Adaptive Stochastic Spectral Embedding (ASSE)

• Adaptive Stochastic Spectral Embedding (ASSE) is a hierarchical surrogate modeling method that constructs piecewise spectral representations using local polynomial-chaos expansions.
• It employs data-driven, recursive partitioning based on error and variance indicators to concentrate sampling in regions with nonlinearity or localization.
• ASSE is applied in uncertainty quantification, reliability assessment, Bayesian inference, and graph clustering, delivering improved accuracy and computational efficiency over global approaches.

Adaptive Stochastic Spectral Embedding (ASSE) is a family of hierarchical surrogate modeling techniques that construct piecewise spectral (typically polynomial-chaos) representations of high-dimensional mappings. ASSE frameworks achieve robust and efficient approximation, uncertainty quantification, rare-event estimation, and clustering in settings where global spectral surrogates (such as classical PCE) fail due to nonlinearity, localization, or multi-modality. The core mechanism is recursive, data-driven partitioning of the input space, embedding local residual expansions only where dictated by adaptive error criteria. ASSE has been deployed in uncertainty quantification, reliability assessment, Bayesian inference, and, via its scalable graph analytic cores, attributed graph clustering.

1. Theoretical Foundations

ASSE builds upon stochastic spectral embedding (SSE), which decomposes a square-integrable function $Y = \mathcal M(X)$, $X \in \mathbb R^M$, as a sum over hierarchical local polynomial expansions,

$$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$

with $\mathcal D^k$ denoting a sequence of adaptively-defined subdomains. Within each $\mathcal D^k$, $Y$ is approximated by a truncated (sparse) polynomial chaos expansion (PCE),

$$\widehat R_S^k(X) = \sum_{j \in \mathcal A^k} a_j^k \Psi_j^k(X),$$

where $\{\Psi_j^k\}$ are orthonormal polynomials with respect to the local measure restricted to $\mathcal D^k$. The global expansion is constructed recursively by computing the residuals at each level and partitioning the domain where error criteria or variance indicators are maximized.

In adaptive (greedy) splitting, the refinement indicators—such as weighted leave-one-out error or local estimator variance—ensure that refinement is concentrated in regions where the global surrogate is least accurate, potentially in small-measure, high-importance subsets (as in rare-event domains or sharply-peaked likelihoods) (Marelli et al., 2020, Wagner et al., 2021, Wagner et al., 2020).

2. Algorithmic Structure and Adaptivity

The principal ASSE algorithm iterates over the following logical steps:

1. Initialization: Build a low-degree global PCE over the full domain.
2. Residual Computation: Evaluate the residual between the target mapping and its current hierarchical surrogate.
3. Error/Variance Estimation: Compute a local refinement indicator, for example,

$$E^{k} = (\text{measure of } \mathcal D^k) \times (\varepsilon_{\mathrm{LOO}}^k)$$

where $X \in \mathbb R^M$0 is the leave-one-out cross-validation error for the local surrogate (Marelli et al., 2020, Wagner et al., 2020).

1. Partition Selection: Select subdomain $X \in \mathbb R^M$1 to refine, maximizing the error indicator.
2. Domain Splitting: Partition $X \in \mathbb R^M$2 along the coordinate (commonly by highest first-order Sobol' index or variance criterion) yielding two child domains that balance statistical mass or misclassification.
3. Enrichment: For each child domain, generate new sample points (possibly focusing on high-error or high-variance regions), recompute local surrogates and refinement indicators.
4. Stopping: Halt when maximal local error indicators drop below tolerance, sample budget is reached, or leaf domains become too small for robust local regression.

This methodology ensures the construction remains linear in sample size $X \in \mathbb R^M$3 under reasonable sampling policies, as the total number of cell expansions is $X \in \mathbb R^M$4 for minimum leaf sample size $X \in \mathbb R^M$5 (Marelli et al., 2020).

3. Implementation Variants and Application Domains

Surrogate Modeling and UQ

In computational uncertainty quantification, ASSE delivers piecewise-PCE surrogates with dramatically improved accuracy and data efficiency relative to traditional global or sparse PCE, especially on highly nonlinear and localized systems. The approach enables analytic computation of statistical observables (moments, quantiles) and supports hyperbolic truncation and sparse regression (e.g., via LARS) within each local expansion. Used in probabilistic optimal power flow (OPF), ASSE outperforms conventional surrogates for responses exhibiting strong local structure or heavy tails, achieving mean errors of order $X \in \mathbb R^M$6 and quantile errors $X \in \mathbb R^M$7, and reductions in validation error by 1–2 orders of magnitude relative to global sparse PCE (Wang et al., 2024).

Reliability and Rare-Event Estimation

In reliability analysis, the ASSE approach underlies adaptive, active-learning reliability estimation methods (e.g., SSER). The domain is recursively partitioned so that accuracy is focused near the limit-state surface or rare-event region, severely reducing the number of expensive model calls required. For benchmark rare-event problems ranging from 2D to 870D, ASSE provides probability of failure estimates matching Monte Carlo reference at $X \in \mathbb R^M$8 of the cost (Wagner et al., 2021).

Bayesian Inference

For Bayesian model inversion, the stochastic spectral likelihood embedding (SSLE) variant applies ASSE to hierarchically approximate the likelihood function, enabling analytic expressions for the posterior evidence, marginals, and moments. Adaptive sample enrichment preferentially targets regions of high posterior concentration, ensuring accurate mass allocation even in high nominal dimension and multi-modal settings (Wagner et al., 2020). In all test cases, only adaptive partitioning successfully recovers multi-modal posteriors at practical computational budgets.

Attributed Graph Clustering

A distinct instantiation of the ASSE mechanism occurs in large-scale attributed graph clustering. The scalable and adaptive spectral embedding (SASE) method (Liu et al., 2024) uses a three-stage pipeline:

1. k-order Simple Graph Convolution for feature smoothing across $X \in \mathbb R^M$9-hop neighborhoods;
2. Scalable Spectral Clustering leveraging random Fourier features (RFF) to approximate dense kernel similarity matrices, avoiding $$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$0 costs;
3. Adaptive Order Selection, where the optimal graph-convolution depth $$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$1 is chosen by minimizing a normalized intra/inter-cluster embedding-distance criterion over $$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$2.

The SASE method achieves linear complexity in both time and space with respect to graph size, and surpasses learned or deep approaches (e.g., S3GC, DGI) both in accuracy and speed, for example achieving a 6.9% improvement in clustering accuracy and a $$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$3 speedup on the ArXiv citation graph ($$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$4K nodes) (Liu et al., 2024).

4. Methodological Details and Computational Characteristics

ASSE employs a set of core technical tools and strategies:

• Domain Partitioning: Partitioning is performed along axes maximizing sensitivity (e.g., Sobol’ indices) or error variance, creating binary, recursively-defined trees or more general domain decompositions.
• Local Surrogate Construction: Within each subdomain, polynomial bases are constructed to be orthonormal with respect to the induced (conditional) distribution, either by direct Gram–Schmidt or via isoprobabilistic transform. Expansion coefficients are learned via local least-squares or sparse regression.
• Adaptive Design: Sample enrichment is concentrated in subdomains contributing most to global error; for Bayesian or reliability applications, this directs resources to high-likelihood or failure regions.
• Error Diagnostics: The local leave-one-out error estimator is central to both adaptivity and global error control. In all instances, adaptation prevents wasteful refinement of regions that are easy to approximate or of negligible importance.
• Computational Scalability: Expected total cost scales as $$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$5 in the number of training samples; evaluation cost is dominated by fast tree descent (logarithmic in the number of splits) and a single local surrogate evaluation.

Table: Key ASSE Application Domains and Characteristics

Application	Domain/Target	Key Adaptivity Aspect
UQ/Surrogate modeling	General $$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$6	Error-adaptive residual expansions
Reliability (SSER)	Limit-state, rare event zones	Variance-driven partitioning
Bayesian Inference (SSLE)	Posterior-concentrated regions	Likelihood-specific adaptive sampling
Attributed graph clustering (SASE)	Large attributed graphs	Adaptive k-order smoothing

5. Empirical Performance and Comparative Analyses

Benchmarking ASSE and its variants demonstrates major empirical advances:

• For surrogate modeling and UQ tasks with high local complexity or nonstationarity, ASSE consistently reduces the relative RMSE by one to two orders of magnitude compared to global or sparse PCE, as measured on both low- and high-dimensional test problems ($$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$7) (Marelli et al., 2020).
• In rare-event reliability assessment, ASSE achieves accurate failure probability estimates at $$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$8 or less of the cost of Monte Carlo simulation, maintaining precision for extremely small probabilities ($$\widehat Y_{\mathrm{SSE}}(X) = \sum_{k \in \mathcal K} \mathbf 1_{\mathcal D^k}(X) \widehat R_S^k(X),$$9–$\mathcal D^k$0) and in high dimensions (Wagner et al., 2021).
• In Bayesian inference, only adaptive SSLE accurately recovers complex, multi-modal posteriors or sharply localized credible regions within a finite sample budget (Wagner et al., 2020).
• For attributed graph clustering on large-scale datasets, SASE (ASSE) matches or exceeds the accuracy of deep graph neural and contrastive methods, while running $\mathcal D^k$1–$\mathcal D^k$2 faster, enabled by fully linear O($\mathcal D^k$3) time and space complexity (Liu et al., 2024).

6. Limitations, Connections, and Future Directions

While ASSE provides mean-square convergence guarantees for uniform (nonadaptive) refinement, fully adaptive (greedy) partitioning does not guarantee that all regions will be properly covered, though weighted error-driven adaptivity is empirically robust against this limitation (Wagner et al., 2020). In high-dimensional settings, the number of subdomains may grow rapidly, but if problem structure is low- or moderately-active dimensional, or samples are focused adaptively, computational costs remain practical.

ASSE is closely related to other piecewise-surrogate or local polynomial methods but distinguishes itself by the explicit use of spectral expansions, hierarchical residual correction, and statistically-informed partitioning. There are ongoing research directions in extending ASSE frameworks to multi-output surrogates, integrating with deep neural surrogates, and improving partitioning strategies for extreme-scale or highly anisotropic domains (Marelli et al., 2020, Liu et al., 2024).

7. References to Primary Literature

• "Stochastic spectral embedding" (Marelli et al., 2020)
• "Efficient Probabilistic Optimal Power Flow Assessment Using an Adaptive Stochastic Spectral Embedding Surrogate Model" (Wang et al., 2024)
• "Rare event estimation using stochastic spectral embedding" (Wagner et al., 2021)
• "Bayesian model inversion using stochastic spectral embedding" (Wagner et al., 2020)
• "Scalable and Adaptive Spectral Embedding for Attributed Graph Clustering" (Liu et al., 2024)