search

dstndstn / tractor

The Tractor: measuring astronomical sources via probabilistic inference
Other
86 stars 24 forks source link

Bias in small stamp size #86

Closed DriftingPig closed 5 years ago

DriftingPig commented 5 years ago

Hi Dustin, I'm just curious about this. If I define a Gaussian distribution, with mean,sigma. Then I draw random points from this Gaussian distribution and add them in a binned histogram. If the binning is large, then the sigma that I can get from this binned histogram will converge to a larger sigma value. And because of this, the half radius of galaxies will be larger than they actually are. Does tractor correct for this effect?

dstndstn commented 5 years ago

I'm not really sure what you're talking about. We use mixtures of Gaussians as approximations of the deV and exp radial light profiles. These are evaluated (after convolution by the PSF) at the centers of the pixels. The PSF is assumed to be the pixel-convolved PSF, so you shouldn't think of pixels as "little boxes"; they are delta-function samples from an underlying smooth surface.

dstndstn commented 5 years ago

@davidwhogg - want to say anything?

davidwhogg commented 5 years ago

If you are asking about the point that we want the seeing-deconvolved sizes then yes, that's the entire point of The Tractor.

davidwhogg commented 5 years ago

Or if you are asking about the point that the finite pixelization is a convolution and we want the pixel-deconvolved sizes, that's more complicated, but yes, I think we do correct for this effectively if the imaging is well sampled. Because the psf in The Tractor is the pixel-convolved PSF, in the well-sampled limit.

DriftingPig commented 5 years ago

My test shows that stamp size might have some effect, but it's not significant.