Multi-level Compressed Sensing Petrov-Galerkin discretization of high-dimensional parametric PDEs
For researchers solving parametric PDEs, this method reduces computational cost while maintaining accuracy, but it is an incremental extension of existing CS-PG methods.
The paper proposes a multi-level compressed sensing Petrov-Galerkin method for high-dimensional parametric PDEs, achieving (almost) optimal convergence order without requiring a-priori assumptions on coefficient index set structure. The work-accuracy trade-off is asymptotically equivalent to a single PG solve on the finest level up to a constant.
We analyze a novel multi-level version of a recently introduced compressed sensing (CS) Petrov-Galerkin (PG) method from [H. Rauhut and Ch. Schwab: Compressive Sensing Petrov-Galerkin approximation of high-dimensional parametric operator equations, Math. Comp. 304(2017) 661-700] for the solution of many-parametric partial differential equations. We propose to use multi-level PG discretizations, based on a hierarchy of nested finite dimensional subspaces, and to reconstruct parametric solutions at each level from level-dependent random samples of the high-dimensional parameter space via CS methods such as weighted l1-minimization. For affine parametric, linear operator equations, we prove that our approach allows to approximate the parametric solution with (almost) optimal convergence order as specified by certain summability properties of the coefficient sequence in a general polynomial chaos expansion of the parametric solution and by the convergence order of the PG discretization in the physical variables. The computations of the parameter samples of the PDE solution is "embarrassingly parallel", as in Monte-Carlo Methods. Contrary to other recent approaches, and as already noted in [A. Doostan and H. Owhadi: A non-adapted sparse approximation of PDEs with stochastic inputs. JCP 230(2011) 3015-3034] the optimality of the computed approximations does not require a-priori assumptions on ordering and structure of the index sets of the largest gpc coefficients (such as the "downward closed" property). We prove that under certain assumptions work versus accuracy of the new algorithms is asymptotically equal to that of one PG solve for the corresponding nominal problem on the finest discretization level up to a constant.