Question about base_SdssShape_instFlux_yy_Cov in the pipeline source catalog

Hi, I’m working on real/bogus classification, and I’m testing different quantities in the LSST pipeline source catalog. I find that the performance of a quantity called “base_SdssShape_instFlux_yy_Cov” seems to be quite difference from “base_SdssShape_instFlux_xx_Cov” and “base_SdssShape_instFlux_xy_Cov”. It is a quantity in both the src catalog of a direct image, or the diaSrc catalog of a difference image.
The information in the FITS table header is like this.

TTYPE54 = 'base_SdssShape_instFlux_yy_Cov' / uncertainty covariance between base                                               
TFORM54 = '1E      '           / format of field    
TDOC54  = 'uncertainty covariance between base_SdssShape_instFlux and base_Sds&'
CONTINUE  'sShape_yy'    
TUNIT54 = 'count*pixel^2'    
TCCLS54 = 'Scalar  '           / Field template used by lsst.afw.table 

I notice that bright transients usually have large values at “base_SdssShape_instFlux_xx_Cov” and “base_SdssShape_instFlux_xy_Cov” but small values at “base_SdssShape_instFlux_yy_Cov”. However, in theory covariances should not be strongly related to direction.
How does the pipeline compute those SDSS shape flux - 2nd moment covariances? I was not able to find much information about that.
Any suggestions will be appreciated.

All of the documentation about SdssShape is here. Unforunately, there’s not much to go on.

http://doxygen.lsst.codes/stack/doxygen/x_mainDoxyDoc/classlsst_1_1meas_1_1base_1_1_sdss_shape_algorithm.html#details

This is the C++ code for the algorithm, most of which goes back to ~2014.

Note that we are moving away from SdssShape to HsmShape. I’ve asked about the documentation for that (it was just rewritten in python). That package is here:

1 Like

Thinking about this some more: what do you mean by “the performance of”, in your question? Because I could believe that the xx and yy shape covariances do encode information about which sources are bogus. y is the column direction, so bad columns and bleed trails would probably have quite different yy_Cov than less pathological sources.

Thank you for pointing me to those webpages. The SdssShape.cc page shows those flux - moments covariances are derived from a 4D Fisher matrix. In my question, “behavior” could be a better word than “performance”. What I found is that bright sources usually had large Flux_xx_Cov and Flux_xy_Cov values, but small/random Flux_yy_Cov values, and I’m trying to understand why. Thanks for pointing out the issues of bad columns and bleed trails – I think that’s probably the reason.