The Behrens-Fisher problem revisited
Nagananda K G, Jong Sung Kim
TL;DR
The paper tackles the Behrens--Fisher problem of testing equal means for two normal samples with unknown, unequal variances by deriving an explicit, computable null distribution for the Behrens--Fisher statistic $T$. It introduces a Mellin--Barnes contour decomposition to decouple the square-root denominator, yielding a tractable one-dimensional integral and a residue series that terminates when either degrees of freedom are odd; an Euler--Beta reduction then expresses the density in terms of Gauss' hypergeometric function ${}_2F_{1}$, with the equal-variance case recovering Student's $t$ and Ramanujan's master theorem providing exact tail coefficients. This structure enables precise tail bounds, Lugannani--Rice corrections, and exact cdf tables, while clarifying the relationship to and improvements over Welch's approximation and Nel et al. (1990). Collectively, the results deliver an analytically transparent, numerically robust framework for exact Behrens--Fisher inference, including practical critical-value tables and stability-aware implementations. The approach unifies hypergeometric, Beta, and MB techniques to produce exact density, cdf, and tail expansions across all $(n_1,n_2,r)$ and supports reliable small-sample inference where traditional approximations falter.
Abstract
We revisit the two-sample Behrens-Fisher problem -- testing equality of means when two normal populations have unequal, unknown variances -- and derive a compact expression for the null distribution of the classical test statistic. The key step is a Mellin--Barnes factorization that decouples the square root of a weighted sum of independent chi-square variates, thereby collapsing a challenging two-dimensional integral to a tractable single-contour integral. Closing the contour yields a residue series that terminates whenever either sample's degrees of freedom is odd. A complementary Euler-Beta reduction identifies the density as a Gauss hypergeometric function with explicit parameters, yielding a numerically stable form that recovers Student's $t$ under equal variances. Ramanujan's master theorem supplies exact inverse-power tail coefficients, which bound Lugannani-Rice saddle-point approximation errors and support reliable tail analyses. Our result subsumes the hypergeometric density derived by Nel et al.}, and extends it with a concise cdf and analytic tail expansions; their algebraic special cases coincide with our truncated residue series. Using our derived expressions, we tabulate exact two-sided critical values over a broad grid of sample sizes and variance ratios that reveal the parameter surface on which the well-known Welch's approximation switches from conservative to liberal, quantifying its maximum size distortion.