Bahadur Slope: Efficiency in Testing

• Bahadur Slope is a measure that quantifies the exponential decay rate of type II error probabilities in hypothesis testing and estimation.
• It unifies large deviation theory, asymptotic power analysis, and information-theoretic optimality to rank statistical tests and estimators.
• Benchmarks based on Bahadur efficiency guide the evaluation of procedures like maximum likelihood estimation and goodness-of-fit tests in various settings.

The Bahadur slope, also referred to as Bahadur efficiency, quantitatively characterizes the exponential decay rate of error probabilities—particularly the type II error—in statistical hypothesis testing or the tail probability for estimation, under fixed or local alternatives. It unifies large deviation theory, asymptotic power analysis, and information-theoretic optimality. The Bahadur slope has become the canonical measure for “exponential sensitivity” of statistical procedures, supporting principled ranking and comparison of tests and estimators in both parametric and non-parametric settings. It also serves as a benchmark for the optimality of maximum likelihood and related procedures across a wide array of inferential paradigms.

1. Formal Definitions and Large Deviations Framework

The classical Bahadur slope for a sequence of tests or estimators is based on large deviation asymptotics. For a sequence of estimators $T_n$ for a parameter $\theta$,

$$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$

is defined as the exact large deviation rate of the deviation probability from the true parameter. For hypothesis tests, if a sequence $T_n$ is used to test $H_0 : \theta = \theta_0$ versus local or fixed alternatives, the Bahadur slope is often defined by

$$S(\theta_0) = -2 \lim_{n \to \infty} \frac{1}{n}\log \beta_n(\theta_0)$$

where $\beta_n(\theta_0)$ denotes the type II error probability at the induced threshold. Bahadur efficiency of a procedure is assessed via its slope relative to the “separation rate” derived from the Kullback–Leibler divergence under the parametric model. For moderate deviations, as in (Ermakov, 27 Apr 2025), the slope generalizes to

$$S_{MD}(\theta_0) = -\liminf_{n \to \infty} \frac{1}{n u_n^2} \log P_{\theta_0}(|\hat\theta_n - \theta_0| > u_n)$$

for a sequence $u_n \to 0$.

For test statistics with known limiting behavior under $H_0$, the (approximate) Bahadur slope $\theta$0 can often be decomposed as

$\theta$1

where $\theta$2 and $\theta$3 are constants tied to the limiting distribution and large-deviation exponent, and $\theta$4 captures the Law of Large Numbers limit or the scaling of the test under the alternative (Milošević et al., 2021, Meintanis et al., 2022).

2. Canonical Applications and Optimality Statements

The Bahadur slope is central in determining the exponential discrimination power for both estimation and hypothesis testing:

• Estimation: For the maximum likelihood estimator (MLE), the Bahadur slope matches the rate derived from the information lower bound (Fisher information), ensuring that MLE is Bahadur-optimal in a large class of parametric settings. For example, in the Cauchy location–scale model, the MLE and certain one-step estimators both achieve the Bahadur separation rate given by the Kullback–Leibler divergence, with small-$\theta$5 slopes satisfying

$\theta$6

and matching the Cramér–Rao lower bound for the mean-square error (Akaoka et al., 2021).

• Goodness-of-fit Testing: The Bahadur slope directly governs the rate at which the $\theta$7-value of a test statistic approaches zero under close alternatives. Classical tests such as the Kolmogorov–Smirnov, Cramér–von Mises, and Anderson–Darling are rigorously compared on this basis, with the Anderson–Darling test typically achieving the largest approximate local Bahadur slope across a variety of alternatives (Milošević et al., 2021). Newer empirical characteristic function (ECF)-based tests, such as the Epps–Pulley and BHEP statistics, admit explicit Bahadur slope calculations and, when tuned appropriately, can exceed the Bahadur efficiencies of many classical procedures (Ebner et al., 2021, Meintanis et al., 2022).

3. Explicit Slopes and Efficiency Benchmarks

The Bahadur slope is available in closed or semi-closed form for many estimators and statistics:

• Power Divergence Statistics: For plug-in statistics in multinomial models, one has

$\theta$8

where $\theta$9 is the limiting value of the statistic under the alternative. Notably, the Kullback–Leibler (KL, $$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$0) divergence statistic is infinitely more Bahadur efficient than any $$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$1, regardless of alternative (Harremoës et al., 2010).

• Empirical Distribution Function (EDF) Tests: For goodness-of-fit tests, slopes can be computed using operator eigenvalues associated with Gaussian process covariance kernels under the null. For instance, the local approximate Bahadur slope for the Anderson–Darling test is

$$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$2

with $$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$3 being the leading eigenvalue, and $$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$4 reflecting the alternative’s effect (Milošević et al., 2021).

• Empirical Characteristic Function (ECF) Tests: For weighted $$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$5-type ECF statistics, the slope is

$$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$6

where $$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$7 is the largest eigenvalue of the test’s null covariance kernel operator and $$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$8 is the LLN limit under the alternative (Meintanis et al., 2022, Ebner et al., 2021). For the Epps–Pulley test, careful tuning of the weight parameter maximizes the Bahadur slope, making it highly competitive or even superior to many classical EDF tests.

Test Statistic/Class	Slope Formula	Key Reference
KL divergence	$$c_T(\theta, \epsilon) = -\limsup_{n \to \infty} \frac{1}{n}\log P_\theta(|T_n - \theta| > \epsilon)$$9	(Harremoës et al., 2010)
Power divergence ($T_n$0)	$T_n$1	(Harremoës et al., 2010)
EDF tests (CV-M/Supremum)	$T_n$2	(Milošević et al., 2021)
ECF-based tests	$T_n$3	(Meintanis et al., 2022)

4. Lower Bound Theory and Fisher Information Optimality

Foundational work (Bahadur, Hajek, Le Cam), and recent formalizations by Ermakov, establish that the Fisher information $T_n$4 dictates the maximal achievable Bahadur slope for regular estimators or test statistics. For moderate deviations,

$T_n$5

and for $T_n$6, the classical large deviation rate

$T_n$7

is recovered (Ermakov, 27 Apr 2025). These bounds are tight; estimators such as the MLE and certain quasi-arithmetic estimators in the Cauchy model provably achieve equality (Akaoka et al., 2021).

5. Local and Global Comparisons of Procedures

The Bahadur slope enables rigorous local (small-alternative) and global comparison of tests:

• Analytical and numerical studies strongly indicate that "integral" (quadratic-form) tests outperform "supremum" (maximum deviation)-type tests in local Bahadur efficiency.
• The information divergence LRT is maximally Bahadur efficient in all regular parametric settings and outclasses power divergence statistics with $T_n$8 by an infinite margin (Harremoës et al., 2010).
• For normality testing, BHEP and ECF-based tests, when tuned, can attain local efficiencies exceeding 90\% of the LRT; the Anderson–Darling test remains the omnibus EDF-based benchmark (Milošević et al., 2021, Meintanis et al., 2022, Ebner et al., 2021).

6. Bahadur Efficiency in Specific Statistical Contexts

• Cauchy Location–Scale Estimation: Bahadur efficiency of both the MLE and quasi-arithmetic one-step estimators is established, matching the KL separation rate and Cramér–Rao bound for the mean-square error. The methodology extends to circular (wrapped-Cauchy) models via Möbius transformation (Akaoka et al., 2021).
• Weighted L²–Type ECF Tests: Bahadur slope calculations for BHEP, energy, and logistic tests offer guidance for optimal tuning in normality and exponentiality testing—higher slope values associating with parameter regions that maximize discrimination under specific alternatives (Meintanis et al., 2022).
• EDF-Based Normality Tests with Parameter Estimation: Explicit local slope and efficiency calculations provide benchmarks for new proposal development and inform practitioners about statistical power under local alternatives (Milošević et al., 2021).

7. Practical Implications and Contemporary Practice

Computation and interpretation of the Bahadur slope are pivotal for the asymptotic evaluation and selection of test statistics and estimators. Tests achieving high (or maximal) Bahadur slopes require fewer samples to attain a desired power or significance under close alternatives, and serve as sharp tools for both theoretical analysis and applied model assessment. Slopes and relative efficiencies yield principled selection criteria:

• For normality, the BHEP test with $T_n$9 and the GE test with $H_0 : \theta = \theta_0$0 (tail contamination) are recommended, while the Anderson–Darling and Epps–Pulley tests, appropriately tuned, rival or outperform classical EDF tests (Ebner et al., 2021, Meintanis et al., 2022).
• For multinomial and divergence-type testing, the KL ($H_0 : \theta = \theta_0$1) statistic is unconditionally preferred due to its infinite Bahadur superiority over other divergences with $H_0 : \theta = \theta_0$2 (Harremoës et al., 2010).
• Estimation procedures that meet or achieve the Fisher information lower bound in slope are optimal in both central-limit and large-deviation regimes (Akaoka et al., 2021, Ermakov, 27 Apr 2025).

The Bahadur slope thus remains integral to theoretical statistics, large deviations, and asymptotic efficiency evaluation, with contemporary research extending its reach to moderate deviation probabilities and nonparametric or semiparametric regimes.