Antony Lewis

1
Antony Lewis Institute of Astronomy,
Cambridge http//cosmologist.info/
2
KK99

• Try to find valid (robust) but non-optimal method
  
• Baseline for comparison that any good valid
  method should be able to beat. May help to
  identify important issues.
• In general need to have consistent galaxy model
  to get the correct answers. All (?) existing
  models except KK99 do not use a consistent galaxy
  model.
• On average galaxies have circular symmetry easy
  to model. Simple priors obvious (e.g. smooth
  probably monotonic function)
• By Central Limit Theorem, measured averaged
  galaxy also has Gaussian distribution know
  statistical model.
• Fit PSF-convolved sheared radially-symmetric
  averaged galaxy to stacked image
• Use splines to model average shape accurately.
  Well determined because lots of galaxies used to
  make average.
• For constant shear only need averaged PSF (varies
  due to pixels even in Great08)
• Sub-optimal but well-defined method for low noise
  that seems to work (possible small calibration
  error)
• Maybe hope gives ideas for new separate galaxy
  method where galaxy modelling errors cancel

3
Problems with noise

• With noise, centroid cannot be determined
  accurately have to account for galaxy-dependent
  error in centroid estimate. Shear -gt centroid
  error aligned with shear -gt calibration
  errorsim bias from PSF, other effects?
• Can estimate centroid error and circularize it to
  eliminate some of the bias.
• But model no longer strictly valid only
  approximate
• Currently no attempt made to tune weights, some
  obvious room for improvement.

4
GREAT08 feedback

• Web sites well set up and lots of useful
  information. Challenge appears to be well defined
  (modulo assumptions about galaxies) without being
  too easy.
• Files too large to be convenient (I had at most
  200GB free on a disk, so had to use two days
  spent downloading and shuffling files around
  file corruptions, ssh connection problems, etc)
• Unable to test well on noisy images due to not
  enough test images with same shear hard to tell
  if differences due to random or systematic
  effects. Making own similar simulations too time
  consuming.
• Properties of PSF worryingly idealized (i.e. no
  dipole, qualitatively different from reality,
  even if known and constant)
• How do you know even in principle how well we
  should be able to do?

5
Retrospective GREAT08 wish list

• Provide results for some known method (e.g.
  KSB) - many good methods will be iterative,
  e.g. provisional estimates useful for weighting
  galaxies in unbiased way. Save lots of time by
  providing results from known method. - can
  speed convergence by starting near known-good
  parameters - useful for sanity checks on things
  going wildly wrong
• At least one set of more galaxies to beat down
  noise for testing
• Source code (minus parameters) for the
  simulations for confidence in validity - when
  youre code doesnt work, very easy to doubt that
  simulations have been done correctly e.g. is
  high-enough sub-pixel resolution used, are stamps
  cut after rather than before PSF convolution, are
  galaxies occurring in mirror-image pairs, etc.
• Some feedback in leaderboard about what errors
  are due to (e.g. in my case probably a
  calibration-like error, gamma too large i.e.
  error dominated by plates with largest
  shears)Ps. Are fixed star images linked from
  site?