Details

Type: Story

Status: Done

Resolution: Done

Fix Version/s: None

Component/s: None

Labels:None

Story Points:8

Epic Link:

Team:Data Release Production

Urgent?:No
Description
See notebook for examples from DP0.2:
VarianceUnderestimateOverlaps.html^{}
Currently, the code that mosaics templates from coadds does so by coadding them. By default the variance plane is computing perpixel by taking the inverse of the sum of inverse variances. This assumes inputs are independent, but the same calexps go into neighboring patches. (It's just counting the same calexps twice) Therefore the variance planes in those overlap regions are underestimated.
A potential $\sqrt{2}$, or $\sqrt{4}$ underestimate sounds scary, but the variance in the science image >> variance of the template. So in the case where you have 4 coadds overlapping, this ends up being a 7% effect for 10 visits and 0.7% effect for 100 visits in practice on the variance places of the image differences...
... which is also why we haven't noticed. Not catastrophic for DP0.2, but we should post in the known issues.
Attachments
Issue Links
 blocks

DM35284 Process the full DECam HiTS dataset with new image differencing
 Done

DM35285 Process the full HSC Cosmos dataset with new image differencing
 Done

DM36715 Reprocess diffim sprint dataset after bug fixes
 Invalid
 relates to

DM33001 Design refactoring of ImageDifferenceTask
 Done

DM35817 Turn off compatibility mode for image differencing
 Done

DM35706 Analyze diaSources in patch centers with new diffim
 Invalid
 mentioned in

Page Loading...
Adding a link to
DM33001, since the variance plane underestimation is a much more serious effect when convolving the science image.