1: \begin{abstract}

2: We introduce a new class of preconditioners to enable flexible GMRES

3: to find a least-squares solution, and potentially the pseudoinverse

4: solution, of large-scale sparse, asymmetric, singular, and potentially

5: inconsistent systems. We develop the preconditioners based on a new

6: observation that \emph{generalized inverses} (i.e., $\boldsymbol{A}^{g}\in\{\boldsymbol{G}\mid\boldsymbol{A}\boldsymbol{G}\boldsymbol{A}=\boldsymbol{A}\}$)

7: enable the preconditioned Krylov subspaces to converge in a single

8: step. We then compute an \emph{approximate generalized inverse} (\emph{AGI})

9: efficiently using a \emph{hybrid incomplete factorization} (\emph{HIF}),

10: which combines multilevel incomplete LU with rank-revealing QR on

11: its final Schur complement. We define the criteria of $\epsilon$-accuracy

12: and stability of AGI to guarantee the convergence of preconditioned

13: GMRES for consistent systems. For inconsistent systems, we fortify

14: HIF with iterative refinement to obtain \emph{HIFIR}, which allows

15: accurate computations of the null-space vectors. By combining the

16: two techniques, we then obtain a new solver, called \emph{PIPIT},

17: for obtaining the pseudoinverse solutions for systems with low-dimensional

18: null spaces. We demonstrate the robustness of HIF and HIFIR and show

19: that they improve both accuracy and efficiency of the prior state

20: of the art by orders of magnitude for systems with up to a million

21: unknowns.

22: \end{abstract}

23: