Le 19/10/2016 à 17:34, katja a écrit :
Thanks for your update Cyrille. This seems a useful approach to find approximations. I should really learn how to do that with libre office. Do you have tutorial links? So far I found this: https://newtonexcelbach.wordpress.com/2015/06/28/using-linest-for-non-linear....
I learn to do this yesterday... Most informations I found on the web was outdated, but it's very simple. for short : create 2 column with X/Y data. select the data. click insert / diagram. select X/Y dispersion, in order to draw a X/Y graph of your data. then double click on the diagram and click on the curve to select it. then right click and "insert a tendency curve" then enjoy all option available. (there is no "odd only polynomial", but even value was almost 0 in the last example).
Don't forget to click "draw the equation" The only problem is that i did not find a way to cut/paste the equation...
(my libre office is in french, so menu translation may not be accurate)
For the filter recipe the curve must go through coordinates [0 1] and [pi -1], to get the expected behavior when cutoff frequency is set to DC or Nyquist. [hip~ 0] should not block DC, which is only possible when the coefficient is exactly 1 like you get it with (1-sin(0))/cos(0). I tuned my factors 'by hand' to do so. That may be possible with your libre office result as well.
yes, in my example the result is slightly different from 1 / -1. i tried again, adding about 20 points for X=0 and X=1 to get more weight for this coordinate. the coef where better, but not perfect.
still it's a good starting point for manual adjustment.
However, when done in
double precision the calculation will give slightly different result. Above or below 1, I don't know yet. In any case this has to be considered when writing the C.
shouldn’t it be clipped? (what is a filter with negative frequency, or frequency above Nyquist?)
cheers c
Katja
On Wed, Oct 19, 2016 at 4:18 PM, cyrille henry ch@chnry.net wrote:
since you manually adjust the coefficient, i wanted to see the difference with coefficient adjusted/optimised by a computer. I update the calc using X from -0.5 to 0.5 with X = (x-0.5)*Pi The coefficient i get are really different from yours (when considering the Pi factor between them).
I update you patch with this coefs. I don't understand why, but your coefs works better for low frequency. so, nothing better here, but I still think it's worth sharing, in case i'm not the only one wondering...
also, there was a typo in my original mail, it's not a linear regression, but a polynomial regression (since the result was obviously polynomial)
cheers c
Le 19/10/2016 à 15:06, katja a écrit :
Changing the thread title to reflect the new approach. Extract of the original thread;
- I suggested using iemlib's hi pass filter recipe to improve
frequency response of [hip~]
- Christof Ressi pointed to formula in
http://www.arpchord.com/pdf/coeffs_first_order_filters_0p1.pdf
- this formula calculates feedback coefficient k = (1 - sin(a)) /
cos(a) where a = 2 * pi * fc / SR
- the filter implementation is y[n] = (x[n] - x[n-1]) * (1 + k) / 2
- k * y[n-1]
- following convention in d_filter.c (and pd tilde classes in
general), trig functions should best be approximated
- Cyrille provided libre office linear regression result for
(1-sin(x))/cos(x)
Thanks for the useful infos and discussion. My 'math coach' suggested using odd powers of -(x-pi/2) in an approximation polynomial for (1-sin(x))/cos(x). The best accuracy/performance balance I could get is with this 5th degree polynomial:
(-(x-pi/2))*0.4908 - (x-pi/2)^3*0.04575 - (x-pi/2)^5*0.00541
Using this approximation in the filter formula, response at cutoff frequency is -3 dB with +/-0.06 dB accuracy in the required range 0 < x < pi. It can be efficiently implemented in C, analogous to an approximation Miller uses in [bp~]. So that is what I'll try next.
Attached patch hip~-models.pd illustrates and compares filter recipes using vanilla objects:
- current implementation, most efficient, accuracy +/- 3 dB
- implementation with trig functions, least efficient, accuracy +/- 0.01
dB
- implementation with approximation for trig functions, efficient,
accuracy +/- 0.06 dB
A note on efficiency: coefficients in [hip~] are only recalculated when cutoff frequency is changed. How important is performance for a function rarely called? I'm much aware of the motto 'never optimize early', yet I spent much time on finding a fast approximation, for several reasons: it's a nice math challenge, instructive for cases where performance matters more, and I want to respect Miller's code efficiency when proposing a change. Today pd is even deployed on embedded devices so the frugal coding approach is still relevant. After 20 years.
Katja
On Tue, Oct 18, 2016 at 10:28 AM, cyrille henry ch@chnry.net wrote:
Le 18/10/2016 à 00:47, katja a écrit :
The filter recipe that Christof pointed to was easy to plug into the C code of [hip~] and works perfectly. But when looking further in d_filter.c I came across an approximation function 'sigbp_qcos()' used in the bandpass filter. It made me realize once more how passionate Miller is about efficiency. I'm not going to make a fool of myself by submitting a 'fix' using two trig functions to calculate a filter coefficient when a simple approximation could do the job. So that is what I'm now looking into, with help of a math friend: an efficient polynomial approximation for (1-sin(x))/cos(x).
according to libre office linear regression, for x between 0 and Pi, (1-sin(x))/cos(x) is about : -0.057255x³ + 0.27018x² - 0.9157x + 0.99344
the calc is in attachment, if you want to tune the input source or precision. cheers c
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list
Pd-list@lists.iem.at mailing list UNSUBSCRIBE and account-management -> https://lists.puredata.info/listinfo/pd-list