On Wed, 03 Dec 2014 15:18:25 -0600, Tim Wescott <seemywebsite@myfooter.really> wrote:>On Wed, 03 Dec 2014 04:55:39 -0800, Rick Lyons wrote: > >> Greetings Earthlings, >> Every now and then I run across a paper in >> the DSP literature that is SO different, >> so "odd", that I don't know how to evaluate the technical material in >> the paper. "Is the author exceedingly clever, or an unfortunate victim >> of wild misconceptions?" >> >> When you have nothing better to do, I'd sure like to hear anyone's >> opinion on the validity of the DSP paper found at: >> >> http://jocoleman.info/pubs/papers/Spew2K/Spew.pdf >> >> Thanks, >> [-Rick-] > >If he lectures as he writes, I'm very glad that I never had to take a >class from him, or be his teaching assistant. > >The points that I see here are: > >1: What's so new about this approach? With the exception of how he's >weighting his impulses, this looks like how I was taught DSP. > >2: The writing is exceedingly terse. Way terse. He's got at least 20 >pages of concepts that he's trying to present in 5 pages, and he's dealt >with the issue by removing four paragraphs out of five. > >3: The writing is disorganized. This is perhaps a natural result of it >being terse, but -- damn I'm glad I never took a class from this guy. > >So to sum up: > >The only divergence that I see here from normal practice is how he weights >his impulses. He does not even attempt to show how his "wonderful new" >method relates to the existing chain-o-impulses method of teaching DSP, >nor does he, in my mind, do a good job of showing how this makes multi- >rate DSP so loverly. I think I just wasted 20 minutes of my life for you, >Rick, and now that I'm over 50 I'm keeping track of that sort of thing!Hi Tim, Thanks for your thoughts. If you were able to understand the author's Figure 1 then you're a better man than I. [1] I could not figure out what the word "realization" meant. [2] I also had trouble with the x(t) equation at the bottom of the left column on the first page. The author wrote: x(t) = x(t)*e^(j*2*pi*t/T). How can he have x(t) on both sides of the equation? [3] What really bothered me was where the "Decimation" section of the paper implied that time-domain decimation is equivalent to frequency-domain convolution. THAT I believe is simply not true. Anyway Tim, thanks again. As for your 20 minutes, sometime I'll stop by your place and mow your front lawn as repayment. [-Rick-]
Looking for opinions on a DSP paper
Started by ●December 3, 2014
Reply by ●December 4, 20142014-12-04
Reply by ●December 4, 20142014-12-04
On Wed, 03 Dec 2014 23:53:10 GMT, eric.jacobsen@ieee.org (Eric Jacobsen) wrote: [Snipped by Lyons]>So far I've read the abstract, intro, and conclusion, and just from >those I can tell you that it is very, very poorly written. There's >almost no info in any of those about what he wants to say about the >topic.Hi Eric, I read the Abstract carefully, twice. It was far too cryptic for me to understand. Thanks, [-Rick-]
Reply by ●December 4, 20142014-12-04
On Wed, 3 Dec 2014 23:32:03 +0000 (UTC), glen herrmannsfeldt <gah@ugcs.caltech.edu> wrote:>Rick Lyons <R.Lyons@_bogus_ieee.org> wrote: > >> Greetings Earthlings, >> Every now and then I run across a paper in >> the DSP literature that is SO different, >> so "odd", that I don't know how to evaluate >> the technical material in the paper. "Is the >> author exceedingly clever, or an unfortunate >> victim of wild misconceptions?" > >(snip) > >> http://jocoleman.info/pubs/papers/Spew2K/Spew.pdf > >Not that I claim to understand all of it, but he starts out >showing that a discrete time signal can be considered a train >of appropriate amplitude delta functions in continuous time. > >I think most of us know that in the back of our head, but don't >think about it often. > >Mostly we like to keep our continuous time signals and discrete >time signals separate, and not think of them at the same time. > >In the recent discussion about the maximum reconstructed >amplitude from a sampled signal, I thought again about how the >result at any time depends on samples from -infinity to +infinity. > >That is obvious from the Fourier transform with the same limits >on the integral, but again, we don't think about it so often. > >Consider a sampled signal that is zero for all times before t=0, >then +1 for zero and all even samples after t=0, and -1 for all >odd samples after t=0. > >If I have it right, there is a huge peak a little before zero. > >But OK, back to the paper. > >He starts out explaining that you can think of sampled signals >in continuous time, and then goes on to show some examples of >problems that might be hard in sampled time, but easier in >continuous time. > >If you think of DSP as sample band-limited analog signal at a uniform >sample spacing, process it, convert back to analog signal, then >it probably doesn't help. > >I think he shows that some resampling problems are easier to think >of as trains of delta functions in continuous time, instead of just >as samples it time.Hi Glen, Unless I'm missing something significant, I don't believe there exists any way to think of decimating a discrete-time sequence of sample values by a factor of two as any kind of continuous-time operation. It seems to me that continuous-time operations can only be performed on continuous-time (analog) signals. Thanks, [-Rick-]
Reply by ●December 4, 20142014-12-04
Rick Lyons <R.Lyons@_bogus_ieee.org> wrote: (snip)> [2] I also had trouble with the x(t) equation > at the bottom of the left column on the first > page. The author wrote:> x(t) = x(t)*e^(j*2*pi*t/T).> How can he have x(t) on both sides of the > equation?Hmm, that is an interesting way to say it. I believe he is comparing continuous and sampled data. That seems to be the complicated way of saying that t/T is an integer. Seems to me that being able to think in both sampled and continuous time, even for the same problem and at (almost) the same time is useful. For me, I have a reasonably easy time thinking about Fourier transforms and the frequency domain, but a much harder time thinking about z-transforms and IIR filters. I can't look at the equation for an IIR or FIR filter and imagine the frequency response of the filter. I presume some here can do that. I suspect that getting students used to thinking in both domains early helps them get used to thinking about there problems. -- glen
Reply by ●December 4, 20142014-12-04
On Thu, 04 Dec 2014 14:56:08 -0800, Rick Lyons wrote:> On Wed, 3 Dec 2014 23:32:03 +0000 (UTC), glen herrmannsfeldt > <gah@ugcs.caltech.edu> wrote: > >>Rick Lyons <R.Lyons@_bogus_ieee.org> wrote: >> >>> Greetings Earthlings, >>> Every now and then I run across a paper in >>> the DSP literature that is SO different, >>> so "odd", that I don't know how to evaluate the technical material in >>> the paper. "Is the author exceedingly clever, or an unfortunate >>> victim of wild misconceptions?" >> >>(snip) >> >>> http://jocoleman.info/pubs/papers/Spew2K/Spew.pdf >> >>Not that I claim to understand all of it, but he starts out showing that >>a discrete time signal can be considered a train of appropriate >>amplitude delta functions in continuous time. >> >>I think most of us know that in the back of our head, but don't think >>about it often. >> >>Mostly we like to keep our continuous time signals and discrete time >>signals separate, and not think of them at the same time. >> >>In the recent discussion about the maximum reconstructed amplitude from >>a sampled signal, I thought again about how the result at any time >>depends on samples from -infinity to +infinity. >> >>That is obvious from the Fourier transform with the same limits on the >>integral, but again, we don't think about it so often. >> >>Consider a sampled signal that is zero for all times before t=0, then +1 >>for zero and all even samples after t=0, and -1 for all odd samples >>after t=0. >> >>If I have it right, there is a huge peak a little before zero. >> >>But OK, back to the paper. >> >>He starts out explaining that you can think of sampled signals in >>continuous time, and then goes on to show some examples of problems that >>might be hard in sampled time, but easier in continuous time. >> >>If you think of DSP as sample band-limited analog signal at a uniform >>sample spacing, process it, convert back to analog signal, then it >>probably doesn't help. >> >>I think he shows that some resampling problems are easier to think of as >>trains of delta functions in continuous time, instead of just as samples >>it time. > > Hi Glen, > Unless I'm missing something significant, > I don't believe there exists any way to think of decimating a > discrete-time sequence of sample values by a factor of two as any kind > of continuous-time operation. > > It seems to me that continuous-time operations can only be performed on > continuous-time (analog) > signals.There are some aspects of sampling that only seem to make sense if you replace the Dirac deltas with functions of area = 1 but finite extent in time, and then take things in the limit as that time extent goes to zero. I'm sure that with careful application of the above, you could shoe-horn sub-sampling into that model. But -- why? -- Tim Wescott Wescott Design Services http://www.wescottdesign.com
Reply by ●December 4, 20142014-12-04
On Thu, 04 Dec 2014 13:54:46 -0800, Rick Lyons wrote:> On Wed, 03 Dec 2014 15:18:25 -0600, Tim Wescott > <seemywebsite@myfooter.really> wrote: > >>On Wed, 03 Dec 2014 04:55:39 -0800, Rick Lyons wrote: >> >>> Greetings Earthlings, >>> Every now and then I run across a paper in >>> the DSP literature that is SO different, >>> so "odd", that I don't know how to evaluate the technical material in >>> the paper. "Is the author exceedingly clever, or an unfortunate >>> victim of wild misconceptions?" >>> >>> When you have nothing better to do, I'd sure like to hear anyone's >>> opinion on the validity of the DSP paper found at: >>> >>> http://jocoleman.info/pubs/papers/Spew2K/Spew.pdf >>> >>> Thanks, >>> [-Rick-] >> >>If he lectures as he writes, I'm very glad that I never had to take a >>class from him, or be his teaching assistant. >> >>The points that I see here are: >> >>1: What's so new about this approach? With the exception of how he's >>weighting his impulses, this looks like how I was taught DSP. >> >>2: The writing is exceedingly terse. Way terse. He's got at least 20 >>pages of concepts that he's trying to present in 5 pages, and he's dealt >>with the issue by removing four paragraphs out of five. >> >>3: The writing is disorganized. This is perhaps a natural result of it >>being terse, but -- damn I'm glad I never took a class from this guy. >> >>So to sum up: >> >>The only divergence that I see here from normal practice is how he >>weights his impulses. He does not even attempt to show how his >>"wonderful new" method relates to the existing chain-o-impulses method >>of teaching DSP, nor does he, in my mind, do a good job of showing how >>this makes multi- rate DSP so loverly. I think I just wasted 20 minutes >>of my life for you, >>Rick, and now that I'm over 50 I'm keeping track of that sort of thing! > > Hi Tim, > Thanks for your thoughts. If you were > able to understand the author's Figure 1 then you're a better man than > I. > > [1] I could not figure out what the word "realization" meant."Making real", or "thing made real". As in, he defines a signal as an abstract thing, and it's "realization" as the actual measured thing (such as a voltage).> [2] I also had trouble with the x(t) equation at the bottom of the left > column on the first page. The author wrote: > > x(t) = x(t)*e^(j*2*pi*t/T). > > How can he have x(t) on both sides of the equation?I didn't even get into the math, I was just looking at what he was saying. You're making me waste more of my life. Fortunately, since the recent reclamation of my front yard from the blackberry patch, mowing it takes more than 20 minutes.> [3] What really bothered me was where the "Decimation" section of the > paper implied that time-domain decimation is equivalent to > frequency-domain convolution. THAT I believe is simply not true.I couldn't find that particular part of the paper. However, multiplying two signals together in the time domain is the same as convolving their frequency-domain counterparts, and you can easily model decimation as multiplication by a pulse train of zeros and ones. I'm sure that's where whatever claim he's making comes from.> Anyway Tim, thanks again. As for your 20 minutes, sometime I'll stop by > your place and mow your front lawn as repayment.I'll hold you to that. -- Tim Wescott Wescott Design Services http://www.wescottdesign.com
Reply by ●December 4, 20142014-12-04
On Thu, 04 Dec 2014 14:14:26 -0800, Rick Lyons wrote:> On Wed, 03 Dec 2014 23:53:10 GMT, eric.jacobsen@ieee.org (Eric Jacobsen) > wrote: > > [Snipped by Lyons] > >>So far I've read the abstract, intro, and conclusion, and just from >>those I can tell you that it is very, very poorly written. There's >>almost no info in any of those about what he wants to say about the >>topic. > > Hi Eric, > I read the Abstract carefully, twice. > It was far too cryptic for me to understand.I can read his abstract and think I have a clear idea of what he's thinking (even if I don't agree with it). So, is that a sign of superior mental processes, or impending insanity? -- Tim Wescott Wescott Design Services http://www.wescottdesign.com
Reply by ●December 4, 20142014-12-04
On Thu, 4 Dec 2014 23:02:36 +0000 (UTC), glen herrmannsfeldt <gah@ugcs.caltech.edu> wrote:>Rick Lyons <R.Lyons@_bogus_ieee.org> wrote: > >(snip) > >> [2] I also had trouble with the x(t) equation >> at the bottom of the left column on the first >> page. The author wrote: > >> x(t) = x(t)*e^(j*2*pi*t/T). > >> How can he have x(t) on both sides of the >> equation? > >Hmm, that is an interesting way to say it.Hi glen, The author's continued failure to define the nature of the variables in his equations (whether those variables are continuous or discrete numbers) means we could debate the meaning of his material for a long time. I thought his 't' was an integer. But Gosh, who knows. [-Rick-]>I believe he is comparing continuous and sampled data. > >That seems to be the complicated way of saying that t/T >is an integer. > >Seems to me that being able to think in both sampled >and continuous time, even for the same problem and at >(almost) the same time is useful. > >For me, I have a reasonably easy time thinking about Fourier >transforms and the frequency domain, but a much harder time >thinking about z-transforms and IIR filters. > >I can't look at the equation for an IIR or FIR filter and >imagine the frequency response of the filter. I presume some >here can do that. > >I suspect that getting students used to thinking in both domains >early helps them get used to thinking about there problems. > >-- glen
Reply by ●December 4, 20142014-12-04
On 12/4/14 4:54 PM, Rick Lyons wrote:> > > [2] I also had trouble with the x(t) equation > at the bottom of the left column on the first > page. The author wrote: > > x(t) = x(t)*e^(j*2*pi*t/T). > > How can he have x(t) on both sides of the > equation? >maybe t is always an integer multiple of T. if that's the case, he should say so. -- r b-j rbj@audioimagination.com "Imagination is more important than knowledge."
Reply by ●December 5, 20142014-12-05
On Thu, 04 Dec 2014 17:04:08 -0600, Tim Wescott <seemywebsite@myfooter.really> wrote:>On Thu, 04 Dec 2014 14:56:08 -0800, Rick Lyons wrote: > >> On Wed, 3 Dec 2014 23:32:03 +0000 (UTC), glen herrmannsfeldt >> <gah@ugcs.caltech.edu> wrote: >>[Snipped by Lyons]>>> >>>I think he shows that some resampling problems are easier to think of as >>>trains of delta functions in continuous time, instead of just as samples >>>it time. >> >> Hi Glen, >> Unless I'm missing something significant, >> I don't believe there exists any way to think of decimating a >> discrete-time sequence of sample values by a factor of two as any kind >> of continuous-time operation. >> >> It seems to me that continuous-time operations can only be performed on >> continuous-time (analog) >> signals. > >There are some aspects of sampling that only seem to make sense if you >replace the Dirac deltas with functions of area = 1 but finite extent in >time, and then take things in the limit as that time extent goes to zero. > >I'm sure that with careful application of the above, you could shoe-horn >sub-sampling into that model. > >But -- why?Hi Tim, The author's "Decimation" section was on page 4 of the paper. If there's a way to show decimation (sub-sampling) of a discrete sequence as some kind of 'multiplication in the time domain' I'd sure like to see it. In any case, it looks like we've beaten this subject half to death. Thanks for your opinions Tim. [-Rick-] PS. I hope your lawn mower is not an old-fashioned "push mower". (Remember those?)
