Separation-Utility Pareto Frontier: An Information-Theoretic Characterization

We study the Pareto frontier (optimal trade-off) between utility and separation, a fairness criterion requiring predictive independence from sensitive attributes conditional on the true outcome. Through an information-theoretic lens, we prove a characterization of the utility-separation Pareto frontier, establish its concavity, and thereby prove the increasing marginal cost of separation in terms of utility. In addition, we characterize the conditions under which this trade-off becomes strict, providing a guide for trade-off selection in practice. Based on the theoretical characterization, we develop an empirical regularizer based on conditional mutual information (CMI) between predictions and sensitive attributes given the true outcome. The CMI regularizer is compatible with any deep model trained via gradient-based optimization and serves as a scalar monitor of residual separation violations, offering tractable guarantees during training. Finally, numerical experiments support our theoretical findings: across COMPAS, UCI Adult, UCI Bank, and CelebA, the proposed method substantially reduces separation violations while matching or exceeding the utility of established baseline methods. This study thus offers a provable, stable, and flexible approach to enforcing separation in deep learning.

💡 Research Summary

The paper tackles the fundamental fairness‑utility dilemma that arises when machine‑learning models must satisfy the “separation” (or equalized odds) criterion while maintaining high predictive performance. The authors adopt an information‑theoretic viewpoint, defining predictive utility as the mutual information between the model output U (and its deterministic version bY) and the true label Y, i.e., u = I(U;Y). Separation violation is quantified by the conditional mutual information (CMI) between the output and the sensitive attribute Z given Y, i.e., v = I(U;Z | Y). This mapping places every feasible predictor on a two‑dimensional “information plane” where the horizontal axis is the separation violation v and the vertical axis is utility u.

The authors first characterize the set of attainable (v,u) pairs for deterministic predictors (S_det) and for randomized predictors (S_rand). They prove (Theorem 2.2) that the randomized attainable set is exactly the convex closure of the deterministic set. Consequently, the Pareto frontier of the randomized case is concave and can be achieved by mixing at most two deterministic predictors. This concavity implies an increasing marginal cost: each additional reduction in separation violation requires a larger sacrifice in utility.

Next, the paper justifies CMI as the precise scalar measure of separation. Proposition 2.3 shows that I(U;Z | Y)=0 if and only if U⊥Z | Y, i.e., the separation condition holds exactly. Moreover, Lemma 2.4 and Theorem 2.5 demonstrate that CMI upper‑bounds the conditional correlation between any bounded functions of the output and the sensitive attribute. Thus minimizing CMI uniformly suppresses every possible statistical test an adversary could use to detect dependence, providing a strong, distribution‑free guarantee.

The authors also derive an “information budget” identity (Lemma 2.6): u + v = I(U;(Y,Z)) ≤ I((X,Z);Y) + H(Z | Y). This shows that, in general, there is no inevitable trade‑off; perfect utility and perfect separation can coexist in degenerate cases (e.g., when Y already encodes Z). To focus on realistic settings, they impose a non‑degeneracy condition: X⊥Z | Y and Y ⊥̸Z | X. Under this assumption, Theorem 2.8 proves that the maximal utility achievable with zero separation violation equals the utility of the best predictor that uses only X (denoted u*_X). If Z carries any additional predictive information about Y beyond X (i.e., I(Z;Y | X)>0), then any predictor achieving utility greater than u*_X must incur a strictly positive separation violation. Hence, a strict, unavoidable trade‑off emerges.

From theory to practice, the paper proposes a simple empirical regularizer for discrete tasks: directly estimate CMI from empirical joint frequencies (a plug‑in estimator) and add λ·ĥv to the loss, where λ controls the fairness‑utility balance. This avoids the need for adversarial networks, variational bounds, or auxiliary density models, and retains the theoretical guarantees because the estimator is unbiased under i.i.d. sampling (Proposition 3.1).

Extensive experiments on four benchmark datasets—COMPAS, UCI Adult, UCI Bank, and CelebA—validate the theory. By sweeping λ, the authors trace smooth Pareto curves that dominate those obtained by reduction‑based methods, adversarial debiasing, or variational CMI approximations. Their method consistently reduces CMI (hence separation violations) while matching or improving standard utility metrics such as accuracy or cross‑entropy loss. Notably, even in high‑cardinality output spaces (CelebA), the simple statistics‑based regularizer remains stable and effective.

In summary, the paper makes three major contributions: (1) a general, model‑agnostic characterization of the separation‑utility Pareto frontier, proving its concave shape and the necessity of randomization; (2) a rigorous justification of conditional mutual information as the canonical scalar for separation violation, with uniform dependence guarantees; and (3) a practical, theoretically‑grounded CMI regularization technique that works with any gradient‑based deep model. These results bridge a gap between fairness theory and scalable deep‑learning practice, offering a principled, transparent, and computationally efficient pathway to enforce separation in modern high‑dimensional models.