Neyman-Pearson Optimal Decision Rules

• Neyman-Pearson optimal decision rules are a framework for constructing hypothesis tests that maximize statistical power at a fixed type-I error using likelihood-ratio thresholds.
• Recent generalizations extend these rules to convex risk measures and composite hypothesis settings, ensuring robust performance in complex statistical models.
• Applications span finance, classification, and robust decision-making, with randomization offering optimal solutions when deterministic thresholds are insufficient.

Neyman-Pearson optimal decision rules formalize the construction of hypothesis tests that maximize statistical power at a fixed level of type I error. The classical Neyman-Pearson lemma establishes that, for testing between two simple hypotheses, the most powerful test is a threshold rule on the likelihood ratio. Extensive recent research has generalized this principle to convex or sublinear expectation frameworks, composite hypothesis settings, and nonlinear testing objectives. Below, the principal mathematical structure, existence results, generalizations, and functional, statistical, and practical consequences are presented.

1. Mathematical Foundations and Problem Statement

Let $(\Omega, \mathcal{F}, \mu)$ be a probability space. The classical Neyman-Pearson problem tests $H_0: X \sim P_0$ against $H_1: X \sim P_1$ for $X \in L^\infty(\mu)$ via randomized tests $\phi: \Omega \to [0,1]$. The aim is to maximize $P_1\{\phi=1\}$ (power) subject to $P_0\{\phi=1\} \leq \alpha$ for significance level $\alpha \in (0,1)$. The optimal rule is a likelihood-ratio threshold: $$\phi^*(\omega) = \begin{cases} 1, & \Lambda(\omega) > \eta \ \gamma, & \Lambda(\omega) = \eta \ 0, & \Lambda(\omega) < \eta \end{cases}$$ where $\Lambda = dP_1/dP_0$, $\eta$ is chosen such that $$P_0\{\Lambda > \eta\} + \gamma P_0\{\Lambda = \eta\} = \alpha$$, and $\gamma \in [0,1]$ when $P_0\{\Lambda = \eta\} > 0$ (Dulek et al., 2018, Pena, 2019).

In modern settings, this formulation is extended to convex (or sublinear) risk functionals $p$ on $L^\infty(\mu)$, with

$p(X) = \sup_{P \ll \mu}\{E_P[X] - p^*(P)\},$

where $p^*$ is a convex, penalty-type functional that yields a robust expectation (Sun et al., 2019, Chuanfeng et al., 2019). The generalized NP problem is: $$\min_X\; p_2(K_2 - X)\quad \text{subject to}\quad p_1(X) \leq \alpha,\;\; K_1 \leq X \leq K_2,$$ with $K_1,K_2 \in L^\infty(\mu)$ supplying general lower and upper bounds.

2. Existence and Structure of Optimal Tests

For convex expectations on $L^\infty(\mu)$ (satisfying monotonicity, translation invariance, convexity, and continuity from below), existence of an optimal solution $X^*$ is guaranteed. The proof leverages the weak$^*$-sequential compactness of the feasible set and Komlós' subsequence theorem: any minimizing sequence admits a subsequence whose Cesàro averages converge almost surely to a limit $X^*$ that remains within the original constraints and is feasible for the primal optimization (Sun et al., 2019, Chuanfeng et al., 2019).

Moreover, the optimal value can be characterized via a minimax (saddle-point) theorem: $$\inf_{X \in \mathcal{X}_\alpha}\sup_{Q \in \mathcal{Q}} \{E_Q[K_2 - X] - p_2^*(Q)\} = \sup_{Q \in \mathcal{Q}}\inf_{X \in \mathcal{X}_\alpha} \{E_Q[K_2 - X] - p_2^*(Q)\}.$$ The supremum over $Q$ is attained (by a "least-favorable" $Q^*$), and likewise, there exists $P^* \in \mathcal{P}$ achieving the $\alpha$-constraint (Chuanfeng et al., 2019).

3. Generalized Neyman-Pearson Form and Randomization

The optimal test under convex expectations retains the familiar likelihood-ratio thresholding structure, but between representative "worst-case" priors $P^*$ and $Q^*$ chosen via saddle-point arguments. For $K_1=0, K_2=1$, there exist densities $g = dP^*/d\mu$ and $\ell = dQ^*/d\mu$ so that the optimal test is

$$\phi^*(\omega) = \begin{cases} 1, & \ell(\omega)/g(\omega) > z \ B(\omega), & \ell(\omega)/g(\omega) = z \ 0, & \ell(\omega)/g(\omega) < z \end{cases}$$

with randomization $B(\omega) \in [0,1]$ on the boundary and threshold $z$ chosen to satisfy the type-I constraint, paralleling the classical NP rule (Sun et al., 2019, Chuanfeng et al., 2019, Sun et al., 2016).

Randomization is only necessary when the level-set $\{\ell/g = z\}$ has nonzero probability under the null; otherwise, the optimal test is deterministic. This structure persists for generalized sublinear expectations and in the presence of finitely additive (non-countably additive) measures, via decompositions in the Yosida-Hewitt and Mazur-Orlicz theorems.

4. Beyond Simple Hypotheses: Composite and Nonlinear Cases

For composite hypothesis testing, where $H_0$ and $H_1$ correspond to sets of distributions, the optimal test can still be realized as a single-threshold rule on an appropriately weighted likelihood ratio: $$\Lambda_w(y) = \frac{\int_{\Theta_1} w(\theta) f_\theta(y) d\theta}{f_{\theta_0}(y)},$$ with data-driven weights $w(\theta)$ derived from the specific performance criterion and priors (Song et al., 23 May 2025). The null can be simple, average-composite, or worst-case composite; optimality is achieved by using the relevant mixture density in the denominator and calibrating the threshold to the false-alarm constraint.

If the objective is a nonlinear function $g(p(\theta; \delta))$ of the detection probability, subject to a type-I constraint, Lagrange/KKT arguments yield that the optimal rule is a threshold on a weighted likelihood ratio constructed from derivatives $g'(p^*(\theta))$ (Song et al., 23 May 2025). In practical settings (e.g., exponential families), the sufficient statistics allow explicit description of the rejection regions.

5. Extensions to Novel Decision Criteria and Surrogate Losses

The NP lemma extends to settings with tunable convex surrogates for the $0$-$1$ loss. For instance, tests minimizing type-II error subject to a type-I error defined by a smooth $\nu$-loss yield randomized "soft-threshold" rules on the likelihood ratio, recovering the hard-threshold classical NP rule as $\nu \to \infty$ (Kamatsuka, 2022). Importantly, the type-II exponent (error rate) in these regimes coincides with the classical KL divergence, regardless of the surrogate, confirming the optimality of (possibly randomized) likelihood-ratio-type tests in the large-sample regime.

Convex analytic approaches further ensure that any criterion depending on error probabilities or type-I/type-II trade-offs admits an optimal rule as a mixture of at most two deterministic (likelihood-ratio) rules (Dulek et al., 2018). Even in complex performance functions (including prospect theory or nonconvex trade-offs), the sufficiency of the likelihood ratio is preserved.

6. Applications and Impact

Neyman-Pearson optimal rules appear ubiquitously: in robust hypothesis testing with convex risk measures (Sun et al., 2019, Chuanfeng et al., 2019), outlier detection under exponential constraints (Zhou et al., 2020), universal or empirical-likelihood classification with unknown alternatives (Boroumand et al., 2022), or power-maximizing conformal selection for FDR calibration (Qin et al., 23 Feb 2025). In all cases, the core insight is reduction to a threshold rule—often on an appropriately "weighted" likelihood ratio reflecting the objective, uncertainty set, or empirical construction.

In financial mathematics, the same structure yields optimal shortfall-risk portfolios and quantile hedging strategies: the optimal replicating payoff under convex shortfall risk and budget constraint is the likelihood-ratio threshold between "least-favorable" martingale measures (Chuanfeng et al., 2019).

7. Theoretical and Methodological Consequences

The Neyman-Pearson framework, when generalized to convex or sublinear risk functionals, preserves two fundamental features: (i) existence of optimal tests in the space of randomized rules (using compactness/closure arguments in weak$^*$ topologies), and (ii) necessity and sufficiency of (randomized) likelihood-ratio thresholding on representative priors/distinct measures. All nonlinearity—in the form of convexity, risk-averse preferences, or composite uncertainty—is absorbed into the representative pairs $(P^*,Q^*)$ via saddle-point (minimax) or duality constructions (Sun et al., 2019, Chuanfeng et al., 2019, Sun et al., 2016).

A succinct summary is provided in the table below:

Setting	Optimal Test Structure	Representative Measures
Simple hypotheses, linear exp.	Hard/randomized LRT on $\Lambda$	$(P_0, P_1)$
Convex/sublinear expectations	LRT on $\ell/g$ for $(Q^*, P^*)$	Saddle-point representatives
Composite/average/worst-case	Weighted/mixed LRT, least-favorable priors	Mixture densities or mixtures
Surrogate/convex loss	Soft-threshold on $\Lambda^\nu$	As above
Classification/conformal selection	LRT or similar statistics w/ empirical mixture	Empirical or plug-in substitutions

This theoretical generality unifies a substantial range of statistical inference and robust optimization problems under the logic of Neyman-Pearson thresholding, conditioned on analytically identifying or estimating the representative priors to reflect the required error-rate control or risk sensitivity. Such generalizations have been shown to be both practically realizable and statistically minimax optimal in a wide range of contemporary settings (Sun et al., 2019, Chuanfeng et al., 2019, Sun et al., 2016, Dulek et al., 2018, Song et al., 23 May 2025, Kamatsuka, 2022, Zhou et al., 2020, Qin et al., 23 Feb 2025).