NARX and stochastic frequency transformation. Finally, for latent-variable SPCEs of stochastic simulators, the asymptotic and statistical properties of the maximum-likelihood estimator are not yet fully studied, high-dimensional sparse structure beyond the mean function remains an open direction, and active-learning convergence curves remain noisy enough that rigorous stopping criteria are still difficult to formulate (Young et al., 2012, &&&^^^^4all:\4all:\4^^^^&&&, &&&^^^^4 OR ti:\44^^^^&&&, &&&^^^^4 OR ti:\4^^^^5&&&, &&&^^^^4all:\4^^^^&&&, &&&^^^^4query4^^^^&&&).