The question is whether the bias plus the overclock values can correctly track and subtract the bias. The answer from our analysis to date is: yes, the bias maps are accurate to better than 1 electron or so in the mean, and the variance introduced by the bias correction process is negligible compared to the hardware system noise.
Details:
I examined the distribution of pulseheights in the event neighborhood as a diagnostic of the validity of the current bias subtraction methodology.
The data used were all the cal. source observations since the temperature was dropped to -120C (excluding the data in which the temperature was not stable). I only considered G0 events and excluded the top pixel from my analysis because it is effected most strongly by CTI. Including G02346 or the top pixel does not change the results significantly. I fit a single Gaussian to the neighbor pixel histogram.
| Mode of neighboring pixels: | 0 ADU | * |
| Centroid of Gaussian: | within 0.3 ADU of 0 | ** |
| Standard Deviation of Gaussian: | 2.2 - 2.6 ADU | |
| For the imaging array alone | 2.2 - 2.3 ADU | |
| All FI CCDs: | 2.2 - 2.4 ADU | |
| All BI CCDs: | 2.5 - 2.6 ADU | |
| * except for S1 (1 ADU) and S4 (-1 ADU) | ||
| ** except for S1 (0.8 ADU), S3 (0.5 ADU), and S4 (0.8 ADU) | ||
Given that CTI-related re-emission of big cosmic rays would show up in the neighboring pixels as increased noise, I'm pretty convinced by these numbers that the standard bias-taking methodology is fine.
| Chip | Mode | Mean | Stdev | G.Cent. | G.Sigma |
|---|---|---|---|---|---|
| I0 | 0 | -0.02 | 2.48 | -0.15 | 2.26 |
| I1 | 0 | -0.07 | 2.46 | -0.19 | 2.22 |
| I2 | 0 | -0.12 | 2.51 | -0.25 | 2.27 |
| I3 | 0 | -0.13 | 2.53 | -0.25 | 2.29 |
| S0 | 0 | -0.07 | 2.64 | -0.20 | 2.44 |
| S1 | 1 | +1.07 | 3.62 | +0.82 | 2.56 |
| S2 | 0 | -0.15 | 2.51 | -0.26 | 2.27 |
| S3 | 0 | +0.76 | 2.87 | +0.47 | 2.49 |
| S4 | -1 | -0.68 | 2.68 | -0.82 | 2.43 |
| S5 | 0 | -0.10 | 2.59 | -0.21 | 2.31 |