kareem wrote:
> Hello.
> 
> Generally we do up sampling by inserting zeros and passing through Image
> rejection filter. Instead of inserting zeros we can duplicate the samples
> and pass through a filter. the advantage of going for second mathod is i
> can choose a less order filter compared to the first methode. similarly is
> it possible to do the same methode for sampling rate conversion by non
> integer factor. if so what is is the cut off frequency of the filter and
> which design  method i should fallow.
> 
> -kareem
> 
> 

This was covered in depth recently.  The apparent economy that this
technique offers breaks down quickly on close inspection.

http://groups.google.com/group/comp.dsp/browse_frm/thread/e44b4a9e529996fc/05a740edfab354cb?tvc=1&q=comp.dsp+lyons+upsampling+zeros+hold#05a740edfab354cb
or
http://tinyurl.com/eqqax


-- Mark B

kareem wrote:
> Hello.
>
> Generally we do up sampling by inserting zeros and passing through Image
> rejection filter. Instead of inserting zeros we can duplicate the samples
> and pass through a filter. the advantage of going for second mathod is i
> can choose a less order filter compared to the first methode.

That is true. You can view the duplicating samples as the action of an
FIR with kernel

h1 = [1 1]

on the zero-inserted upsampled sequence. h is a lowpass filter with a
zero at Nyquist. Notice that you have -3dB gain at Nyquist/2, which
means that you affect your original signal as well as attenuate the
images.

Now assume that the second filter which gets rid of the rest of the
images is h2, and it has length N. Because the upsampling with
duplication and filtering occur in series, you might as well upsample
by zero-inserting and filter the the resulting stream with h3 = h1 *
h2, the resulting output the same. The length of h3 is N+1. When you
compute the output of h3, you only need to calculate (N+1)/2 MACs,
because every second value in the filter state delay is equal to zero.
So you have:

1. Cost for upsampling with duplicating samples and filtering with h2:
N multiply/accumulates.
2. Cost for upsampling by zero-inserting and filtering with h3: (N+1)/2
multiply/accumulates.

For N > 2, the second way produces less computation cost - you don't
save anything by duplicating samples! This and the fact the h1 has
considerable attenuation in the passband is the reason for using
upsampling by zero-inserting.

Regards,
Andor

Hello.

Generally we do up sampling by inserting zeros and passing through Image
rejection filter. Instead of inserting zeros we can duplicate the samples
and pass through a filter. the advantage of going for second mathod is i
can choose a less order filter compared to the first methode. similarly is
it possible to do the same methode for sampling rate conversion by non
integer factor. if so what is is the cut off frequency of the filter and
which design  method i should fallow.

-kareem