Near the cut, the same stroke flickers between two readings

What it is

When lift-off speed sits on the threshold, the same person repeating almost the same motion will sometimes page and sometimes bounce back. The critical band makes the classifier chatter. People cannot form a model of “how hard is a flick”; they gamble. The bug is not that the number is slightly wrong; it is a hard decision with no dead zone under noise.

Why it happens

Velocity estimates have variance: sample jitter in the last frames, and a collapsing contact area at lift, throw the estimate across the cut. A hard threshold turns a continuous quantity into two symbols, so tiny noise becomes opposite outcomes (next page versus snap-back). Hysteresis—“faster to enter swipe, slower to fall back to pan”—stretches the critical point into a band, so repeats inside the band keep the last reading, or refuse to commit and ask for a clearer follow. A pager with no hysteresis looks random on borderline flicks.

Studying it

Train people (metronome or on-screen speedometer) to lift near the threshold and repeat tens of times; measure how often the label switches. Add hysteresis and measure switch rate versus error. The measure is how often the label changes across two identical intents, not overall accuracy—accuracy is already poor in the band.

Where it stops holding

Clear flicks and clear drags far from the cut never see this. Hysteresis makes “just a bit slower than the cut but I still wanted a page” harder; that trades sensitivity for stability. If the tracking preview already looks as if the next page has won, chatter at lift is visually masked. Inertial deceleration curves are not this jitter.

Applying it

• Add hysteresis to the swipe decision, or refuse to commit a page inside the band and bounce, forcing a more definite flick.
• Do not let the tracking preview declare “next page has won” before the threshold.
• Repeat the same motion twenty times at ±10% of the threshold speed; more than two or three switches is chatter.

Related

• Same group: C3.29.1 The swipe-versus-pan velocity cut decides a jump or a follow · C3.29.2 Velocity must be judged together with distance · C3.29.4 Velocity thresholds must be recalibrated across size and resolution
• Adjacent: C3.06 Fling and inertia · C3.19 Gesture conflicts and disambiguation
• Search: hysteresis · decision chatter · threshold jitter