Before clipping a tiny negative interpolated variance at an input knot
Check knot reproduction separately from variance nonnegativity. Even tiny values need an error policy appropriate to their use.
A representable number versus an admissible variance
Define knot reproduction and nonnegative-variance requirements separately before clipping small negatives. Otherwise a correction can hide which path failed to retain the original position. Record the task’s numerical tolerance separately from the physical lower bound. A table joining original knots and actual interpolated values at matching positions can start the Truthound reconciliation. [1, 2, 3]
The reported multidimensional legacy path
SciPy #24281 reports a small negative at an original knot in the multidimensional linear interp1d path, with alternative linear-spline controls. Those reported controls do not establish a universally exact replacement at every interior point and boundary. [1]
Retain the original knot reference
Preserve original knots, y shape, interpolation path, and evaluation positions. Reconcile actual knot values with the originals, holding square-root or probability calculations that consume inadmissible negative variances. Operators prepare actual and independent references under the same curve and knot IDs. [1, 3]
Compare original 0 with a negative counterexample
The illustrative input defines variance exactly 0 at knot x=2, so the required value is the original y=0. An observed 0 passes; -0.0000000000000000000542101086 differs and is negative, failing this conceptual contract. Real tasks still define their own knot tolerances; this is not exact equality for every curve. [1, 3]
The interpolation-value boundary
ColumnPairValidator in Truthound Python SDK 3.3.2 reconciles actual_knot_value extracted by the operator with original_knot_value. Operators run interpolation and supply error and sign policies. Core does not establish stability of the entire curve or justify automatic clipping. [1, 2, 3]
Test boundaries in alternative linear paths
Compare a verified make_interp_spline(k=1) or other path, checking knots, intermediate points, boundaries, and tolerances together. Existing regressions for the same interpolation domain reduce the need for another audit. A passing control at the knot does not validate extrapolation. [1, 3]
Application conditions
Retain original knots, y shape, evaluation positions, and prior error and nonnegativity policies.
Alternatives
Test knots, intermediate points, and boundaries of a verified linear-spline alternative under the same criteria.
Limitations
Exact comparison at the conceptual knot does not establish whole-curve stability or justify automatic clipping.
Evidence
- Event checked
- Subject checked
- Modified
