A reported distance of zero can conceal weighted tail mass
A tiny weight can retain a material contribution over a large distance. Check the meaning of zero using transport definitions and reflected controls.
Small weights can still contribute
Derive the contribution of tail mass and displacement independently, then check reflection invariance before interpreting zero as identical distributions. Small weights alone do not justify that decision: their contribution depends on the distance over which mass moves. Derive the reference distance supplied to Truthound separately from original masses and positions. [1, 2, 3]
The reported artificial boundary cases
SciPy #26254 reports zero energy or Wasserstein distances for highly skewed valid weights and a W1 result that changes under reflection. The author discloses AI-assisted exploration and personal mathematical and execution checks. Those reported extreme float64 cases remain distinct from the small illustrative transport problem here. [1]
Retain positions and masses
Retain positions, raw weights, and the normalization rule. Prepare a hand-checkable reference distance and reflected controls, excluding a mismatching zero from similarity decisions. Operators join actual and independently defined values by distribution-pair ID, keeping missing values and failed calculations separate. [1, 3]
Move half the mass by 4
In the illustrative source distribution, half the mass is at 0 and half at 4; the comparison distribution has all mass at 0. Moving mass 0.5 through distance 4 gives expected W1=2. An observed 2 passes; 0 fails. This does not reproduce the extreme-weight report, and the energy-distance formula is separate. [1, 3]
Responsibility for the reference distance
ColumnPairValidator in Truthound Python SDK 3.3.2 reconciles operator-computed observed_w1 and independently_derived_w1 values. Operators supply weight normalization and the independent transport reference. Core does not establish distribution equality or numerical precision automatically; passing this fixture is not execution evidence for the original report. [1, 2, 3]
What reflected controls add
Combine high-precision cumulative-distribution or tail calculations with reflected controls and measure the affected range before revising distance-based decisions. If the same input ranges are already covered by independent regressions, another table may add little. Reflection agreement alone cannot replace the reference calculation. [1, 3]
Application conditions
Retain positions, raw weights, normalization rules, and an independent transport reference.
Alternatives
Evaluate high-precision tail calculations together with reflected controls.
Limitations
Passing the conceptual W1 fixture does not establish the extreme-weight execution or distribution equality.
Evidence
- Event checked
- Subject checked
- Modified
