1: \begin{abstract}

2:   We present a multi-level solver for drawing constrained Gaussian

3:   realizations or finding the maximum likelihood estimate of the CMB

4:   sky, given noisy sky maps with partial sky coverage. The method

5:   converges substantially faster than existing Conjugate Gradient (CG)

6:   methods for the same problem. For instance, for the 143 GHz Planck

7:   frequency channel, only 3 multi-level W-cycles result in an absolute

8:   error smaller than 1$\,\mu\textrm{K}$ in any pixel. Using 16 CPU

9:   cores, this translates to a computational expense of 6 minutes wall

10:   time per realization, plus 8 minutes wall time for a power 

11:   spectrum-dependent precomputation. Each additional W-cycle reduces 

12:   the error by more than an order of magnitude, at an additional computational

13:   cost of 2 minutes.  For comparison, we have never been able to

14:   achieve similar absolute convergence with conventional CG methods

15:   for this high signal-to-noise data set, even after thousands of CG

16:   iterations and employing expensive preconditioners. The solver is

17:   part of the Commander 2 code, which is available with an open source

18:   license at \verb@http://commander.bitbucket.org/@.

19: 

20: \end{abstract}

21: