Price of Non-Discrimination

Updated 30 January 2026

Price of non-discrimination is a metric that quantifies the trade-off between unconstrained optimization and fairness constraints in pricing and resource allocation.
Empirical analyses reveal that enforcing fairness can incur moderate losses, such as revenue drops of 3–8% or poly-logarithmic efficiency declines.
The framework spans mechanism design, predictive modeling, auctions, and regulation, offering actionable insights into balancing economic efficiency with societal fairness.

The price of non-discrimination—frequently abbreviated as PoND (Editor's term)—is a quantitative metric and conceptual framework for assessing the utility loss, performance degradation, or economic cost incurred when enforcing non-discrimination constraints in pricing, resource allocation, or decision-making processes. It is defined as the trade-off between an unconstrained optimum (using all potentially discriminatory information for maximum profit, accuracy, or efficiency) and the constrained outcome that meets specified fairness or parity requirements. This concept arises across mechanism design, actuarial science, dynamic pricing, regulation, and algorithmic fairness, reflecting a central tension between economic or predictive efficiency and societal requirements for group-based fairness.

1. Formal Definitions and Measurement

The price of non-discrimination is most often operationalized as a ratio or difference of objectives under constrained (fair) and unconstrained scenarios:

For maximization objectives (e.g., profit, welfare):

$\mathrm{PoND} = \frac{\mathrm{Optimum\ Unconstrained\ Value}}{\mathrm{Optimum\ Fair/Constrained\ Value}} \geq 1.$

For loss/minimization settings (e.g., MSE in insurance or classification error):

$\Delta = \text{Constrained Loss} - \text{Unconstrained Loss}.$

In dynamic environments, expected regret bounds quantify the exploration cost and strategic manipulation induced by fairness constraints.

The choice of fairness constraint (demographic parity, equalized odds, statistical parity of prices, etc.) defines the feasible set for the constrained optimum and thus the realized PoND.

2. Mechanism Design and Economic Analysis

Multi-Agent Contracts

In principal-agent models with $n$ agents and submodular project success functions, non-discriminatory contracts require all agents to receive the same payment. The tight bound is PoND $=\Theta(\log n)$ : in the worst case, enforcing equal pay can cause the principal's utility to drop to a $1/\log n$ -fraction of the unconstrained optimum (Ding et al., 23 Jan 2026). If one allows bounded wage ratios (relaxed $\beta$ -ND: max/min contract ratio $\leq \beta$ ), the PoND drops to $O(1/\delta)$ with $\beta = n^\delta$ , smoothly interpolating from logarithmic to constant loss.

Posted-Price Mechanisms

In markets with convex production costs and combinatorial buyer preferences, non-discriminatory posted-price mechanisms (buyer-independent prices) achieve at least a $1/4$ approximation to optimal welfare with XoS buyers; for general subadditive valuations, the guarantee is $O(\log m)$ , where $m$ is the number of goods (Sekar, 2016). In prior-free environments, the approximation is $O(\log N \log m)$ . Thus, removing discriminatory flexibility induces at most a poly-logarithmic loss in rich settings.

Personalized and Uniform Pricing

In models with consumer heterogeneity by protected characteristic, the revenue-optimal, non-discriminatory personalized pricing rules are characterized via optimal transport between value distributions. The constrained optimum leaves "information rents" (consumer surplus) and enforces statistical parity of price distributions across groups (Strack et al., 26 Jun 2025). Under model conditions, the non-discrimination constraint never reduces revenue below $50\%$ of the unconstrained maximum, and in calibrated examples achieves $95$– $99\%$ of full surplus. Stricter constraints on outcomes (rather than just prices) can further reduce profit and systematically redistribute surplus between groups.

3. Algorithmic and Statistical Trade-Offs

Predictive Modeling and Insurance Pricing

Group fairness in predictive analytics can be enforced by adversarial learning, multi-task modeling, or data preprocessing. The empirical and theoretical PoND is evident in performance metrics (Gini, MSE, KL divergence) as one interpolates between unconstrained and fairness-constrained solutions:

In adversarial autoencoder insurance pricing, tightening demographic parity or equalized odds constraints via increasing a Lagrange multiplier $\lambda$ manifests as a strictly increasing trade-off curve between predictive accuracy (e.g., AUC) and fairness metrics (p-rule, FairQuant). For instance, as $\lambda$ increases from $0$ to $2.0$, accuracy drops from $0.78$ to $0.63$ while the p-rule improves from $65\%$ to $98\%$ —empirically quantifying each step in the price-fairness plane (Grari et al., 2022).
Multi-task neural architectures for discrimination-free pricing achieve discrimination-free premiums (marginalizing over protected attributes), with the additional KL divergence ("loss") typically $25\%$ higher than unawareness-only schemes in synthetic health insurance data (Lindholm et al., 2022).
For privatized sensitive attributes (local differential privacy), statistical generalization error bounds tie directly to the privacy parameter and group counts: the price of non-discrimination diverges as privacy increases, reflected in increased MSE or cross-entropy—for example, moving from full information ( $\pi=1$ ) to strong privacy ( $\pi=0.6$ ) raises MSE by $50$– $80\%$ (Zhang et al., 16 Apr 2025).

Dynamic Pricing and Strategic Manipulation

In contextual dynamic pricing with unobserved group status and strategic buyers, fairness constraints on price gaps induce regret of $O(\sqrt{T}+H(T))$ over horizon $T$ , where $H(T)$ reflects buyers' price-fairness learning error. The lower bound is $\Omega(\sqrt{T})$ , and empirical reductions in regret (35\%) are possible when actively managing strategic responses versus naïve methods (Liu et al., 25 Jan 2025).

4. Audit, Regulation, and Empirical Quantification

Black-box algorithmic audits in regulated markets, such as auto insurance, provide concrete estimates of monetary PoND. An audit of Italian car insurance pricing found that non-Italian drivers pay surcharges of €100–400 and younger drivers €64–285 purely because of protected attributes, with quote availability varying sharply by profile. Removing birthplace surcharges alone would reduce premiums by €200–400/year, illustrating the direct financial impact of fairness constraints (Rondina et al., 10 Feb 2025).

5. Data Collection, Statistical Limits, and Learning Theoretic Effects

When pricing policies or discrimination are necessarily learned from finite samples, the "statistical" price of non-discrimination emerges. Data-driven (third-degree) discriminatory pricing has $O(n^{-1/2})$ rate of convergence in revenue deficiency, while uniform pricing attains $O(n^{-2/3})$ . For practical sample sizes, non-discriminatory uniform pricing can outperform discrimination due to the curse of dimensionality, with the classical advantage of discrimination manifesting only at sufficiently large $n$ (Xie et al., 2022).

6. Bidding, Auctions, and Utility Costs of Group Parity

In online ad auctions, enforcing gender parity (via strict or ratio constraints on ad impressions) yields a quantifiable PoND: for tight parity (e.g., difference of 1 impression), utility drops by $10$– $20\%$ in highly skewed markets, but for moderate slack, the utility loss is $2$– $5\%$ . Implementing parity constraints has minimal impact on overall auction revenue and is comparable to the price of demographic parity observed in classification (Nasr et al., 2019).

7. Legal, Regulatory, and Societal Considerations

European non-discrimination law bans both direct and indirect price differentiation on protected grounds. Enforcing statistical parity in pricing mandates a transition from segment-specific pricing (third-degree discrimination) to uniform pricing, with explicit revenue loss:

$\Delta R = R_{PD} - R_{U} \geq 0$

In real retail settings, this "compliance cost" is empirically 3–8% of profit, plus ongoing algorithmic fairness auditing and legal risk overhead. Notably, gaps and loopholes in current regulatory regimes allow hidden algorithmic discrimination unless robust ex ante audits and transparency mandates are enforced (Borgesius, 28 Sep 2025).

The price of non-discrimination is thus a multi-dimensional concept encompassing worst-case economic ratios, empirically observed utility or risk gaps, and practical barriers to reconciling accuracy, profit, privacy, and fairness. The literature consistently shows that, while the cost varies by setting and constraint severity, achieving non-discriminatory outcomes rarely comes for free, but in many operational domains, the theoretical and empirical penalties are moderate and tractable.