[evlatests] Sensitivity of 3-bit vs. 8-bit
Rick Perley
rperley at nrao.edu
Wed May 11 14:53:19 EDT 2011
I had a go at this, but there are problems ...
The phase scatter method requires a strong source -- preferably a
calibrator. The one used in the April tests is quite strong -- 4.1 Jy.
For a 2 Mhz wide channel and 1 second integration, the phase scatter
(rms) is about 3 degrees.
Atmospheric phase winds appear to be slow, so I used the 1-minute
calibration table entries that I had already made from yesterday's
calibration.
Phase histograms were made. It was immediately apparent that the
phase widths at the edge of the bandpass is quite a bit higher than in
the middle. I attribute this to the 'delay clunking'. This requires us
to use central channels only. But if we use too few channels, we get
poor statistics -- not enough independent (b,t) samples -- to get us to
the 1% target. Or, we should use a weaker source, so the rms noise per
channel is higher. Something around 1 Jy would be better if we adhere
to 2 MHz channels. Or we can use a setup with narrower channels, or
shorter integrations.
I tried the method of measuring the phase rms over the bandpass at a
given integration. This will also suffer from the 'delay clunk' problem
adding noise at the edges, but avoids the need to remove the atmospheric
phase drift. But I find here there always seems to be some 'feature' or
slope in the detailed spectra which makes the rms invalid. (Eric's
program doesn't provide the rms phase about a mean slope. Perhaps we
should ask him to code this in?)
I personally believe that the amplitude dispersion in the real or
imaginary parts, when observing blank sky, is the most robust method.
Array amplitude stability is much better than phase stability, and the
noise distributions on cold sky are independent of the latter.
Rick
Barry G. Clark wrote:
> As I've said before, I consider the phase noise in a single channel
> as the gold standard. This method probably doesn't, but might, couple
> in spectral anomalies through the calibration process. We are going
> for 1% after all.
>
> Anybody got any ideas about the 'noisy' 8 bit digitizers?
>
>
>> I applied a 'fresh set of eyes' to 3 and 8-bit data taken at
>> C-band
>> on April 21.
>>
>> The observations were of 3C147, J0217+7349 (our favorite northern
>> 'dot' source), and of a nearby 'noise' field, devoid of strong
>> emission. The 3-bit observations were taken first, with all
>> appropriate
>> levels set correctly (says Vivek, who ran the test). The 8-bit
>> observations followed about 45 minute later. I utilized only the
>> northern data, so there will be no important change in elevation.
>> The data, for both 3 and 8 bits, were taken with 8 subbands of 128
>> MHz width each, with 2 MHz channel resolution, spanning 5488 through
>> 6512 MHz. Unfortunately, subbands 4 and 5 both have contamination
>> from
>> the Mangas-Grayhill telecom link (centered at 6004 MHz, 30 MHz wide),
>> so
>> these subbands were not utilized in the analysis.
>>
>> The data are of excellent quality. Calibration followed the usual
>> path, except that I did not attempt to transfer the gain calibration
>> from 3C147. We are interested in ratios between 3 and 8 bits, so I
>> merely used the approximately correct flux of 4.1 Jy for J0217+7349.
>> Once the levels were set, the 3-bit data showed excellent
>> amplitude
>> and phase stability.
>>
>> Examination of the bandpass solutions showed very good flatness,
>> except in subband 7, which has (for reasons unknown to me) a strong
>> slope of 4 dB over 128 MHz in 15LCP and 22 LCP. Maximum (power)
>> differentials within all other subbands are less than 2 dB.
>> Curious sharp features, seen in antennas 12 and 22 in subbands 6,
>> 7,
>> and 8 were found to be due to an *enormous* internal birdie on those
>> antennas, which I recall is due to the sampler intself (we had these
>> features in the previous design also). Autocorrrelation spectra are
>> available for two of the four antennas (12 and 22), so I could see the
>> characteristics of this tone. The maximum power density (in radio
>> astronomy units) is about 2000 Jy within the 2 MHz channel resolution.
>> The tone is unresolved, but is not of perfectly constant frequency --
>> it
>> is seen to 'wobble' a bit in the 'waterfall' plots. The tone
>> frequencies are different for each antenna and IF -- hence we don't
>> expect a strong cross-correlation between any pair of antennas.
>> But we do see faint imprints of these 'tones' in the
>> cross-correlation data, enough to cause the sharp features in the
>> bandpass spectra noted above. Some sort of coupling would seem to be
>> happening.
>>
>> It would be good if autocorrelation spectra were made available
>> for
>> all four antennas, in future tests...
>>
>> The main purpose of this test was to measure the sensitivity
>> degradation between 3 and 8 bit data. To do this, I formed histograms
>> of the noise scatter in the imaginary part of the visibility on the
>> 'noise' field. Checks with the real part showed excellent agreement
>> in
>> all cases. I also checked the parallel hands against the crossed
>> hands,
>> with again excellent agreement.
>>
>> Below are the 'degradation matrices' for four of the subbands: 2,
>> 3, 6, and 8. Subband 1 was not utilized since we know the 8-bit path
>> is
>> severely degraded. Subbands 4 and 5 have RFI. Subband 7 has the
>> strongest internal birdies, and a strong spectral slope, as noted
>> above.
>>
>> For the four chosen subbands, I measured the noise histograms,
>> using
>> the central 45 channels. All showed gorgeous Gaussian shapes, with no
>> outliers or sharp features. The matrices below give the percentage
>> degradation of the noise in 3-bit to that in 8-bit modes. A value of
>> 5%
>> would be considered expected. A value of 10% is a little worrisome.
>> RCP is in the upper right, LCP in the lower left.
>>
>> Subband 2 (5680 MHz,center)
>>
>> 12 15 22 28
>> -------------------------------------
>> 12 | X 5 8 3
>> 15 | 3 X 7 3
>> 22 | 12 12 X 7
>> 28 | -2 -3 9 X
>>
>> Subband 3 (5808 MHz)
>>
>> 12 15 22 28
>> -------------------------------------
>> 12 | X 7 4 0
>> 15 | 3 X 9 2
>> 22 | 6 12 X 8
>> 28 | -4 0 1 X
>>
>> Subband 6 (6192 MHz)
>>
>> 12 15 22 28
>> -------------------------------------
>> 12 | X 11 19 13
>> 15 | 7 X 9 6
>> 22 | 23 23 X 11
>> 28 | 5 2 17 X
>>
>> Subband 8 (6448 MHz)
>>
>> 12 15 22 28
>> -------------------------------------
>> 12 | X 3 10 7
>> 15 | -1 X 10 -2
>> 22 | 18 12 X 4
>> 28 | -2 -8 6 X
>>
>> The overall average degradation 6.5%, which is only slightly
>> greater
>> than that anticipated. The dispersion of this average is 6.5%.
>> It's clear that antenna 22 on the RCP side has an anomalously high
>> loss in 3-bit mode. It is also notable that antenna 28, on the RCP
>> shows effectively no loss at all -- presumably this is telling us
>> something about the 8-bit side.
>> Subband 6 shows the highest loss ratios -- this is caused by a
>> *decreased* rms noise from the 8-bit side, but I have no easy
>> explanation for this. Perhaps something in the calibration.
>>
>> The strong spectral slope seen in subband 7 on some antennas gives
>> an opportunity to see if the noise in the 'spectrally weak' channels
>> is
>> higher than on the spectrally strong ones. The slope is notably
>> strongest for antennas 22 and 15 in LCP, where there is a 3 dB
>> difference between channels 12 and 52. The expected trend in the
>> noise
>> is easily seen on this baseline:
>>
>> Channel 12: rms = 135 mJy
>> Channel 32: rms = 116
>> Channel 52: rms = 113.
>>
>> All other baselines to these two antennas show the effect as well,
>> but not as strongly (as expected, since the slopes in antennas 12 and
>> 28
>> are not as steep).
>>
>>
>> _______________________________________________
>> evlatests mailing list
>> evlatests at listmgr.cv.nrao.edu
>> http://listmgr.cv.nrao.edu/mailman/listinfo/evlatests
>>
>>
>
>
More information about the evlatests
mailing list