New Gain and Read Noise Values

New Analysis

The gain values for each amp that are inserted into the SISPI headers and currently propagate through DESDM are incorrect because they were obtained through the standard method of gain = <mean>/<variance>, however the variance is suppressed from the shot-noise level by the brighter-fatter effect. This led to a ~10% over-estimate of the gains and hence a ~5% under-estimate of the shot noise in all measured quantities.

Furthermore the read noise of all CCDs was calculated (in e units) using these gains and is hence slightly incorrect. Plus the read noise has been elevated for (at least) all of season 1 because of some excess electronic noise. So it would be useful at this point to update the RDNOISE and GAIN values that SISPI inserts into DECam image headers, and also have DESDM insert new values into all Y1 exposures for Y1A1 and subsequent reductions.

New gain values were derived by two distinct methods from photon transfer curve data taken on 6 Feb 2014. Each image had overscan and bias subtracted, and the ADU values linearized using linearity_table_v0.5.fits. Daniel Gruen derived gain estimates from pairs of images taken with exposure times between 3s and 14s using the standard mean/variance formula (with correction for read noise). The gain vs signal data were regressed to zero signal, which the b/f model predicts should yield the true gain.

As a sanity check, Gary Bernstein derived gains independently by measuring the variance at each pixel of values in a stack of ~20 exposures of 2s duration, which were interspersed in time with the pairs that Daniel used. The 2s exposures are still slightly affected by the b/f effect so we expect these gain values to come in slightly low. This plot shows the ratio of Daniel/Gary-derived gain values for each amplifier (the dead CCDs 2 and 61 are omitted from the plot). It shows that, apart from the expected difference in mean, these two methods yield very consistent gain values.

For comparison, here is the ratio of the new values (from Daniel) to the GAIN[AB] values currently in use. Note the ~10% shift:

New read-noise values (in ADU) were derived from a set of ~20 bias exposures taken on 8 Feb 2014. Each has overscan subtracted, then at each pixel we measure the variance of the values of the ~20 exposures. For each amplifier, we take an outlier-clipped mean value of these variances in a central region of the amplifier. The square root of this is the read noise, in ADU.

These values are similar to those currently in use, except that we find elevated noise levels in all amplifiers read from backplane 6. See ReadoutNoise showing these elevated levels to have been present all of Year 1 (and maybe earlier).

New Values

Attached is a file newgain.dat containing newly-derived gain values. These come from the Gruen analysis, except for:
  • CCD31 = S7, which has a drifting nonlinearity on one amplifier, making the Gruen analysis unreliable. We take the Bernstein value, times the factor 0.985 that is the mean of the Gruen/Bernstein values for other amps.
  • CCD2=S30 and CCD61=N30, which are not currently functional, but have data in earlier exposures which we may want to use. In these cases we scaled the original FNAL lab-derived values by the factor 0.906 that is the mean of Gruen/old values for other CCDs.

Also attached is a file newrn.dat giving the recently measured read noise in each amplifier. Again the values from FNAL have been placed into the currently non-functional S30 and N30 lines.

The files noise.txt and gains.txt give the noise (in ADU) and gain in a format meant to facilitate ingestion into the SISPI constants database.

The file keywords.txt gives values for the GAIN[AB] and RDNOISE[AB] keywords that can be directly inserted into past images. The RDNOISE values are in electrons and are simply the product of the rn and gain values described above.