Neural nonstationarity and signal drift

Neural nonstationarity is the fact that the statistics of a chronic recording change over time, so a decoder fit at one moment is progressively mismatched to later data. On short timescales, electrode micromotion, changing noise, and shifts in which units dominate a channel alter the observed features within a session; across days and weeks, the specific neurons captured turn over, waveform amplitudes decay, baseline firing rates shift, and the tissue interface remodels. This drift, not a lack of information, is the leading practical reason motor-BCI performance decays without intervention.

Nonstationarity is what forces recalibration, adaptation, and alignment methods to exist. It is usefully separated into changes at the electrode/waveform level (some channels gain or lose units) and changes at the representational level (the population code itself reorganizes, e.g. through plasticity). The first can be mitigated by tracking features or re-fitting the observation model; the second is subtler because the map from intention to activity has genuinely moved. A key modern observation is that although individual channels are highly nonstationary, the low-dimensional latent structure they span is comparatively stable, which motivates latent alignment approaches.