1: \begin{abstract}

2: Algebraic multigrid (AMG) is one of the fastest numerical methods for solving large sparse linear systems. For SPD matrices, convergence

3: of AMG is well motivated in the $A$-norm,  and AMG has proven to be an effective solver for many

4: applications. Recently, several AMG algorithms have been developed that are effective on nonsymmetric linear systems.

5: Although motivation was provided in each case, the convergence of AMG for nonsymmetric linear systems

6: is still not well understood, and algorithms are based largely on heuristics or incomplete theory.

7: 

8: For multigrid restriction and interpolation operators, $R$ and $P$, respectively, let $I - \Pi:= I - P(RAP)^{-1}RA$ denote the projection

9: corresponding to coarse-grid correction in AMG. It is invariably the case in

10: the nonsymmetric setting that $\|\Pi\| = \|I - \Pi\| > 1$ in any known norm. This causes an interesting dichotomy: coarse-grid correction is

11: fundamental to AMG achieving fast convergence, but, in this case, can actually increase the error. Here, we present a detailed

12: analysis of nonsymmetric AMG, discussing why SPD theory breaks down in the nonsymmetric setting, and developing a general

13: framework for convergence of NS-AMG. Classical multigrid weak and strong approximation properties are generalized

14: to a \textit{fractional approximation property}. Conditions are then developed on $R$ and $P$ to ensure that $\|\Pi\|_{\sqrt{A^*A}}$

15: is nicely bounded, independent of problem size. This is followed by the development of conditions for two-grid and multilevel

16: W-cycle convergence in the $\sqrt{A^*A}$-norm.

17: \end{abstract}

18: