The lower bound on target size is set by accuracy, not precision

What it is

How small an eye-controlled button can be is set by accuracy (whether mean offset walks the landing off the target), not by precision (how much the landing jitters inside it). Poor precision makes edge hits flicker, but if the mean still sits inside and the target is a bit wider than the jitter, most frames still hit. Poor accuracy moves the whole cloud into a neighbor’s house; no amount of smoothing the target saves it. Set the floor from accuracy, then add a little edge margin from precision.

Why it happens

Treat the target as a circle of radius R, mean offset A, and a characteristic jitter P (say 1 RMS). The center of the landing distribution sits A from truth. Keeping the bulk of the distribution inside the target needs R ≳ A + kP, with k around 2. When A is much larger than P—as is common in gaze, where post-calibration P can be 0.2° and A 1°—the inequality is dominated by A. Sizing a button to “just larger than RMS” looks scientific; the cloud is already outside.

Spacing is likewise set by A: center-to-center distance must exceed about 2A, or offset will stably send samples meant for the left into the right. Precision only asks that the edge not flicker between two hit regions, a smaller term.

Studying it

On measured A and P, sweep target width and spacing and plot hit rate and neighbor-hit rate. Hit rate typically improves sharply once width crosses about 2A, while a width only slightly above P barely moves neighbor errors. That plot puts accuracy, precision, and Fitts-style acquisition on one figure. Use A measured on the real device, not the brochure; brochure numbers are often taken in ideal pose, bare eye, central field. Assistive users may have larger A and P; the floor should follow their distribution, not the laboratory median.

Where it stops holding

If the system heavily smooths every frame, P is artificially lowered and sizing from P is more dangerous, because A has not moved. If implicit correction pulls A down, the floor can shrink during the session, but must return to the uncorrected A when correction fails. Head pointing has different A and P; it does not share this floor. Interfaces that only highlight a region and never point-select can size by region rather than by button, relaxing the accuracy constraint one notch.

Applying it

• Set minimum target visual angle from mean offset measured on this device and this population, then add about two RMS as edge; do not size from RMS alone.
• Space neighboring centers at about twice mean offset, so landings are not stably sent to a neighbor.
• Verify after measuring A and P with a “width ≈ P” band versus a “width ≈ 2A + 2P” band, confirming only the latter knocks neighbor-hits down.

Related

• Same group: C8.09.1 Accuracy is the mean offset between estimated and true gaze · C8.09.2 Precision is the spread of repeated measures; the two can vary independently · C8.09.3 High precision with low accuracy looks stable but systematically offset
• Adjacent: C2.02 Minimum touch target size · C8.02 Dwell selection
• Search: target size · spatial accuracy · gaze pointing