1: \begin{abstract}

2:   Compressive sensing is a powerful technique for recovering sparse

3:   solutions of underdetermined linear systems, which is often

4:   encountered in uncertainty quantification analysis of expensive and

5:   high-dimensional physical models.  We perform numerical

6:   investigations employing several compressive sensing solvers that

7:   target the unconstrained LASSO formulation, with a focus on linear

8:   systems that arise in the construction of polynomial chaos

9:   expansions. With core solvers of \lls{}, \sparsa{}, \cgist{},

10:   \fpcas{}, and \admm{}, we develop techniques to mitigate overfitting

11:   through an automated selection of regularization constant based on

12:   cross-validation, and a heuristic strategy to guide the

13:   stop-sampling decision. Practical recommendations on parameter

14:   settings for these techniques are provided and discussed.  The

15:   overall method is applied to a series of numerical examples of

16:   increasing complexity, including large eddy simulations of

17:   supersonic turbulent jet-in-crossflow involving a 24-dimensional

18:   input.  Through empirical phase-transition diagrams and convergence

19:   plots, we illustrate sparse recovery performance under structures

20:   induced by polynomial chaos, accuracy and computational tradeoffs

21:   between polynomial bases of different degrees, and practicability of

22:   conducting compressive sensing for a realistic, high-dimensional

23:   physical application.  Across test cases studied in this paper, we

24:   find \admm{} to have demonstrated empirical advantages through

25:   consistent lower errors and faster computational times.

26: \end{abstract}

27: