Turn your head left and right: a headset must redraw digital outlines exactly on top of real objects. Add latency or gyro drift and the outline slides away; add camera correction or motion prediction and it locks back on.
The science
Pinhole projection. A point at bearing β seen by a camera yawed at ψ lands at pixel offset x = f·tan(β − ψ) from the screen centre, with f = (W/2) / tan(FOV/2) (W = 1000 px here). A hologram anchored to a real object should sit at the same x as that object.
Latency. If the pose used for drawing is τ old, the error is Δx = f·[tan(β−ψ(t−τ)) − tan(β−ψ(t))] ≈ f·ω·τ for an object near the centre, where ω is the head angular speed. For a sinusoidal sweep ψ = A·sin(2πν t) the peak is f·A·2πν·τ. Extrapolating the pose by τ (ψ ≈ ψ(t−τ) + ω·τ) removes the first-order term, leaving only the acceleration term.
Drift. A gyroscope with bias b integrates a wrong heading; a camera/SLAM correction pulls the estimate back at rate k: d(drift)/dt = b − k·drift, so it settles at b/k (with k = 0 it grows without limit). White jitter σ models pose noise from feature tracking.
Simplifications (honest). Yaw-only (one axis), perfect knowledge of the real scene, pinhole camera without lens distortion, no depth/parallax, ideal velocity for the predictor (real predictors use noisy gyro data), optical see-through with zero latency for the real world (video passthrough would delay both). All numbers are illustrative, not specifications of any product.