For different type of communication, there is different type of algorithms to acquire CSI. For instance 4G (LTE) network have SRS or DMRS to estimate the channel based on pilots.
I am designing a similar algorithm based on probes (or pilots) in order to have the time based channel estimate of groundwave HF channels (this is for my equalization block). The Doppler spread measuring the rapidity of channel time variations seldom exceeds 1 Hz on HF channels.
Let $y(t) = h(t) \ast s(t) + n(t) $, where $s(t)$, the pilots symbols, have an autocorrelation given by $R_s (t) = \delta (t) $.
The pilot (or training) symbols are chosen so that the intercorrelation between $s(t)$ and $y(t)$ is given by: $R_{y,s} = R_s (t) \ast h(t) \approx \delta (t) \ast h(t) $ if one neglects the noise,
I have 15 kHz of bandwidth and I need 300 estimates. The sampling frequency is $F_e = 30\,\text{kHz}$. That's one channel coefficient every 100 Hz. One estimate takes at least 300/(30 kHz) = 0.01 s.
The channel coherency time (channel span) is about 0.5s up to 1.5s (mean value) .
For a 1800 Bd symbol rate and a maximum time spread of 5 ms, N = 22 coefficients are needed to (almost) fully represent the channel.
I will make simulation for noise and propagation conditions, with SNR varying between 5 and 15 dB.