Noise fluctuation when combining data in velocity space

[emonts]

Hi,

In my 7mm spectral-line data (33.509 GHz), I see significant noise fluctuations across the band when I combine data sets in velocity space, following:
standard calibration
- uvlin (setting the 'line' parameter in channels)
- puthd in=setX/restfreq value=33.509
- uvredo options=velocity velocity=barycentric
- invert vis=set1,set2,set3,.... line=felo,1835,-8000,8.5235,8.5235
- imstat to plot the noise across velocity

Interestingly, when I combine the channels bluntly and swap to velocity space at the end, the noise looks smooth:
- standard calibration
- uvlin (setting the 'line' parameter in channels)
- invert vis=set1,set2,set3,.... out=total (line-parameter not set)
- puthd in=total/restfreq value=33.509
- velsw in=total axis=optical,barycentre
- imstat to plot the noise across velocity

I attached a plot that shows the difference.

Unfortunately, when data is taken across a long period, because of earth rotation etc it is generally more accurate to combine spectra in velocity space rather than channels, so the approach that created the noise fluctuation is in many circumstances the preferred one.

I wonder whether this has something to do with how miriad re-weights the channels in uvredo or invert (because the velocity width that you give is not exactly the same as the original channel width)? In this respect, when taking wider velocity-width for the channels in invert (i.e. 'averaging' across 4 channels), the effect is reduced, but still there.
Also, at other locations within the 7mm band, the 'sinusoidal' noise fluctuations very so rapid across velocity that I initially did not pick this up (but they seem to be there to some degree as well).

Does anyone have an idea what causes this? In the example that I show this effect corresponds to ~50% additional integration time that is needed to reach the target noise level....

Thanks,
Bjorn

[Mark.Wieringa]

Hi Bjorn,

I had a look at this with some simulated noise data, and I get the same effect. However in my case the noise level in the line=channel image is flat at the level of the peaks in the line=felo image. The peak to trough ratio is very close to sqrt(2), making me suspect it is just an averaging effect - if two channels contribute equally to the velocity channel you get a noise level 1.4 times lower than when the channel lines up exactly with the velocity.

In your case the channel image noise is midway between the peaks and troughs in the felo image, which is more difficult to explain - unless it is due to the fact the velocity channels are wider than the real channels. (The effect seems too large though).

I might investigate some more later.

Cheers,

Mark

velocity noise cube

Screen shot 2011-09-16 at 2.17.22 PM.png (35.96 KiB) Viewed 35457 times

channel noise cube

Screen shot 2011-09-16 at 2.18.09 PM.png (35.99 KiB) Viewed 35457 times

[Mark.Wieringa]

Playing with invert some more seems to confirm this is an averaging effect.

If you set the 'width' value of the line parameter to much less than the channel width (e.g.,0.1) the resulting noise variation with velocity channel is very similar to that in the channel cube. If you set the width a few times larger than the channel width, the variations in the noise level are much reduced. Thus it seems to depend mainly on the number of input channels contributing to each output channel.

If the data had significant velocity drift (several channels across the observation) the ripples should be much reduced as well.

Cheers,

Mark