Hi,

We know that high-order delta-sigma modulators with a single-bit quantizer have limited overload capabilities, and a large enough input can destabilize these modulators. Typically, one does not expect better than -6 dBFS peak input for overload.

However, if one uses a CIFF (feedforward branches based) structure, one has the option of scaling the integrator outputs such that the last integrator will saturate first and the first integrator will saturate last (or not at all). So, for a given signal level (say - 6 dBFS), the first integrator has the smallest swing, the next one a bit more, the next one a bit more etc.

Now, if one adds signal limiting here which will naturally arise due to limited output swing of real integrators, then it is possible to see that the modulator gracefully degrades , reducing from a 4th order to progressively become a 2nd order, and maybe even a 1st order modulator as input signal levels are raised. All this is known.

Now, I am able to make a model in MATLAB with all this and see no instability even when I exceed fullscale for the modulator, in fact by a large factor, say 10x and more, simply because my integrators are defined to clip beyond fullscale.

So, where is the catch? Why do we speak of limited input signal handling capability for single-bit, higher-order modulators when it is possible to use the feedforward scheme to bypass this limitation? Naturally, my model is very simple, using sinewave inputs and the clipping is instantaneous with no large time for recovery from clipping (which will usually be the case in reality), but the principle seems to work.

Am I missing something?

I am able to see the effect of the clipping on the state histograms, so the last integrator is practically always operating under clipped conditions but the SQNR is good, the states are all well within reasonable levels.

Thanks,
Vivek

Berti wrote on Nov 4^th, 2008, 3:36am:


Hi Berti,

I agree that the useful range of the modulator is limited by the overload limit. However, with this scheme of scaling states appropriately an using feedforward, what one is achieving is that the modulator order automatically drops down from say 4 or 5 to 3 and then to 2 and then to 1 as the signals get larger and larger, keeping the modulator stable.

The useful range of the modulator is probably still something like -6 dBFS but atleast the modulator doesn't go crazy when an unexpectedly large signal occurs. This is assuming of course that SNR is not much of an issue when the signal is beyond fullscale, and neither is THD which will be also awful, only that the modulator can be stabilized inherently with this choice of architecture rather than worrying about making other tradeoffs (such as opting for poorer NTFs) which might impact SNR at smaller signal levels.

Of course, practically, one has to guarantee that the opamps which get saturated here are able to recover quickly, otherwise this reverse recovery transient will act as a further "pole".

So I think the fuss that people make about trying to make higher order modulators stable is assuming that the NTF is always the same and then there is instability. If the NTF varies, adapting itself to the signal strength, then there is no issue.

What do you think?

Regards,
Vivek