Amortized Fairness for Repeated Resource Allocation

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 270 papers99 conceptsfrozen 2026-10-04

In traditional algorithmic fairness, we often treat decisions as isolated events. If we are assigning tasks to workers or bandwidth to households, we try to make that specific moment as equitable as possible. However, as Tom Sühr, Abhijnan Chakraborty, and their colleagues argue in Two-Sided Fairness for Repeated Matchings in Two-Sided Markets, this "instantaneous" fairness often comes at a massive cost to efficiency. If every single ride-hail match must be perfectly fair to every driver, the platform’s operational performance can collapse.

The emerging research direction of amortized fairness (or temporal fairness) suggests a different path: instead of forcing fairness into every transaction, we should balance the scales over time. By expanding the horizon from seconds to weeks, we can achieve both high efficiency and long-term equity.

The Efficiency-Fairness Tradeoff

The fundamental tension in resource allocation is that the most "efficient" choice—the one that maximizes total utility or minimizes cost—is rarely the most "fair" one in the short term. In vehicle routing, Piotr Matl and Thibaut Vidal note in Workload Equity in Vehicle Routing Problems that balancing driver workloads on a daily basis is unnecessarily restrictive.

When we insist on "per-step" fairness, we lose the ability to exploit favorable conditions. For example, a GPU scheduler like THEMIS, developed by Kshiteej Mahajan and Arjun Singhvi, recognizes that raw resource sharing doesn't guarantee fair completion times for long-running ML jobs. By allowing short-term deviations—letting one job "hog" resources to finish quickly—the system can actually improve overall cluster efficiency by up to 250% while still ensuring that no user is worse off in the long run than they would be on private hardware.

The Mechanism of Temporal Memory

To make amortized fairness work, a system must have a "memory" of past inequities. This is often modeled as a credit or debt system.

In Mind Your Own Bandwidth, Felix Ming Fai Wong and Mung Chiang propose a virtual credit economy for broadband. Users "spend" credits to get priority during peak hours and "earn" them back by yielding during off-peak times. Similarly, Chen Xu and Xu Chen introduce BankFair for recommendation systems. When user traffic is low, the system prioritizes accuracy (user satisfaction) and accumulates "exposure debt" to under-represented providers. When traffic peaks, the system "pays back" this debt by boosting those providers' visibility.

This stateful approach allows the system to implement "turn-taking" strategies. As Toryn Q. Klassen and Sheila A. McIlraith describe in Pluralistic Alignment Over Time, an AI doesn't need to find a single "average" compromise between conflicting stakeholder values; it can achieve alignment by reflecting different values at different points in a sequence of decisions.

Why Repetition Enables Fairness

The mathematical beauty of this direction lies in the fact that repetition dissolves the "impossibility results" of fair division. In Repeated Fair Allocation of Indivisible Items, Ayumi Igarashi and Martin Lackner show that while it is often impossible to divide a single set of items fairly and efficiently, doing so over multiple rounds allows agents to compensate for "bad" rounds with "good" ones.

This works because of combinatorial variety. Over a 5-day work week, there are many ways to combine routes to ensure every driver has an equal total workload. As Najmeh Nekooghadirli and Michel Gendreau demonstrate, aggregating metrics over a multi-day horizon allows companies to achieve near-perfect equity without increasing the total distance traveled. The imbalances of Monday are canceled out by the assignments of Tuesday and Wednesday.

From Static Constraints to Dynamic Optimization

The shift toward amortized fairness requires moving from static solvers to dynamic frameworks like Reinforcement Learning (RL) or Online Optimization.

  1. LoTFair: Ruijie Du and Yanning Shen developed LoTFair, which treats fairness as a cumulative constraint. This acts as a "buffer," preventing noisy data spikes from forcing the system into suboptimal decisions.
  2. Nash Welfare: Rupert Freeman and Vincent Conitzer explore maximizing the product of agents' total utilities over time, showing that greedy online algorithms can achieve high long-term efficiency by accounting for historical utility accrual.
  3. Regularity: Bin Li and Xiaoyi Wu link "reward debt" to the "age" of the last reward, ensuring that users receive service at predictable intervals—a concept known as service regularity.

By viewing fairness as a longitudinal commitment rather than a momentary hurdle, researchers are creating systems that are more resilient, more efficient, and ultimately more equitable in the ways that matter to real-world participants.

Papers in this version

Showing 60 of 70 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 98 papers
  2. Version 1 · 2026-10-04 · 70 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_amortized_fairness_cdbfcde2,
  title = {amortized fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/amortized-fairness/rdr_cdbfcde208874add8cf913bc044aa818},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_amortized_fairness_for_repeated_resource_cdbfcde2_v1,
  title = {Amortized Fairness for Repeated Resource Allocation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/amortized-fairness-for-repeated-resource-allocation/rdr_cdbfcde208874add8cf913bc044aa818/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}