On its seminal paper, Dr. Thomas Marzetta, considers the the complex propagation cofficients into the frequency domain, however, for my study I need to know how the coefficients would be into time domain. If we take the IFFT of the coefficients on Marzetta's paper we would have coefficients with the same distribution, i.e., an independent of time index large scale term and a time-index dependent small-scale term (with variance proportional to NFFT). My question is related to the independence of the large scale fading term regarding the time index of a channel. Is it the same for all the paths (or taps of a tap-delayed line filter) comprising a channel?
Many of the paper on Massive MIMO consider frequency-flat channels (1-tap). This can model can represent a subcarrier in an OFDM system. Since the number of taps typically is much smaller than number of subcarriers, the channel coefficient at each subcarrier will be correlated. So if you want to model an OFDM system you should definitely begin with an L-tap time-domain model and then transform it into a frequency domain model.
Your question is not directly related to OFDM or Massive MIMO, but is more general. How can we model a random L-tap channel? There are different models for this. It is common to assume that the taps have independent small scale fading. The variances of the taps are called the power delay profile (PDP). Some models assume a constant PDP (= same large-scale fading at each tap) and some model assume a decaying PDP (=different large-scale fading at each tap). I would recommend the latter.
At the end of the day, if you measure the performance of a Massive MIMO-OFDM system, you will see that the PDP itself is not important, but only the sum of the variances of the L taps.
Thanks for your answer Dr. Björnson, however, I am still in doubt regarding the large scale component of such time-domain channels. Are they the same for all the paths (or taps of a tap-delayed line filter) comprising the channel? I mean, even if there are paths that are longer than others is the large scale term the same for all of the paths? Something like: g[n] = beta*h[n], i.e., beta is invariant regarding the taps of the channel.
No, in general the variances of the taps will be different (the variances are the parameters that you call large-scale components). And there are different models for how to select the variances. This is what is called PDPs.
Note that the channel filter is time-continuous, while the tap-model is time discrete (after sampling at the receiver). Hence, the L taps do not correspond to L different signal paths. It is more complicated than that...