Filters

Hi -

My question is about filter functionality.
My understanding is that filters work in the frequency domain.
So the typical discrete time domain samples have to be first converted into
frequency domain and then passed through the filter to get the right
filtered output.
Is my understanding correct?

Or the filters themselves inherently convert time domain samples to
frequency domain and then perform filtering?

Thanks, 

On Aug 12, 11:42&#4294967295;am, "manishp" <manishp.p18@n_o_s_p_a_m.gmail.com>
wrote:
> Hi -
>
> My question is about filter functionality.
> My understanding is that filters work in the frequency domain.
> So the typical discrete time domain samples have to be first converted into
> frequency domain and then passed through the filter to get the right
> filtered output.
> Is my understanding correct?

No.

> Or the filters themselves inherently convert time domain samples to
> frequency domain and then perform filtering?

No. Time domain and frequency domain are alternate representations of
the same data. No conversion is needed. A signal, say an audio signal,
is a stream of information. (Think of the squiggles in the groove of a
phonograph record. Remember those?) Filtering if can change the
relative strengths of various frequencies, but in order to hear the
result, a stream of information must be delivered to the loudspeaker.
Stream.1-->Filter-->Stream.2 It's that simple.

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

Jerry

Jerry Avins <jya@ieee.org> wrote:
> On Aug 12, 11:42&#4294967295;am, "manishp" <manishp.p18@n_o_s_p_a_m.gmail.com>

>> My question is about filter functionality.
>> My understanding is that filters work in the frequency domain.
>> So the typical discrete time domain samples have to be first 
>> converted into frequency domain and then passed through the 
>> filter to get the right filtered output.
>> Is my understanding correct?

> No.

Well, there are some that do FFT, filter, IFFT, but that
is pretty rare.

In describing analog RLC filters (the first two comments didn't
mention sampled data) it seems a little more obvious to describe
as working on the frequency domain, but the R, L, and C don't
know that.

>> Or the filters themselves inherently convert time domain samples to
>> frequency domain and then perform filtering?

> No. Time domain and frequency domain are alternate representations of
> the same data. No conversion is needed. A signal, say an audio signal,
> is a stream of information. (Think of the squiggles in the groove of a
> phonograph record. Remember those?) Filtering if can change the
> relative strengths of various frequencies, but in order to hear the
> result, a stream of information must be delivered to the loudspeaker.
> Stream.1-->Filter-->Stream.2 It's that simple.

The analysis of filters is often done in the frequency domain,
but the actual filtering usually is done in the time domain.

-- glen

On Fri, 12 Aug 2011 10:42:51 -0500, manishp wrote:

> Hi -
> 
> My question is about filter functionality. My understanding is that
> filters work in the frequency domain. So the typical discrete time
> domain samples have to be first converted into frequency domain and then
> passed through the filter to get the right filtered output.
> Is my understanding correct?
> 
> Or the filters themselves inherently convert time domain samples to
> frequency domain and then perform filtering?

Your understanding is incorrect.

Everything that happens in this crazy old Universe of ours happens in the 
time domain (or at least our perceptions are solidly fixed in the time 
domain).  Even though physicists can't explain time, we're still stuck 
with (and in) it.

So all filters must operate in the time domain.

Filters that happen to be _linear_ (or, more accurately, happen to behave 
_as if_ they are linear) are much easier to describe, specify and design 
in terms of the frequency domain -- so you see much attention paid to 
frequency-domain characteristics of filters.  But the frequency domain is 
just a tool that we use to understand the behavior of linear systems -- 
not someplace that we can buy airplane tickets to and visit.

Let's try that again: Filters that happen to behave as if they are linear 
systems can be analyzed, etc., using the Fourier transform and its 
relatives.  These are frequency domain techniques, and they are nothing 
more than convenient ways to solve linear differential (or difference) 
equations.  _Nothing_ more.  Even though that "nothing more" is plenty of 
"more", it's still nothing more than that.  No alternate realities, no 
deep philosophical significance, just a really deep and wide bag of 
really useful tricks for solving differential/difference equations.  But 
still at root -- just a bag of tricks.

-- 
www.wescottdesign.com

On Fri, 12 Aug 2011 17:23:00 +0000, glen herrmannsfeldt wrote:

> Jerry Avins <jya@ieee.org> wrote:
>> On Aug 12, 11:42&nbsp;am, "manishp" <manishp.p18@n_o_s_p_a_m.gmail.com>
> 
>>> My question is about filter functionality. My understanding is that
>>> filters work in the frequency domain. So the typical discrete time
>>> domain samples have to be first converted into frequency domain and
>>> then passed through the filter to get the right filtered output. Is my
>>> understanding correct?
> 
>> No.
> 
> Well, there are some that do FFT, filter, IFFT, but that is pretty rare.

Even those, while heavily couched in frequency-domain terms, are still 
being carried out in the time domain.

>>> snip the rest <<<

-- 
www.wescottdesign.com

Tim Wescott wrote:


> Let's try that again: Filters that happen to behave as if they are linear 
> systems can be analyzed, etc., using the Fourier transform and its 
> relatives.

What is the Fourier transform of a coffee filter or an IP address filter?

Thank you all.
To be frank, I had some inkling that transformtation is not needed and that
is the reason why I asked this question.

But can you tell me cases where the transformation is a must.

Currently, my colleagues are working on an algorithm which does fft
transformation, not once but twice. I don't remember specifics since I am
not directly involved in that work.

Thank you very much ...

On 8/12/2011 10:59 AM, Vladimir Vassilevsky wrote:
>
>
> Tim Wescott wrote:
>
>
>> Let's try that again: Filters that happen to behave as if they are
>> linear systems can be analyzed, etc., using the Fourier transform and
>> its relatives.
>
> What is the Fourier transform of a coffee filter or an IP address filter?
>

A coffee filter is a 2 dimensional highpass filter.  It blocks large 
particles (grounds) and passes small ones (oils).

-- 
Rob Gaddi, Highland Technology
Email address is currently out of order

On Fri, 12 Aug 2011 12:59:00 -0500, Vladimir Vassilevsky wrote:

> Tim Wescott wrote:
> 
> 
>> Let's try that again: Filters that happen to behave as if they are
>> linear systems can be analyzed, etc., using the Fourier transform and
>> its relatives.
> 
> What is the Fourier transform of a coffee filter or an IP address
> filter?

I don't think that either of those filters happen to behave as linear 
systems.

-- 
www.wescottdesign.com

On Fri, 12 Aug 2011 13:08:15 -0500, manishp wrote:

> Thank you all.
> To be frank, I had some inkling that transformtation is not needed and
> that is the reason why I asked this question.
> 
> But can you tell me cases where the transformation is a must.

Probably never.

> Currently, my colleagues are working on an algorithm which does fft
> transformation, not once but twice. I don't remember specifics since I
> am not directly involved in that work.

There are some cases where filtering is more convenient if you do a FFT 
into the 'frequency domain', do a point-by-point multiply, then do an IFFT 
back into the time domain.

Even there, however, you and your equipment are still stuck in the time 
domain, doing time domain operations on data that persists to exist in 
the time domain.

-- 
www.wescottdesign.com