Limitations of Polynomial Chaos#
Polynomial chaos is most effective when a model response depends smoothly on a moderate number of uncertain variables. The method becomes less attractive when any part of that regime fails.
The first limitation is dimension. A common isotropic total-degree truncation of order \(p\) treats all input directions equally. For \(d\) uncertain variables, it contains
basis functions (Xiu and Karniadakis, 2002). Increasing either the dimension or the polynomial degree can therefore make coefficient estimation and model evaluation expensive.
Several refinements reduce the candidate set. Least-angle regression can construct a sparse expansion from a hyperbolically truncated candidate set (Blatman and Sudret, 2011). Anisotropic polynomial index sets use direction-specific order parameters (Hampton and Doostan, 2018), while a greedy admissible-neighbor algorithm grows a downward-closed index set adaptively (Loukrezis et al., 2020). These methods can delay combinatorial growth, but they still require enough data to identify the relevant directions and interactions.
The second limitation is regularity. Spectral convergence relies on a smooth dependence of the quantity of interest on the uncertain inputs. Discontinuities in random space lead to slow convergence of a global polynomial expansion. Local or multi-element expansions can recover accuracy when the nonsmooth regions can be isolated by partitioning the random-input domain and fitting local expansions (Wan and Karniadakis, 2005). A single global polynomial basis is then a poor representation.
Long-time dynamics create a related difficulty. Even a smooth dynamical system can develop increasingly oscillatory dependence on uncertain inputs over time. For periodic solutions with random frequencies, amplified phase differences can cause a fixed global polynomial expansion to lose accuracy rapidly during long-time integration (Wan and Karniadakis, 2005). Statistical quantities or shorter prediction horizons may remain accessible, but they require validation tailored to those targets.
These limitations define the regime in which polynomial chaos can be expected to work efficiently. When dimension, cost, or irregularity dominates, the following sections replace a fixed global expansion with learned surrogates, multiple information sources, adaptive data collection, and structural constraints.