Rigol DS2000 series High-resolution mode bandwidth, how is it determined

[ab initio]

Hi all,

I can't seem to understand exactly what my ds2072a is doing when operating in "HighRes" acquisition mode. In theory, it is quite simple to understand: the scope samples a signal at a rate *hopefully* much higher than the actual signal's bandwidth (i.e., we have an oversampled  signal) and then uses a low-pass filter and decimation (e.g., boxcar averaging) to produce a downsampled signal with increased resolution. Here, if the number of samples in the average is N, then the expected increase in resolution given in units of bits is

Increased resolution in bits = 1/2 * log₂( N )

For the ds2072a, the high resolution acquisition should provide effective 12 bit resolution at the expense decreased bandwidth.

This concept is completely fine with me; however, I fail to understand how it is actually implemented, because the results I see on the scope imply something else is going on under the hood. Namely, the theory of operation and the actual bandwidth in highres mode seem to be in disagreement.

I'm sure you all are about to post this link, so let me beat you to it:

Here are my first impressions and review of the Rigol DS2072 (plus added Sample Rate, Segment Number, Fastest Update Rates, 56MPt Acquistion Time Table, FFT Ranges, and High Res Bandwidths)
....

Edit9: Rigol DS2000 series - Bandwidths of DSO when in High Res mode:


....

Here's my issue: Hooking up the probe to the probe 1kHz probe compensation signal and setting the time base for 20ms/s, and acquire to "normal", the scope samples at 50MSa/s and creates a nice intensity graded display indicating a signal predominately at 0V and at 3V.
[see Figure1]

Now, setting the scope to "high res" acquire mode, the signal is badly aliased. Why?
[see Figure2]


The signal is 1kHz square wave. The sample rate is 50MSa/s and with 256 point averaging, therefore, the effective bandwidth should be

Effective bandwidth = Min( Analog bandwidth, (SampleRate/2) / N )

Since the scope's analog BW is well above everything else, all we need to worry about is the BW of the sampling/averaging combination. Here, 50 MHz/2 / 256 = 96 kHz

So what is going on? 96 kHz >> 1 kHz such that I expect the 1kHz square wave and plenty of its harmonics to be correctly captured using High-res acquisition mode! I expect a nice intensity graded display indicating most of the signal is at 0V and at 3V, not the absolute garbage I'm seeing. Why does the bandwidth appear to be really poor and the display horribly aliased?

Furthermore, it doesn't seem like the data is simply being decimated first to 700*256 samples, then being boxcar averaged, and then decimated down to 700 display points. At the 50 ms/div time base, the rigol outputs something like Figure3
[see Figure3]
However, if it was just pre-decimating, averaging, and the plotting the 700 averages, it should look like Figure4
[see Figure4]

It looks like neither!? How are the bandwidths in that tabled determined, and why are they so low compared to what the should be given 256 sample boxcar averaging and the scope's sample rates?

What is going on? 

Cheers