Ryan

It might be a nice feature request. In the meantime I got something working by hand using the output of `gmmposterior` (which was very helpful) so I'm good to go.

In the meantime, if you're interested, you can find my hand-rolled version with weights in [Celeste.jl/blob/master/src/PSF.jl:fit_psf_gaussians](https://github.com/jeff-regier/Celeste.jl/blob/master/src/PSF.jl)

I think it would be useful for GaussianMixtures.jl to expose a `set` method for the sigma matrix. In my opinion, that was the only fiddly part of rolling my own...

I understand Steve is working on this now.

Looking at the galsim cases today, it appears that the FFT version is not finding a good optimum. It would be useful if the ```omitted_ids``` functionality were restored (https://github.com/jeff-regier/Celeste.jl/issues/480).

With the mysterious fix mentioned in https://github.com/jeff-regier/Celeste.jl/issues/534, the FFT is now doing pretty well on GalSim again. I'm running stripe 82 and will check it tomorrow. So far, I've seen...

Yes, I agree it's doing something strange. If you increase the function convergence tolerance, the FFT code drives the galaxy brightness even further to zero. Let's keep this issue open....

On this source, at least, changing the kernel parameter to the (more principled) -0.5, together with setting ```f_tol_rel``` to ```1e-10``` and increasing ```max_iters``` to 200, fixes this problem. (The advantage...

Recall that as soon as one TR step is below the threshold, the algorithm terminates. At the beginning, the Newton steps don't appear to be principally in the brightness direction....

David's interested to have me work on this and has convinced me that it's important.