Fairness–utility trade-offs in binary decisions collapse to simple group-specific probability thresholds: the entire Pareto frontier is formed by deterministic threshold rules (sometimes favoring lower-risk applicants), driven by population characteristics and objective weights rather than how the algorithm is engineered.

Main Finding

The Pareto frontier of binary prediction-based decision systems (trade-off between decision-maker utility and group fairness) — taken over the full space of possible decision rules, including group-dependent rules — consists of group-specific deterministic threshold rules applied to individuals’ success probability p(x) = P[Y=1 | x]. These thresholds can be lower-bound, upper-bound, or combinations thereof. The frontier is determined solely by (i) the population distributions of p conditional on group (g(p | a)), (ii) the decision-maker (DM) and decision-subject (DS) utility functions, and (iii) the chosen fairness score (FS) derived from group-wise expected DS utilities. It does not depend on the technical design (pre-, in- or post-processing) of the algorithm. The paper generalizes prior optimality results (which focused on specific confusion-matrix parity constraints and lower-threshold rules) to a wide class of fairness metrics (including egalitarian, prioritarian, Rawlsian, sufficientarian principles) and to partial fairness regimes, and highlights that Pareto-optimal solutions can include “upper-threshold” rules (cherry-picking) when DS utility structures make it optimal.

Key Points

• Problem setup:
  • Binary decision D ∈ {0,1} based on features x (which may include protected attribute a), with binary true outcome Y ∈ {0,1}.
  • DM utility U specified by 2×2 payoff matrix (u_ij) for (D=i, Y=j).
  • DS impact/fairness captured by DS utility matrix V (v_ij) and aggregated into group-wise expected DS utilities; fairness score FS is a function of those expected utilities across groups. FS can represent equality measures (differences) or other justice-inspired objectives (maximin, prioritarian, sufficientarian).
• Main theorem:
  • Over all decision rules (including those conditional on a), any Pareto-optimal rule is deterministic and takes the form: for each group a, accept iff p(x) is inside some interval (possibly a half-line) — i.e., thresholding on p with group-specific lower and/or upper cutoffs.
  • Depending on FS and DS utility matrix, Pareto-optimal thresholds can be:
    • Lower-bound only (classical: accept above threshold),
    • Upper-bound only (prefer low-p individuals — “cherry-picking”),
    • Mixed (accept only if p in [t_low, t_high]).
• Relation to prior work:
  • Recovers classical results that parity constraints like statistical parity, TPR/FPR parity produce group-wise lower-bound thresholds as special cases.
  • Extends to: (a) generalized DS-utility-based fairness metrics (not only confusion-matrix metrics), (b) other justice principles (Rawls, prioritarianism), and (c) partial enforcement of fairness (not only strict parity constraints).
• Technology-agnostic benchmark:
  • The Pareto frontier gives an outer bound for achievable utility-fairness trade-offs independent of model architecture. Thus different algorithmic approaches that optimally trade off the same objectives will converge to these threshold rules.
• Practical observation:
  • For locating the frontier it suffices to estimate the distributions g(p | a) of p = P[Y=1 | x] within groups, not a perfect Bayes predictor for each x.
• Empirical / implementation note:
  • The authors provide code (GitHub) and an example showing that an in-processing method can produce group-specific thresholds even without explicit access to the protected attribute.

Data & Methods

• Nature of work: theoretical/analytical with illustrative examples and code (no original field dataset).
• Formal model:
  • Binary decision framework, joint distribution of (Y, x); p(x) = P[Y=1 | x] and feature density h(x).
  • DM utility matrix (u_ij) and DS utility matrix (v_ij). Expected utilities per group computed via integrals over g(p | a).
  • Fairness score FS = F( E[V | group] ), where F can encode difference, min, weighted sum, thresholds, etc.
• Optimization viewpoint:
  • Multi-objective optimization in two dimensions: maximize DM expected utility and optimize (minimize or maximize depending on principle) FS across all measurable decision rules f: x ↦ {0,1} (including those that use a).
  • Characterization of Pareto-optimal set via variational/ordering arguments on p — show pointwise optimality implies thresholding on p within each group.
• Generality:
  • Results proven for arbitrary DM/DS payoff matrices and arbitrary group distributions of p; hold for pre-, in- and post-processing approaches because proof is over the space of all decision rules.
• Reproducibility:
  • Python code implementing constructions and examples is shared: https://github.com/mcwilms/Pareto-Frontier-of-Binary-Decision-Systems
• Limitations of method:
  • Framework limited to binary outcomes & binary decisions; focuses on group fairness (not individual, causal, or counterfactual fairness); assumes access to/estimability of g(p | a) and specified utility matrices.

Implications for AI Economics

• Benchmarking and compliance:
  • Provides a technology-agnostic benchmark for regulators, auditors, and firms to judge whether a deployed system is (near) Pareto-optimal in utility-fairness trade-offs. This can be used in “Less Discriminatory Alternative” (LDA) assessments required by law: if a model lies strictly inside the Pareto frontier, an LDA exists.
• Cost-of-fairness quantification:
  • The frontier quantifies the minimal DM utility loss needed to achieve any desired fairness level (or, conversely, maximal fairness achievable at given utility). Economists can translate these trade-offs into monetary costs, prices, or firm profit impacts.
• Policy and regulation design:
  • Regulators can use the frontier to set feasible fairness targets or to require disclosures (e.g., estimated g(p | a), assumed utility functions) so stakeholders can assess achievable trade-offs. Because the frontier is independent of algorithmic architecture, policy can focus on outcomes rather than specific technical constraints.
• Market behavior and incentives:
  • Firms deciding on adoption of fairness constraints can use the frontier to decide governance/investment: e.g., whether to accept profit loss for fairness gains, or to redesign products to shift g(p | a) (e.g., feature collection strategies) to obtain better frontier points.
  • The possibility that Pareto-optimal solutions include upper-bound thresholds (cherry-picking) implies regulation should consider within-group impacts and specify which DS utility notions are socially acceptable. Simple parity constraints may be insufficient to prevent types of discriminatory selection that nonetheless satisfy some fairness metrics.
• Empirical strategy for economists:
  • Estimation focus: estimate distributions g(p | a) (density of score p within groups) and elicit or model DM/DS utility matrices to compute the frontier. This reduces data demands compared with trying to learn full optimal classifiers.
  • Welfare analysis: combine DM profits and group DS utilities into social welfare functions (e.g., weighted sum, Rawlsian) to select socially preferred points on the frontier; use counterfactual experiments to evaluate market-wide effects (entry, pricing).
• Mechanism and mechanism-design implications:
  • When protected-attribute–conditional decision rules are legally prohibited, the frontier under such constraints can be compared to the unconstrained frontier to quantify efficiency loss from prohibition. This informs trade-offs between anti-discrimination legal norms and economic efficiency.
• Cautions for applied work:
  • Translating theoretical payoffs into monetary utilities requires careful modeling of DS impacts (v_ij). Different DS utility specifications can reshuffle the frontier and the nature of optimal rules (e.g., creating upper-threshold optima). Policymakers and analysts must make normative choices explicit (egalitarian vs. prioritarian vs. maximin).
• Research directions for AI economics:
  • Empirical estimation of g(p | a) in real markets (credit, health triage, hiring) to map practical frontiers.
  • Integrate frontier analysis into firm-level cost-benefit and compliance models.
  • Study multi-period/dynamic extensions (learning, strategic behavior) and non-binary settings (multi-class, continuous decisions).
  • Investigate implications of prohibiting group-dependent rules vs. allowing them under transparency/oversight.

References / provenance - Wilms, M. & Heitz, C. (2026). Fairness vs Performance: Characterizing the Pareto Frontier of Algorithmic Decision Systems. FAccT ’26; arXiv:2605.10604. Code: https://github.com/mcwilms/Pareto-Frontier-of-Binary-Decision-Systems

If you want, I can: - Sketch the constructive procedure the authors use to compute the group-wise thresholds from g(p | a) and utilities, or - Outline an empirical protocol (data + steps) for estimating g(p | a) and producing an empirical Pareto frontier for a particular application (credit, hiring, health).