Joey,

Nice job on your final presentation.  I would like a copy of the
slides when you get a chance to email them to me.

Talk organization:

 - generally great

 - I would have liked to hear about SDSS/SDSS-I earlier in the talk so
   that it is more obvious, sooner, about why automating galaxy
   classification is important.

Discussion of curvelets:

 - would be nice to know what basis functions are being used for the
   fourier transform underlying curvelet work.

 - your geometric description (structure within rotated rectangular
   boxes) was great

 - a couple of algebraic displays explaining about threshholding might
   have been nice

 - the visual displays illustrating threshholding were great

 - like to hear more about the threshholding fraction but I know you
   haven't done much with that yet.

Galactic model:

 - good description, built up well from simple small model to full
   8-dim model

Need some measure of uncertainty in your estimates! 

 - after class we talked about adding a parameter for the diameter of
   the convolving gaussian but now I don't think that's the right
   measure (because the convolved gaussian is part of the galactic
   model, not part of the error structure).

 - You are basically fitting a model of the form

   Data = curveletsmooth + error1
        = (galactic model + error2) + error 1

   and if we ignore the curvelet smooth for a moment, you are fitting

   Data = galacticmodel + error                               (*)

   by least squares.  This is a normal-errors nonlinear regression
   model with eight parameters (the "galacticmodel" here is the final
   galactic model you fitted, including the convolved gaussian).  

   You can estimate the resudual variance as simply

   S^2 = \sum_{\rm pixels} (Data - galacticmodel)^2          (**)

   and then you have two "easy" choices for standard error estimates
   for the eight parameters:

   1. you can get SE's that look like weighted-OLS se's essentially by
      applying the delta method - see for example Seber and Wild's
      text on nonlinear regr models, or probably a ton of websites
      from nonlinear regression classes on the www...

   2. you could try bootstrap estimates of SE's based on (*).  One
      simple idea would be a nonparametric bootstrap where you just
      bootstrap the residuals in (*) to create new data sets to which
      you refit the galactic model (probably including presmoothing
      with curvelets) and then compute empirical SE's from the set of
      bootstrapped param estimates you get.  

   3. A "parametric bootstrap" would do the same thing, but replace
      permutations of the real residuals in (*) with draws from a
      N(0,S^2) [where S^2 is from (**)]...

   I think I like #2 the best, because it seems to account best for
   both the smoothing and the galactic model fitting operations.

-BJ