Quasi-Monte Carlo methods achieve faster convergence than standard Monte Carlo for numerical integration — rates approaching O(n^{−1}) instead of O(n^{−1/2}). But they require the integration domain to be bounded. In uncertainty quantification for PDEs, the parameters are often Gaussian or otherwise unbounded. The standard fix — truncating the domain or applying a change of variables — can destroy the smoothness properties that quasi-Monte Carlo exploits.
Kazashi, Nobile, and Zech propose boundary-damping importance sampling: multiply the integrand by a carefully chosen weight function that suppresses contributions from the tails, then apply standard scrambled nets on the resulting bounded-like function. The weight function is not arbitrary — it must decay fast enough to make the tails negligible but slowly enough to preserve the integrand's smoothness structure.
The payoff is dimension-independent convergence rates. In elliptic PDE problems with hundreds or thousands of uncertain parameters, the mean squared error rate O(n^{−1−α+ε}) depends on the parametric regularity α, not the dimension. The curse of dimensionality — the usual enemy of numerical integration in high dimensions — is sidestepped because the importance sampling exploits the problem's anisotropic structure: most dimensions contribute little to the integral and need few quadrature points.
The through-claim is about the role of transformation in numerical analysis. The integration problem is hard because the domain is unbounded. The transformation makes it bounded. But the transformation is not neutral — it reshapes the integrand's regularity, determining whether quasi-Monte Carlo can exploit its structure. The quality of the answer depends on the quality of the reformulation, not the quality of the solver.