<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<meta http-equiv="content-type" content="text/html; charset=ISO-8859-1">
</head>
<body bgcolor="#ffffff" text="#000000">
Using the 'C' code I originally developed over 10 years ago, I ran some
software correlator simulation tests to determine the distortion of a
spectral line, with some width (created by shoving a broad-band noise
source thru a narrow passband FIR filter). I didn't generate two lines
at slightly different frequencies, rather I correlated the same data
with a 3-level phase rotator, and a floating-point phase rotator, both
4-bit quantization, for the purposes of comparing them. <br>
<br>
The frequency shift difference was ~8.7 spectral channels (X=15.6
channels, Y=6.9 channels) in the plots, or, if the sub-band width were
128 MHz, equivalent to ~270 kHz, and with the line width ~1 MHz. There
was no continuum correlation component.<br>
<br>
I ran a simulation with 10e6 samples, and see "spiky" artifacts at
about the 10e-3 level, and about 4X the noise level, spread ~20X the
fshift frequency around the line.<br>
<br>
Here are the plots:<br>
<br>
<img src="cid:part1.05040701.05030506@nrc-cnrc.gc.ca" alt=""><br>
<br>
What I believe is going on is that the harmonics of the 3-level phase
rotator are multiplying the strong line, and producing copies of
themselves imprinted in the noise (i.e. since there is only one 3-level
mixer the harmonics don't correlate, but ghosts of them haunt the data
as higher noise), but only in the region of the strong line, and only
with a spread about the line of, say ~20 or so times the phase rotator
fundamental. <br>
<br>
I think this is likely the cause of the Zeeman problems seen on the sky
in recent testing.<br>
<br>
In looking back at my original simulations in NRC-EVLA memo#001 (e.g.
page 95, 96), I compare a strong line with the same two correlators,
and show the amplitude difference in percentage terms, but with a
fairly strong continuum signature as well, and such a difference didn't
particularly show up, although there was likely not enough spectral
resolution to see. In the case where there was a strong line with no
continuum source (pages 106 and 107), it doesn't appear I did this kind
of comparison, as I was primarily looking for the effect on
quantizer-generated harmonics from a strong line.<br>
<br>
I believe the spiky noise signature is not a correlation, but rather is
imprinted in the noise and integrates down with time, but is still
larger than the noise. Of course, as the fshift frequencies drop, the
spread of the spiky stuff decreases, and at some point it could be less
than a spectral channel. <br>
<br>
This weekend I'm running a 10e7 sample comparative correlation to see
if this is the case (i.e. integrates down with time), and will
hopefully report on it early next week.<br>
<br>
The workaround for these kinds of line observations, is to use the SSB
mixer in stage 2 of the Filter FPGA. It has 10 bits of phase
resolution, and so harmonics of the mixer will be at ~-60 dB, and will
correlate at that level. In talking to Dave, it is all ready to go,
can run at all required sub-band bandwidths including 128 MHz. What
needs to happen is the 32-bit phase models going to the Baseline Boards
need to be set to 0 (for the sub-bands that want to do this), and
instead the coefficients go into the SSB mixer phase synthesizers, and
some bit gets set in the filter fpga to direct PHASERR to the SSB mixer
to set it to zero going to the Baseline Boards (you could set the corr
chip phase enable to off, but the RXP phasing logic is still using the
models verbatim, so best to generate them with 0-valued coefficients).<br>
<br>
As a fallback to the above, I've checked the resources in the Recirc
FPGA, and 8 SSB mixers should fit, but with 8-bit arithmetic, and
32-tap Hilbert FIRs, good for -50 dB mixer harmonics, and losing
~1/16th of the sub-band at either end. The SSB mixer in the filter
fpga is definitely the better performing solution.<br>
<br>
In both SSB mixer cases, if there is no artificial fshift, and it is
just earth-rotation phase that is removed, the harmonics of the mixer
will likely be << channel width, and no artifacts at any level
should be seen.<br>
<br>
In either case, if the SSB mixer is used instead of the 3-level phase
rotator, phase is undefined near the sub-band edges, so seamless
stitching of sub-bands is not possible. Presumably for targetted
spectral-line observations, this would not have to be done.<br>
<br>
Apologies for yet another correlator signal processing skeleton. Sigh.<br>
<br>
--Brent<br>
<pre class="moz-signature" cols="72">--
Brent R. Carlson
<a class="moz-txt-link-abbreviated" href="mailto:Brent.Carlson@nrc-cnrc.gc.ca">Brent.Carlson@nrc-cnrc.gc.ca</a>
Tel: 250-497-2346 | Fax: (250) 497-2355
Design Engineer | Ingenieur Concepteur
National Research Council Canada | Conseil national de recherches Canada
Dominion Radio Astrophysical Obs. | Observatoire federal de radioastrophysique
P.O. Box 248, 717 White Lake Rd | C.P. 248, 717 Rue White Lake
Penticton, BC, Canada V2A 6K3 | Penticton, (C.-B.), Canada V2A 6K3
Government of Canada | Gouvernement du Canada
"When and where humans are involved, mistakes inevitably happen"</pre>
</body>
</html>