Do not dismiss an integer matrix’s 0.5 power as a dtype preference
Use a squared residual to check preservation of the fractional exponent. A plausible dtype can conceal a different operation.
The exponent belongs in the contract
Check preservation of a fractional exponent and the squared residual before dismissing a changed result as output precision preference. An integer input representation cannot justify silently reducing a noninteger exponent to 0. The mathematical operation precedes the representation policy. Truthound reconciles externally squared results with matching diagonal positions in the original matrix. [1, 2, 3]
Read the specific numerical audit case
The explicit 2026-10-13 audit in SciPy #25964 describes integer fractional_matrix_power losing a noninteger exponent. The broad linalg list is not one failure rate or evidence that every function is defective. This article addresses the specific numerical case and keeps other dtype policies separate. [1]
Retain the matrix and exponent together
Retain the original matrix, exponent, dtype, and returned matrix. Reconcile the square of a 0.5-power candidate with the original input before handoff, testing explicit promotion on failure. Preserve the definitions losslessly and confirm that the chosen root and product criterion apply to the matrix. [1, 3]
A square-root reference from 4
For illustrative integer diagonal entry 4 and exponent 0.5, the principal root candidate 2 squares to required input 4 and passes. If the exponent is reduced to 0, candidate 1 squares to 1 and fails. A general matrix with multiple branches needs more than this diagonal comparison. [1, 3]
The bounded diagonal comparison
ColumnPairValidator in Truthound Python SDK 3.3.2 compares squared_output_diagonal calculated externally with original_input_diagonal. Operators define exponent preservation, products, and branch criteria. Core does not run fractional_matrix_power; one diagonal cannot establish the full matrix, complex branch, or every linalg dtype policy. [1, 2, 3]
Retest an explicit computation type
Test explicit promotion to an approved floating computation type and verify actual squared residuals, ranges, and resource costs. Existing independent checks under the same input conditions reduce another audit’s benefit. Promotion is a tested alternative, not an automatic proof that the original operation survived. [1, 3]
Application conditions
Preserve the matrix and exponent losslessly and verify the applicable principal-root and product criteria.
Alternatives
Test actual squared residuals, range, and cost after explicit promotion to an approved computation type.
Limitations
One squared diagonal does not establish the full matrix, complex branch, or every linalg dtype policy.
Evidence
- Event checked
- Subject checked
- Modified
