Interpolation response overshoot, why?
in [DSP]
|
Prev: Possible new faster algorithms for FFT of 3^n and 6^n sizes.
Next: conventional wisdom how to upsample very large arrays, accurately?
From: Vladimir Vassilevsky on 6 Jan 2010 10:04 Laron wrote: > Hi, > When simulate the FIR filter response, run interp(Matrix,n) in matlab, > the maximum of Matrix is 1,but the response is larger than 1? > I wonder know why this could be happen and how to degrade this > effect? Even if the original data is limited by X, the interpolated data can be higher then X. And this is correct. Example Original data: 1 2 2 1 2x interpolated data: 0.5 1 1.5 2 2.5 2 1.5 1 0.5 Vladimir Vassilevsky DSP and Mixed Signal Design Consultant http://www.abvolt.com
From: glen herrmannsfeldt on 6 Jan 2010 14:28 Tim Wescott <tim(a)seemywebsite.com> wrote: (snip) > Reconstruction filters can ring. Filters that ring, > when presented with inputs with sharp edges, can ring > strongly enough to exceed the bounds of the input. That is true, but there are also sampled values that, when reconstructed, result in a sine with amplitude greater than full scale of the D/A conversion. Easiest to see is the signal with period of four samples [ 1 1 -1 -1 ]. If the source is analog data through an A/D conversion then that isn't so likely, though. -- glen
From: Jerry Avins on 6 Jan 2010 15:07 Mark wrote: > On Jan 6, 3:19 am, "Laron" <jason.pi...(a)inbox.com> wrote: >>> At a guess, Matlab is using a stiff bandpass filter and you're seeing >>> Gibbs phenomenon. Does Matrix have sharp edges? >>> Read the fine manual, see if you can chose a different filter for >> interp. >> >>> -- >>> www.wescottdesign.com >> Matrix do have sharp edges like [1 0 1 1 0 1] etc, for rasied cosine >> filter application. >> >> Thanks. > > search around over at rec.audio.pro. > > there have been discussions in the past about how/why a reconstructed > waveform can exceed the max digital value. It's obvious to me. Any waveform so sampled that the peaks are missed will have a peak reconstruction that exceeds any sample. For example, consider the sequence 0,1,1,0,-1,-1,,.... That's a sine of amplitude 1.55 -- 2/sqrt(3) -- sampled every 60 degrees. Jerry -- Engineering is the art of making what you want from things you can get. �����������������������������������������������������������������������
From: glen herrmannsfeldt on 6 Jan 2010 18:04 Jerry Avins <jya(a)ieee.org> wrote: (snip on amplitude of reconstructed waveforms) > It's obvious to me. Any waveform so sampled that the peaks are missed > will have a peak reconstruction that exceeds any sample. For example, > consider the sequence 0,1,1,0,-1,-1,,.... That's a sine of amplitude > 1.55 -- 2/sqrt(3) -- sampled every 60 degrees. Yes, but for sampling an actual signal you have to be very lucky to have that happen, and not actually clip. Well, I once had a recording of a concert that clipped at five samples. (I did't check to see if they were contiguous or not.) So I would say that in that case I was already lucky (not to have more clipped) and so there could have been some where the peak was between samples as mentioned. One could take actual CD recordings and do what the OP suggests and find where the peaks are. Then we would know... -- glen
From: Laron on 6 Jan 2010 18:40
>On Wed, 06 Jan 2010 02:19:55 -0600, Laron wrote: > >>>At a guess, Matlab is using a stiff bandpass filter and you're seeing >>>Gibbs phenomenon. Does Matrix have sharp edges? >>> >>>Read the fine manual, see if you can chose a different filter for >> interp. >>> >>>-- >>>www.wescottdesign.com >>> >> Matrix do have sharp edges like [1 0 1 1 0 1] etc, for rasied cosine >> filter application. >> >> Thanks. > >Reconstruction filters can ring. Filters that ring, when presented with >inputs with sharp edges, can ring strongly enough to exceed the bounds of >the input. > >If you understand how the Matlab function works then you will know >whether or not you can change the properties of the reconstruction filter >it uses. If you don't, well, Matlab's documentation should tell you. > >If you know how to change Matlab's reconstruction filter, you're home >free. > >If you don't, then if you understand how the reconstruction process works >it's pretty easy to do the whole thing yourself, with whatever filter you >want to use. > >Do you understand how reconstruction works? > >-- >www.wescottdesign.com > I just know there would be a lpf, not quite sure about the reconstruction process. I got an idea that the overshoot caused from the lpf, but the detail "how" is still not clear. Thanks. Thanks. |