Capability bounds move with signal-processing progress; they are not fixed physical limits

What it is

That today’s scalp channel cannot separate a continuous fine trajectory does not weld the line to Maxwell’s equations. Volume conduction and skull filtering are physical constraints; they limit information in the absence of noise models and priors. Dry electrodes, individual head models, transfer learning, and better stimulus codes push the reachable operating point outward. The bound moves. What moves is the engineering-reachable region, not a cancellation of “thought becomes text.”

Why it happens

An ill-posed inverse becomes somewhat more tractable once anatomical priors, temporal smoothing, and task structure are added: the solution is still non-unique, the acceptable set shrinks. Deep models can borrow statistical strength across people and cut per-person calibration, trading a data prior for training cost. Stimulus coding (denser frequency tags, shorter trials) raises observable bits per unit time. All of that changes estimator quality, not the skull’s role as a low-pass—SNR from terrible to merely poor can take class count from 2 to 8, and struggles to take it from 8 to 26 independent continuous dimensions. Writing current failure as “physically forever impossible” misses the movable stretch; writing a laboratory record as the physical cap lifted misses that the skull is still there.

Studying it

Systematic reviews by year or method generation: how ITR, class count, and calibration time move on the same task. Prospective work ablates: drop the head model, drop transfer, drop trial averaging, and see where the bound retreats. The essential report is what was assumed about the physical constraint. Showing only a new model’s gain on an old benchmark, without electrode lift-off and cross-day tests, sells movability as unconditional.

Where it stops holding

Some physics barely moves: conductivity and geometry of an intact skull, short of surgery. Swapping wet electrodes for dry ones can pull the bound inward (higher impedance); progress is not monotone outward. Ethical and regulatory bounds (who may be implanted, whether data may leave the body) are independent of signal processing and do not open automatically when accuracy rises. Product claims should mark current reach “under which electrodes, which calibration, which task,” and expect the numbers to age within two years.

Applying it

• Write capability as a dated, method-tagged operating point (“2026, dry electrodes, 8 classes, 10 min calibration”), not as an eternal ceiling or an eternal lack of one.
• Leave room in the architecture for more classes and continuous dimensions, but keep default interaction on discrete choices that are stable today.
• Refuse unreplicated single-paper peaks as roadmap milestones.
• Verify by retesting the same task on the same user cohort each year; if the numbers move out clearly, update the task set promised to users, not only the paper list.

Related

• Same group: C9.07.1 Non-invasive EEG can reliably separate only a few coarse intents, not continuous fine control · C9.07.2 Invasive electrodes have substantially higher bitrate, but surgical risk confines them to medically necessary settings · C9.07.3 Current BCIs fit discrete choice, not replacing continuous pointing or text input
• Adjacent: C9.02 Brain-Computer Interfaces · C9.10 Individual Calibration and Baseline Drift
• Search: BCI capability bound · EEG inverse problem · transfer learning BCI