The test data

Linearity data were taken at 1 V bias on 2002 Sep 21, and at 0.75 V bias on 2002 Sep 21. Flamingos was in the instrument lab, and the illumination at Ks was provided by a piece of brown cardboard between the window and the window cover.

Forty reference images at 3 second itime were taken at the beginning and end of the sequence. Twenty exposures were taken with itime stepped by 5 seconds, starting with 5 seconds and ending with 5 seconds, with 20 reference images at 3 second itime interspersed. Darks were taken at each itime, at the end of the sequence. Summarized in tabular form:

Data Reduction Technique for 1 V data (presented below)

The bad pixel mask was then multiplied on each image in the linearity sequence.

For each itime, the dark images taken at the end of the linearity sequence were median combined, and then subtracted from the corresponding linearity images. At least one image from each set was excluded from the average.

Image statistics for each image were then computed with iterstat, with 5 sigma rejection, and 10 iterative loops.

1 V Bias 'High' Illumination

Exposure Times vs. Time of Acquisition

This shows the mean signal in each image at each itime, and the order they were taken. A complete set of darks was taken right after this sequence, in the same exposure time order, but not plotted.

Drift of All darks vs Time of Acquisition

Only the final set of 3 second darks seemed different, so the darks at each itime were combined, and then subtracted from their respective images.

Drift of 3 second Reference Frames vs. Time of Acquisition

The first image of each reference taken after a series of images with high signal is depressed, as usual. It's not clear what happened at (MJD - 52538) = 0.61; the glitch at (MJD - 52538) = 0.75 is when MCE 4 was restarted after switch from utility power to generator.

The overall drift in the signal, first to lower levels, and then to higher levels is a little weird. The green curve is a 4th order polynomial fit generated by the iraf routine curfit (in the utilities package). The fit is f(x) = a0 + a1 x + a2 x² + a3 x³ +a4 x⁴, where a0 = 2.55910E4, a1 = -1.449404E5, a2 = 3.169708E5, a3 = -3.076710E5, a4 = 1.119618E5.

Gain Curve & Read Noise

1) Running Difference Images

For each itime the first image was discarded, and a difference image was formed by subtracting two consecutive images. Iterstat was run on the difference image; the resultant mean was squared and multiplied by 0.5 (becuase of the root 2 times higher noise in the difference image), and plotted against the mean counts in the first good image.

Gain = 4.9 e/ADU

Read Noise = 78 e

I realized later that the drift correction actually affects the mean counts that the sigma-squared should be plotted against, in this case moving the points to lower x-values. Here's the plot and fit, which really doesn't change any conclusions for the gain and read noise:

2) Images at Each Itime Median Combined

The first image was discarded, and the remaining images were imcombined, generating an average image and a sigma frame. The square of the mean from the sigma frame is plotted against the mean from the average frame. The result is not really any different from the running difference method. The fit is to the (5, 10, 15, 20, 25) second itime data. The fit gives:

Gain = 5.1 e/ADU

Read Noise = 71 e

Conclusion: Gain & Read Determination

The numbers from the running difference method are probably better, as they use more data for the fit. So I'd quote a gain of 4.9 e/ADU for this array at 1 V bias. I don't know why we got 75 +/- 5 e of read noise for this data set. We should check it again, once we get a chance to test the new A/D filters.

Linearity Curves

ADU rate vs. Time of Acquisition

The mean count rate for each image at each itime (5 to 60 seconds, excluding 3 second references) is plotted against the time of acquisition. Also plotted, in green is the 4th order polynomial fit to the drift in the 3 second references.

The blue curve is the fit shifted downards, to match the 20 second itime data, which is coincident with the minimum in the fit to the 3 second data. Additionally, the fit has been shifted downwards in the blue and purple curves, to show that the rate of change in the 35 and 45 second itime data matches the slope of the fit.

ADU rate vs. ADU

This ones confusing. The red curve shows the data rate vs. counts for every image at each itime. The points for each itime are generally vertical, showing the drift seen in the previous plot.

The green curve shows what happens to the rate curve after applying the drift correction implied by the reference data (see the previous plot). It actually looks less flat.

For the uncorrected data, over the 10,000 to 30,000 ADU data range the data varies by ~0.5%; if we include the 5 second itime data at 4,000 ADU, the rate curve varies by ~1.3%.

For the drift corrected data, over the 10,000 to 30,000 ADU data range the data varies by ~1%; if we include the 5 second itime data at 4,000 ADU, the rate curve varies by 1.5%.

Any ideas why this doesn't look better?

Similarly to the gain curve plot, I realized that the rate curve corrected for drift should have the points shifted both in the rate value and the correspoinding mean value, ie parallel to both axes. The addition of the shift along the x-axis mean values steepens the curve, but not enough to change anything concluded from the plot above. Here's the new plot: