Predictive compensation overshoots when direction changes abruptly
Aliases: predictive compensation · overshoot · abrupt direction change
What it is
To mask touch latency, a system can predict the next finger position from recent velocity and render ahead. Predictive compensation reduces lag in stable motion, but when a person suddenly turns, stops, or reverses, prediction continues along the old path and overshoots the real position.
Why it happens
A predictor estimates velocity, acceleration, or trajectory from prior samples and displays a future position, then corrects when real input arrives. A direction change invalidates the old model; larger correction creates stronger visual jump or reverse pull. There is a second-order trade-off easy to overlook here: how the correction itself is done actually reproduces the exact problem prediction was meant to solve. Snapping instantly to the true position gives the person a visible, discontinuous jump; smoothing the correction over a few frames reintroduces the same lagging feel the original latency caused, just relocated from "always present" to "a brief window right after each misprediction." Prediction, in other words, has not actually eliminated the trade-off between latency and accuracy — it has only moved that trade-off from "whether to predict at all" down to "how to correct once prediction is wrong," and the correction phase still has to choose between a jump and a lag.
Studying it
Compare no prediction and prediction strengths on straight drag, abrupt stop, corners, back-and-forth adjustment, and edge docking. Record average lag, overshoot distance, corrections, hit rate, and perceived directness; uniform straight motion is the case most favourable to prediction. One specifically targeted task is "reverse direction suddenly during the last short stretch of approach to a target," since that is exactly when overshoot does the most damage to final placement accuracy — overshoot far from the target is usually harmless, but a prediction model taking one extra step in the old direction right at the moment of docking directly hurts the hit result.
Where it stops holding
Prediction is not always harmful: stable rapid motion can benefit greatly. The problem is treating prediction as true position, especially around small targets, frequent direction change, or consequential placement. This also sets the task boundary for where prediction belongs: operations that only care about the final settled position, where momentary deviation along the way is irrelevant (most drag-and-drop), can safely use stronger prediction; operations where every instant's visual position must be trustworthy (precise cropping, point-by-point tracing) treat any prediction-induced momentary deviation as a direct error that "it'll get corrected anyway" cannot excuse.
Applying it
- Adjust prediction strength from motion stability, rapidly reducing look-ahead on stops and turns.
- Correct smoothly without hiding true position so objects do not jump back unexpectedly.
- Prefer accuracy and conservative compensation for fine placement and edge docking.