new killer scope in town - a true game changer from R&S - RTB2002 & RTB2004

[bayjelly]

It is independent of the window. Any of these: FlatTop, Hanning, Rect etc. gives the same false harmonic result. Each RTB2000 owner can verify this on his own.

For comparison, Siglent SDS2000X+ and Keysight DSO3034 provides an adequate spectrum for all settings and situations.

Can you show us that? Maybe there is an actual bug on how the RTB applies the window function, I'd sure want to know that. Be sure to include what window was chosen on either scope.

[maxspb69]

The second situation where a false harmonic is observed is a simple 433.9 MHz generator, which produces a fairly clean sine signal and spectrum. RTB2000, in addition to the fundamental frequency, always shows a false harmonic at exactly 382 MHz, the other two my scopes (Siglent and Keysigth) - only one  peak at 433.9 in the 200 MHz span.

[maxspb69]

Can you show us that? Maybe there is an actual bug on how the RTB applies the window function, I'd sure want to know that. Be sure to include what window was chosen on either scope.

Ok, but tomorrow. ))

But you yourself can easily reproduce the situation. RTB2000 and a 50Ohm cable with an BNC connectors are enough for this

[Zlotnik]

After playing around a bit myself, I think we're seeing aliasing artefacts. I'm not sure yet if this is just due to the limited oversampling, or something more funky due to the AWG and ADC clocks or distortion from the AWG creating a signal with higher frequency content than expected.
Haven't tried yet with a separate frequency generator, that will be helpful to differentiate.

To see what I mean, try out a more simple scenario than your sinc example, eg a square wave or even a simple sine wave, set the frequency to eg 1/4 of the sample rate, and tune the frequency of the generated signal up and down a bit.
The jumpiness of the effect you see in your video is just the sample rate changing to accommodate the changing stop frequency of the fft when you pan the fft.

[maxspb69]

I tried with an external generator - the same result.

[maxspb69]

Extended version of yesterday's video about FFT. The built-in Siglent generator is used so that there is no dependence on RTB2000. False harmonics on RTB2000 strongly depend on the span (the narrower the more there are). They do not depend on the selected window in anyway. Demonstrating on screens of two different scopes. Siglent always shows the expected adequate spectrum picture. RTB often adds harmonics from itself

https://youtu.be/cjJ-yCplxBU

Or am I doing something wrong?

[nctnico]

Or am I doing something wrong?

Yes. Look at the samplerate on the RTB2004!

[maxspb69]

On RTB2000 you cannot control the samplerate in FFT mode. Scope  chooses samplerate  himself depending on the  center frequency, span and RBW settings.
If the cause of phantom harmonics is a low samplerate, then the scope is choosing it incorrectly for current span/RBW and this is an error in any case.

[nctnico]

On RTB2000 you cannot control the samplerate in FFT mode. Scope  chooses samplerate  himself depending on the  center frequency, span and RBW settings.
If the cause of phantom harmonics is a low samplerate, then the scope is choosing it incorrectly for current span/RBW and this is an error in any case.

No. You are analysing a signal with the wrong settings. The samplerate is sufficient to fulfill the Nyquist criteria for the FFT settings.

[Zlotnik]

If the cause of phantom harmonics is a low samplerate

We're not seeing "phantom harmonics", but aliasing. The test case you're looking at is much too complex to see this easily though.
A better example is to set a sine wave to a frequency near e.g. slightly below 1/3 of the sample rate, and tune the frequency up until slightly above 1/3 of the sample rate.
For example:
fft stop frequency 2.4MHz has the RTB choose a sample rate f_sample=6.25MSa/s. Let's look at signals around f_signal=f_sample/3=2.0833MHz. I attach a sequence of screenshots with a sine wave of f_signal=2.077MHz, 2.080MHz, 2.083MHz, 2.086MHz 2.089MHz. You can see how the aliasing artefacts move in and then out again around f_sample/3. In the time domain you can also see the aliasing very clearly as beating.







When you chose a more complex waveform, the harmonics of the wave have the same aliasing, at the same absolute frequency differences, but this results in different relative frequency differences. The result is a nice comb of the harmonics folded down around our base frequency:


I'm not quite certain yet why we see aliasing for a band-limited waveform entirely below the Nyquist frequency (f_Nyquist=3.125MHz and f_signal=2.089MHz in our example), when we place the signal frequency close to an integer fraction of the sampling frequency. Has been some time since I ran the math myself or played around in octave, I have to admit...
Is this fundamental? Due to finite record lengths (but windowing doesn't really improve it)? Are we effectively downsampling somewhere? Some weird harmonics creating signal components above f_Nyquist? I have to think a bit more about it, but if somebody can already see the solution, let us know :-)

What's clear is that f_sample of 6.25MSa/s is less than ideal for signals around 2MHz. Just to give a gut-feeling, here is the situation in time domain, without interpolation:

I think it's pretty obvious this samplerate is not the best basis for a good analysis...

The question now is, why is the samplerate so low?

On RTB2000 you cannot control the samplerate in FFT mode. Scope  chooses samplerate  himself depending on the  center frequency, span and RBW settings.
If the cause of phantom harmonics is a low samplerate, then the scope is choosing it incorrectly for current span/RBW and this is an error in any case.

No. You are analysing a signal with the wrong settings. The samplerate is sufficient to fulfill the Nyquist criteria for the FFT settings.

@Nico, you're right, the samplerate is sufficient to fulfil Nyquist at the FFT settings. But we still clearly see aliasing.
And max is right as well: the RTB2k does not give the option to choose a different sample rate. Actually, this is infuriating:
The sample rate is chosen automatically to sit a step or two above Nyquist for the chosen stop frequency. Even without this weird aliasing, there are situations where you know you have frequency components above the automatically chosen Nyquist, but you're zooming the FFT to look at a lower frequency component. You cannot avoid downconverting some stuff into your lower-freq fft band of interest in that scenario!

Unless I just missed how to set the sample-rate. @Nico: how does this work in the RTM?

[nctnico]

The sample rate is chosen automatically to sit a step or two above Nyquist for the chosen stop frequency. Even without this weird aliasing, there are situations where you know you have frequency components above the automatically chosen Nyquist, but you're zooming the FFT to look at a lower frequency component. You cannot avoid downconverting some stuff into your lower-freq fft band of interest in that scenario!

Unless I just missed how to set the sample-rate. @Nico: how does this work in the RTM?

On the RTM3004 it works the same. You have to set the RBW stop frequency to be above the highest frequency in the signal to avoid aliasing. But in the end there is a limited number of FFT points which limits the frequency resolution. The only way around is to limit the frequency spectrum of the incoming signal to make sure there is no aliasing. This is true no matter what the FFT depth is.

Edit: not RBW but 'stop frequency'. Long day...

[bayjelly]

I suppose the scope could oversample and digitally lowpass filter the signal before the FFT, but I don't know if that's commonly done. And I don't know if I would consider that intuitive/expected (unless it was an explicit option) either.

[Zlotnik]

On the RTM3004 it works the same. You have to set the RBW to be above the highest frequency in the signal to avoid aliasing. But in the end there is a limited number of FFT points which limits the frequency resolution. The only way around is to limit the frequency spectrum of the incoming signal to make sure there is no aliasing. This is true no matter what the FFT depth is.

Huh? Do you really mean the RBW?
On the RTB changing the RBW just changes the "Time Range", i.e. the how long a window from the time domain capture is used for the FFT. The actual acquisition is not touched, the sample rate remains constant. The only way to increase the sample rate is to increase the max frequency displayed in the fft.
Is this different on the RTM?

I suppose the scope could oversample and digitally lowpass filter the signal before the FFT, but I don't know if that's commonly done. And I don't know if I would consider that intuitive/expected (unless it was an explicit option) either.

Maybe I'm being daft, but I don't yet see the need to digitally lowpass filter beyond just cutting the top end of the fft as the user asks. Pretty much what happens with the low-frequency end of the fft when you don't display down to DC, unless I'm missing something.
If the user is fine with a lower RBW, the scope could up the sample rate and just display the fft only up to the chosen stop frequency.
In the example in the below screenshot, I keep the fft start and stop frequency as before, but relax RBW to 16kHz. This reduces the time window to 90.24us (down from 3.97ms in the previous examples with an RBW of 362Hz). This fft uses just 564 points.
The scope could display the same range of the fft, but increase sampling to the full 2.5GSa/s, and still not hit the 128kpoint fft limit. It would be a lot slower, but I would like to have the option.

@nico, is this how it's on the RTM?

[bayjelly]

Maybe I'm being daft, but I don't yet see the need to digitally lowpass filter beyond just cutting the top end of the fft as the user asks. Pretty much what happens with the low-frequency end of the fft when you don't display down to DC, unless I'm missing something.

If you sample at a low sample rate (or really any sample rate, but illustratively for here), then any component of the signal with a frequency above half the sample rate (for real signals) will get folded into the range from 0 to half the sample rate: Aliasing.

If you'd instead, for example, sample at the full sample rate of the scope, then lowpass filter, and then downsample, then there are no components left at frequencies over half the sample rate that could alias back into it.

For anything that's over the scope's bandwidth, that's of course already filtered by... well, the frontend, the very definition of the scope's "analog bandwidth", so that shouldn't be an issue.

But it's entirely possible that *I* am the one being daft right now.

[Zlotnik]

... okay. There definitely is more going on here than simple undersampling.
I've tried an experiment with a definitely not undersampled signal, and there's spurious galore which move like aliasing as shown in the more modestly sampled example above.

The first screenshot shows a square-wave example like before. f_sample=6.25MSa/s, f_signal is very close to 1/3 f_sample at f_signal=2.08MHz. Same frequency ratios as before. RBW is reduced to 1.81kHz to be able to keep it constant. For that reason I've also chosen a square wave, so that we can see the spurious better.


Now, the same signal at a stop frequency of 10MHz, for which the RTB choses f_sample=31.2MHz. The signal now lies close to f_sample/15. If it was just undersampling, we'd expect to see the 2.08MHz signal and the 6.24MHz 3rd harmonic. Instead, we see some spurs near the second harmonic, and loads of spurs both around the fundamental as well as the third harmonic.

Those spurious move about like aliases when the fundamental is moved around a integer fraction of the sample rate, as before in the more simple sine example. I'll not bother to create a gazillion screenshots to demonstrate that though, as it clearly is not just undersampling and aliasing.

I start to concur with maxspb69. Something fishy is imho going on. Those spurs should not appear.

@Rich, are you there?

[Zlotnik]

If you sample at a low sample rate (or really any sample rate, but illustratively for here), then any component of the signal with a frequency above half the sample rate (for real signals) will get folded into the range from 0 to half the sample rate: Aliasing.

If you'd instead, for example, sample at the full sample rate of the scope, then lowpass filter, and then downsample, then there are no components left at frequencies over half the sample rate that could alias back into it.

Trudat.
However, the most simple low-pass filter could be simply "don't display the calculated fft beyond f_stop". I think this is what happens to the poor fft points below f_start (but not being a DSP wizard I might be mistaken).

It would be wasteful of resources, but there can be scenarios where it makes sense.

... but anyway, that's not the main point. The spurious seen in the fft are not from undersampling, see the experiment in my last post.

[bayjelly]

However, the most simple low-pass filter could be simply "don't display the calculated fft beyond f_stop". I think this is what happens to the poor fft points below f_start (but not being a DSP wizard I might be mistaken).

Oh, you mean always sample at the full sample rate, and never downsample? But that means that you fit much less of the waveform in memory... wouldn't that be an issue here as well?

Because as soon as you downsample/sample at a lower rate without filtering beforehand, well, there's your aliasing again.

[Zlotnik]

Oh, you mean always sample at the full sample rate, and never downsample? But that means that you fit much less of the waveform in memory... wouldn't that be an issue here as well?

Not necessarily the full sample rate, but selectable sample rate up to full or whatever fits into the fft buffer.

The way it's implemented now, if you reduce the RBW, but keep the same stop frequency, the scope just wastes samples. They're up there in the time domain plot, but not used for the fft. This can be useful to check eg the fft of a short event visible in the time domain plot, or it can increase the frame rate as you have a lot less points to transform.
But it would be handy to have the choice...
Take a look at my RBW 16kHz screenshot a few posts above.

... but again: while this would be a nice feature (which I think other scopes do have?), this is besides the point for the issue maxspb69 uncovered. Higher sampling rate does NOT seem to help, the issue appeared at f_signal=f_sample/15, which should be plenty for a square wave. It does appear however when there's frequency components close to an integer fraction of the sample rate, which seems fishy to me.

[nctnico]

On the RTM3004 it works the same. You have to set the RBW to be above the highest frequency in the signal to avoid aliasing. But in the end there is a limited number of FFT points which limits the frequency resolution. The only way around is to limit the frequency spectrum of the incoming signal to make sure there is no aliasing. This is true no matter what the FFT depth is.

Huh? Do you really mean the RBW?

No, I made a typo. It has to be stop frequency. I fixed my posting.

[Zlotnik]

@nico, do you see spurs like this appearing for signals close to integer fractions of the sample rate on the RTM as well?

[Zlotnik]

@maxspb69 same question for the SDS2k+. Could you try to place a signal close to the sample rate the SDS2k+ chooses, and see if you see spurs?

In your videos it looked like the sample rate was 100MSa/s, so the situation will be very different. Maybe it will be useful to limit the sample rate to create a more comparable situation.

[tautech]

@maxspb69 same question for the SDS2k+. Could you try to place a signal close to the sample rate the SDS2k+ chooses, and see if you see spurs?

In your videos it looked like the sample rate was 100MSa/s, so the situation will be very different. Maybe it will be useful to limit the sample rate to create a more comparable situation.

Typically Siglents do better FFT with a slow timebase setting with lots of periods displayed however you're up against it with the SDS2000X Plus 2 Mpts FFT.
Just been playing with their new economy SDS1104X-U and while it's limited to just 128 Kpts FFT it does a decent job in the short while I've been fiddling with it.

[Zlotnik]

Thanks tautech.
Could you try explicitly setting a signal very close to an integer fraction of the sample rate? I think this is what's causing the spurs and it would be interesting to see if the problem does not occur in other implementations.

[2N3055]

Problem is that realtime FFT based spectral analysis cannot have all the combinations of sample rate RBW and bandwidth, like classic SA.
Without going too much in detail, it simply means this: if you want to look at 10 MHz with span of +- 1 MHz, that means you need to calculate FFT from 0 to 12 MHz, and then only show from 9-11Mhz. To do that , you need to sample at least at 25 MHz, and to get wanted RBW you need to get enough FFT bins to achieve it. It is not like with standard SA that can have constant RBW anywhere in a full span, and capture only bandwidth of interest. In reality you also might want to oversample a bit to avoid aliasing.

Spectrum analysis option on RTM3000 should use more advanced algorithms borrowed from realtime SA (R&S should know how that's done!) like overlaped FFT and many other tricks. I would be surprised if that one shows same problems.

I agree with Zlotnik (golden coin  ), problem is that you don't have control over sampling rate. I like how Keysight implemented this, timebase fixes sampling rate, FFT will show you what is possible with that sample rate.  Also Picoscope has interesting implementation, they don't pretend it is SA, they simply do FFT from 0 to 1/2 sample rate, number of bins defines RBW as percentage of sample rate, and if you wan't to see something in the middle you zoom in on the screen... It seems weird to someone coming from SA, but is consistent with how FFT works, so no surprises.

[Zlotnik]

Problem is that realtime FFT based spectral analysis cannot have all the combinations of sample rate RBW and bandwidth, like classic SA.
Without going too much in detail, it simply means this: if you want to look at 10 MHz with span of +- 1 MHz, that means you need to calculate FFT from 0 to 12 MHz, and then only show from 9-11Mhz. To do that , you need to sample at least at 25 MHz, and to get wanted RBW you need to get enough FFT bins to achieve it. It is not like with standard SA that can have constant RBW anywhere in a full span, and capture only bandwidth of interest. In reality you also might want to oversample a bit to avoid aliasing.

Yep, I get that.
That's why I meant that it would be useful in some situations to have more control over the sample rate that the RTB2k gives, eg to avoid aliasing at the expense of RBW and update rate, when you know there are some components that would alias down into your frequency range of interest.

Spectrum analysis option on RTM3000 should use more advanced algorithms borrowed from realtime SA (R&S should know how that's done!) like overlaped FFT and many other tricks. I would be surprised if that one shows same problems.

I'm hanging on the edge of my seat to see what Nico reports from his RTM3k. And expect to be jealous.

I agree with Zlotnik (golden coin  ),

Hehe. Indeed one of my non-electronics hobbies is goldsmithing.
It's a really old handle named for Pitr from the old User Friendly webcomic, though.

problem is that you don't have control over sampling rate. I like how Keysight implemented this, timebase fixes sampling rate, FFT will show you what is possible with that sample rate.  Also Picoscope has interesting implementation, they don't pretend it is SA, they simply do FFT from 0 to 1/2 sample rate, number of bins defines RBW as percentage of sample rate, and if you wan't to see something in the middle you zoom in on the screen... It seems weird to someone coming from SA, but is consistent with how FFT works, so no surprises.

In principle I like the SA style approach a lot. When I look at the fft I'm not interested in the technicalities of the transform, but in the spectrogram. So setting parameters in the spectrogram makes more intuitive sense to me than thinking about Fourier transforms.
(Yes. You do need to understand what's going on under the hood).
I would prefer to be able to simply tap the sample rate, and increase it if I think it's needed. Then the update rate should drop, and at some point the RBW should increase.

But again: the sample rate does not seem to be the problem! There's spurs that appear in some scenarios. Sometimes they're barely noticeable, but in some scenarios they're all over the screen, at the level of the signals real harmonics. Eg when you place a square wave close to an integer fraction of the sample rate. (Actually: probably also for other rational fractions)
That way you cannot trust the fft, and unless I'm missing something very fundamental I think this is a major flaw.

Actually this would be exactly the kind of flaw that made me opt to shell out for an A brand scope instead of eg the SDS2k+. I was sick of my old Rigol, and a bit turned back by the idiosyncrasies discussed in the SDS2k+ thread. The RTB2k seemed to have less features and sticker figures, but seemed more reliable and polished. Hope I'm not proven wrong.