LPF butterworth filter question
in [DSP]
|
From: Jerry Avins on 18 May 2010 12:22 On 5/18/2010 5:44 AM, Rune Allnor wrote: > On 17 Mai, 10:24, zs<zsolt.garamvol...(a)gmail.com> wrote: > >> This seems to be an implementation issue of sptool. > > No, it isn't. It's the quantization noise associated with > double precision floating point numbers. > >> I've designed a >> filter with the same specification using fdatool, and it's correct (no >> noisy response below -350 dB). > > Are you sure? > > It might just as well be that the matlab's support team were so > pissed off by incompetent users, who asked this very question, > that they 'doctored' the spectrum display tool such that it > shows a smooth curve. > > Find the coeffs of the filter and hack up your own spectrum > display from raw matlab, *not* using any of the purpose-made > tools that come with the toolboxes. It seems to me that the whole thing is a crock. How much numerical precision (to say nothing of physical isolation) would one need to get 350 dB suppression with an umpty-ump stage filter? The fuzz is arithmetic noise. (What does "single section" mean in the spec?) Jerry -- "I view the progress of science as ... the slow erosion of the tendency to dichotomize." --Barbara Smuts, U. Mich. �����������������������������������������������������������������������
From: zs on 19 May 2010 02:25 On máj. 18, 11:44, Rune Allnor <all...(a)tele.ntnu.no> wrote: > On 17 Mai, 10:24, zs <zsolt.garamvol...(a)gmail.com> wrote: > > > This seems to be an implementation issue of sptool. > > No, it isn't. It's the quantization noise associated with > double precision floating point numbers. > I intended to say that the filter response calculation is implemented with insufficient precision. > > I've designed a > > filter with the same specification using fdatool, and it's correct (no > > noisy response below -350 dB). > > Are you sure? > > It might just as well be that the matlab's support team were so > pissed off by incompetent users, who asked this very question, > that they 'doctored' the spectrum display tool such that it > shows a smooth curve. > This might be the case but doesn't sound very realistic to me. Matlab has support for variable precision arithmetic. If I were at Mathworks I would use that for correct response calculation instead of some 'symptomatic treatment'. > Find the coeffs of the filter and hack up your own spectrum > display from raw matlab, *not* using any of the purpose-made > tools that come with the toolboxes. > To further increase credibility, the response calculation could be tested by implementing it on a different platform (C/C++ or a custom digital circuit). However, the 1-bit information of deciding whether fdatool is correct under -350 dB does not sound very motivating to me. > Rune
From: Rune Allnor on 19 May 2010 13:37 On 19 Mai, 08:25, zs <zsolt.garamvol...(a)gmail.com> wrote: > On máj. 18, 11:44, Rune Allnor <all...(a)tele.ntnu.no> wrote: > > > On 17 Mai, 10:24, zs <zsolt.garamvol...(a)gmail.com> wrote: > > > > This seems to be an implementation issue of sptool. > > > No, it isn't. It's the quantization noise associated with > > double precision floating point numbers. > > I intended to say that the filter response calculation is implemented > with insufficient precision. Do you have the slightest idea what kind of dynamic range 350 dB represents? > > > I've designed a > > > filter with the same specification using fdatool, and it's correct (no > > > noisy response below -350 dB). > > > Are you sure? > > > It might just as well be that the matlab's support team were so > > pissed off by incompetent users, who asked this very question, > > that they 'doctored' the spectrum display tool such that it > > shows a smooth curve. > > This might be the case but doesn't sound very realistic to me. Incompetent users is a fact. Support teams being pissed off by incompetent users is a fact. Developers hiding trivial insignificants from incompetent users in order to ease the pressure on the support teams is a fact. Especially if the same people who develop the tools have to face the incompetent users for support. > Matlab > has support for variable precision arithmetic. If I were at Mathworks > I would use that for correct response calculation instead of some > 'symptomatic treatment'. You aren't at matlab. The comments you have made on the issue suggests that you are an incompetent user of matlab. > > Find the coeffs of the filter and hack up your own spectrum > > display from raw matlab, *not* using any of the purpose-made > > tools that come with the toolboxes. > > To further increase credibility, the response calculation could be > tested by implementing it on a different platform (C/C++ or a custom > digital circuit). However, the 1-bit information of deciding whether > fdatool is correct under -350 dB does not sound very motivating to me. Again, you seem not to have the slightest clue what you are talking about. Rune
From: zs on 21 May 2010 02:59 On máj. 19, 19:37, Rune Allnor <all...(a)tele.ntnu.no> wrote: > On 19 Mai, 08:25, zs <zsolt.garamvol...(a)gmail.com> wrote: > > > On máj. 18, 11:44, Rune Allnor <all...(a)tele.ntnu.no> wrote: > > > > On 17 Mai, 10:24, zs <zsolt.garamvol...(a)gmail.com> wrote: > > > > > This seems to be an implementation issue of sptool. > > > > No, it isn't. It's the quantization noise associated with > > > double precision floating point numbers. > > > I intended to say that the filter response calculation is implemented > > with insufficient precision. > > Do you have the slightest idea what kind of dynamic range 350 dB > represents? > > > > > I've designed a > > > > filter with the same specification using fdatool, and it's correct (no > > > > noisy response below -350 dB). > > > > Are you sure? > > > > It might just as well be that the matlab's support team were so > > > pissed off by incompetent users, who asked this very question, > > > that they 'doctored' the spectrum display tool such that it > > > shows a smooth curve. > > > This might be the case but doesn't sound very realistic to me. > > Incompetent users is a fact. Support teams being pissed off by > incompetent users is a fact. Developers hiding trivial insignificants > from incompetent users in order to ease the pressure on the > support teams is a fact. Especially if the same people who > develop the tools have to face the incompetent users for support. > > > Matlab > > has support for variable precision arithmetic. If I were at Mathworks > > I would use that for correct response calculation instead of some > > 'symptomatic treatment'. > > You aren't at matlab. The comments you have made on the issue > suggests > that you are an incompetent user of matlab. > > > > Find the coeffs of the filter and hack up your own spectrum > > > display from raw matlab, *not* using any of the purpose-made > > > tools that come with the toolboxes. > > > To further increase credibility, the response calculation could be > > tested by implementing it on a different platform (C/C++ or a custom > > digital circuit). However, the 1-bit information of deciding whether > > fdatool is correct under -350 dB does not sound very motivating to me. > > Again, you seem not to have the slightest clue what you > are talking about. > > Rune I'm not talking about filter realization. From that viewpoint, I agree, that the response below, say, -100 dB is indifferent for most applications due to thermal noise, error of the ADCs, etc. Even if we wouldn't have this dynamic range limit stemming from the analog world, we still have finite word length effects, and why would someone need 350 dB suppression in the passband...and so on. What I'm talking about is filter design and visualization: when designing a filter, I think it's useful to see its actual response first, without taking 'secondary' effects into account. E.g, the (visualized) response being corrupted below -150 dB won't imply any problems during realization, but having a correct response in a larger dynamic range gives insight on the filter design process: the response shown by fdatool indicates that the filter is an analog Butterworth filter digitized using bilinear transform. Zsolt
From: Rune Allnor on 21 May 2010 05:38
On 21 Mai, 08:59, zs <zsolt.garamvol...(a)gmail.com> wrote: > On máj. 19, 19:37, Rune Allnor <all...(a)tele.ntnu.no> wrote: > > Again, you seem not to have the slightest clue what you > > are talking about. > What I'm talking about is filter design and visualization: when > designing a filter, I think it's useful to see its actual response > first, without taking 'secondary' effects into account. That's utter and total BS coming from somebody who haven't got the slighets clue what he is talking about. How do you think this works? Somebody *first* computes the 'true' response as would be seen in a perfect world and *then* adds some quantization effects and noise just for the heck of it? The finite precision effects are there because there are no way to avoid them in a digital computer. It's as simple as that. Rune |