The results from October analysed using the latest code gives these results. There are a few more outliers because I found a bug in my code that was hiding some before. The majority of the outliers come from MQD10BFF which had more variation than the others.

Outliers

Large moves during the orbit logging seem to explain the outliers in the IQ rotation angle from the jitter subtracted calibration. The reason for the anomalies from the first calibration file below is a mystery.

Filename	Orbit	Scale Residual/%			Angle Residual/rads			Comments
		Norm	JS	SVD	Norm	JS	SVD
bpmAllMoverCalibration_20111027_070330.dat	bpmAllLog_20111027_070257.dat	27.42	-9.64	0.90	0.0932	0.561	0.0969	Looks normal
bpmAllMoverCalibration_20111027_065054.dat	bpmAllLog_20111027_065040.dat	49.49	29.82	9.12	0.114	0.0778	0.0396	±25μm
bpmAllMoverCalibration_20111026_033625.dat	bpmAllLog_20111026_033552.dat	89.12	37.98	-1.444	-0.0320	0.298	-0.0149	big move in orbit file, beam motion large
bpmAllMoverCalibration_20111027_065711.dat	bpmAllLog_20111027_065657.dat	59.15	-13.04	4.27	0.126	-0.272	0.0173	±25μm
bpmAllMoverCalibration_20111027_045945.dat	bpmAllLog_20111027_045911.dat	61.06	-3.589	-0.331	0.000187	0.213	0.000253	Big move in orbit file

BPM Number Problem

As the number of BPMs in the regression was increased, a problem was seen where the scale jumped by 9 orders of magnitude

If the singular values go to zero, it means that the matrix is non-invertible. This does not happen for the BPMs upstream.

Removing the singular values for which s/max(s)<1e-5 seems to solve the problem. It also improves the statistics for the upstream BPMs.

The following histograms show the statistics for calibrations of the problematic downstream BPMs without singular value cuts (left), with cuts (middle) and without cuts but using less than 25 BPMs upstream for the regression (right). The most any one BPM was calibrated was four and some BPMs were only calibrated once. There were also two very extreme outliers for which the mover was not responding at all.