It's not (only) a bandwidth issue. It's the failure to interpolate the sample data points correctly.
EDIT: Rephrased the question.
@shapirus, assume that you probe an almost ideal square wave and a scope would show you either figure7 or figure8.
Which one would you prefer subjectively? [ I'll explain the difference later in order to get an unbiased vote. ]
Neither pictures are correct. They both show Gibbs ears which do not exist in the real world. If you want to prevent seeing Gibbs ears, you need to use extra bandwidth limiting so the sin x/x reconstruction does not add the Gibbs ears. IOW, for the best view of a square wave you need to bandwidth limit it in respect to the oscilloscope's bandwidth.
Nobody said they are "correct". Yes, both are imperfect (similarly imperfect as the 3rd screenshot here).
My question was just "which one [of the two imperfect ones] would you prefer subjectively".
You need to keep in mind that sin x /x interpolation on the Rigol is broken resulting seriously distorted results. So there is little use in selecting either imperfect one.
The difference between my two plots is in fact the interpolation filter. Figure7 was calculated with a sharper filter (Siglent-like, about 0.4*fs...0.6*fs), and figure8 uses a filter with a softer transition band (Rigol-like, about 0.3*fs...0.7*fs). The sharper interpolator does not help here. Contrary, I find figure8 a little bit more pleasing. Both suffer severely from aliasing of course.
Maybe you overlooked where I stated the obvious, a scope is a scope is a scope.
They all operate the same, exactly the same !
V/div and s/div, really how hard is that ?
Certainly not rocket science.
The modern DSO is no different...
Have you been paying attention to all the discussions in this thread about the
theory and
practice of sampling?
In general there are many significant erroneous beliefs about sampling, held by far too many people - beginners and professionals alike.
Sampling is not an easy topic to understand; at one of my previous employers we used a
very simple question to weed out those that understood it from those that thought they understood it.
You need to keep in mind that sin x /x interpolation on the Rigol is broken resulting seriously distorted results. So there is little use in selecting either imperfect one.
At least analogue scopes are repairable
This is why most DSOs have a "Default" button that completely (almost, anyway) restores the scope to an original configuration. I use it routinely when I can't remember what I was using the scope for last and just want to start over. This prevents me missing something like having turned up the holdoff time that won't immediately be obvious but might cause some issues.
I must admit to being lazy, and occasionally using the "beam find" button
It's not (only) a bandwidth issue. It's the failure to interpolate the sample data points correctly.
EDIT: Rephrased the question.
@shapirus, assume that you probe an almost ideal square wave and a scope would show you either figure7 or figure8.
Which one would you prefer subjectively? [ I'll explain the difference later in order to get an unbiased vote. ]
Neither pictures are correct. They both show Gibbs ears which do not exist in the real world. If you want to prevent seeing Gibbs ears, you need to use extra bandwidth limiting so the sin x/x reconstruction does not add the Gibbs ears. IOW, for the best view of a square wave you need to bandwidth limit it in respect to the oscilloscope's bandwidth.
Nobody said they are "correct". Yes, both are imperfect (similarly imperfect as the 3rd screenshot here).
My question was just "which one [of the two imperfect ones] would you prefer subjectively".
You need to keep in mind that sin x /x interpolation on the Rigol is broken resulting seriously distorted results. So there is little use in selecting either imperfect one.
The difference between my two plots is in fact the interpolation filter. Figure7 was calculated with a sharper filter (Siglent-like, about 0.4*fs...0.6*fs), and figure8 uses a filter with a softer transition band (Rigol-like, about 0.3*fs...0.7*fs). The sharper interpolator does not help here. Contrary, I find figure8 a little bit more pleasing. Both suffer severely from aliasing of course.
The problem isn't in the interpolation filter at all. Rigol implemented the sin x/x reconstruction wrong. That's it. There is no trade-off to be made here.
The real problem is that you can only use Fourier series to construct continuous functions. You can't use Fourier series to construct functions with a step in them like a square wave. However, when sampling a square wave
like signal (which in the real world can never be a step function) it can turn into a step function in the digital world due to insufficient samples to follow the edges. And as a result you'll get Gibb's ears when applying the sin x / x filter to the sampled signal.
BTW @awakephd, Dave Jones has received a SDS800X HD (along with one of the 1000x HD series) for review. I would suggest to hold your purchase for a little bit more until he gets better and films that review. I belive it may help in making a decision. He also reviewed the Rigol some time last year, make sure you watch that one too.
Yes, I voted for him to do the 800 first! Hopefully he will be able to get to it soon. And yes, I have watched his review of the Rigol - that's actually what led me to this forum!
Have you been paying attention to all the discussions in this thread about the theory and practice of sampling?
In general there are many significant erroneous beliefs about sampling, held by far too many people - beginners and professionals alike.
Sampling is not an easy topic to understand; at one of my previous employers we used a very simple question to weed out those that understood it from those that thought they understood it.
I am happy to report that I do not misunderstand it. I simply don't understand it at all! But I'm learning ...
The problem isn't in the interpolation filter at all. Rigol implemented the sin x/x reconstruction wrong.
The sin(x)/x (in practice an approximation to it) reconstruction
is the interpolation filter. Or more strictly it is the impulse response of the interpolation filter. The tap weights in the FIR interpolation filter come directly from that approximated sin(x)/x. Or vice versa if you prefer to look at it that way.
The problem isn't in the interpolation filter at all. Rigol implemented the sin x/x reconstruction wrong. That's it. There is no trade-off to be made here.
In the screenshots that have been posted so far, I could not see that it were "wrong". They just seem to truncate the sin(x)/x to a shorter length than others do, which leads to wider transition band (0.3*fs...0.7*fs) and to a lower reconstruction limit of about 0.3*fs. No one can do it "right" and implement an ideal brickwall at 0.5*fs. So anything below 0.5 is just a compromise at the end.
The difference between my two plots is in fact the interpolation filter. Figure7 was calculated with a sharper filter (Siglent-like, about 0.4*fs...0.6*fs), and figure8 uses a filter with a softer transition band (Rigol-like, about 0.3*fs...0.7*fs). The sharper interpolator does not help here. Contrary, I find figure8 a little bit more pleasing. Both suffer severely from aliasing of course.
That is theory.
In practice one scope has specified BW of 250MHz, softer transition reconstruction interpolator, and 312.5 MS/s sampling rate.
The other scope has specified BW 200 MHz, (slightly) sharper transition interpolator and minimum sample rate of 500 MS/s.
Which one will alias in practice and have worse waveform representation?
The problem isn't in the interpolation filter at all. Rigol implemented the sin x/x reconstruction wrong. That's it. There is no trade-off to be made here.
In the screenshots that have been posted so far, I could not see that it were "wrong". They just seem to truncate the sin(x)/x to a shorter length than others do, which leads to wider transition band (0.3*fs...0.7*fs) and to a lower reconstruction limit of about 0.3*fs. No one can do it "right" and implement an ideal brickwall at 0.5*fs. So anything below 0.5 is just a compromise at the end.
I disagree. All DSOs I've had through my hands have no trouble to show a sine wave to little over fs / 2.5. There is no upside to having a lower upper frequency simply due to less effective reconstruction as it basically reduces the useable bandwidth of a DSO as can be clearly seen from the screenshots.
I'm trying to understand the discussion of the two square-ish wave forms, and I'm probably getting it wrong ... I think you all are saying that neither of the wave forms represent the actual signal exactly; both are an approximation of the signal. Yes? Or did I get that totally wrong?
If I'm on the right track ... are the "extra bumps" in the wave form (the little squiggle when it changes from high to low or vice-versa) artifacts caused by the approximation, rather than features of the actual signal?
Again, I may be getting this totally wrong ... I'm trying to follow the discussion, but not sure I am on the right track.
If I'm on the right track ... are the "extra bumps" in the wave form (the little squiggle when it changes from high to low or vice-versa) artifacts caused by the approximation, rather than features of the actual signal?
Sorry to be adding even more confusion, but they can be
both :)
You can't use Fourier series to construct functions with a step in them like a square wave.
not sure what you mean but... with some condition, such as 10 sampling points per rise/fall time.... you can..
However, when sampling a square wave like signal (which in the real world can never be a step function) it can turn into a step function in the digital world due to insufficient samples to follow the edges. And as a result you'll get Gibb's ears when applying the sin x / x filter to the sampled signal.
the problem is not Sinc or interpolation, the problem is lack of sample points, IOW Nyquist limit is violated. remove the BW limiter in siglent, and you'll see gibb's ear in siglent scope. however perfect the Sinc implementation is...
You can't use Fourier series to construct functions with a step in them like a square wave.
not sure what you mean but... with some condition, such as 10 sampling points per rise/fall time.... you can..
However, when sampling a square wave like signal (which in the real world can never be a step function) it can turn into a step function in the digital world due to insufficient samples to follow the edges. And as a result you'll get Gibb's ears when applying the sin x / x filter to the sampled signal.
the problem is not Sinc or interpolation, the problem is lack of sample points, IOW Nyquist limit is violated. remove the BW limiter in siglent, and you'll see gibb's ear in siglent scope. however perfect the Sinc implementation is...
That is near what I wrote
But the problem isn't Nyquist limit; the problem is the real world square wave becoming a step function in the digital domain.
That is near what I wrote But the problem isn't Nyquist limit; the problem is the real world square wave becoming a step function in the digital domain.
and you see a step in digital domain is because lack of sampling rate, we can go round and round egg and chicken thing. but the way i see it, lack of sampling rate is the cause, and step function/sampling (and also violation of nyquist) is the effect, when we violate nyquist, no interpolation theory can work. cause first, effect comes later. ymmv.
That is near what I wrote But the problem isn't Nyquist limit; the problem is the real world square wave becoming a step function in the digital domain.
and you see a step in digital domain is because lack of sampling rate, we can go round and round egg and chicken thing. but the way i see it, lack of sampling rate is the cause, and step function/sampling (and also violation of nyquist) is the effect, when we violate nyquist, no interpolation theory can work. cause first, effect comes later. ymmv.
Not quite. You really need to see two seperate problems here. As DSOs have anti-aliasing filters which prevent violating Nyquist, you basically can't violate Nyquist in a DSO (assuming maximum possible sampling frequency). What is left is the effect the signal reconstruction algorithm has on the sampled signal. In the end sin x / x is a fairly crude method for reconstruction a signal so artefacts are to be expected where sine waves and square waves lead to different behaviours. With a better, higher order reconstruction filter you can achieve much better results compared to sin x/x at the expense of computational power (I did some experiments with this a long time ago).
With a better, higher order reconstruction filter you can achieve much better results compared to sin x/x at the expense of computational power (I did some experiments with this a long time ago).
i bet your higher order reconstruction filter is one of sin x/x derivative:
https://en.wikipedia.org/wiki/Reconstruction_filterhttps://en.wikipedia.org/wiki/Whittaker%E2%80%93Shannon_interpolation_formulaabout "anti-aliasing filters which prevent violating Nyquist", we went through that rigol is not properly BW limited at 4CH turned on, and Sinc reconstruction implementation is a bit broken. but your earlier post is too general, so i'm agitated to reply... such as this...
The real problem is that you can only use Fourier series to construct continuous functions. You can't use Fourier series to construct functions with a step in them like a square wave. However, when sampling a square wave like signal (which in the real world can never be a step function) it can turn into a step function in the digital world due to insufficient samples to follow the edges. And as a result you'll get Gibb's ears when applying the sin x / x filter to the sampled signal
some part of it imho is misleading, esp the bolded line. a properly bw limited scope and correct implementation of Sinc derivative filter will not produce gibbs ear, thats not me saying, thats from theory saying (at least what i understand). ymmv.
That is near what I wrote But the problem isn't Nyquist limit; the problem is the real world square wave becoming a step function in the digital domain.
and you see a step in digital domain is because lack of sampling rate, we can go round and round egg and chicken thing. but the way i see it, lack of sampling rate is the cause, and step function/sampling (and also violation of nyquist) is the effect, when we violate nyquist, no interpolation theory can work. cause first, effect comes later. ymmv.
Not quite. You really need to see two seperate problems here. As DSOs have anti-aliasing filters which prevent violating Nyquist, you basically can't violate Nyquist in a DSO (assuming maximum possible sampling frequency). What is left is the effect the signal reconstruction algorithm has on the sampled signal. In the end sin x / x is a fairly crude method for reconstruction a signal so artefacts are to be expected where sine waves and square waves lead to different behaviours. With a better, higher order reconstruction filter you can achieve much better results compared to sin x/x at the expense of computational power (I did some experiments with this a long time ago).
sin(x)/x
is the best reconstruction filter wrt accuracy. It
is the highest order - infinite. Ideal low pass. Brick wall. Any realisable filter requires
less computational power.
Have you been paying attention to all the discussions in this thread about the theory and practice of sampling?
In general there are many significant erroneous beliefs about sampling, held by far too many people - beginners and professionals alike.
Sampling is not an easy topic to understand; at one of my previous employers we used a very simple question to weed out those that understood it from those that thought they understood it.
I am happy to report that I do not misunderstand it. I simply don't understand it at all! But I'm learning ...
Excellent
That is an ideal beginning to a never-ending quest for deep understanding.
Fortunately that quest never ends; there are
always new and interesting topics. Hence, in a real sense, I will continue to be a beginner - and love it
at one of my previous employers we used a very simple question to weed out those that understood it from those that thought they understood it.
Is that why it's a previous employer? 😉😉
Nah!
I left them because after >10 years I was in a fur-lined rut, and because the new CEO struck me as someone I didn't believe would be good for the company.
Both judgements turned out to be prescient
3 & 5 - One of the responses rightly points out that appropriate defaults can offset these issues. On the other hand, I have experienced the overwhelming frustration (with other sorts of instruments) of somehow changing one of those settings, and since I don't know enough to even know the setting is there - up to now the default has conveniently hidden it - I spend hours trying to figure what I'm doing wrong.
This is why most DSOs have a "Default" button that completely (almost, anyway) restores the scope to an original configuration. I use it routinely when I can't remember what I was using the scope for last and just want to start over. This prevents me missing something like having turned up the holdoff time that won't immediately be obvious but might cause some issues.
The user definable Default is even more useful so with the press of just one button to have the DSO set up just as you like it.
Further, some DSO's allow internal and/or external saving of a Setup file where you might have the scope configured for specific tasks.