Discrepancy in LMMSE Performance Between Time-Domain and Frequency-Domain Channel Implementations

[jiaruixu]

I am experimenting with the example code 'MIMO OFDM Transmissions over the CDL Channel Model'. My goal is to compare the performance of LMMSE with both perfect CSI and estimated CSI. I noticed that when using the time-domain channel implementation, LMMSE with perfect CSI performs worse than LMMSE with estimated CSI in the high Eb/No regime, which is counterintuitive. However, this issue does not occur when I switch to the frequency-domain channel implementation. Could this discrepancy be due to a problem with my simulation settings, or is there a potential issue with the time-domain channel implementation?

Here is the code for reproducing my observation. Colab code

[jhoydis]

Hi @jiaruixu,

I would rather say that this could be a problem of the way the perfect CSI for time-domain simulations is computed. It could be related to this line:

 h_freq = cir_to_ofdm_channel(self._frequencies, a_freq, tau, normalize=True)

This channel frequency response might be slightly different from the one you would obtain when computing the DFT of the complex-baseband equivalent channel impulse h_time. In other word, using h_freq as perfect CSI might be slightly wrong. Maybe you could give this a try.

[jiaruixu]

Thank you! I changed to use the DFT of the h_time. The problem is solved.