Prior Hypothesis Bias (PHB)

• Prior Hypothesis Bias (PHB) is a systematic distortion in decision-making due to the quantization of true prior probabilities under cognitive and resource constraints.
• The analysis reveals that excess Bayes risk decays inversely with the square of the number of quantization levels, highlighting key metrics such as MBRE and optimal point density.
• Differential allocation of quantization cells results in group disparities, with majority groups benefiting from finer resolution while minority groups face increased error rates.

Prior Hypothesis Bias (PHB) refers to systematic distortions in hypothesis testing or decision-making that arise specifically from the representation, implementation, or misallocation of prior probabilities—particularly when such priors are quantized, under- or over-specified, or reflect cognitive constraints rather than true underlying distributions. This bias manifests either as excess Bayes risk in statistical inference due to misaligned or coarsely categorized priors, or as systematic disparities in decision accuracy across different populations when the cognitive or algorithmic mechanism for handling priors is capacity-limited or asymmetric.

1. Quantization of Priors and the Emergence of PHB

In classical Bayesian hypothesis testing, it is assumed that the decision-maker applies the true prior probability $p_0$ (or prior vector in the multi-hypothesis case) directly in the likelihood-ratio test, leading to the minimum Bayes risk $J(p_0)$. However, practical decision-makers—especially humans—suffer cognitive or resource constraints and therefore cannot continuously represent all possible priors. Instead, the true prior is quantized to one of $K$ discrete levels $\{a_1, ..., a_K\}$ and applied in the test, resulting in a mismatched Bayes risk $\widetilde{J}(p_0, a_k)$. The excess risk due to quantization is captured by the mean Bayes risk error (MBRE):

$d(p_0, a) = \widetilde{J}(p_0, a) - J(p_0)$

This distortion measure is nonnegative and strictly convex as a function of $p_0$ for fixed $a$. The design of the prior quantizer involves two key criteria:

• Nearest Neighbor Condition (Eq. 2):

$$b_k = \frac{c_{01} (p_{E}^{(II)}(a_{k+1}) - p_{E}^{(II)}(a_k))}{c_{01} (p_{E}^{(II)}(a_{k+1}) - p_{E}^{(II)}(a_k)) - c_{10} (p_{E}^{(I)}(a_{k+1}) - p_{E}^{(I)}(a_k))}$$

which locates cell boundaries by intersecting tangents to the mismatched risk functions;

• Centroid Condition (Eq. 3):

$$c_{10} I_k^{(I)} \left.\frac{dp_E^{(I)}(a)}{da}\right|_{a=a_k} + c_{01} I_k^{(II)} \left.\frac{dp_E^{(II)}(a)}{da}\right|_{a=a_k} = 0$$

which determines the optimal quantizer levels $a_k$ for each cell through minimization of aggregate excess risk.

The high-rate analysis (large $K$) results in a quadratic distortion in each cell:

$d(p_0, a_k) \approx B(a_k) (p_0 - a_k)^2$

where $B(p_0)$ (Eq. 5) involves derivatives of the error probabilities with respect to $p_0$. The overall distortion decays as $1/K^2$, with the optimal point density:

$$\lambda(p_0) = \frac{[B(p_0) f_{p_0}(p_0)]^{1/3}}{\int_0^1 [B(p) f_{p_0}(p)]^{1/3} dp}$$

2. Cognitive Constraints, Social Disparities, and PHB

The phenomenon of PHB is accentuated in human contexts where individuals make decisions (e.g., referees adjudicating fouls, police evaluating suspects) relying on categorical representations of underlying continuous priors. Human memory and processing constraints (famously “7 ± 2” categories) result in coarse quantization of prior probabilities.

When decision-makers interact more frequently with one population than another (e.g., the majority vs. minority groups), they are able to allocate more quantization levels—and thus achieve lower excess Bayes risk—for the familiar (majority) group, while the minority group is assigned fewer, broader categories. The two-population MBRE (Eq. 9):

$$D^{(2)} = \frac{w}{w+b} \mathbb{E}[\widetilde{J}(P_0, v_{K_w}(P_0))] + \frac{b}{w+b} \mathbb{E}[\widetilde{J}(P_0, v_{K_b}(P_0))] - \mathbb{E}[J(P_0)]$$

where $K_w$ and $K_b$ are the quantization levels for majority and minority groups, quantifies this asymmetry. Since MBRE decays monotonically as $K$ increases, the majority group sees systematically lower error, and the minority group greater error due solely to quantization resource allocation. This forms a mechanism for PHB: systematic bias in the error rates of out-group vs. in-group individuals, even when the true prior distributions are identical.

3. Technical Characterization and High-Resolution Analysis

The decision-maker’s quantized prior introduces additional risk strictly bounded above by the curvature of the Bayes risk function with respect to the prior and the density of true prior probabilities (through $B(p_0)$ and $f_{p_0}(p_0)$). The quantizer’s performance is upper bounded by:

$$D_L = \frac{1}{12K^2} \int_0^1 B(p_0) f_{p_0}(p_0) \lambda(p_0)^{-2} dp_0$$

This analysis leads to two core insights for PHB:

• The allocation of quantization points should be denser in regions where both the prior is more likely (high $f_{p_0}(p_0)$) and the Bayes risk is more sensitive to prior errors (high $|B(p_0)|$).
• The magnitude of excess error is a universal function of both the number of categories and the local properties of the classification task (error type curvature and prior density).

4. Implications for Experimental Phenomena and Error Disparities

Quantized cognition-based PHB helps explain observed “own-group” effects such as own-race bias in face recognition, increased foul calls against out-group players, or increased arrest rates among minority suspects when evaluated by majority officers. Even with ideal, unbiased error costs and likelihoods, the coarse quantization imposed by cognitive limits leads to increased false positives or negatives for the out-group solely due to the allocation of finite decision “cells.” This mechanism persists even when priors for both groups are drawn from the same distribution.

5. Mitigating PHB and Theoretical Limits

Mitigation of PHB is possible by increasing quantization resolution—allocating more categories, thereby reducing MBRE for all groups. However, practical constraints (human working memory, computational resources) cap the achievable $K$. The analysis also clarifies that fine quantization, optimally allocated according to $B(p_0)f_{p_0}(p_0)$, minimizes overall excess risk but cannot fully eliminate PHB unless $K \to \infty$ for all populations. In most practical (and especially human) systems, this is unattainable, marking PHB as an inherent and, to some degree, unavoidable phenomenon wherever decisions are based on limited prior representations.

Summary Table: Consequences of Prior Quantization

Factor	Effect on PHB	Mechanistic Origin
Number of quantization cells	Inverse quadratic decay of MBRE with $K$	High-rate quadratic distortion
Allocation of quantizer cells	Asymmetric error rates across populations	Differential cognitive exposure
Sensitivity of Bayes risk	Finer allocation needed in high curvature regions	Local sensitivity $B(p_0)$

6. Broader Impact and Connection to Disparities

PHB, as described through quantization analysis, establishes a rigorous, quantifiable link between cognitive limitations, resource-constrained computation, and systematic group disparities in decision-making. The mathematical framework demonstrates that such disparities can arise independently of overt prejudice or explicit bias in cost or likelihood specification. This theoretical result is significant for understanding and addressing fairness in human and algorithmic decision systems, providing a pathway for evaluating and potentially remediating error disparities rooted in prior representation constraints.