TDL channel normalization about sionna

[faycalaa]

The observed artefact is expected when normalizing the channel frequency responses of each batch example separately. Starting from version 0.11, CDL and TDL power delay profiles are normalized to have unit total power. For system level models (UMi, UMa, and RMa), the power delay profiles were already normalized. Note that the per-example normalization is still very useful to get a clean SNR defintion since codewords are contained within one slot.

[faycalaa]

Your understanding of how (optional) normalisation is done in Sionna is correct. Optional per-batch-example normalisation was implemented as it allows convenient control of the SNR of a resource grid or block by fixing the transmit power or the noise power.

Note that the TDL channel model provides access to the path powers through the property mean_powers. Ensuring average power of one could be achieved by (i) not normalising with Sionna through cir_to_ofdm_channel() or cir_to_time_channel() and (ii) normalising the channel impulse response by dividing the paths by sqrt(sum(mean_powers)).

[Inv4lidn4m3]

Thanks for your answer. For the TDL I have actually created a custom version of Sionna's TDL channel model, so that I can normalize the channel as you propose (and also set custom json model). But I wasn't sure it was fully theoretically correct to do it that way. I add a _normalize_taps parameter and simply add this code at the end of the _load_parameters() method:

      if self._normalize_taps: 
          self._mean_powers = mean_powers / tf.reduce_sum(mean_powers) # do not need sqrt() as it is power
      else:
          self._mean_powers = mean_powers

Sionna's per-batch-example way to normalize should be ok most of the time, but I think it may create some strange behavior if the sample size is to small. For instance the noise power is not normalize on a per sample base: the noise distribution is set once and the Monte Carlo simulation will provide results corresponding to the desired noise distribution. You do not need to normalize the noise power for each sample and you should be able to do the same with the channel distribution.

From this code I obtain the following figure (sorry for the custom class, you should be able to reproduce the first curve with a normalized TDL model A json and standard Sionna's TDL channel model):

Sionna's normalization seems here a little unbalanced in the spectrum. But this is not always the case depending on the fft size. Again, I think this is not a major issue or not even an issue, but I fill like it could be done more correctly if that makes sense.

Anyway, thanks for the great job!