High-Dimensional Uncertainty Quantification for Stochastic Physical Systems

This direction is split into other directions: curse of dimensionality (stochastic dimensionality scaling). Its last version stays readable and citable here.

Version 1 of 1150 papers166 conceptsfrozen 2026-10-04

When we model a bridge, a power grid, or blood flow through an artery, we rarely know the exact physical parameters. Material stiffness, wind velocity, and fluid viscosity are not single numbers but distributions. To account for this, researchers solve Stochastic Partial Differential Equations (SPDEs), which treat these uncertainties as additional dimensions in the problem.

However, this introduces the curse of dimensionality. If a system has 100 independent random variables, a standard grid-based simulation would require an astronomical number of points. Even Monte Carlo methods, which are dimension-independent in convergence rate, require thousands of expensive high-fidelity runs to achieve precision. The research direction explored here focuses on breaking this bottleneck by exploiting the underlying structure of stochastic spaces.

The Spectral Shortcut: Polynomial Chaos Expansion

The foundational strategy for moving beyond simple sampling is to treat uncertainty as a coordinate system. In Polynomial Chaos Expansion (PCE), the solution to a stochastic equation is projected onto a basis of orthogonal polynomials. This transforms a single stochastic PDE into a system of coupled deterministic PDEs.

As Melvin Leok and Gautam Wilkins demonstrate in An Empirical Chaos Expansion Method for Uncertainty Quantification, the challenge is that the number of required polynomial terms grows exponentially with the number of random variables. To combat this, they propose using data-driven, time-evolving empirical bases rather than fixed ones, allowing the simulation to adapt to the system's actual dynamics.

Compressing the Dimension with Low-Rank Tensors

If the stochastic space is a massive, high-dimensional "cube" of data, we often find that the most important information lives on a much smaller, lower-dimensional "surface." Researchers use tensor decompositions to exploit this.

Zheng Zhang, Luca Daniel, and colleagues argue in Tensor Computation: A New Framework for High-Dimensional Problems in EDA that traditional matrix math is insufficient for modern electronic design. By using Tensor Train (TT) and Canonical Polyadic (CP) decompositions, they compress high-dimensional design spaces, enabling simulations that were previously computationally impossible. Similarly, Peter Benner and Akwum Onwunta show in Low-Rank Solution of Unsteady Diffusion Equations with Stochastic Coefficients that the solutions to massive stochastic systems often possess a low-rank structure that can be exploited by specialized iterative solvers for order-of-magnitude speedups.

Bi-fidelity and Multi-level Strategies

Another way to bypass the curse of dimensionality is to stop treating every simulation as equally important. Bi-fidelity methods use a large number of "cheap," low-accuracy models to explore the parameter space and a tiny number of "expensive," high-accuracy models to anchor the results.

In A bi-fidelity surrogate modeling approach for uncertainty propagation in three-dimensional hemodynamic simulations, Han Gao and Jianxun Wang apply this to blood flow. By combining 1D simplified models with 3D high-resolution simulations, they achieve 100x speedups. This "shortcut" logic is also seen in the work of Irene M. Gamba and Shi Jin 0003, who provide the mathematical error estimates for these bi-fidelity approaches in kinetic equations.

Physics-Informed and Spectral Neural Networks

The most recent shift involves using deep learning to learn the mapping from uncertain parameters to physical outputs. Unlike standard black-box models, Physics-Informed Neural Networks (PINNs) embed the governing equations directly into the loss function.

George Em Karniadakis and his team have pushed this further with BO-fPINN (Bi-orthogonal fPINN), which uses bi-orthogonal decomposition to handle complex stochastic fractional PDEs. By leveraging transfer learning, they can speed up training by 60x.

To achieve even higher efficiency, Nithin Chalapathi and Aditi S. Krishnapriyan introduced Neural Spectral Methods. Instead of learning in the spatial domain, these models learn in the spectral domain, using orthogonal bases to solve parametric PDEs with 500x faster inference than traditional PINNs.

Finding the Active Subspace

Finally, some researchers argue that we can beat the curse of dimensionality by simply rotating our perspective. Guang Lin 0001 and Xiu Yang describe a method in Enhancing sparsity of Hermite polynomial expansions by iterative rotations that identifies "active subspaces"—the specific directions in a high-dimensional input space where the output changes most rapidly. By rotating the coordinate system to align with these directions, they transform dense, impossible problems into sparse, solvable ones.

Papers in this version

Showing 60 of 150 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_high_dimensional_uncertainty_quantificat_b7c7fe65,
  title = {High-Dimensional Uncertainty Quantification for Stochastic Physical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-dimensional-uncertainty-quantification-for-stochastic-physical-systems/rdr_b7c7fe65e21e4a21b8bf3dcb0edd40f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_high_dimensional_uncertainty_quantificat_b7c7fe65_v1,
  title = {High-Dimensional Uncertainty Quantification for Stochastic Physical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/high-dimensional-uncertainty-quantification-for-stochastic-physical-systems/rdr_b7c7fe65e21e4a21b8bf3dcb0edd40f5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}