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

[Mechatrommer]

16 MHz square wave (with 30ps edges) will have harmonics into gigahertz range. Any of these harmonics above nyquist will alias.

In theory yes, but given the spacing between harmonics and their levels, the chance of harmonics aliasing back as a visual artifact are not very high as you have to consider the anti-aliasing filter in a DSO. Especially when dealing with a square wave which has mostly odd harmonics which are spaced at twice the fundamental frequency.

a perfect square wave is just that... from 0 suddenly 1 (or whatever Vmax it is, dont be pedantic), sampling rate is just that... sampling at certain dt interval. too slow sampling rate, the machine will miss the details during the step. plotting straight line point to point will just looks like trapezoid, or slowly risetime square, or at worst a bunch of triangles. add signal reconstruction/interpolation to it, the imperfect one, thats where the devil is where people misunderstand it  "aliasing", in fact it is just a digital post processing artifacts. maybe people got deluded due to they think in frequency domain, not in time domain, where the samples capturing happens. as long as "fundamentals" is not exceeding nyquist limit, there will be no aliasing. granted the displayed squares will be distorted here and there, but not aliased, maybe due to jittery square wave on circuit, or just that.... artifacts from imperfect interpolation. even if it is indeed an aliasing, if harmonics amplitudes are not too much, it could be unimportant in digital circuit, otherwise the dut square wave got problems, too much oscillations and all, in this case, to verify we must go up the DSO ladder to 10-100GSps class. on the subject of adding LP filter for the sake of antialiasing, the true signal's shape will get hidden in the process, this is what low end DSO is about, we wont know if the signal is good or oscillates in real life.fwiw.

[wasedadoc]

Of course, the output waveform of a "fast" square wave (e.g., the output of a good binary flip-flop) passed through an elementary low-pass filter (e.g., a simple one-pole R-C circuit) should not exhibit the Gibbs phenomenon on a high-bandwidth analog oscilloscope or a fast-enough sampling digital oscilloscope.

True, but a more complex LC filter with linear phase response can be physically constructed which will show the phenomenon.

[Mechatrommer]

Of course, the output waveform of a "fast" square wave (e.g., the output of a good binary flip-flop) passed through an elementary low-pass filter (e.g., a simple one-pole R-C circuit) should not exhibit the Gibbs phenomenon on a high-bandwidth analog oscilloscope or a fast-enough sampling digital oscilloscope.

True, but a more complex LC filter with linear phase response can be physically constructed which will show the phenomenon.

do siglent and rigol enough room at front end for that? or do they have that actually already? experienced people should already spotted that from tear down photos. and why the effect disappears on screen at max sampling rate? and reappears on much lower sampling rate?

[mawyatt]

Of course, the output waveform of a "fast" square wave (e.g., the output of a good binary flip-flop) passed through an elementary low-pass filter (e.g., a simple one-pole R-C circuit) should not exhibit the Gibbs phenomenon on a high-bandwidth analog oscilloscope or a fast-enough sampling digital oscilloscope.

Will throw this out for discussion, altho deviating from the DHO800, but seems lots of folks are involved, so maybe worthwhile.

Think it can be stated that ANY system described by continuous domain IIR Response of ANY kind, be that low pass, band pass, high pass, ANY combinations of poles and zeros, or whatever can NOT exhibit the Gibbs effect on an ideal squarewave.

This is stated without any proof (we left the DSP world long long ago at it's early beginnings).

Will be interesting to see what some of the True DSP gurus have to say!!!

BTW this is not to start/continue an argument, but to maybe add some clarity to the on going discussions.

Best,

[mawyatt]

Of course, the output waveform of a "fast" square wave (e.g., the output of a good binary flip-flop) passed through an elementary low-pass filter (e.g., a simple one-pole R-C circuit) should not exhibit the Gibbs phenomenon on a high-bandwidth analog oscilloscope or a fast-enough sampling digital oscilloscope.

True, but a more complex LC filter with linear phase response can be physically constructed which will show the phenomenon.

Please elaborate on such, see above.

Very interested in seeing if this is possible with an IIR system.

Best,

[ebastler]

To the moderators (I will report this post to aks for their consideration):

Any chance the discussion from reply #1602 onwards could be moved to a separate thread? This thread is getting very long and hard to keep track of anyway, and the several pages of philosophical discussion on pulse response have nothing to do with the DHO800.

EDIT: And it's done. A big "Thank-you" to the moderators! 

[mawyatt]

Probably a good idea. The discussions involving Nyquist, Gibbs, and Causality perked our interest from way back when we turned Nyquist around into an ally.

Best,

[nctnico]

Of course, the output waveform of a "fast" square wave (e.g., the output of a good binary flip-flop) passed through an elementary low-pass filter (e.g., a simple one-pole R-C circuit) should not exhibit the Gibbs phenomenon on a high-bandwidth analog oscilloscope or a fast-enough sampling digital oscilloscope.

Will throw this out for discussion, altho deviating from the DHO800, but seems lots of folks are involved, so maybe worthwhile.

Think it can be stated that ANY system described by continuous domain IIR Response of ANY kind, be that low pass, band pass, high pass, ANY combinations of poles and zeros, or whatever can NOT exhibit the Gibbs effect on an ideal squarewave.

I agree. Gibbs effect is about approximating a square wave by adding a series of sine waves which is what the sin x /x function starts to struggle with at some point (say a math problem). Showing a square wave in a bandwidth limited form is different; there simply is not enough bandwidth to display all the harmonics (say a physics problem). IMHO these are two entirely different issues and should not be confused.

[Mechatrommer]

i should have provided a "literature" instead of putting it out my broken english mouth...
https://resources.altium.com/p/how-gibbs-phenomenon-produces-measurement-artifacts

Therefore, the Gibbs phenomenon is not something that actually happens in a band-limited channel, it’s a mathematical problem that could be easily mistaken for a real effect or could mask a real effect.
...
To summarize, Gibbs ringing is entirely an artifact of your measurement instrument and Fourier/Laplace transform operations on bandlimited network parameters

To the moderators (I will report this post to aks for their consideration):

Any chance the discussion from reply #1602 onwards could be moved to a separate thread? This thread is getting very long and hard to keep track of anyway, and the several pages of philosophical discussion on pulse response have nothing to do with the DHO800.

that will be a "task" for the mods... i believe there is already long discussion on these subject in a siglent thread, i thought these all are sorted out among siglent enthusiasts, but i guess not, they came here with false "attack" on signal accuracy and aliasing and all etc...

[Serg65536]

5) you yourself demonstrated the gibbs ear by showing 2c captures, one is using X, another is sinc interpolation, but it turned out you are looking at the wrong side (the actual

That's definitely the sinc interpolation effect! On this scale it shows only 100 samples per screen. Ripple waves are exactly between the real samples.

[Fungus]

To me that looks exactly like a bandwidth limited square wave:


Image from here: https://www.nayuki.io/page/band-limited-square-waves

Nothing to do with sinc function.

[Mechatrommer]

To me that looks exactly like a bandwidth limited square wave:

DIGITALLY bandwidth limited.. it never happened on the actual circuit. but if you like "made up representation", then thats your personal preference i cant do anything about. https://resources.altium.com/p/how-gibbs-phenomenon-produces-measurement-artifacts

That's definitely the sinc interpolation effect! On this scale it shows only 100 samples per screen. Ripple waves are exactly between the real samples.

thats why dot or vector (sinc interpolation disabled) rendering can be useful so we know what we are dealing with. if these "educational" scope with "Default" button really meant for educational, they should have dot or vector plotting. imho.

ps: wow! its possible to move a bunch of posts into another thread! amazing! i didnt know.

[tautech]

ps: wow! its possible to move a bunch of posts into another thread! amazing! i didnt know.

A rarely given privilege. Enjoy.

However as OP you can use the Move Topic to shift a whole thread anywhere in the forum just don't go to the Supporters Lounge or you need a Mod to rescue you....don't ask how I know !

[ebastler]

ps: wow! its possible to move a bunch of posts into another thread! amazing! i didnt know.

A big thank-you to the helpful moderator who made this happen! 

[rf-loop]

as long as "fundamentals" is not exceeding nyquist limit, there will be no aliasing.

In my opinion, this statement is a bit false.

But of course.
When the system does not produce an alias frequency from fundamental (aka 1st harmonic) then fundamental is not aliasing.

When harmonics are downconverted by ADC these down conversion products are there. And these down converted "false" signals we call a bit carelessly aliases. This is true with all harmonics, including also 1st. 1st harmonic aka fundamental.

When you drive some waveform that contains harmonics that cross the fNyquist "wall" these are converted down and these all down conversions are alias frequencies from these harmonics (every harmonic is harmonic but also we can handle higher than 1st harmonic as "its own fundamental". But down conversion products they are in this case aliases from these harmonics.

Take 11MHz "square" and look harmonics in system. 1st harmonic (aka fundamental) is 11MHz, 3rd harmonic is 33MHz and 5th is 55MHz, etc. (and now important thing, all harmonics, including also 1st (aka fundamental) can alias.

Now sample it using 50MSa/s.
There is fundamental aka 1st harmonic. 11MHz in this case not aliasing
There is now 17MHz which is alias (conversion) from 3rd
There is also 5MHz which is alias from 5th.
Fundamental is not aliasing. Yes this is now of course true. Still system produces aliases from these harmonics. In this case all other harmonics are aliasing but not 1st harmonic. (but then totally other case is how we see these aliases and there is big difference if we look in time domain or or in frequency domain.)
Naturally if also fundamental is aliasing then we see directly this aliasing and dramatic figures on time domain display. When harmonics are aliasing we see different square wave what is input. If we filter out harmonics what are higher than fNyquist also then we see different "square" what is our input but also different if we compare it to image what we get if other harmonics than fundamental is generating.
Fundamental (1st harmonic) aliasing looks very very different than higher harmonics aliasing. Specially in oscilloscope time domain (digital oscilloscope of course because analog oscilloscope do not have any ADC based aliasing.

Then... very different thing is what Sinc "reconstruction" do. It must not mix here with all fun Gibb's etc what must not mess with aliasing.

Then we come to the oscilloscope time domain display.
When fundamental is aliasing we can see its alias frequency. Example if fundamental is 15MHz and our fNyquist is 10MHz (sampling 20MSa/s) we can see 5MHz alias.

I think we all know that ADC can be used for frequency conversion. (as it is also used somewhere) And now where is way to think it works differently for "fundamental" and harmonics. Because there is nothing but harmonics. Every harmonic can thing as it is "fundamental" and it behaves just as fundamental. (just filter fundamental out and use this harmonic as fundamental.)
We can use 1GSa/s ADC so that we convert 995MHz signal to 5MHz but yes USB and LSB swapped.
If we are doing even step decimation it also works same way.

This is not for Siglent, not for Rigol, not for... long list of names. Brand names do not change physics or math, even if they want or hope. (except if they do mistakes or other way handle things wrong way)

[Mechatrommer]

as long as "fundamentals" is not exceeding nyquist limit, there will be no aliasing.

In my opinion, this statement is a bit false...

sorry if the "s" confuses you. i meant "fundamental" the 1st harmonic. i know your other arguments about 2nd, 3rd etc harmonics can alias, yes i agree, esp if signal has excessive ringing and overshoots. but few points to note so you can understand how i think:

1) try drawing a square wave on paper, (a realistic signal you can think) with a little bit distortion in 2nd, 3rd etc harmonics... and then think yourself as sampling machine creating sampling dots a little higher rate than fundamental frequency * 2.5 and connecting lines among dots. fundamental will not alias, i think we both agree, you'll see trapeziods or triangles that still can (barely) capture data (if its like digital serial data transmission). now increase a little bit sampling rate just below 2nd harmonics / 2... yes 2nd harmonics will alias, but when you connect the dots, you can still see square waves closer to original signal and if its serial transmission, you can still recover its data content.

2) now on to 2nd, 3rd etc harmonics. if they alias so badly up to a point rendering fundamental (1st harmonics) unrecognizable, then we have problem with signal integrity, reflection, mismatch and all... or analog signal distortion or oscillations so badly, yes we need higher priced DSO. if we dont have money, we can low down our cheap DSO to single channel only at max sampling rate to tackle problematic signal one by one. but since discussion is on a "fairly good" square waves, then imho aliasing will not affect much our judgement (given the bad Sinc interpolation is turned off)

your points are valid on generally all type of analog signals (digital also analog), but we focus on digital signal like mcu clock presented use case earlier, and verified Leo Bodnar pulse with fairly good square wave and no problem with ringing and mismatch. and btw as you also agreed, Sinc can get thing much worse by fabricating what never happened there, amplifying alias that should not be, masking actual signal integrity and all as the article stated... ymmv.

[Fungus]

sinc can get thing much worse by fabricating what never happened there, amplifying alias that should not be, masking actual signal integrity and all as the article stated... ymmv.

But that's not sincs fault. The problem is undersampling.

Turning off sinc is NOT a good idea. Learning to interpret what you see IS.

[nctnico]

To me that looks exactly like a bandwidth limited square wave:

DIGITALLY bandwidth limited.. it never happened on the actual circuit. but if you like "made up representation", then thats your personal preference i cant do anything about. https://resources.altium.com/p/how-gibbs-phenomenon-produces-measurement-artifacts

IMHO that article is wrong about what the Gibbs effect is in DSOs. This Wikipedia article is more accurate and shows it is a math problem; when approximation a square wave using fourier series, you get a signal which doesn't exist: https://en.wikipedia.org/wiki/Gibbs_phenomenon

[Mechatrommer]

To me that looks exactly like a bandwidth limited square wave:

DIGITALLY bandwidth limited.. it never happened on the actual circuit. but if you like "made up representation", then thats your personal preference i cant do anything about. https://resources.altium.com/p/how-gibbs-phenomenon-produces-measurement-artifacts

IMHO that article is wrong about what the Gibbs effect is in DSOs. This Wikipedia article is more accurate and shows it is a math problem; when approximation a square wave using fourier series, you get a signal which doesn't exist: https://en.wikipedia.org/wiki/Gibbs_phenomenon

and thats what a Sinc (or whatever digital filter there is) in scope does! Try to approximate (interpolate) what points in between actual sampled points.

Turn off sinc, turn on dot mode, calculate how many points and sample rate there is. If you are lucky, and the scope is not transparent enough, you'll be surprised dho800 or low entry siglent can do 10GSps on screen on single channel smallest time div. Where do all those extra points come from? Sampled from actual event? Or simply.. approximated? If your eyes sharp enough, Lecroy sda6000 snapshot i linked earlier did 100GSps.. wow? No!

But if you are unlucky, the dots are raw based on sample rate used by dso, you wont see any gibbs anyway. Try it and prove me wrong

[Mechatrommer]

sinc can get thing much worse by fabricating what never happened there, amplifying alias that should not be, masking actual signal integrity and all as the article stated... ymmv.

But that's not sincs fault. The problem is undersampling.

Turning off sinc is NOT a good idea. Learning to interpret what you see IS.

i give up! You win! do whatever you like!

[rf-loop]

Turn off sinc, turn on dot mode, calculate how many points and sample rate there is. If you are lucky, and the scope is not transparent enough, you'll be surprised dho800 or low entry siglent can do 10GSps on screen on single channel smallest time div. Where do all those extra points come from? Sampled from actual event? Or simply.. approximated? If your eyes sharp enough, Lecroy sda6000 snapshot i linked earlier did 100GSps.. wow? No!

Please, explain this part in bold.
Do you know what you are saying or was it just what you just believe. My opinion is that religion is not for this forum.

I have used and examined many (nearly all) Siglent SDS1000 models including the 1000X HD and all 2000 (2000, 2000X, 2000X Plus, 2000X HD) models... also some 6000 models.
Based on these my experience and knowledge, your statement is not a general truth.
Be it anyone, it would be nice to stay as close to the truth as possible.

[rf-loop]

Turning off sinc is NOT a good idea. Learning to interpret what you see IS.

In many cases turning Sinc off (least momentarily and in some cases) is very good idea. But also, sadly, impossible in some oscilloscopes.

But hey, just keep your bullshit generator running, you'll get good numbers in the message counter if you get any pleasure out of it.

[mawyatt]

1) try drawing a square wave on paper, (a realistic signal you can think) with a little bit distortion in 2nd, 3rd etc harmonics... and then think yourself as sampling machine creating sampling dots a little higher rate than fundamental frequency * 2.5 and connecting lines among dots. fundamental will not alias, i think we both agree, you'll see trapeziods or triangles that still can (barely) capture data (if its like digital serial data transmission). now increase a little bit sampling rate just below 2nd harmonics / 2... yes 2nd harmonics will alias, but when you connect the dots, you can still see square waves closer to original signal and if its serial transmission, you can still recover its data content.

2) now on to 2nd, 3rd etc harmonics. if they alias so badly up to a point rendering fundamental (1st harmonics) unrecognizable, then we have problem with signal integrity, reflection, mismatch and all... or analog signal distortion or oscillations so badly, yes we need higher priced DSO. if we dont have money, we can low down our cheap DSO to single channel only at max sampling rate to tackle problematic signal one by one. but since discussion is on a "fairly good" square waves, then imho aliasing will not affect much our judgement (given the bad Sinc interpolation is turned off)

Note a proper squarewave has no even harmonics, even a terrible rise/fall time "squarewave" at limit, a triangle wave if you will, has little or no even harmonics (triangle has no even harmonics). A pulse waveform can have even and odd harmonics of varying intensity depending on duty cycle, in the limit of "zero" pulse-width becomes a repetitive "impulse".

We mention this because it seems many folks are a bit confused about this topic.

Best,

[Mechatrommer]

Turn off sinc, turn on dot mode, calculate how many points and sample rate there is. If you are lucky, and the scope is not transparent enough, you'll be surprised dho800 or low entry siglent can do 10GSps on screen on single channel smallest time div. Where do all those extra points come from? Sampled from actual event? Or simply.. approximated? If your eyes sharp enough, Lecroy sda6000 snapshot i linked earlier did 100GSps.. wow? No!

Please, explain this part in bold.
Do you know what you are saying or was it just what you just believe. My opinion is that religion is not for this forum.

i dont own siglent, i only assume. But please!  Show me gibbs effect in dot mode in siglent scope, a good scope should tell the truth about where its sampled points are located. What religion got to do here? Are you trolIing? empty talk without proof is that..

[nctnico]

In dot / vector mode, you won't get the Gibbs effect because it comes with sin x /x reconstruction. Still, the sin x/x reconstructed signal is the closest you'll get to seeing something that resembles the real signal when you are getting near the nyquist limit. There is a tradeoff to be made by the user.