actionet.tools.scale differs from R’s base::scale with disabled or explicit centering #2

actionet.tools.scale is documented as mirroring R’s base::scale(). With scale=True, it currently divides by the sample standard deviation around the column mean, even when centering is disabled or an explicit centre is supplied.

R instead divides each column, after any centering, by its root mean square. These calculations agree when center=True, but generally differ when center=False or a centre vector is supplied. Calls with scale=False or an explicit scale vector are unaffected.

Reproduction

import numpy as np
from actionet.tools import scale

scale(np.array([1.0, 2.0, 3.0, 6.0]), center=False)
# Current ACTIONet: [0.4629, 0.9258, 1.3887, 2.7775]
# R's documented formula: [0.2449, 0.4899, 0.7348, 1.4697]

x = np.array([[1, 10], [2, 20], [3, 60], [6, 30]], dtype=float)
scale(x, center=np.array([1.0, 10.0]))[:, 0]
# Current ACTIONet: [0, 0.4629, 0.9258, 2.3146]
# R's documented formula: [0, 0.3162, 0.6325, 1.5811]

The same discrepancy occurs with one-dimensional and sparse input. The expected numbers above were calculated from R’s documented scale formula; R itself was not run for this reproduction.

Against a reference implementation of that formula, upstream main (e9e33a7) matches in the default and scale=False matrix cases, but differs in the center=False and explicit-centre cases.

Impact

The in-repository run_actionet pipeline calls scale() with its defaults, so this issue does not claim a change to its layouts. It affects direct callers using scale=True with disabled or explicit centering, whose outputs can differ from R.

Expected behavior

With scale=True, compute the scaling factor from each column after centering, using sqrt(sum(x²) / (n - 1)) for the cases above, as specified by R’s base::scale documentation.