SDSS photo-z with model templates

1
SDSS photo-z with model templates
2
Photo-z

• Estimate redshift ( physical parameters)
• Colors are special projection of spectra, like
  PCA

3
PCA
3-10 DIMENSION
5-10? DIMENSION
LIGHT Spectrum 1M objects
3000 DIMENSIONAL POINT DATA
5 DIMENSIONAL POINT DATA
BROADBAND FILTERS
MAGNITUDE SPACE 270M objects
REDSHIFT
4
Photo-z techniques

• Empirical
• Polyfit
• Neural net
• Nearest neighbor
• Tempate fitting
• Empirical templates
• Repair
• Model templates
• All the same
• generate a reference set (from observed
  photometry, synthetic photometry of observed or
  model spectra)
• Linear (weighted sum) or nonlinear function of
  neighbors redshift
• The key issue get a good reference set
• Easy to get good results for a good reference set
• Extrapolation only hope is better models

5
Catalogs

• Test set DR6 spectro set , 666697 galaxies (few
  outliers removed)
• Charlot et al. 100k stochastic SFH model library
  
• u,g,r,i,z synthetic magnitudes, 200 redshift bins
  in z0-1
• Using colors only for redshift estimation

6
Fast kd-tree based NN in SQL server

• Index the color space with a search tree
• Find k-nearest neighbors quickly
• Implemented in SQL server (SQLCLR)
• Local polyfit, average, weighted (photo errors)
  sum
• Time to calculate photoz for DR5 (200M object)
• Tempate fitting 150 processor-day
• Kd-fit 10 processor-day

7
Spectro training set

• Local linear fit 150 NN
• ?z0.0294

• Average 150 NN
• ?z0.0306

8
100k Stochastic library

• Local linear fit 150 NN
• ?z0.2275

• Average 150 NN
• ?z0.1429

9
Best subset

• Conclusion too rich reference set with probably
  un-physical templates
• Subset 100 of 100k
• Iteration
• Closest templates in colorredshift space
• Removing templates those cause systematic errors

10
Best subset 1st step

• Average of 150 NN
• ?z0.05477

11
Best subset final iteration
Average of 150 NN Dz 0.0467
Local linear fit of 150 NN Dz 0.1937
12
Open questions

• Cannot reach as good estimation as with the
  empirical reference set
• Are the templates in the best set physical?
• Systematic calibration mismatch
• SDSS filter curves?
• Model dust ?
• Different sampling
• How to compare the sets and find the offset in
  colors
• Correcting for the average difference vector
• ?(ug,gr,ri,iz) (-0.0649,0.0036,-0.0109,0.0212)
• does not improve the results
• Find optimal subset with matching models and
  observations in spectral space (Laszlo)