Smoothed Fairness Constraints

• Smoothed fairness constraints are techniques that substitute discontinuous fairness metrics with smooth surrogate functions for gradient-based optimization.
• They enable stable optimization in complex, nonconvex, high-dimensional prediction tasks using proxy-Lagrangian frameworks and parameter-space smoothing.
• The approach provides theoretical guarantees and calibration bounds, ensuring surrogate constraint satisfaction closely mirrors actual fairness metrics across various learning settings.

A smoothed fairness constraint is a methodological approach in algorithmic fairness wherein discontinuous, non-differentiable, or otherwise intractable fairness metrics (such as step-function-based group disparity, demographic parity, or risk differences) are approximated by smooth surrogate functions. This smoothing enables stable and efficient optimization, especially in non-convex, high-dimensional prediction and resource allocation tasks. Smoothed fairness constraints now encompass a broad array of technical formulations across supervised learning, reinforcement learning, online optimization, and econometrics, with rigorous theoretical, algorithmic, and empirical foundations in the current literature.

1. Formalizing Smoothed Fairness Constraints: Principles and Canonical Forms

Many group or individual fairness metrics originally depend on step or indicator functions, rendering direct optimization infeasible due to non-differentiability. For instance, demographic parity may be enforced by a constraint

$|\pi_0(\theta) - \pi_1(\theta)| \leq \epsilon,$

where each group rate is an empirical average of indicator functions: $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$ To facilitate learning, the indicator $1_{g\geq0}$ is replaced by a smooth surrogate $\sigma_s(g)$, such as a sigmoid, soft-ramp, or smoothed hinge function, giving

$$\tilde{\pi}_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} \sigma_s(g(x_i;\theta)),$$

and a smoothed constraint: $$\tilde{r}(\theta) = |\tilde{\pi}_0(\theta) - \tilde{\pi}_1(\theta)| - \epsilon \leq 0.$$ This paradigm underpins the proxy-Lagrangian and two-player game architectures developed for fairness-aware model optimization (Cotter et al., 2018).

Further canonical smooth surrogates include piecewise-linear lower/upper bounders (e.g., SLIDE: $\nu_\tau(z)$), concave/convex functions for convexification (e.g., $\kappa$, $\delta$ surrogates), Gaussian parameter smoothing in function space (averaging $f(x;W+\Delta)$ over $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$0), and “softmax-based” ordering or entropy surrogates in sequential decision making.

2. Optimization Methodologies and Algorithmic Strategies

Proxy-Lagrangian/Stochastic Game Formulation

Optimization under smoothed fairness constraints typically takes the form of a constrained minimization of predictive loss plus weighted smoothed constraint violations: $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$1 with Lagrange multipliers $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$2 adapted through projected (or mirror-descent) dual updates based on the unsmoothed constraint evaluations. Cotter et al. (Cotter et al., 2018) formalize the resulting learning dynamics as a two-player nonzero-sum game, for which semi-coarse correlated equilibria with provable optimality and near-feasibility are guaranteed under mild boundedness conditions.

Convex, Nonconvex, and Piecewise Surrogates

For practical optimization, surrogate choices include convex functions (e.g., hinge, log-loss, or square loss), but recent research demonstrates advantages to using nonconvex, tightly controlling surrogates. Goh et al. and others propose surrogates such as $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$3 (a smooth, nonconvex approximation to the Heaviside step), enabling direct hard constraints rather than soft penalization. Such surrogates satisfy tightness theorems, controlling the gap between surrogate and “actual” fairness violation (Khatti et al., 21 May 2025).

Parameter-Space Smoothing

Gaussian parameter-space smoothing constructs classifiers of the form $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$4. This operation induces a Fréchet-differentiable Nemytskii operator mapping parameters to function space, and allows fairness constraints to be certified in a way that is robust to data-distribution shift (Jin et al., 2022).

Entropy and Softmax Techniques in Sequential Decision Settings

Lipschitz or soft fairness constraints over policies enforce requirements such as

$$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$5

where $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$6 is the selection probability of arm $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$7 and $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$8 is the total variation distance between reward distributions. Sampling policies are constructed using posterior Thompson sampling or softmax-weighted value differences, with the degree of smoothing controlling the fairness–performance trade-off (Liu et al., 2017, Li et al., 2022).

3. Theoretical Guarantees and Surrogate Calibration

Most smoothed fairness frameworks establish calibration bounds relating surrogate constraints to their original discontinuous metrics. Tightness theorems state that, under mild regularity (e.g., symmetric, well-scaled surrogates), surrogate-constraint satisfaction up to $$\pi_k(\theta) = \frac{1}{|I_k|} \sum_{i:A_i=k} 1_{f(x_i;\theta)\geq\tfrac{1}{2}}.$$9 ensures real fairness violation is bounded by $1_{g\geq0}$0, for arbitrarily small $1_{g\geq0}$1 on proper scaling (Khatti et al., 21 May 2025).

For proxy-Lagrangian games, under $1_{g\geq0}$2 iterations and proper step-size schedules $1_{g\geq0}$3, the expected optimality gap and constraint violation are $1_{g\geq0}$4, and feasibility under strict constraint margin $1_{g\geq0}$5 can be improved further (Cotter et al., 2018).

Furthermore, input-agnostic guarantees are available in parameter-space smoothing models, where Lipschitz continuity is governed entirely by the norm of parameter differences and the smoothing radius $1_{g\geq0}$6—rendering the fairness certification independent of the data distribution (Jin et al., 2022).

4. Representative Instantiations and Comparative Surrogates

A variety of smoothing surrogates and frameworks are adopted in the literature, tailored to both empirical and theoretical desiderata:

Surrogate/Framework	Main Construction	Calibration/Guarantee
Proxy-Lagrangian (Cotter et al.) (Cotter et al., 2018)	Surrogate $1_{g\geq0}$7 for indicators; two-player game	$1_{g\geq0}$8 optimality and feasibility
SLIDE (Kim et al., 2022)	Piecewise-linear (lower-bounding) surrogate	Asymptotic fairness consistency, nonparametric rate
Nonconvex Smooth Step (Khatti et al., 21 May 2025)	Bounded, tight, nonconvex smoothed Heaviside	Surrogate tightness to within $1_{g\geq0}$9
Gaussian Parameter Smoothing (Jin et al., 2022)	Smoothing in $\sigma_s(g)$0-space (Nemytskii operator)	Input-agnostic, certified $\sigma_s(g)$1-fairness
Softmax Fairness in Bandits (Li et al., 2022)	Policy probabilities via softmax Q-differences	Entropy–performance trade-off, suboptimality control

In addition, convexification via $\sigma_s(g)$2 surrogates (Wu et al., 2018) and other lower/upper bounds are analyzed with explicit calibration functions.

5. Applications in Supervised, Online, and Sequential Learning

Supervised Classification

The landscape is dominated by empirical risk minimization subject to group-disparity constraints (demographic parity, equal opportunity, equalized odds, risk difference, etc.), each cast in a smoothed form. Cross-entropy, hinge, and novel surrogates (SLIDE, smooth nonconvex) enable both convex and constrained nonconvex optimization, with guarantees aligned to the surrogate’s calibration (Cotter et al., 2018, Kim et al., 2022, Khatti et al., 21 May 2025, Wu et al., 2018).

Online Convex Optimization

In online resource allocation—e.g., data center provisioning—the “long-term” fairness regularizer is decomposed into per-round smooth convex costs via auxiliary variables. FairOBD constructs a dynamic Lagrangian structure with switching and hitting costs, and achieves asymptotic competitive ratios against fairness-constrained benchmarks (Li et al., 11 Dec 2025). Smoothing is essential to the tractability and performance guarantees of the method.

Bandits and Reinforcement Learning

Multi-armed and restless bandits benefit from smooth probabilistic ordering constraints to prevent starvation and ensure that allocation probabilities respect underlying arm quality. Algorithms such as subjective/objective-smooth Thompson sampling and SoftFair enforce soft fairness or Lipschitz constraints, with provable upper bounds on cumulative fairness regret and entropic allocations (Liu et al., 2017, Li et al., 2022).

Structural Econometrics

Fairness is encoded as operator constraints in inverse problems, with a quadratic smoothing penalty on fairness violations. Analytic forms for fair and approximately fair projections are derived, with convergence governed by the choice of smoothing parameter (Centorrino et al., 2022).

6. Limitations, Trade-offs, and Empirical Insights

Smoothed constraints introduce an approximation–strictness trade-off. Choices of smoothing parameter, surrogate function, and regularization determine how closely the imposed constraint tracks the actual “unfairness” metric. Under-influential surrogates—such as covariance for demographic parity—may control a proxy without guaranteeing desired fairness on held-out data (Khatti et al., 21 May 2025). Empirical results consistently demonstrate that using tightly controlled/nonconvex surrogates, hard constraint enforcement, or parameter smoothing yields high fairness levels with strongly mitigated loss in accuracy or efficiency (Cotter et al., 2018, Kim et al., 2022, Jin et al., 2022, Cruz et al., 2022).

7. Future Directions and Open Questions

Recent developments suggest robust surrogate design, parameter-space smoothing, and input-agnostic certification are at the frontier of smoothed fairness constraint research. Open challenges remain in mapping surrogates tightly to nuanced fairness goals under context shift, extending smoothing to complex structural and temporal settings, and unifying theoretical calibration arguments across domains (Jin et al., 2022, Khatti et al., 21 May 2025). The need for rigorous surrogate validation in new settings and for interpretable trade-offs between utility and fairness remains a central research theme.