NumPy corrcoef on constant channels: do not rank a finite 1 as defined correlation
Pearson correlation is undefined for a constant channel.
The domain to inspect before finiteness
Pearson correlation is undefined for a constant channel. Check constancy directly and do not approve a finite coefficient as evidence of association. Supply Truthound with constancy and handoff decisions as well as the library output; a finite value within range does not establish a defined denominator. [1, 2, 3]
A discussion that did not establish a correction
NumPy #32446 reports finite ±1 from corrcoef on exactly constant inputs. Follow-up discussion distinguishes floating-point effects from API policy, and the maintainer closed it with an invitation to reopen for actionable work. Closure is not a shipped correction or a final classification of a library defect. [1]
Check constancy directly
Record raw unique values, constancy and valid observation counts for each channel. Exclude constant pairs from correlation ranking and report the undefined reason. Verify exact constancy directly, distinguishing near-constant tolerance and aligning missing-value handling and sample scope. [1, 3]
The handoff condition for denominator 0
Illustrative X=[7,7,7] and Y=[2,2,2] both have variance 0. The Pearson denominator is 0, so expected ranking admission is false. false respects that domain; a finite correlation 1 leading to true fails. Replacing it with 0 would not make it a defined Pearson relationship. [1, 3]
Compare analytical admission
Truthound Core Python SDK 3.3.2 can compare prepared approved_admission and admitted_to_correlation_ranking with ColumnPairValidator using relationship="==". The operator supplies constancy policy and non-missing flags. The SDK does not compute Pearson correlation or decide whether upstream behavior is a bug, and does not turn unresolved discussion into a repair guarantee. [1, 2, 3]
Explain the reason for exclusion
Report constant channels with separate descriptive statistics or obtain observations supporting a meaningful alternative. Explain computational success separately from analytical admissibility so users understand exclusion from ranking. Reuse existing verified domain checks on the same consumer path. [1, 3]
Application conditions
Check exact constancy directly, separating missingness and near-constant tolerance policies.
Alternatives
Report separate descriptive statistics or obtain observations supporting a scientifically meaningful alternative.
Limitations
Core compares prepared constancy and admission flags, does not calculate Pearson correlation or classify an upstream defect, and does not claim a repair.
Evidence
- Event checked
- Subject checked
- Modified
