Rigol DS1000E series: Possible error/fail in Sin(x)/x interpolation?

[Mark_O]

Colin,

actually, I've enjoyed reading both viewpoints.  Both are correct, yet come from differing perspectives.  Janne's excellent mathematical analysis clearly shows how/why we wind up with the kind of waveforms we do, with reduced amplitudes, and not "hitting" the true sample points that concern rf-loop so much.  Your comments are more pragmatic, and make the excellent point that it's all too easy to wind up with aliasing and artifacts, due to undersampling, and not even realize it.  

In the specific case of the Rigol's, it's even easier to wind up in this boat, since the actual sampling rate isn't even shown on-screen under normal conditions.  You have to go looking for it (in the Acquire menu on the older units, moved to the Horz timebase menu in the newer ones).  But no scope is immune from these types of considerations, and at least the newer Rigol's give the option to disable sinx/x, so the poor linear interpolations expose these limitations.

I appreciate the time everyone has taken to contribute, and I think we've all benefited from the discussions.  In particular, rf-loop's testing has shown that the Rigol's (and likely many other economy DSOs) sinx/x does not provide an accurate picture of the waveform, past a certain frequency less than its claimed BW... which will differ based on how many channels are in operation, as a result of the bandwidth and sampling limits.  Even though it still looks good.  This is something that's important to be aware of.

- Mark

[jahonen]

Colin,

What I use the scope for at work most often, is the high-speed digital/mixed signal design debug/verification phase, for almost any kind of signal you can imagine (although most often it is SMPS or digital bus signals, but I think that the world is analog at the end, nevertheless what the people like to call it). I often need to make a characterization on a signal and that means if I need to measure something like a signal risetime or overshoot, the honeymoon is over. You simply need to have enough bandwidth to do that kind of task. Of course you can stretch it a bit by subtracting the scope rise time squared from the measured rise time square and then taking a square root, but that is a stretch at a best.

I imagine that repair work is a bit different, you don't very often actually need to get exact reading of any those properties I mentioned earlier. You just have to find out what has gone wrong. For example, it is not often necessary to adjust the termination on a signal when doing repairs. I also mean that trying to measure overshoot or almost any time-domain property of a signal near the scope specified bandwidth is a ill-conditioned operation (BW = 0.35/t with gaussian signal or BW = 0.5/t with trapezoid signal). But again, if you know what it should look like with a reduced bandwidth, then I that it is ok. But if the signal is completely unknown or the DUT has never worked yet, you can't use that.

What I mean by aliasing, are the gross cases where the frequency (or time axis more exactly) is completely wrong, but the waveshape seems to be okay. Actually, if analog bandwidth if sufficient, undersampled signal is just fine for examining the waveform, the time axis is just wrong but the signal shape is there (like I showed with Tek TDS3034A at the page I mentioned some posts ago). One of the most difficult signals for DSOs to measure is the envelope measurements on a high-frequency carrier (AM-modulated signal), although peak detect tends to solve this problem.

I didn't mean that I have never seen alias (if fact I see that all the time on some DSOs), but what I meant that I have never seen such a aliasing with MSO6034A, regardless of the time/div setting, the signal just turns to "bar" like on an analog scope, although I have deliberately tried to reproduce the aliasing. My working theory is that Agilent runs their ADC at full sample rate all the time, and then they just digitally decimate and filter the data to get the suitable "display sample rate". That makes it possible to avoid aliasing and reproduce the "analog look" even with long time/div-settings. So there is obviously a way to suppress the aliasing so that it does not confuse the user. I don't know why other scope manufacturers don't do the same thing. Even with the relatively old HP 54645A we have at work (which is far below the "$30,000+" price you mention at ebay, although quite more expensive than Rigol), it is very difficult to get any aliasing (certainly not anything that would confuse the user), although sample rate is just 200 MSa/s and specified bandwidth of 100 MHz.

As Mark mentioned, the Matlab idealized plots were made just to try understand why there is a loss of signal fidelity when using sin x/x-feature. Whenever I notice something odd, then I'll try to understand it through theory to have clear view what is actually wrong by varying the parameters to analyze what are the implications for my measurements. But anyway I'm going to investigate this effect rf-loop mentioned with my friend next week with his DS1102E. I'll post my results when I have something conclusive to show.

Regards,
Janne

[jahonen]

Ok, here are some measurements using the casual setup I use (abuse of the HP8753D VNA at work as a signal generator, some coaxial adapters to finally get the BNC output from PC7, and inline BNC terminator just at the scope end):



Here is the measured frequency response (actually, the roll-off or amplitude flatness is not required for characterizing the sin x/x roll-off!). I performed the measurement by using the scope's RMS measurement function, switching the sin x/x on and off in the Acquire menu. Sorry, this time I didn't bother building a LabView software to control the scope and perform the measurement automagically, so there are not so many frequency points. But I believe there are enough for making conclusions about sin x/x:



There indeed is quite a additional attenuation of high frequency signals when using sin x/x. Up to ~10 MHz, the results are essentially identical, whether the sin x/x is used or not. Above that, the discrepancy among two grow steadily.

An another side-effect I noticed is that if you have sin x/x turned on, and when zooming out the waveform, the sin x/x interpolation turns off at some point and the amplitude seemingly jumps up or vice versa. Kind of annoying effect.

We can calculate the additional attenuation caused by sin x/x by using the non-interpolated curve as an reference:



It seems that -6 dB bandwidth of the sin x/x is approximately 100 MHz. The roll-off is also quite shallow, which means that sin x/x windowing is probably quite short. There clearly is a room for an improvement.

And finally, here is a screenshot of that same 100MHz signal using a slow timebase, quite nice undersampling, isn't it? My point is that if the undersampling is severe, neither linear or sin x/x interpolation would give any hint that we are actually undersampling.



Regards,
Janne

[colinbeeforth]

Hi Janne,

Thanks for the reply and description of your work. Repair work can be far more tricky than simply finding what is faulty.  Many old test instruments as still quite useful, so much of the repair work becomes restoration, and replacing obsolete output transistors require careful confirmation that a pulse generator's rise and fall times are still ok.  Many faults are intermittent and the ability to get some idea of a problem, set a complex trigger and leave it armed for a day or so, maybe using sequence mode, is a great way to identify tricky faults.  Even simple repair work is quite demanding when dealing with digital video.

On a more philosophical level, I feel that "digital scope" is an incorrect name.  In the rush to sell ever more test equipment, the manufacturers have built high speed digitisers with complex trigger and a display and called them a "digital oscilloscope".  I don't think this is a fair description of what they do, they are both much more and less than an analogue scope.  Both analogue and digital scopes have their place in electronic work.  The classical ham radio application of the trapezoid display, RF on one axis, modulating audio on the other, will always be a problem for the digital instrument.  I know also that many makers have struggled to make DSOs behave like analogue scopes to show things like eye diagrams of fast serial data.  I don't think the analogue scope can be replaced by digital for some applications.  In terms of results for your dollar, the analogue scope is a valuable instrument for many electronic jobs.

I wish maker's would give up on trying to make DSOs look like analogue scopes and just call them something else.  Unfortunately, the window of opportunity has passed for that and we are stuck with it.

You said, "Even with the relatively old HP 54645A we have at work (which is far below the "$30,000+" price you mention at ebay, although quite more expensive than Rigol), it is very difficult to get any aliasing (certainly not anything that would confuse the user), although sample rate is just 200 MS/s and specified bandwidth of 100 MHz."

I'm not surprised, the spec says it has 1M of acquisition memory per channel!  Long memory ensures far higher digitiser speeds over the middle range of timebase settings.  If DSO has short memory, then starting from the fastest time base settings down, you have to throttle back the digitiser to stay within the limits of the memory you have available.  If the memory is 1M, that is very long, and will the major factor in why you don't see alias as often.

You also said,  "My working theory is that Agilent runs their ADC at full sample rate all the time, and then they just digitally decimate and filter the data to get the suitable "display sample rate". That makes it possible to avoid aliasing and reproduce the "analog look" even with long time/div-settings. So there is obviously a way to suppress the aliasing so that it does not confuse the user. I don't know why other scope manufacturers don't do the same thing."

I'm sorry to be disagreeable, but I don't believe that using this technique contributes anything towards avoiding aliasing.  What is different between a single shot sampled at 1GS/s and decimating 1000 to 1, so the effective sample speed is 1MS/s, and a single shot taken at a digitser speed of 1MS/s?  Sample gates usually use the same fixed sample period, so that is no different.  There is no difference in the contents of the acquisition memory, so no edge against alias exists.  I think this is another spurious argument put out by salesmen keen to empty your wallet.

I don't wish to create discord, I just can't see how this can work.  If you have a DSP chip that can keep a running record and min/max bin data every 1 nanosecond, then it might be a workable technique.  I think this is another problem like the screen alias with low cost scopes decimating acquisition memory to squeeze it onto a 300 horizontal pixel display - you can 't see a glitch in the middle of the memory, so how is that any different to a 300 word acquisition memory?  I know you can take that single shot and zoom and scroll for a long time, but in reality, most folk look at the screen, don't see any reason for further investigation, and move on to the next job.   It might as well be a 300 word memory...  Without actually processing the data and compressing it at that maximum sample rate, it is of little practical use.  I have yet to see any DSP processor that can do this.  The other difficulty is that makers are unwilling to describe how they do whatever magic is in their wondrous box.  If they don't describe precisely how it works, how can you know it is good?  I'm not going to trust anything said by someone who is paid on commission!

Even moderate price scopes like the Tek DPO2000 series don't use any display algorithm.  I was shocked to learn this recently.  For low cost scopes you expect compromise, but for a Tek worth AU$3000, I would expect better.  It makes no sense, so there must be some missing information - I hope someone can explain why LeCroy managed this over 15 years ago, and modern Tektronix can't. *shakes head*

My best wishes, Colin
Melbourne, Australia

[jahonen]

Colin,

My statements/speculations about Agilent DSO sampling/processing behavior was based just on my personal hands-on experience, not their marketing BS (but even those look pretty realistic for me based on the experience on that instrument) by any means. One of my personal favorite torture tests to any digital oscilloscope is the aliasing test. Just set the time base to convenient slow sweep time per div, and then start ramping up the input frequency, up to the point that signal vanishes (BW is reached). IMO, good DSOs do not show any kind of alias, at any input frequency or time/div setting combination. Unfortunately, I have not found many DSOs that pass this test. Actually, very expensive 6 GHz Agilent at my work didn't pass this test

Sorry for large images, but didn't bother to build another web-page to just show these.

Here is a case of a measurement of modulated RF (at low end of UHF band, about 480 MHz) on a MSO6034 with 8 Mpts of sample memory and upto 2 GSa/s, 300 MHz DSO, so carrier is clearly beyond its specified bandwidth.



We can even zoom to the field sync region without changing the timebase and re-sweeping:



Or even to scan line level (not so nice anymore but perfectly readable still):



Of course that is nicer when sampled directly:



And finally, here is the measurement of the carrier frequency:



Of course, bottom line is that can't much argue about the price point. But just wanted to show that there exists technology today to measure something similar what you described. I would have tried exactly what you described, but unfortunately I don't have a appropriate signal generator around. Although personally I think that the spectrum analyzer is much better for RF work (DSO FFT is not very good substitute for various reasons), since specifications are usually frequency domain based.

Regards,
Janne

[jahonen]

Colin,

On second thought, I did try the trapezoid you mentioned. Did you mean something like this? (RF on Y-axis and modulating video on X-axis):



here is a B/W version, you can clearly see the luminance levels:



Somewhat dark image at least on my monitor, but much better visible on the scope itself.

(I think this thread should be splitted to something else, as this has nothing to do with the Rigol sin x/x problem anymore)

Regards,
Janne

[colinbeeforth]

Hi Janne,

Very interesting results from the MSO6034.  That scope costs $12,400 in Australia, I suppose I could sell my car...  Wait, my car is not worth that much...  Even my very nice motor bike cost less than that, and I know which I'd rather ride down the road, although the picture of me sitting on top of a digital scope holding reins and revving down the road has me laughing for a bit!
 
The DSO behaves very well with RF, and far better than any DSO I have seen before.  When they come down in price, I'll push my $800 analogue scope to the back of the bench.

It would be interesting to know how they can arrange for the scope to understand what is being thrown at it, and arrange itself to give a fairly analogue type display.  How does the scope know it is receiving modulated RF?  Did you trigger externally from the video signal?

Also, with the XY pictures, one picture showed a sample rate at 500MS/s, is that operator adjustable?  How did you get the video to be coherent, so it painted a series of levels?

I suspect that my earlier comments would still apply for Rigol and the others in that price range.  However, I am pleased to see that HP/Agilent have improved their earlier products.  It's an interesting thing, Tek has always had the reputation for scopes, but even their $3000 dollar scopes don't have display compression, they have long memory, which is horribly decimated to squeeze it onto their LCD screens.  I spoke to a Tek dealer yesterday about it and was assured that they don't do anything clever with displays.  For that sort of money, I would have expected some attempt to improve such a glaring problem.  They keep on saying, but you can use PeakDetect mode.  Yeah, if you know there is something there to use it on.  Many DSOs can do many clever things, if you know about it and can set up for it.  In research and even fault finding, it is always the unexpected that wastes your time, and that is one reason why I harp on about display algorithms - done right, you see what you <didn't> expect.   The cheap scopes don't often address this either, although all the responses aren't in yet, I'm still discussing this with some makers to ensure clear communication.  One maker recently replied that how they display their waveform was company proprietary information!  I said if they don't explain the broad principle of how their test gear worked it was a sad day.  Would you buy and fly an aeroplane when the maker refused to tell you whether his undercarriange was electric or hydraulic, or tell you where the emergency release was???

Cheers, Colin

[jahonen]

Hi Colin,

It would be interesting to know how they can arrange for the scope to understand what is being thrown at it, and arrange itself to give a fairly analogue type display.  How does the scope know it is receiving modulated RF?  Did you trigger externally from the video signal?

No special setups were done. Actually, I did not use any kind of triggering on the RF plots, just untriggered single sweeps. It would not be very good if you had to setup the scope to show the signal just like it does.

Also, with the XY pictures, one picture showed a sample rate at 500MS/s, is that operator adjustable?  How did you get the video to be coherent, so it painted a series of levels?

The video was from old color bar generator what I happened to have lying around, thus the levels. I believe the sample rate and memory depth (certain channel combinations reduce that to half) is internally set and can't be set manually. Although I can't know if the ADC runs at constant rate and the data is processed afterwards. But unlike with Tek ones, the long memory does not reduce the waveform capture rate, there is no harm (turn the 10M record length on Tek 4000-series and you'll get seconds/waveform instead of waveforms/second!).

In fact I was a bit surprised too on the results since XY-plot looks really good, and even better on the scope itself. The update rate on the display is so good that it is easy to forgot that it actually is a DSO (which is common drawback on most DSOs). I guess one could look up the Agilent patents on the MegaZoom and how it works, but it really is enough for me if I don't need to worry about alias etc. Although you still need to use peak detect mode if chasing really narrow spikes, but usually casual spikes show nicely on the screen with normal display. But those would not be easy to see either on the analog scope, since they would be very faint.

If you have a chance, I suggest that you 5000/6000/7000-series by yourself, don't believe me I guess too that we have to wait still for some time to these techniques to become common. I know that Tek has DPO but that works sweep-wise, unlike Megazoom which seems to work even for single sweeps. I tried to reproduce the nice Agilent display with matlab but not much success there. According to Agilent Appnote 1604, MegaZoom uses an dedicated 0.13 µm ASIC chipset, thus it can achieve relatively cost effectively much things what would be very challenging and expensive (for manufacturing cost) using OTS parts.

I believe that cheapest scope having the same MegaZoom display processing on the Agilent lineup is the 2-channel 100 MHz DSO5012A, which should cost much less than MSO6034A. Their 1000-series is made by Rigol (essentially same machine branded as Agilent and some firmware tweaks). But I guess 5012A is still much above Rigol etc. and certainly no match for analog scope bang for the buck.

And finally, quite a coincidence that you mentioned motorbikes, since I like those too and intend to start the hobby again soon after some years of pause

Regards,
Janne

[EEVblog]

I miss my Agilent 6000 series scope I had at my previous company 
I used to have a 54621D at home and that was pretty darn good, but not a fast sample rate or high bandwidth, it had Megazoom too.
I thought it was "InfiniiVision" technology that gave the Agilent that kind of performance?, the megazoom I thought was just the deep memory architecture? But I could be mistaken.

Tek claim DPO is better than Megazoom:
www.tek.com/Measurement/App_Notes/41_18987/41W_18987_0.pdf

Dave.

[rf-loop]

Becouse off topic. And also talking about expensive oscilloscopes.
Bacround: I have tens of years experience with many kind of oscilloscopes and also I have many opinions what are good and so on... also in history I have change my opinions. Future I do not know if I need agen change my opinions. But if we talk cheap "oscilloscopes" and reasonable good if compare price. Rigol DS1000E is "not bad".

I have keep strong opinion that Tektronix and also some Agilent (and HP) are good equipments. Also today I think that ( also price question becouse today electronics for me is only hobby) never I do not loose my some Tek analog oscilloscopes... they are like Rolls-Royces.

But now this world is fast changing and China go so fast that...maybe we can not understand. It is not yet in full speed but it is going also fast now. Maybe I need in future change all my opinions what are good and bad...

"Who is this" and also I can more underline question... who is agilent... and who is Agilent or Tektronix tomorrow. My forecast is that some day Agilent wake up to this kind of  morning that all have changed.. how fast it come.. this can not know.

So if talk medium expensive scopes.. Who is this, what he can do? :

[jahonen]

I miss my Agilent 6000 series scope I had at my previous company  
I used to have a 54621D at home and that was pretty darn good, but not a fast sample rate or high bandwidth, it had Megazoom too.
I thought it was "InfiniiVision" technology that gave the Agilent that kind of performance?, the megazoom I thought was just the deep memory architecture? But I could be mistaken.

Tek claim DPO is better than Megazoom:
www.tek.com/Measurement/App_Notes/41_18987/41W_18987_0.pdf

Dave.

That is basically true what is shown at the PDF, although Agilent introduced later a "Freeze display" function which retains the display as it is, retaining the shades.

There has been a number of generations of MegaZoom. Agilent 6000-series has 3rd generation. We have at work one of the early ones, 54645D, which has 1 Mpoints of memory but does not perform shading of modulated signals. That must have been introduced later.

MegaZoom works nicer on modulated signals (i.e. RF), whereas DPO shows nicer sweep-wise shades, although the difference in practice is not so dramatic as the PDF shows. The only thing that 6000 misses is the variable persistence. I also guess that sweep-wise shading is much easier to implement than shading with the modulated RF, like in the pictures I shown. For that, the DPO fails miserably due to alias problems. And the nice DPO shades are gone instantly (and cannot be recovered) if you touch the time/div knob, whereas MegaZoom shading is perfectly retained even if you zoom around the captured waveform.

Having used quite a lot both Tek TDS3000 and Agilent 6000-series, my conclusion is that in practice, I like the MegaZoom much more. For some odd reasons, the DPO seems only to work as one expects on fairly limited sweep speeds (fast ones). That "loss of DPO" seems to happen just when you would have wanted to see the shades. And if you turn on the roll mode, the displayed trace sometimes flickers like a badly coded computer game, something what one wouldn't expect to see on expensive professional equipment. Although the PDF implies that 1 Mpoint memory is available only at special conditions, even halving that (512k) is very much more than Tek 3000's 11k per channel, what is not even expandable. I believe Agilent now ships all their models with all the installed 8 Mpoints of memory enabled

I think that InfiniiVision term was coined somewhat later on, since my early (autumn 2005 I think) model of MSO6034A is clearly labeled as MegaZoom.

Someone said that Agilent never know how to make a good scope and Tektronix couldn't make a decent spectrum analyzer. Well, I don't know about Tektronix spectrum analyzers but certainly current Agilent scopes are very good ones.

It is certainly possible that someone in China will make something similar in the future. I don't know if Agilent chipset is on sale for a price if anyone wishes to use this same technology, or if someone finds a way to do same cost-effectively using OTS components. I would certainly like to see same development in lower price range too, although that would not change things for me for quite a while.

(I'm not in any way affiliated with the Agilent, but maybe I should start selling the Agilent scopes since I seem to like them )

Regards,
Janne

[chscholz]

Still the same. By default sinx/x is off, the user can enable sinx/x if he wants to.

[...] LeCroy knew this and never had sinx/x interpolation on any of their earlier high powered scopes - but I have not looked at their range for 5 years, so I don't know what they do know.  [...]

[Mechatrommer]

well. i have alot of pending stuff to be done but i dont know why i got the chance to read this year old thread (maybe a link from another thread). i remembered passing by here time ago, but only for a quick look, this time i read the whole thread (except the war stuff). kinda interesting stuff, and i believe the knowledge is still relevant. some important point to highlight

1) the OP msg is lost in somekind of DSO performance, analog vs digital war.
2) this thread is unconcluded, most people argued with OP that sinc(x) is supposed to be that way. ie, not plotting through the real sample, but the OP still hold strong to his belief.
3) jahonnen as always, proved that its possible to curve fit the real data using sinc(x). both picture repeated below for comparison. first is i believe closer to OP (rf-loop) reported as Rigol implementation of sinc(x). and the second picture is what the "should" theory said. (who am i to argue with Nyquist )

point (1) i believe its due to issue arise on "undersampling" which lead to the war (a healthy one i believe), and OP msg got lost from there. So i have to pick side and i pick the OP himself and Janne pictures below, and... Mr Nyquist

My conclusion is based on a postulate (based on Nyquist "2 sample" Theorem and got it inversed) that if we have a set of N sampled data at consistent dt (time interval at each consecutive point), we should be able to reconstruct acceptable sinc(x) graph at half frequency (N/2dt) closer to the original data, the more point the more accurate it will be. And another postulate is based on my study years back on curve fitting algorithm, that there is (or are!) method to curve fit the real given points (spline IIRC) regardless of what the actual original data look like. and we can tune the parameters to get different but nicely shape curve (points) fitted graph.

And its apparent Rigol is not doing that (curve or point fitting). will be a good subject for my study on this later. Well, this is just my 2cnts and reply is not needed. And to be clear, i'm NOT talking about "accuracy of the plotted data compare to the original". I'm just talking, its possible to do "nicely shaped" curve fitting, just as the OP concern... even if its, not accurate and undersampled.




ps: All these theorems and algorithms, are just a gross approximation/speculation of what the real science is

[saturation]

A simpler thing is if you're not sure, turn off sinc, turn off vector, and see what the actual sample dots look like, then connect the dots yourself and do the reconstruction in your head.

[Mechatrommer]

...and do the reconstruction in your head...

well thats funny human brain can generate all kind of curves. maybe the point in this is that we somehow come out with some (automated) solution (curve fitting) that we think best (backed by proven [or best] theory) and settle with that, and the newbie will have a soothing feeling of what the display shows, and for more advance people, they will know how to find the true point (samples) from there (by means of measuring at some consistent interval etc, or simply turn the nice graph off). but the main point is to the former (newbie) from the manufacturer point of view (to have a nice curve display) and for the advance people have respect at them if they do it correctly (points lies within the curve). but we all know, there is no way that the nice curve is the true representation of the continuous data (at the limit of Nyquist frequency), just dont screw up already good known data, that was the OP's point, i think.

and this kind of fancy communication theorem, is very important in situation that reach the limit of human knowledge and capability, such as when try to communicate with someone at Mars etc. without it, we can get much worst data, if not accurate. now, why we should not use this "exploit" in the lower budget low BW scope? so the newbie wont get cheated by fake data? well, thats their problem, soon they will learn. but i agree with OP that rigol has erased the good sampled data from their sinc(x), and for me thats worst than the "true sinc(x)" or "point fitting algorithm". we tried to get data as accurate as possible, not rendering 3D model in PC using Bezier curve.

[saturation]

In repetitive waveforms and other common things, you can confirm what the actual waveform appears vs the reconstruction, I think your eyeball will see its less distorted than the reconstruction but I wouldn't do a formal harmonic analysis with something like this:

https://www.eevblog.com/forum/index.php?topic=2323.msg32022#msg32022

[Mechatrommer]

A simpler thing is if you're not sure, turn off sinc, turn off vector, and see what the actual sample dots look like, then connect the dots yourself and do the reconstruction in your head.

i tried this on 2ns scale, set to dot, but the graph still connect itself. unless i press "stop" is it normal or just me?

[dfnr2]

I don't own a Rigol, so I'm curious to know what is the flaw in the sinc interpolation.  Does anyone know any details about the routine?

I am not inclined to believe that the originally demonstrated error is due to buggy sinc interpolation.  I rather suspect it's due to the samples not being evenly spaced, which violates an assumption of not only the sinc interpolation, but also the "human eye" interpolation of the uninterpolated points, since the scope displays the waveform samples with uniform spacing.

Higher bandwidth signals are more susceptible to sampling errors from jitter or DAC interleave timing errors.  This can be compounded if more DACs are interleaved at the higher sample rates.  For lower bandwidth signals, the signal simply doesn't change as much between the actual and expected sample times.  Not so for higher BW signals.

If the errors are due to random jitter, then it's better to limit the use of the scope to sufficiently low BW to not produce the distortions.  If there is a systematic error (e.g., one DAC always samples early or late) then it is possible to use a an interpolation algorithm that does not assume a uniform grid, but I doubt any scope manufacturers do this.  I suspect most focus on uniform sample timing.

I noticed this thread, because I observed something similar with my Hantek 200 MHz scope, with pics posted to the Hanktek/Tekway thread.

BTW, Sinc is handy to have sometimes, but I do agree that it's better, as a rule, to just sample much higher than the bandwidth without interpolation.

[saturation]

Yes, turning off sinc or vector is only visible only when you press 'stop'.  Not sure why that is!

A simpler thing is if you're not sure, turn off sinc, turn off vector, and see what the actual sample dots look like, then connect the dots yourself and do the reconstruction in your head.

i tried this on 2ns scale, set to dot, but the graph still connect itself. unless i press "stop" is it normal or just me?

[saturation]

For the  Rigol, its detailed in the whole thread, pictures and analysis are superbly put by jahonen.  

I think the post you made on the Hantek thread is a similar issue, its likely from the number of sample points effect on the reconstruction algorithm.

I don't own a Rigol, so I'm curious to know what is the flaw in the sinc interpolation.  Does anyone know any details about the routine?

[dfnr2]

For the  Rigol, its detailed in the whole thread, pictures and analysis are superbly put by jahonen.  

I think the post you made on the Hantek thread is a similar issue, its likely from the number of sample points effect on the reconstruction algorithm.

I did read the thread, and I do think it's the same issue, but I do not believe that this artifact is due to too small a window for the sinc interpolation, as proposed by Jahonen.  The artifact he describes at length is real; in fact this is why on LeCroy scopes the sinc interpolation is available only in the math menu, not the acquire menu, and when selected, the outermost points are not displayed.  With too small a window, you'd expect a diminished artifact at the center of the screen, and I haven't observed it.  That's why I asked if anyone has actual knowledge of the sinc interpretation implementation on the rigol.

I think it is more likely that two separate manufacturers would have problems with timing of 5 interleaved DACs, to achive 1Gs/s than that they would have perfect DAC timing, and would both use too small a convolution window, especially for a relatively limited saving in processing time.

[rf-loop]

Mainly it is not timing problem with interleaved ADC's. Not in Rigol, and specially not in Tekway/Hantek.

I have do some amount of work with Rigol for inspect this issue and issue (what I talk in this thread is absolutely "error" in Sin(x)/x (there is not any real Sinx, there is only simple filter nearly as "smooth".) If do Sinx there is not any reason to "hide" real sample point. Example some Tektronix model highlight real points in sinx display (and this highlight can select off/on. Why? Becouse Tektronix know what is oscilloscope and what is Sinx)

This Rigol issue can see very easy if you plau littlebit with scope itself. This comparation is included my original pics.

Also I have done some inspection for finding sampling timing jitter... of course as we know.. ADC is also "modulator."  You can modulate output with sample timing...

But with simple method it can look if situation is very bad or not. All you need for this is read littlebit about how these machines work and example some nice articles from example Agilent library.

For Hantek I have not yet make enough isnpections. Only what I have look is spectrum analyze sampling clock (9288 signal ENCx) and it is not bad. Phasenoise and "jitter" I have not check but with my "eyes" and simple method look... not very bad.


Mechatrommer show nearly just what what Rigol do in Figure 1.
If this is not issue in Sin(x)/x then what is... and this pic is not from sleeping dreams... it very good estimate just about what Rigol do in real world.  Most bad is that they do not correct it never. Who care... scope sell good so no problem...just this.

But it is not bad if you tightly stay <50MHz



Just same: Hantek DSO5102B is 100MHz oscilloscope. Stay below 100MHz and no big problem at all.

What I have think... maybe these economy class scopes give less problems if they do analog front end BW as "brick wall". There go away many small problems.

(do not open it more wide... reject it better. )

[saturation]

I never read much on this until now, but it does explain why very high frequency spurs occur on an FFT at high frequencies, far above the frequency under examination, and at an inconsistent amplitude:



http://www.analog.com/library/analogdialogue/archives/37-08/post_processing.html

A practical question is what do we do about it since its a hardware issue?  I guess it comes down to the useful frequency response of the DSO independent of the input analog BW limit.

Thus, hacking the analog filter off, as with both the Rigol and Hantek threads, won't make this artifact go away, so one has to weigh the benefit of the hack given the limits imposed by the sampling rate and other artifacts.

I think it is more likely that two separate manufacturers would have problems with timing of 5 interleaved DACs, to achive 1Gs/s than that they would have perfect DAC timing, and would both use too small a convolution window, especially for a relatively limited saving in processing time.

[Mechatrommer]

...If this is not issue in Sin(x)/x then what is... and this pic is not from sleeping dreams...

if what you are showing is true, then it a Bezier Curve to me, not sinc(x).

jitter problem is another issue and will need lengthy discussion in different place.
in this thread, lets just assume there is no jitter. its already simulated by Jahonen.

maybe some people get confused with this Nyquist Theorem. From what "he taught" me, if say rigol sampling at 1Gs/s, we can pretty much confident that sinc(x) will render accurately "the 500MHz harmonics (or a little bit lesser) and including the FFT anaysis. pls note i said "harmonics", NOT "fundamental". So if you want to analyze up to 10th harmonics, then you only can measure 50MHz system (fundamental) with it.

in other word, rigol will only be good for 50MHz measurement at maximum sample rate if your design requirement need stability up to 10th harmonics. at long memory double channel (ie 250Ms/s) = 125Mhz 10th harmonics, ie only good for 12.5MHz design.

thats it. i stand to be corrected.

[Mechatrommer]

...but it does explain why very high frequency spurs occur on an FFT at high frequencies...
...A practical question is what do we do about it since its a hardware issue?

from my fast read on the link, it seems about the discussion on inconsistency of ADC reading (spur?), that aspect i cannot help, and imho is why there is dso thats more expensive than a house. for the sake of simplicity of this thread, lets just assume the hardware capture mechanism is perfect (spur, jitter, noise etc). the only question is how the implementation of post processing is done (after capture/sampling, which previously i assume to be perfect), ie things like sinc(x), and fft if you want to be off topic. if we want to include hardware side (or even if sinc and fft is done in hardware), then we are in deep shit if we want to measure serious design (high speed high accuracy) with rigol ds1052e. just my humble 2cnt opinion.