Andy mentioned in the circuit-modeling thread that he might be going to explain

this stuff at some point, but since I thought it might be insightful for any newbie
thinking of how to get different modes from a filter, and I have this in fresh
memory having worked with it recently, I thought I'd explain the thing.

The basic filter configuration we are dealing with is exactly the same as the Moog
ladder, though I believe the non-linearities in the Xpander filter are different,
but since we will be dealing with linear stuff only, that makes no difference
whatsoever. As long as you keep your stage gains right, you can mess with whatever
non-linearities you want.

We are going to proceed with the concept of adding together 5 outputs from the
filter: one after the resonance feedback, and 4 after each 1-pole LP stage. I'm
going to name these outputs as A,B,C,D and E. (We'll ignore any Xpander specific
magic here, since I'm explaining the concept and not how to model an Xpander
filter.)

The concept is simple (and hinted at in the service manual of the Oberheim thing,
though it's somewhat obscured by analog implementation issue there): we work out
the transfer function of each output, then solve for constants we need to multiply
them to get a desired transfer function. We'll ignore the resonance from the
feedback loop, as it happens that it makes no useful difference in the discussion,
and there is nothing in the original Oberheim about it anyway (IIRC and AFAIK and
whatever qualifiers).

Now a one-pole analog lowpass can be taken to have the transfer function 1/(1+s)
which is about as much as we need to understand about Laplace transforms to be able
to get this done. The other thing we need to understand is that if we put two
lowpasses in a row, it happens to become multiplication in the transfer domain
(this follows from the relationship of convolution and multiplication, and the fact
that both of them are associative operations).

So, the outputs therefore get transfer functions as follows:

A = 1
B = 1/(1+s)
C = (1/(1+s))^2
D = (1/(1+s))^3
E = (1/(1+s))^4

Now to be able to combine the outputs in a useful way, we are going to assign
coefficients [a,b,c,d,e] for them, giving us a totally generic (as far as the
filter can provide) transfer function:

a*A + b*B + c*C + d*D + e*E

From that you obviously get a 4-pole LP by letting e=1 and rest of them 0. For two
pole LP, you take just the C output, and so on.

Anyway, just different LPs aren't all that interesting, so using your favourite
symbolic mathematics software (I usually use Maxima) or pencil and paper if you
really feel like it, you'll want to expand and then simplify that to get something
like:

as^4 + (4a+b)s^3 + (6a+3b+c)s^2 + (4a+3b+2c+d)s + a+b+c+d+e

If your stages are inverting (like I think they were in the Xpander IIRC) then you
need to negate every other output, which will give you all positive (or negative if
you get it wrong way) coeffs later, but for trivial digital 1-pole lowpasses that's
not necessary. Notice that this is sensitive to individual stage gains though (if
your stages have equal gain and it starts self-oscillating at feedback of 4 you're
at least close; if not, you can either modify the stages or the coeffs you solved).

Anyway, now that we have the general transfer function, we can solve the
coefficients to get a particular transfer function. As an example, we'll take 4-
pole hipass.

One can get 1-pole high-pass from a 1-pole lowpass, by taking the difference
between the input and the output. This obviously (I hope it's obvious anyway) is no
good for multi-pole high-pass though, so we'll multiply (or "put in a row" in the
time domain sense) a couple of single-pole high-passes instead, giving us:

(1 - 1/(1+s))^4 or (s/(1+s))^4 in case you simplify first

Using your favourite symbolic mathematics software again, you might manage to
simplify it to something like:

s^4 / (1+s)^4

ok, in this case it would be trivial to do it on paper as well, but in more

So now we have the ugly polynomial above, and then this simple polynomial, and we
wanna make them the same:

s^4 as^4+(4a+b)s^3+(6a+3b+c)s^2+(4a+3b+2c+d)s+a+b+c+d +e
------- = -------------------------------------------------- -
(1+s)^4 (1+s)^4

The denominators are obviously useless at this point, and if you just solve for
each power of s in the nominator, you'll get a trivial set of linear equations:

/ a = 1 a = 1
| 4a+ b = 0 b = -4
< 6a+3b+c = 0 => c = 6
| 4a+3b+2c+d = 0 d = -4
\ a+ b+ c+d+e = 0 e = 1

Here we go, we've got a 4-pole high-pass response when we multiply the outputs by
the constants we just solved, then sum them together.

Same principle can be used for other modes, say 2+2 pole BP could be solved from:

(s/(1+s))^2 * (1/(1+s))^2 = two-pole highpass going into two-pole lowpass

For notch filter you need to do something more clever, and add hipass and lowpass
responses together (at least 2-pole for each such that they happen to cancel each
other near the cutoff).

You can also make more fancy modes, like notch going into lowpass, or whatever, and
even if we're dealing with the analog transfer functions here, it translates "as-
is" into the digital domain as long as you've got the indivial lowpass-stages
translated properly.

Which ofcourse you won't have, because somehow you'll have to compensate for the
fact that you have an extra delay in the feedback loop. This fortunately causes
only minor problems (there seems to be some anomality with notch and resonance; the
notch doesn't exactly cancel the resonance because the frequencies don't match
exactly). I personally haven't bothered fixing it yet, though potential remedies
would be structuring the one-poles such that correct phases are available inside
them (no idea if this is workable), or by using all-passes (fractional delay
filters should do the job?) for phase correction (more things I'll have to do
before release... blah).

In any case, hope that was useful for someone, and please report any typos/grammar
problems that my proof reading was unable to find.

================================================================================

asomers wrote:
I don't understand how to get from an s-domain function to a transfer function (or
better yet a difference eqn, or um CODE).

if you have your transfer-function in the s-domain (in form of poles and zeros in
the s-plane) - then you are actually almost done. you just need to carry it over to
the z-domain via the bilinear transform. that is, if S is a pole (or zero) in the
s-plane, you obtain the corresponding pole (or zero) in the z-plane, denoted as Z,
via:

Z = (1 / (2*sampleRate)) * (1+S)/(1-S)

the difference equation is then obtained by multiplying out the numerator and
denominator of z-plane transfer function and using the resulting polynomial
coefficients as feedforward or feedback coefficients (numerator->feedforward,
denominator->feedback). you can also pair complex conjugate poles/zeros in the z-
plane to quadratic factors and then implement your filter as cascade of biquads
(which is more advisable for higher order filters).

the hard part of filter design is generally to obtain the poles and zeros in the s-
plane. but once you have them, the above procedure should be straightforward.
finding the poles and zeros in the s-plane is also straightforward for butterworth-
filters, but more involved for chebyshev filters and a serious mathematical
challenge for elliptic filters. ...but as always: the devil is in the detail, even
in the butterworth case: this filter has s-plane zeros at infinity - you need to
catch such special conditions and place z-plane zeros at z=-1 for them.

Quote:

I don't see any connection between your example and the basic IIR stuff I know, is
there any?

this approach here is somewhat different from classical filter design. if you want
to get your head around classical filter design, don't let this here
distract/confuse you

================================================================================
Indeed, the stuff here is mostly interesting when you need a synth-like filter with
properties such as stable response in time-varying case, easy tuning, independent
resonance control, interesting sorts of non-linearities... and you then want to use
the same model to also allow multiple filter modes.

Classical structures like direct form biquads are rather easy to design and CPU
efficient in the time-invariant case, but they lack most of the other features one
would want from a VCF-wannabe.

If you want some example code of what might be a filter suitable for what has been
discussed here, you might want to look into the music-dsp source-code archive,
which contains a couple of "Moog" filters many of which would probably work for
what has been discussed.

=================================================================================

One issue that I have found is that the s-plane equations do not work with the
Stilson-style Moog filter. Tim ended up using 1-pole 1-zero sections in his filter,
with the zeros placed in such a location that the filter was roughly constant-Q in
a useable range without oversampling or lookup tables. However, the zeros also
screw up the summing gains for the different filter modes.

Antti's Moog-style filter may be a better candidate for this kind of multimode
filter, as I think he used one-pole sections exclusively.

=================================================================================

Worry not, said Stilson model can still be made to work. You just have to notice
that the feedback gain for self-oscillation is now (approximately) 1, instead of 4;
there's now extra total gain of 4 in the poles. Since moving some gains around in a
linear filter has no effect, we can return the self-oscillation gain back to 4 by
dividing by sqrt(2) after each pole as sqrt(2)^4 = 4. Since the poles are all
equal, it is reasonable assumption that we should now get the right gains for the
output combinations to work properly. The alternative (if you don't need the gains
right for other purposes) is to just divide the pole output coefficients by
whatever extra gains have been accumulated so far.