Validity of the Behrens--Fisher solution
Abstract
The Behrens--Fisher fiducial interval for the difference of two normal means is known from the numerical work of Robinson (1976) to be conservative, but a proof has been lacking. The results here prove that the fiducial interval for any linear combination of k normal means has coverage probability at least its nominal level for all values of the unknown variances. Taraldsen (2026) explains this in more detail using the original data of Behrens (1926). The proof presented in the following rests on a novel plug-in inequality: if a threshold function controls a scaled sum of pivots pointwise in the scale vector, then it also controls the sum with the observed scales plugged in, provided the location errors satisfy a concentration condition formulated through a weighted geometric mean. Independent normal errors satisfy the condition.
Related
- Fisher's Ideas and the Design of Field Experiments in Agronomy and Plant Breeding
- A Kernel Fisher Discriminant Analysis-Based Tree Ensemble Classifier: KFDA Forest
- On the discrete Fisher information
- Convexity of a ratio function involving polygamma functions and monotonicity of the sectional curvature of the Fisher--Rao manifold
- Uniform efficiency of the Ulrich-Wood sampler for the von Mises-Fisher distribution
BibTeX
@misc{taraldsen2026validity,
title = {{Validity of the Behrens--Fisher solution}},
author = {Gunnar Taraldsen},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.06987},
eprint = {2611.06987},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.06987},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Validity of the Behrens--Fisher solution
AU - Gunnar Taraldsen
PY - 2026
DA - 2026/11/04
PB - arXiv
DO - 10.48550/arxiv.2611.06987
UR - https://arxiv.org/abs/2611.06987
AB - The Behrens--Fisher fiducial interval for the difference of two normal means is known from the numerical work of Robinson (1976) to be conservative, but a proof has been lacking. The results here prove that the fiducial interval for any linear combination of k normal means has coverage probability at least its nominal level for all values of the unknown variances. Taraldsen (2026) explains this in more detail using the original data of Behrens (1926). The proof presented in the following rests on a novel plug-in inequality: if a threshold function controls a scaled sum of pivots pointwise in the scale vector, then it also controls the sum with the observed scales plugged in, provided the location errors satisfy a concentration condition formulated through a weighted geometric mean. Independent normal errors satisfy the condition.
ER -