Sub: Rigol's DHO800 Oscilloscope (Gibbs Effect & Aliasing Misunderstanding)

[wasedadoc]

mind to tell what filter is that? or if that feasible to put on DSO front end?

Photo shows filter output when sine wave input frequency repeatedly swept linearly from 100kHz to 3MHz.  (That range does not occupy entire screen width.  It should be obvious where the two frequency extremes are.)

[wasedadoc]

If you are low-pass filtering, you are removing higher harmonics. The ringing on a bandwidth limited square wave are not Gibbs ears. Just less harmonics. Because Gibbs ears are a digital signal processing artefacts from using sin x /x reconstruction, you can't get these on an analog scope because the whole digital signal processing step isn't there.

If the ears are as you claim "a digital signal processing artefacts from using sin x /x reconstruction" please explain why they are only present on one trace.

[Mechatrommer]

If you are low-pass filtering, you are removing higher harmonics. The ringing on a bandwidth limited square wave are not Gibbs ears. Just less harmonics. Because Gibbs ears are a digital signal processing artefacts from using sin x /x reconstruction, you can't get these on an analog scope because the whole digital signal processing step isn't there.

If the ears are as you claim "a digital signal processing artefacts from using sin x /x reconstruction" please explain why they are only present on one trace.

you misunderstand him... to him the real signal that you captured is not Gibbs, since Gibbs only is the one calculated out of Sinc interpolator. however similar they look like. you do know what a pedantic and literalist are right? there is no point continue arguing round and round. i'm afraid 2N3055 will miss my challenge posts earlier..

[ebastler]

to him the real signal that you captured is not Gibbs, since Gibbs only is the one calculated out of Sinc interpolator. however similar they look like. you do know what a pedantic and literalist are right? there is no point continue arguing round and round.

Yes, that seems to be the problem. We just get circular arguments here: "No, what you show is not Gibbs, since according to my definition it is not Gibbs, but I won't tell you my definition."

I feel strongly reminded of the "Is digital always binary?" thread in the Beginners section, with nctnico playing the radiolistener part. And I think I should draw the same conclusion as in that other thread and stay out of this.

[nctnico]

to him the real signal that you captured is not Gibbs, since Gibbs only is the one calculated out of Sinc interpolator. however similar they look like. you do know what a pedantic and literalist are right? there is no point continue arguing round and round.

Yes, that seems to be the problem. We just get circular arguments here: "No, what you show is not Gibbs, since according to my definition it is not Gibbs, but I won't tell you my definition."

There is no 'my' or 'others' definition. Again, the Wikipedia article is super clear on this! You simply can't create a perfect square wave using fourier series. Not even with an infinite number of harmonics.

This page shows the effect of sin x/x and the Gibbs ears related to the Nyquist limit very clear: https://www.embedded.com/filling-in-the-blanks-in-digital-oscilloscope-waveforms/
https://www.embedded.com/filling-in-the-blanks-in-digital-oscilloscope-waveforms/

[TimFox]

An important detail about Gibbs' phenomenon in the Fourier series of a square wave is that, although the height of the Gibbs' ear grows with the number of harmonics included, the area under the ear decreases.

[wasedadoc]

For those interested in the filter I used, the photo shows log magnitude and phase of the through response measured on a VNA.  Horizontal is linear from 100kHz to 5MHz.  Vertical is 10 dB per div and 90 degrees per div.  The VNA is 50 Ohms in and out but I did not take the trouble to make up complete matching pads.  Just added 25 Ohms in series with filter in and out.

Filter was designed by a specialist colleague.  I wanted to use it as a presampling filter for sampling the luminance of 625/50 video at 5MHz.  Nyquist frequency is 2.5MHz.  Beginning frequency of roll-off and steepness are a tradeoff between blur, ringing and aliasing.  We never selected those parameters just from a theoretical or measured frequency response plot.  Always eyeballed the result on both test signals and real picture material.

Only about 6dB down at half sample rate.  Good phase linearity in the passband is evident.

AIUI the design technique is a combination of a low-pass to give the shape plus an all-pass which has frequency dependent delay to balance that of the low pass.

Apologies for the photo quality but it plus the text above give the salient details.

[tautech]

A selection, mostly dots but some vectors when we need compare.
Toggled captures between Dot and Vector mode.
Most in Stop mode but a few running, all easy to identify.
Any in Stop mode are also toggled between Dot and Vector mode.
Max at 200ps/div. Dot count in the Timebase tab.
All in Auto memory management mode.

Does run mode superimpose multiple traces on the screen, even though Persistence is off?
Apparently it does. How many traces?

Several.
Related to refresh rate I believe which will also be related to timebase settings.
WFPS and display refresh rate obviously has been optimized for Run mode Print not to just show a single sweep of dots. eg, useless for a Single trigger capture at fast timebase settings.
But there is always interpolated Vector display mode.....

In run mode (e.g. ..._9.jpg) there still seems to be some interpolation going on. Not interpolation in the traditional sense, but trigger point interpolation, aligning the traces horizontally by fractional sample time offsets. Since this signal is not undersampled, the trigger point interpolation is expected to work quite well here, and apparently it does. But I wonder how well it works if the edges were undersampled?

Dot mode = true sample points before any interpolation.

In the acquisition mode I use these DSO's* everything that is reported in the info boxes is just one screens width so when pts are listed it's for the 10 horizontal divisions.
Here we have complete control of what we see and how we might capture it.
Run and Print shows some overwrites in dot mode whereas Stop and Print shows a single sweep of data points.
Count them, what's reported and the physical count match.

* We could instead chose Fixed Mem mode and zoom out capability of several orders of magnitude if that is your thing. 

[Mechatrommer]

I give up...

sorry i keep missing your post, because whenever you post, i only see "show the post" for some reason... and some posters like to post pedantic stuffs creating distractions. why give up? the challenge its not so hard, but it will reveals what i want to explain that some of you people very hard to comprehend. this is only basic stuff i dont claim i have phd... why give up? because what you'll find is a fact that i already explained here? https://www.eevblog.com/forum/testgear/split-from-rigols-new-dho800-oscilloscope-unbox-amp-teardown/msg5140329/#msg5140329

now if you still dont get what i want to prove? 1st i already explained in 1st post and various places here (i have to keep repeating again and again) this is for awareness... if you admit the facts i was presenting, simply admit it or simply do not reply (admission by silence) otherwise, present proofs to prove my points are wrong, in maturely and wisely and most relevant manner. forum is for healthy debate to avoid misinformations in scientific,  technical community alike.

now since you (and others) dont want to take the task, may i turn to tautech, to do it for me? i need to ask favor again. because so far, he is the one the most transparent on this... and also since he stated earlier this interpolated data and stuffs are BS... here mr tautech... https://www.eevblog.com/forum/testgear/split-from-rigols-new-dho800-oscilloscope-unbox-amp-teardown/msg5140950/#msg5140950

here what i suggested for you to do earlier.... check my challenge posts earlier. dial back to where you got the gibbs effect like in this picture you showed here...
[img]

screen capture it again to show the gibbs effect as reference... and then.... this is the most important part you need to get right.... change your scope to Sinc OFF, dot mode ON, and single trigger it, until it triggered and stopped. so we can see few points on screen paused, not overlapping around. and do the 2nd capture, and post here again for comparison and analysis... cheers.

here an example of the 2nd capture i need, like tautech have made... (i would love to see same time scale as the one you got the gibbs effect above 200ns/div)
[img]

i was asking it here earlier but maybe not clear enough... https://www.eevblog.com/forum/testgear/split-from-rigols-new-dho800-oscilloscope-unbox-amp-teardown/msg5139081/#msg5139081

anyone? not just tautech and 2N3055..

btw: when we try to make a claim or swear about this device not being good enough, not accurate enough, make sure we have good background knowledge to back it up, otherwise people of knowledge can easily see what is wrong with your claim and continuos attempt at spreading misinformations to people who just got into the arena. swearing about others without good knowledge is an INSULT, to the designer and brand, and to the users too... and not trying to defend your point in maturely and calmly manner is RUDE at least thats the culture around here (my place, not this forum)... imho.

[gf]

An important detail about Gibbs' phenomenon in the Fourier series of a square wave is that, although the height of the Gibbs' ear grows with the number of harmonics included, the area under the ear decreases.

That's IMO the important point. Strictly speaking, the "Gibbs Phenomenon" is less about the ringing itself, which happens near discontinuities when Fourier series are truncated, but it is more about the convergence when more and more harmonics are included. As you said, the strange thing making it a "phenomenon" is that the mean square error of the approximation does converge to zero, but the overshoot (or maximum deviation between true function and approximation) remains constant and does not converge to zero.

There seems to be some disagreement in the literature as to whether the ringing itself (due to Fourier series truncation) should also be called "Gibbs". Some say yes, while some purists just call it overshoot/ringing and associate "Gibbs" only with the convergence phenomenon.

EDIT: In this thread, "Gibbs" seems to be mostly associated with the ringing itself, and not with the convergence phenomenon.

[Mechatrommer]

btw i was stumbled last night with my own code as to why my data introduces Gibbs Effect in "Color Grade" mode where in "Vector Mode" there is none! just now i debug my code just to find out i actually implemented Lanczos Resampler Library which is a derivative or implementing Sinc function in it (a library that i copied from internet that i have no clue about the internal math implementation in it) https://en.wikipedia.org/wiki/Lanczos_resampling i've just forgotten my own code 7 years ago Lanczos Resampler (interpolator) will be activated whenever data (sample points) count on display is less than screen pixel width size. when there are enough or more points, Lanczos will not be activated. so here i think also a good example of the resampling algorithm that produces Gibbs Effect, where in reality there is none. fwiw...

ironically pretty similar result as in here https://www.eevblog.com/forum/testgear/split-from-rigols-new-dho800-oscilloscope-unbox-amp-teardown/msg5140329/#msg5140329 where i thought "the "imaginary" part of Gibbs disappear, maybe got folded back (aliased) during interpolation process?" i wonder if those scopes are actually implementing Lanczos Resampler? who knows? you can get it online free.

[Martin72]

I had just discovered pictures of my Bodnarpulser on the Rigol DHO4204 which I had on loan, I put them here.
One channel active, 4GSa/s maximum sample rate, the signal looks "good".
Two channels active, 2GSa/s sample rate, the signal deforms a bit - but is "faster"(rise time)...
4 channels active, but still 1GSa/s samplerate, the signal looks a bit "blurred" in two places and the rise time becomes slower again.
Why is the bandwidth highest at 2GSa/s....

[bdunham7]

So to start from end: Overshoots you see on that Siglent are result of analog Brickwall filter when you feed it with very fast signal, much faster than it's transition band is. 

Which models of Siglent have an analog 'brick wall'?  Mine seem to trail off for quite some distance.  I've never attempted to accurately characterize them, but I'm shocked to hear them described as 'brick wall'.

I was a bit surprised at this discussion because unless I missed something, the various concepts seemed to be mixed together--perhaps like they are in real life.  Here's a few thoughts.

Gibbs is one thing, and it can be eliminated from the picture by simply turning off interpolation and looking at the dots.  The trigger is still interpolated, of course, but that's not too important for this issue.

If you consider undulations after the transition to be 'ringing' (whatever the cause) and before to be 'pre-ringing', then the question arises as to the causality issue, which I think was central to Mechatrommer's question.  The simple answer is that at a very basic theoretical level, causality is a complete non-issue simply because the display does not appear on the screen until a long time after the actual event, long meaning compared to the timescale of the signal.  You can demonstrate pre-ringing in an entirely analog oscilloscope!  The obvious culprit in a CRO is the delay line along with the delay line compensation circuit, but group delay is a feature of many analog filters and can show up as pre-ringing.  So after you've eliminated Gibbs if you still see a bit of pre-ringing, there you are.

Finally, an RLC filter may be underdamped without really making that much of a difference in the ultimate BW and rolloff.  That will cause actual ringing due to oscillation, but not pre-ringing.  Practically I have no idea whether you can separate this out from group delay for a real-life example. 

[Mechatrommer]

Why is the bandwidth highest at 2GSa/s....

i guess because it produces cleaner less jittery Gibbs (Sinc interpolation) at 2GSa/s (200pts/screen 500ps/pts)? what surprises us is why the properly bw limited scope can still produce gibbs artifact at 10x sampling rate iirc DHO800/DS1000Z dont produce this artifact at max sampling rate 1GSa/s. btw real machine got jittery here and there, internal clock not locked what not resulting sampling at not perfectly consistent interval. making practical application deviates from theory, last night i read about yaigol problem.

[Fungus]

"Yaigol" was fixed by a firmware update about six months after launch.

[Mechatrommer]

"Yaigol" was fixed by a firmware update about six months after launch.

yeah but who have time verifying new scope model?

[rf-loop]

I had just discovered pictures of my Bodnarpulser on the Rigol DHO4204 which I had on loan, I put them here.
One channel active, 4GSa/s maximum sample rate, the signal looks "good".
Two channels active, 2GSa/s sample rate, the signal deforms a bit - but is "faster"(rise time)...
4 channels active, but still 1GSa/s samplerate, the signal looks a bit "blurred" in two places and the rise time becomes slower again.
Why is the bandwidth highest at 2GSa/s....

Bodnar_Single.png

I have one "question"... 

When there is 10ns/div and 4Gsa/s (40 sample points / div) do we really need Sin(x)/x (aka Sinc) specially when oscilloscope is also 200MHz BW.

Think about it.

(yes I know where one other oscilloscope turn Sinc off (without indicating it) when some t/div is selected...and Sinc do not give any real advantage. Even when Sinc is still selected)

Also if I design oscilloscope I will really think carefully when Sinc give some advantage and when not...

Just for thinking..

[gf]

Why is the bandwidth highest at 2GSa/s....

i guess because it produces cleaner less jittery Gibbs (Sinc interpolation) at 2GSa/s (200pts/screen 500ps/pts)? what surprises us is why the properly bw limited scope can still produce gibbs artifact at 10x sampling rate

Keep in mind that any filter in the signal path with a sharp enough cut-off can introduce overshoot/ringing near edges, and if the filter happens to be a linear phase filter, then this ringing is symmetric (pre + post). This is not limited to sinc reconstruction filters, and it is not limited to undersampled signals.

In Bodnar_Dual.png (2GSa/s) I do not think that the pre- and post-ringing comes from an interpolation filter. An upsampling filter from 2GSa/s has a cutoff at 1GHz (or slightly below), and its ripples are supposed to have a period in the order of 1ns. But the period here is rather 2ns (or even a little bit more), which suggests that we see the result of a FIR filter with a cut-off < 500MHz. However, don't ask me what's the objective And since the trace is pretty clean, I'm quite confident that it doesn't suffer from aliasing (at least not significantly), i.e. very likely it was properly bandwidth limited before sampling.

OTOH, I'm pretty confident, that the signal in Bodnar_Quad.png (1GSa/s) was not properly bandwidth limited before sampling and that the apparent (let me call it) "reconstruction jitter" is caused by aliasing.

[Mechatrommer]

Why is the bandwidth highest at 2GSa/s....

i guess because it produces cleaner less jittery Gibbs (Sinc interpolation) at 2GSa/s (200pts/screen 500ps/pts)? what surprises us is why the properly bw limited scope can still produce gibbs artifact at 10x sampling rate

Keep in mind that any filter in the signal path with a sharp enough cut-off can introduce overshoot/ringing near edges, and if the filter happens to be a linear phase filter, then this ringing is symmetric (pre + post). This is not limited to sinc reconstruction filters, and it is not limited to undersampled signals.

show me in any rigol DSO, or any other "better" competitive DSO from teardown photos that have this "sharp enough cut-off can introduce overshoot/ringing near edges". one member here proved that filter exists, but not for DSO. even if it is, it will not fit from teardown photos evident, we dont see fancy inductors capacitors 3x 5x parellel networks over there. prove me wrong, and err... also care to explain why at 4GSa/s snapshots above got no overshoot/undershoot.pre or post ringing?

In Bodnar_Dual.png (2GSa/s) I do not think that the pre- and post-ringing comes from an interpolation filter.

nope! until evident provided. i've provided a few, its your turn... here's the challenge below if you miss it (on any brand scope thats capable both Sinc ON and vector/dot (Sinc OFF) mode. its like nobody understand this thread and i'm just talking to the air in fact the evident that there is no pre-ringing is just right above the 2GSa/s image, ie capture at 4GSa/s. how amazing people can miss it!

https://www.eevblog.com/forum/testgear/split-from-rigols-new-dho800-oscilloscope-unbox-amp-teardown/msg5142249/#msg5142249

I have one "question"... 

When there is 10ns/div and 4Gsa/s (40 sample points / div) do we really need Sin(x)/x (aka Sinc) specially when oscilloscope is also 200MHz BW.

Think about it.

(yes I know where one other oscilloscope turn Sinc off (without indicating it) when some t/div is selected...and Sinc do not give any real advantage. Even when Sinc is still selected)

Also if I design oscilloscope I will really think carefully when Sinc give some advantage and when not...

Just for thinking..

not just when developing DSO, but PC SW that utilize sampled data, or even in graphics 2D 3D rendering will deal with this decision matters. like my SW just above. i let Sinc to take place when sampled points is less than screen size divide by 2.. this is developers decision.. maybe if (data_count < screen_size / X) then do Sinc, else do vector plot.... X can be 2,3,4,5? it depends on developers/designers... but user should decide if need Sinc ON? or want Sinc OFF? in any condition...

[Mechatrommer]

here again i upgraded VisaDSO v1.1.3 to enable interpolated Sinc/Lanczos on single shot capture...see the difference. Leo Bodnar Pulse captured by DHO924 then data downloaded to PC... Sinc OFF (no Gibbs Effect) the boring straight line vector drawing mode, and Sinc ON (with Gibbs Effect) for noob's eyes satisfaction curvy plot. granted this is a scope that violates Nyquist at max sample rate and 4 channels, but for DHO4000 above, 2GSa/s single channel... its puzzling.

[rf-loop]

, ie capture at 4GSa/s. how amazing people can miss it!

As you can see not everyone...

[Mechatrommer]

, ie capture at 4GSa/s. how amazing most people can miss it!

As you can see not everyone...

there corrected for me. i know you in arena that keep rf work again and again usually people less talk because doing more work... cheers.

[nctnico]

Why is the bandwidth highest at 2GSa/s....

i guess because it produces cleaner less jittery Gibbs (Sinc interpolation) at 2GSa/s (200pts/screen 500ps/pts)? what surprises us is why the properly bw limited scope can still produce gibbs artifact at 10x sampling rate

Keep in mind that any filter in the signal path with a sharp enough cut-off can introduce overshoot/ringing near edges, and if the filter happens to be a linear phase filter, then this ringing is symmetric (pre + post). This is not limited to sinc reconstruction filters, and it is not limited to undersampled signals.

show me in any rigol DSO, or any other "better" competitive DSO from teardown photos that have this "sharp enough cut-off can introduce overshoot/ringing near edges". one member here proved that filter exists, but not for DSO. even if it is, it will not fit from teardown photos evident, we dont see fancy inductors capacitors 3x 5x parellel networks over there.

If you look carefully at Dave's teardowns you'll see quite a few scopes have LC filters in front of the ADC to previous aliasing. However, these filters are designed so their phase response doesn't cause (excessive) ringing. For obvious reasons as the designer of a DSO wants to introduce as little 'distortion' to the visible signal as possible.

[Martin72]

Think about it.

I'll read through this first before I start thinking(link to a pdf):

https://cdn.teledynelecroy.com/files/whitepapers/wp_interpolation_102203.pdf

Maybe that will help me think.

[gf]

https://cdn.teledynelecroy.com/files/whitepapers/wp_interpolation_102203.pdf

The Simple Test described in this paper was also my reconstruction validity criterion when I wrote

And since the trace is pretty clean, I'm quite confident that it doesn't suffer from aliasing (at least not significantly), i.e. very likely it was properly bandwidth limited before sampling.

OTOH, I'm pretty confident, that the signal in Bodnar_Quad.png (1GSa/s) was not properly bandwidth limited before sampling and that the apparent (let me call it) "reconstruction jitter" is caused by aliasing.

However, there is admittedly a catch: If the sampled signal happens to be filtered (e.g. if a bandwidth limit is applied), then the check described in the paper indicates whether the filtered signal was reconstructed properly. It does not indicate whether a decent reconstruction would have been possible from the unfiltered samples.

In Bodnar_Dual.png (2GSa/s) I do not think that the pre- and post-ringing comes from an interpolation filter.

nope! until evident provided. i've provided a few, its your turn... here's the challenge below if you miss it (on any brand scope thats capable both Sinc ON and vector/dot (Sinc OFF) mode. its like nobody understand this thread and i'm just talking to the air in fact the evident that there is no pre-ringing is just right above the 2GSa/s image, ie capture at 4GSa/s. how amazing people can miss it!

Well, let's assume it comes from the interpolation filter. Why would you design an interpolation filter that cuts off 50% of the recoverable bandwidth? Sure, in practice you have to sacrifice a little, but why so much? Maybe I am missing some information to understand the intention As a result, the interpolation filter does not only reconstruct, but additionally it acts as a bandwidth limiter. And the latter, of course, can still introduce ringing, even if the reconstruction alone at (almost) full bandwidth would possibly not do that.

EDIT: Of course, we do not know if we are dealing with two separate filters here (one for bandwidth limiting, and a second for reconstruction), or a single filter that combines both. At the end, we can only observe the combined effect.