renaudin wrote:
> Are you agreed that for a fixed filter's shape means fixed: cutoff
> frequency and width of transition band, the order of digital FIR filter
> depends upon the sampling frequency?

1. passband corner frequency (cutoff)
2. transition band width

there are other parameters, most notably the stop-band attenuation and
possible the passband ripple.

but, given a particular filter shape in log frequency (so the shape of
the filter is already determined, which means that the ratio of like
dimensioned quantities are constant - i.e. transition BW to passband
corner frequency), then the effective length of the impulse response,
in units of time, is inversely proportional to corner frequency and the
number of taps needed for the FIR will be equal to the product of that
impulse response length and the sampling frequency.

how to calculate the order needed in an FIR to do a particular job
(that is the performance parameters: passband corner frequency,
transition band width, stop-band attenuation, etc. are pre-determined
and fixed) depends on the type or "family" of FIR filter you're
designing.  the top candidates usually are windowed inverse DFT (often
using the Kaiser window), Parks-McClellan algorithm (often called
"remez"), and least-squares fit.  they all have different heuristic
fomulaes for your initial guess at the FIR length.  you make that
guess, use the computer to design your filter, check the frequency
response to see if it satisfies your requirements and, if not, lengthen
the FIR length and repeat or, if your requirements are way more than
satisfied, shorten the FIR length (checking the result after you do
this) and get fewer taps to get the job done.

that's the only way i really know how to optimally design the FIR
length.  i s'pose it can be automated and somebody has likely done
that, but i haven't.

r b-j

renaudin wrote:
> Are you agreed that for a fixed filter's shape means fixed: cutoff
> frequency and width of transition band, the order of digital FIR filter
> depends upon the sampling frequency?
> 
> As in the book scientists and engineers guide to digital signal processing
> by Steven W. Smith, he states that a rough estimation of the FIR filter's
> order in case of window (Hamming, Black man,...) implementation can be
> made as:
> 
> Filter order = 4 /Normalised width of transition band.
> 
> Can you suggest me some more accurate method to estimate the order of FIR
> filter depending upon the sampling frequency as long as the shape of
> filter is fixed.

Why do you want to know? Is your thirst for knowledge practical or 
theoretical?. In any case, the recipes for turtle soup and filters begin 
similarly. It's either "First, catch a turtle" or "First, pick a sample 
rate."

There are several good estimators for filter length, all of them 
heuristic. If your interest is theoretical, you might ask Grant Griffin 
which ones he uses and why. If it is practical, just get his ScopeFIR at 
  http://www.iowegian.com/ and let the program estimate it for you while 
calculating the coefficients. (Avoid http://www.gothic-iowegian.com/)

Jerry
-- 
Engineering is the art of making what you want from things you can get.
�����������������������������������������������������������������������

"renaudin" <alsaeed86@gmail.com> wrote in message 
news:NbGdnWyK4r7a-UHZnZ2dnUVZ_vednZ2d@giganews.com...
> Are you agreed that for a fixed filter's shape means fixed: cutoff
> frequency and width of transition band, the order of digital FIR filter
> depends upon the sampling frequency?
>
> As in the book scientists and engineers guide to digital signal processing
> by Steven W. Smith, he states that a rough estimation of the FIR filter's
> order in case of window (Hamming, Black man,...) implementation can be
> made as:
>
> Filter order = 4 /Normalised width of transition band.
>
> Can you suggest me some more accurate method to estimate the order of FIR
> filter depending upon the sampling frequency as long as the shape of
> filter is fixed.
>
> Thanks in advance for the tips.
>
> Renaudin

Mitra's book Handbook for Digital Signal Processing quotes Kaiser with a 
simple estimate:

N = [-20*log10(sqrt(dp*ds)) - 13]/[14.6(ws - wp)/2pi]

where dp and ds are the passband ripple peak and stopband ripple peak 
values, ws and wp are the stopband and passband edges - which are normalized 
to 2pi as you can see (where 2pi==fs).
So, N is a direct function of the ripple peaks and an inverse function of 
the normalized transition band width)

If dp and ds are both 0.01 and (ws-wp)/2pi = 0.1 then:

N = [-20*log10(sqrt(10^-4)) -13]/[1.46]
  = [-20*-2 -13]/1.46 = 27/1.46 = 18.5 or, rounded up, 19.

The rough rule I usually use is that the reciprocal of the transition band 
width is no less than the filter length in time (which is N/fs) .. if one 
expresses the transition band width in Hz.

N*(1/FS) = 1/(fs - fp) where FS is the sampling frequency and fs is the stop 
band edge.

Note that this is really independent of the sampling rate because it 
expresses length in terms of time.  Where the sampling rate comes in is when 
it's chosen and directly affects the number of samples in the filter impulse 
response.

For the same situation, choosing FS=1kHz, fs=400Hz and fp= 300Hz:

N = FS/(fs - fp) = 1000/(100) = 10

So, it looks like I've been off by a factor of 2 and should say:

N*(1/FS) = 2/(fs - fp)

N = 2*FS/(fs-fp) = 20

Of course, this can be normalized to FS so it's "independent" of FS (tho 
not):

If one is satisfied with a *fractional* transition band specification, then 
the filter length is independent of sample rate.  In that case, if the 
sample rate goes up, so does the transition band width.

So, part of the answer depends on the filter specification - is it expressed 
in Hz or are the band edges / transition width expressed in percent of total 
bandwidth / sample rate?

As Kaiser's formula points out, the order is related to the passband and 
stopband errors as well.  So this approximation is for the *minimum* length. 
And, if you assume that a symmetrical FIR filter (linear phase filter) 
requires roughly twice as many coefficients then 4*FS/(fs-fp) might be a 
reasonable approximation for a *minimum* length.

I think the real power of all this is in estimating a filter's likely length 
before going to a lot of trouble to design a filter that you don't want / 
can't use because the specifications push the length to be too large.  But 
then, designing filters isn't all that hard any more and I've always been 
satisfied to just see what comes out - even if I have to iterate on the 
length to get what I want.  One figures out pretty quickly if the filter is 
going to be too long.

Fred

renaudin wrote:
> Are you agreed that for a fixed filter's shape means fixed: cutoff
> frequency and width of transition band, the order of digital FIR filter
> depends upon the sampling frequency?

No, I don't agree with that. True, I am splitting hairs now, but one
is usually given a sampling frequency, then chooses cut-off
requencies etc, and have to find a filter order that makes everything
work.

The difference from what you said is subtle: The way you phrased
your statement might give the impression that one may decide on
a filter order and then choose a sampling frequency. It is the other
way around.

> As in the book scientists and engineers guide to digital signal processing
> by Steven W. Smith, he states that a rough estimation of the FIR filter's
> order in case of window (Hamming, Black man,...) implementation can be
> made as:
>
> Filter order = 4 /Normalised width of transition band.

Yep, finding the order if FIR filters is not at all an exact science.

> Can you suggest me some more accurate method to estimate the order of FIR
> filter depending upon the sampling frequency as long as the shape of
> filter is fixed.

There were a few papers on some experimental results about how
to estimate filter orders in the Parks-McClellan algorithm. I *think*
Rabiner was one of the authors, but I am not sure. These papers
ought to have been published some time in the early/mid '70s.

Rune

Are you agreed that for a fixed filter's shape means fixed: cutoff
frequency and width of transition band, the order of digital FIR filter
depends upon the sampling frequency?

As in the book scientists and engineers guide to digital signal processing
by Steven W. Smith, he states that a rough estimation of the FIR filter's
order in case of window (Hamming, Black man,...) implementation can be
made as:

Filter order = 4 /Normalised width of transition band.

Can you suggest me some more accurate method to estimate the order of FIR
filter depending upon the sampling frequency as long as the shape of
filter is fixed.

Thanks in advance for the tips.

Renaudin