Non-invasive EEG can reliably separate only a few coarse intents, not continuous fine control

What it is

What scalp EEG can stably tell apart is usually a handful of coarse classes: left versus right motor imagery, looking versus not looking at flicker, a few time-locked responses. It does not yield the continuous two-dimensional velocity a cursor needs, nor a pixel-level path. The bound is the grain of distinguishable intent, not merely “slower than a keyboard.”

Why it happens

A scalp electrode sees a volume-conducted mixture of synchronous activity over a large cortical patch. Spatial blur stacks neighboring representations; reconstructing cortical sources is ill-posed: tiny noise maps to entirely different source configurations. What survives in that noise are rhythms or evoked potentials that form large, repeatable scalp topographies. Continuous fine control wants independent, high-bandwidth dimensions; the scalp’s effective dimensionality is often two to four, further tied together by covariance and fatigue. Interpolating those dimensions into a smooth path looks continuous but is a low-pass between a few classes; a turn exposes the step. Regression of imagined force or joint angle can correlate in the lab; field variance eats the slope and cannot close a servo loop.

Studying it

Compare directly: N-class discrete choice versus a 2-D cursor versus force/velocity regression. Factors: electrode count, number of classes, trial averaging. Outcomes: between-class separability, path efficiency, and the smallest target width that still works. Calling a continuous task a success because “the cursor got there” hides the inability to make small mid-course corrections. Reporting effective degrees of freedom (how many independent dimensions stay above chance) marks this bound better than one binary accuracy.

Where it stops holding

High-density caps and individual head models improve spatial filtering; they rarely take effective class count from single digits to keyboard scale. Invasive arrays are outside this bound. SSVEP can raise the number of selectable targets with frequency tags, but each target is still a discrete flicker, not continuous fine manipulation, and is limited by visual load. When residual EMG leaks into EEG, apparently “continuous control” may not be brain signal at all.

Applying it

• Design non-invasive EEG as a few mutually exclusive commands (left / right / stop), not unconfirmed continuous dragging.
• If a demo shows a continuous cursor, publish the smallest target that remains stable and compare it with a mouse.
• Match apparently continuous motion to real resolution with explicit quantization (snap to zones).
• Verify by shrinking targets until time or errors collapse; treat that collapse as the device’s fine-control ceiling.

Related

• Same group: 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 · C9.07.4 Capability bounds move with signal-processing progress; they are not fixed physical limits
• Adjacent: C9.02 Brain-Computer Interfaces · C1.15 Pointing Device Metrics and Throughput
• Search: ill-posed inverse problem · EEG spatial resolution · coarse BCI classes