Statistical Methods
Audit Coordinate Alignment Without Hiding Scale Differences
Define the task
Use orthogonal alignment when two configurations contain the same matched points in the same units. Verify correspondence before fitting: a row permutation can change the answer. Decide whether reflections are scientifically permissible. An unrestricted orthogonal transformation can reflect as well as rotate. The tool permits both, with no automatic centering, translation or scale adjustment; anatomical registration and patient-specific interpretation are outside its purpose.
Carry out the check
Minimize the squared differences between A R and B with RᵀR=I. Inspect the determinant: +1 denotes a planar rotation and −1 a reflection. If translation removal is justified, center both configurations explicitly and retain their original centroids. Do not confuse SciPy's returned scale, a singular-value sum, with a fitted size multiplier. Degenerate configurations can have nonunique orientations, and a small residual is not a measurement-agreement analysis.
Worked hypothetical example
Hypothetical example: A contains (1,0), (0,1), (−1,0); B contains (0,1), (−1,0), (0,−1). The row-vector matrix with rows [0,1] and [−1,0] maps every A point to B, giving squared residual zero and determinant +1. Doubling B's coordinates introduces size disagreement that an orthogonal map cannot remove. Record residuals, determinant, coordinate units and any preprocessing with the result.
Method checked 22 November 2026 against Primary method documentation: orthogonal_procrustes.