Something seems to be going on with the SVD algorithm in our correction scheme:

https://drupal.star.bnl.gov/STAR/blog/dmawxc/jetcorr-follow-09112018-svd-unfolding-singular-values-and-kreg

So to continue investigating what's going on, I performed a closure test using both the Bayesian algorithm (which did very well) and the SVD algorithm (which didn't).  The closure test was performed as so:

1. The Run 9 dijet embedding sample is broken up into a Full Field (FF) configuration and a Reversed Full Field (RFF) configuration with roughly equal statistics.  So I created separate response matrices and jet-reconstruction efficiencies for each configuration.
2. I then unfolded the detector-level FF jet spectrum using the RFF response (matrix + efficiency).  The RFF response was then re-applied to the unfolded spectrum to produce a backfolded spectrum.
3. The unfolded and backfolded spectra were then compared to the particle- and detector-level FF spectra respectively.

The results are shown below (there's a typo on the backfolded vs. detector-level plots, the label should read "h+- triggers, eTtrg = 9 - 20 GeV").  The bayesian algorithm worked very well:

The difference in the 23 - 30 GeV/c bin should be taken with a grain of salt: the error bars are likely smaller than what they should be, and the particle-level RFF spectrum is systematically below the FF spectrum at high pTjet.  However, the SVD algorithm did not do as well: