Spectral output vs Dshift parameter

10oct07

        The pdev spectrometer dshift parameter (see layout of spectrometer scaling) down shifts the output of the power computation before accumulating the spectral values in the 40 bit accumulators. It allows longer integrations before overflowing the 40 bit accumulator. As an example: a 1 second integrations at 170 MHz with: (A/D=48,pshift=0x1555, dshift=0,ashift=0) has LO feedthrough birdies close to 1.5e9. The maximum allowed value is 4e9. Setting dshift would allow 1 second integrations without rfi overflowing. The maximum allowable downshift value is set by the requirement that you leave enough bits on the noise.      For a single spectral estimate, the amplitude of this noise is comparable to the mean value of the power spectrum.

    The allowable range for dshift had been computed from the spectral mean with dshift = 0 (see page 2 of the plots). It showed that you would have about 4% of the channels with an rms < 40 counts with dshift=15, pshift=0x1555 and aToD Rms=48 counts.
    On 08oct07 data was taken with different dshifts to verify that the actual settings matched the computed values.

    The datataking setup was:

The red noise generator was used for input to box 103 band 0.
The input signal rms had 48 counts at the digitizer.
Pshift=0x1555, pfbgain=.838
fftlen=8192, bandwidth 170 MHz.
300 seconds of data was taken with .1 second integrations (FCNT=2075).
Data was taken with dshift set to: 0,8,12,15.

The plots show the results for various values of dshift (.ps) (.pdf):

Page 1 Avg spectrum.

The .1 second spectra have been averaged for 300 seconds.
The 4 plots are dshift:0,8,12,15 (divide by 1,256,4096,32768).
The average spectral value for a single dump (.1/2075= 25usecs):

dshift=0 : 5.4e6
dshift=8 : 2.1e4
dshift=12: 1.3e3
dshift=15: 166

The mean value decreased by 2^dshift until dshift=15 where is increased more rapidly.

Page 2: 300 second average spectra band pass corrected to the dshift=0 spectra and then normalized to the mean value.

The top is polA and the bottom is polB.
The colors are: black-dshift=8, red dshift=12, green dshift=15.
The edges of the bandpass do not divide out with dshift=15.

Page 3. Rms/Mean along each frequency channel for dshift=0,8,12,15.

The rms/mean was computed for the 300 seconds of data along each frequency channel for each dshift.
Colors specify different dshifts.
Plots 1,3 are the full scale polA and polB. Plots 2,4 are blowups of the left edge showing where the bandpass falls off.
For dshift=15 the rms/mean differs from the expected value for about 2%*2 of the bandpass. This agrees with the computed value using a single dshift.

Conclusions:

For pshift=0x1555, a/d sig=48 counts,

Dshift = 0 to 8 the noise statistics for the bandpass is unchanged.
Dshift=12 the noise statistics starts to fall off in < 1% of the channels
Dshift=15 about 4% of the channels have a modified noise statistic.

We should be able to use a Dshift of 8 or 10 all the time (with 32 bit sampling, 8k xform) and not have any trouble with the noise statistics.

processing: x101/071008/tstdshift.pro

<- page up
home_~phil