[Elecraft] Auto Notch Revisited

[David Woolley]

You need to qualify LMS with "for FIR adaptive filters", as least mean 
square (least squares) has a much longer history in statistics.

If I understand it, it is the simple algorithm used in (early) modem 
equalisation filters, because it is cheap to compute.  The problem with 
it in the context of this thread is that it needs a value for d(n). 
Although d(n) is confusingly defined in the actual article, I believe it 
represents the "correct" (desired) output of the filter.

That is easy in a digital system, where you can determine it by slicing 
the output of the filter (and possibly running forward error 
correction), but the context of this thread was the use of the 
auto-notch on voice signals.  In that case, it takes human intelligence 
to work out the general nature of the correct signal, and even then that 
doesn't give you the exact waveform.

What I guess you must be doing, for voice, is using an expected signal 
of silence (d(n) = 0) and hoping that the voice signal is sufficiently 
uncorrelated that the filter fails to converge on to it, but only onto 
the, simpler, interfering carrier(s).

-- 
David Woolley
Owner K2 06123

On 02/08/14 01:39, wa9fvp wrote:
> If your interested in some technical stuff, there's a very good tutorial
> about the LMS algorithm on the National Instruments web site.  The "Fast
> Block LMS" is the one we used in the Hamblaster.
>
> http://zone.ni.com/reference/en-XX/help/372357A-01/lvaftconcepts/aft_lms_algorithms/