Traditional optimization algorithms are cold utilitarians. Whether routing delivery trucks, scheduling power grid curtailments, or distributing vaccines, standard algorithms seek to maximize a single global objective: total profit, minimal distance, or maximum average coverage.
However, this global maximization often creates severe localized harms. It creates "service deserts" in remote areas, starves low-income neighborhoods of resources, or repeatedly penalizes the exact same participants to keep a broader network stable.
To fix this, a growing body of research focuses on fairness-constrained optimization. The core insight is that equity should not be a passive metric evaluated after an algorithm runs. Instead, fairness must be mathematically woven into the constraints, objectives, and loss functions of the optimization problem itself. By reshaping the mathematical boundaries of what is considered an "optimal" solution, these frameworks force algorithms to balance raw efficiency with social and geographic equity.
Translating Fairness into Mathematical Constraints
How do you turn a social concept like "fairness" into a rigorous mathematical formulation? Researchers use several distinct mechanisms depending on the structure of the problem.
1. Min-Max Optimization and Service Floors
In spatial planning or resource distribution, algorithms often neglect low-density or remote areas because serving them is inefficient. To prevent this, researchers use min-max optimization or epsilon-constraints to establish a mandatory "floor" for service levels.
For example, in urban facility planning, Jinlin Li et al. incorporate min-max optimization into generative models to ensure resources are distributed equitably across diverse residential demands, actively mitigating the formation of service deserts. Similarly, in supply chain logistics for perishable goods, the authors of Joint Optimization of Fair Facility Allocation and Robust Inventory Management for Perishable Consumer Products use an epsilon-constraint to enforce a minimum service threshold for disadvantaged regions. This treats equitable access as a hard baseline rather than a secondary goal.
2. Demographic Parity and Group Budgets
When resources must be distributed across distinct social or demographic groups, algorithms can be constrained to respect group-level quotas or parity metrics.
In large-scale epidemic control, Prathyush Sambaturu et al. introduce group-specific budget constraints into a network vaccination framework. This mechanism allows practitioners to explicitly manage the trade-off between maximizing overall population health and maintaining equitable access across different age or social groups. In a broader context, Abolfazl Asudeh et al. formalize the Fair Maximum Coverage (FMC) problem, which forces the selection of resource sets to reflect specific target demographic proportions, preventing the algorithm from optimizing for total reach by completely ignoring minority sub-populations.
3. Temporal Fairness and Sacrifice Indices
Fairness is not just spatial; it is also temporal. In smart grids, managing volatile solar energy requires "curtailing" (reducing) power load or supply. A naive algorithm might repeatedly penalize households at the end of a distribution line because it is mathematically easiest to stabilize the grid from those nodes.
To prevent this, Sanmukh R. Kuppannagari et al. develop rounding heuristics that track curtailment budgets for each node over time. Taking this further, Han Yu et al. introduce a "historical sacrifice index" into a Lyapunov-based control strategy for industrial power management. By tracking which entities have previously borne the burden of service interruptions, the system ensures that the operational costs of grid stability are distributed equitably over time. This temporal memory is essential for maintaining long-term stakeholder cooperation and preventing participant withdrawal, as noted by Han Yu et al..
Navigating the Efficiency-Fairness Trade-off
Why does adding fairness constraints work, and what does it cost? In optimization literature, this tension is known as the "price of fairness."
MohammadTaghi Hajiaghayi quantifies this exact mathematical trade-off in online matching markets, proving tight bounds on how close an algorithm can get to maximizing total utilitarian social welfare while simultaneously satisfying envy-freeness and proportionality across different agent classes.
The complexity of this trade-off depends heavily on how resources are discovered and targeted. As Hadi Elzayn et al. demonstrate, when an allocation algorithm relies on precision-based targeting rather than random discovery, achieving equitable outcomes requires complex estimation of latent candidate distributions.
Furthermore, the mathematical structure of the underlying problem dictates the strategy. For instance, Moise Blanchard et al. show that while traditional vehicle routing favors consolidation to minimize total path length, minimizing wait times fairly across a population requires a strategy of dispersion. This forces a deliberate compromise between raw routing efficiency and geographical equity.
Application Domains: From Public Health to Supply Chains
This framework is being actively deployed across diverse real-world systems where purely profit-driven or efficiency-driven models fail social sustainability tests.
Public Health Interventions
Restless Multi-Armed Bandits (RMABs) are widely used to allocate limited human resources, such as health workers making "nudge" calls to patients. In the SAHELI maternal health program, Shresth Verma et al. explicitly oriented algorithmic prioritization toward equity by favoring participants with lower socio-economic indicators, shifting the focus from demographic blindness to active support for marginalized groups.
To make these systems more robust, Andrew Perrault introduced risk-aware objectives into RMABs. This ensures that individuals with low responsiveness to interventions are not systematically abandoned by an algorithm chasing maximum expected reward.
Supply Chain Sustainability
Traditional logistics models minimize operational costs, which often results in "service level inequality" where less profitable or remote regions are underserved. Rifny Rachman et al. leverage Multi-Objective Reinforcement Learning to treat service level equality as a core objective alongside profit and carbon emissions, ensuring social sustainability across multi-echelon supply chains.
Similarly, in disaster relief logistics, Jia Luo highlights that incorporating fairness as a hard mathematical constraint minimizes the variance of satisfaction levels among victims, preventing optimization models from prioritizing easily accessible areas at the expense of remote, high-need locations.
Hidden Pitfalls: Data Privacy and Historical Bias
A major challenge in designing fairness-constrained systems is that the data feeding the optimization models is often flawed, noisy, or biased.
For example, Differential Privacy (DP) is frequently used to protect sensitive demographic data before it is used in public policy decisions. However, Ferdinando Fioretto et al. demonstrate that the "post-processing" steps required to make noisy DP data usable introduce severe, systematic unfairness. Because of how noise corrections interact with allocation formulas, smaller demographic groups or specific geographic districts consistently receive less than their fair share of funding.
Furthermore, if the underlying predictive models rely on biased historical data, the downstream resource allocation will naturally reinforce those inequities. As Zixuan Feng et al. note in the context of public health resource allocation, without explicit fairness constraints, models risk inadvertently disadvantaging specific communities by codifying historical data gaps into future resource distributions. Bryan Wilder echoes this concern, emphasizing that integrating predictive modeling directly with decision-making pipelines is required to ensure that AI benefits marginalized communities rather than just those with the most data.
Key Foundational Work
Optimal Net-Load Balancing in Smart Grids with High PV Penetration (2017) — contributed 2 concept elements to this direction.
Stochastic programming approach for earthquake disaster relief mobilization with multiple objectives (2013) — contributed 1 concept elements to this direction.