Preparing trajectories

Functional analysis does not remove eye-tracking preprocessing decisions. It makes some of them more consequential because the whole trajectory is retained.

Common grid

Grid FPCA assumes curves are evaluated on a common grid. Use from_irregular_long_dataframe() or resample_to_grid() only after defining a scientifically appropriate grid.

import numpy as np
from eyetrajectoriespy import from_irregular_long_dataframe

grid = np.linspace(0, 2, 121)
gaze = from_irregular_long_dataframe(
    data,
    curve_columns=["participant_id", "trial_id"],
    time_column="time_s",
    grid=grid,
    max_gap=0.10,
)

max_gap prevents interpolation across long unobserved intervals.

Missingness

A blink, tracker loss, and a true zero coordinate are different states. Missing values remain NaN until the analyst explicitly handles them.

Smoothing

smooth_trajectories() is opt-in and warns because smoothing can attenuate real abrupt changes.

Time normalization

Normalizing every trial to [0,1] changes interpretation from absolute time to trial progress.

Coordinate normalization

Scaling pixels to [0,1] makes resolution-independent coordinates; it does not make semantically different layouts comparable.