[sdiy] jitter analysis

[Magnus Danielson]

From: "Czech Martin" <Martin.Czech at Micronas.com>
Subject: RE: [sdiy] jitter analysis
Date: Mon, 12 Jul 2004 10:19:33 +0200
Message-ID: <D9D56E8FA1A73542BE9A5EC7E35D37FF01C4EFC5 at EXCHANGE2.Micronas.com>

> I always thought that a sliding DTF IS equivalent to a filter bank,
> but with a strange impulse response for each channel.

Rather, you can view a DFT as being a filter bank, but filter bank as such is
not an exact property but rather a conceptual model for any kind of filter.
So, this is where the implication arrows don't go both ways and therefore there
is no equalence since...

(A <=> B) exactly when (A => B) and (B => A)

So, DFT implies (=>) a filterbank but a filterbank does not imply a DFT.

Let's put it this way, my vocoders contains filterbanks, but none of them is
able to acheive DFT.

You could view DFT as a filterbank where each filter has the impulseresponce 
of pure cosines. Such a filterbank would be _exactly_ at selfresonance and
highly unstable. It is quite reasnoble that existing filterbanks in analogue
does not perform DFT but rather have a lower Q-value. Also, you normally want
that anyway since you want the filterbank to host an ensemble of frequencies
and not a single frequency. 

> The discussion shows that even a "simple" problem of measuring and
> analysing the properties of old hardware is getting complicated, the
> more thought is spend on it.

Actually, some of it is about relearning the basics, they may not be what you
think they are, as just shown. This is why I advocate for the understandning
and use of the LaPlace transform in order to contrast from that of the Fourier
transform. The Fourier transform is one VERY powerfull tool, but it is not that
potent that people claim it is, you need the additional power of LaPlace to
grasp some aspects, and then still more things elude you even with LaPlace
(when your curcuit is time variant and/or non-linear and we see both happening
frequently in synthesizer and audio circuits). 

As for stability of oscillators, there is already a huge and deep researched
being done (and ongoing) from which we may take knowledge, measurement methods
and dedicated analysis tools. My recommendation for a high frequency counter
is one example. There is other methods as well.

> No wonder that current emulations are far of the original.

Seeing my point I guess. People arn't even using the right tools for analysis
so no wonder they don't really get a good picture. 

Cheers,
Magnus

> m.c.
> 
> > -----Original Message-----
> > From: Magnus Danielson [mailto:cfmd at bredband.net]
> > Sent: Freitag, 9. Juli 2004 19:10
> > To: Czech Martin
> > Cc: ifrc at iar.se; synth-diy at dropmix.xs4all.nl; jschmitz at schmitzbird.de
> > Subject: Re: [sdiy] jitter analysis
> > 
> > 
> > From: "Czech Martin" <Martin.Czech at Micronas.com>
> > Subject: RE: [sdiy] jitter analysis
> > Date: Fri, 9 Jul 2004 15:51:24 +0200
> > Message-ID: 
> > <D9D56E8FA1A73542BE9A5EC7E35D37FF01C4EFBF at EXCHANGE2.Micronas.com>
> > 
> > > > The ear, being a bandpassesque analyserbank, can naturally 
> > > > sniff the sidebands,
> > > > but some of it will be masked by normal masking procedure. 
> > > > Analysis done by
> > > > among others the late Julian Dunn has provided insight into 
> > > > where a jitter
> > > > tolerance curve would be given the knowledge we have about 
> > > > masking effects and
> > > > hearing the sidebands. (I am actually missing the paper, so 
> > > > if someone digs it
> > > > out, let me know). This tolerance curve goes below the ns 
> > > > level, just to give
> > > > you an indication... I don't recall it in detail right now.
> > > 
> > > 
> > > and the same thing can the DFT do to a sampled version of 
> > the signal,
> > > since the side bands will all be there.
> > > I think the reall con against the spetral measurement is
> > > its slowness. I have to average somehow and then I loose 
> > track of the
> > > short time changes. Heisenberg biting again.
> > > 
> > > But what about Parsevals theorem? Can't that help?
> > 
> > First of all, a DFT is _NOT_ equalent to a filterbank. A 
> > filterbank with
> > filters have a certain transientresponce of their own, but 
> > the detector
> > mechanism weighs in and this is not necessarilly RMS as DFT 
> > provides with
> > propper scaling. A continous time filterbank has detectors 
> > providing continous
> > time monitoring and in the ear also the continous time 
> > correlation of the
> > detectorsignal. This correlation excists and for instance the 
> > masking effect is
> > one of the results of this correlation.
> > 
> > Going straight to the DFT is not supported by what we know 
> > about the hearing to
> > start with. Again, doing models is easy, verify their 
> > validity to real life is
> > much much harder. The quick and dirty engineering assumption 
> > of tossing a DFT
> > on a problem is not allways even a wise thing to do, you can 
> > end up analyzing
> > all the wrong artifacts.
> > 
> > Also, when trying to do more fundamental research one has to 
> > be carefull not to
> > waste the wrong type of data, but the hole purpose is to 
> > establish what makes
> > it tick and only when you do that you know which shortcuts 
> > you can do. Much
> > work actually become useless due to lack of methology or lack 
> > of analysing the
> > methology. Looks darn nice on the paper, but when you really 
> > think about it at
> > least I end up thinking how much efforts where wasted for the 
> > poor addition to
> > knowledge. That accounts for so many articles a year that I 
> > get really scared!
> > 
> > Cheers,
> > Magnus
> > 
> > 
> >