patrick.melet@dmradiocom.fr wrote:

> Hi all,
> 
> I have a little question...
> 
> Is there a trade-off beetween the cut-off frequency of an IIR filter
> and it's sampling frequency ?

It is the matter of the practical implementation.

> In practice I have noticed that we cannot have a relation more than Fs
> = 100 * Fc with 16 bits coeffcients...

It is possible, however the structure of the filter has to be modified 
for that purpose.

> Is there a theoritical proof ?

The gain of the recursive part of IIR filter is approximately 
Q*(Fs/Fc)^2. The precision problems are the direct consequence of that.



Vladimir Vassilevsky

DSP and Mixed Signal Design Consultant

http://www.abvolt.com

On 8 Mar, 15:02, patrick.me...@dmradiocom.fr wrote:
> Hi all,
>
> I have a little question...
>
> Is there a trade-off beetween the cut-off frequency of an IIR filter
> and it's sampling frequency ?

In principle: No.

> In practice

Yes. It has to do with imperfect numerical computations
in real-life CPUs. All number systems have more or less
inherent inaccuracy, depending on the number formats
they use.

> I have noticed that we cannot have a relation more than Fs
> = 100 * Fc with 16 bits coeffcients...
>
> Is there a theoritical proof ?

I would be surprised if there is. These things vary,
depending on the filter structure and a lot of other
factors. Apparently -- I haven't tried these things
myself -- ladder realizations are more accurate, in
certain respects, than naive direct form 1 or 2
structures, given the same numerical formats.

But now we are entering the field of expertice of
the seasoned DSP (as in specialized processor)
programmers...

Rune

Hi all,

I have a little question...

Is there a trade-off beetween the cut-off frequency of an IIR filter
and it's sampling frequency ?

In practice I have noticed that we cannot have a relation more than Fs
= 100 * Fc with 16 bits coeffcients...

Is there a theoritical proof ?

Thanks