Fairness via Bounded Work

• The paper introduces a framework that relaxes full metric access by enforcing fairness through expected Lipschitz conditions over identifiable subpopulations.
• It details a switching subgradient descent (SSD) algorithm that alternates between feasibility and objective updates to guarantee convergence under bounded computations.
• The approach balances trade-offs between fairness guarantees and computational efficiency, highlighting limitations in metric sample complexity and runtime.

Searching arXiv for the specified paper and closely related work to ground the article in current arXiv records. arxiv_search(query="(Kim et al., 2018) fairness bounded work computationally-bounded awareness Dwork metric multifairness", max_results=5) In the supplied literature, “fairness via bounded work” appears in several related forms: bounded metric queries and runtime in fair classification, a single swap in envy reduction, bounded sharing of indivisible goods, bounded rate and leakage in representation design, bounded temporal skew in FIFO markets, bounded equalized work in periodic assignment, and bounded cumulative deficits in perpetual online allocation (Kim et al., 2018, Echenique et al., 12 Aug 2025, Bismuth et al., 2019, Zamani et al., 18 Aug 2025, Mavroudis, 2019, Lieshout et al., 6 Jul 2025, Kahana et al., 19 May 2026). In fair classification, the clearest formalization is “computationally-bounded awareness”: instead of assuming full access to a task-specific similarity metric, the learner has bounded metric queries, bounded runtime, and bounded sample complexity, and fairness is enforced through a relaxation called metric multifairness (Kim et al., 2018). The resulting framework preserves the metric-based intuition of “treat similar individuals similarly,” but makes it estimable and optimizable from aggregate information over comparison sets rather than from exhaustive pairwise metric knowledge (Kim et al., 2018).

1. Computationally-bounded awareness

The 2018 formulation begins from the “fairness through awareness” framework of Dwork et al., in which fairness is encoded by a task-specific similarity metric $d$ over pairs of individuals. In its classic form, metric fairness requires a classifier $f$ to be approximately Lipschitz with respect to $d$: similar individuals, meaning pairs with small $d(x,x')$, must receive similar predictions, meaning small $|f(x)-f(x')|$. The central limitation is that full metric access is impractical: a learner typically cannot know or query $d(x,x')$ for all pairs. The bounded-work alternative therefore assumes only random sample access to the metric, together with computational and statistical restrictions that still allow a strong relaxation of metric fairness to be enforced (Kim et al., 2018).

The formal setting is standard convex ERM with fairness constraints. Individuals live in a feature space $X \subseteq \mathbb{R}^n$, each $x \in X$ satisfies $\|x\|_1 \le 1$, and labeled examples are drawn from a distribution $D$ over $f$0. The hypothesis class is convex; the paper focuses on linear predictors with bounded weights,

$f$1

where for $f$2,

$f$3

projected onto $f$4. The fairness metric is an arbitrary nonnegative symmetric function

$f$5

and the learner does not know $f$6 explicitly. Instead it receives random metric samples $f$7 with

$f$8

together with bounded magnitude. The loss $f$9 is convex and $d$0-Lipschitz over its domain (Kim et al., 2018).

Under this viewpoint, bounded work has three components. First, there are bounded metric queries: only a small number $d$1 of random samples from the metric are available. Second, there is bounded runtime: algorithms should have polynomial, and ideally polylogarithmic, dependence on the number of fairness constraints. Third, there is bounded sample complexity: the learner must estimate all fairness constraints in a rich but finite family of comparisons $d$2 from few samples. The objective is not to recover the entire metric, but to enforce a fairness notion that is compatible with this information bottleneck (Kim et al., 2018).

2. Metric multifairness

Classical approximate metric fairness requires that for $d$3,

$d$4

Metric multifairness replaces this universal pairwise constraint by a family of expected constraints over comparison sets. Let $d$5 be a collection of possibly overlapping comparison sets over pairs, and let $d$6 be a sampling distribution over $d$7. For $d$8, sampling $d$9 means sampling from $d(x,x')$0 conditioned on $d(x,x')$1. The collection $d(x,x')$2 is $d(x,x')$3-large if every $d(x,x')$4 has mass at least $d(x,x')$5 under $d(x,x')$6:

$d(x,x')$7

Then $d(x,x')$8 is $d(x,x')$9-metric multifair if

$|f(x)-f(x')|$0

This enforces Lipschitzness in expectation over many efficiently identifiable subpopulation comparisons rather than for every pair in the population (Kim et al., 2018).

The comparison class $|f(x)-f(x')|$1 is intended to be expressive but estimable. The paper gives conjunctions of a few features and short decision trees as examples. The point is not merely statistical convenience: the framework guarantees that similar subpopulations are treated similarly, as long as these subpopulations are identified within $|f(x)-f(x')|$2. Because the learner sees only bounded metric information, aggregate constraints of this form are both information-theoretically and computationally estimable from few samples when the sets are $|f(x)-f(x')|$3-large (Kim et al., 2018).

For linear predictors, the fairness constraints are encoded by residuals

$|f(x)-f(x')|$4

Then $|f(x)-f(x')|$5 is $|f(x)-f(x')|$6-metric multifair iff $|f(x)-f(x')|$7 for all $|f(x)-f(x')|$8. This residual form is crucial because it turns fairness enforcement into convex constrained optimization, with one convex constraint per comparison set (Kim et al., 2018).

A central structural implication is that sufficiently rich $|f(x)-f(x')|$9 can recover strong individual protections on large-measure subsets. If $d(x,x')$0 satisfies

$d(x,x')$1

then for any $d(x,x')$2, a $d(x,x')$3-metric multifair predictor is $d(x,x')$4-metric fair on at least a $d(x,x')$5-fraction of pairs in $d(x,x')$6, by Markov’s inequality. Thus multifairness is not simply a coarse group relaxation; when $d(x,x')$7 identifies many low-average-distance comparisons, it yields strong individual-level consequences on most pairs in those comparisons (Kim et al., 2018).

3. Optimization under bounded metric access

The population optimization problem is

$d(x,x')$8

In practice, the objective is estimated from labeled samples and the constraints from metric samples. The algorithm proposed for this setting is “switching subgradient descent” (SSD), which alternates between feasibility steps and objective steps (Kim et al., 2018).

Residual estimation is split across the two terms of $d(x,x')$9. For each $X \subseteq \mathbb{R}^n$0 and current $X \subseteq \mathbb{R}^n$1, the learner computes an estimate $X \subseteq \mathbb{R}^n$2 with tolerance $X \subseteq \mathbb{R}^n$3. Metric samples are used to estimate $X \subseteq \mathbb{R}^n$4, and unlabeled pair samples are used to estimate $X \subseteq \mathbb{R}^n$5. For the objective, the learner uses a stochastic subgradient $X \subseteq \mathbb{R}^n$6 with conditional expectation in $X \subseteq \mathbb{R}^n$7, obtained from one labeled sample. For constraints, it uses a stochastic subgradient $X \subseteq \mathbb{R}^n$8 with conditional expectation in $X \subseteq \mathbb{R}^n$9. In the linear case, a valid subgradient coordinate is

$x \in X$0

and an unbiased estimate is obtained from a single draw $x \in X$1 (Kim et al., 2018).

The SSD update rule is explicit. Starting from $x \in X$2, the algorithm maintains a set $x \in X$3 of empirically feasible iterates. At iteration $x \in X$4, if there exists $x \in X$5 with $x \in X$6, the algorithm picks such an $x \in X$7 and performs a feasibility update

$x \in X$8

Otherwise it records $x \in X$9 in $\|x\|_1 \le 1$0 and performs an objective update

$\|x\|_1 \le 1$1

The final output is the average

$\|x\|_1 \le 1$2

The threshold $\|x\|_1 \le 1$3 and the step sizes are chosen to ensure both feasibility and utility convergence (Kim et al., 2018).

The bounded-work interpretation is sharp in the sample complexity. Under $\|x\|_1 \le 1$4-large $\|x\|_1 \le 1$5, with probability at least $\|x\|_1 \le 1$6, a total of

$\|x\|_1 \le 1$7

metric samples suffice to estimate $\|x\|_1 \le 1$8 within tolerance $\|x\|_1 \le 1$9 for all $D$0. Estimation of $D$1 from unlabeled pair samples has similar bounds. The dependence on $D$2 is logarithmic, which is one of the central quantitative advantages of relaxing pairwise fairness to multifairness over a rich but finite comparison family (Kim et al., 2018).

4. Guarantees, efficient violation search, and lower bounds

The main learning theorem states that if $D$3 and $D$4 is $D$5-large, then with probability at least $D$6, SSD outputs $D$7 that is $D$8-metric multifair and $D$9-optimal in expected loss in

$f$00

iterations, using

$f$01

metric samples. Each iteration uses at most one labeled example and runs in time

$f$02

A supporting fairness lemma shows that if all residual estimators have tolerance $f$03, then the averaged iterate $f$04 is $f$05-metric multifair. A utility lemma gives

$f$06

where $f$07 is an optimal feasible $f$08-metric multifair hypothesis in $f$09 (Kim et al., 2018).

The principal computational bottleneck is violation search: given current $f$10, one must determine whether some $f$11 violates the fairness constraint. The paper reduces this search to agnostic learning of the concept class corresponding to $f$12. The reduction defines labels

$f$13

and uses the fact that if some $f$14 violates, then the indicator of $f$15 correlates with $f$16. An agnostic learner can return a hypothesis with comparable correlation, which induces a soft comparison with residual $f$17 and therefore enables progress. This connects efficient fairness enforcement to the learnability of the comparison class rather than to brute-force search over all constraints (Kim et al., 2018).

The framework is also accompanied by lower bounds showing that bounded work is not merely a proof artifact. Any algorithm that outputs $f$18-metric multifair predictions from random metric samples must either take

$f$19

metric samples or else output predictions whose loss approaches the loss achievable with no metric queries. The paper also gives an informal cryptographic hardness proposition: assuming one-way functions exist, there is no generally efficient algorithm that computes $f$20-optimal $f$21-metric multifair predictions for arbitrary $f$22 and constant $f$23; stronger assumptions yield explicit runtime lower bounds of $f$24 for some $f$25 (Kim et al., 2018).

These results delimit the scope of the bounded-work approach. The framework achieves fairness guarantees that scale logarithmically in the expressiveness of $f$26, but it cannot, in general, recover universal pairwise fairness from very few samples, nor can it provide universally efficient algorithms without learnability assumptions on the comparison family. The bounded-work regime therefore represents a principled compromise between the full-metric ideal and computational feasibility (Kim et al., 2018).

5. Relation to prior fairness notions, practical use, and limitations

Relative to classical metric fairness, multifairness is a relaxation motivated by partial observability rather than by a rejection of metric-based reasoning. Metric fairness requires

$f$27

for all pairs and presumes full access to $f$28. Multifairness instead enforces expected Lipschitzness over many comparisons $f$29, which preserves the modularity and interpretability of metric-based fairness while avoiding the impractical full-metric requirement. The paper explicitly contrasts this with group and subpopulation fairness frameworks such as equalized error rates and calibration across subgroups: those approaches guarantee specific statistical parity notions across rich subgroup families, whereas metric multifairness enforces “treat similar subpopulations similarly” with respect to a chosen metric $f$30 and comparison family $f$31 (Kim et al., 2018).

The framework also admits a model-agnostic post-processing interpretation. Any set of pre-trained predictions can be projected onto the multifair feasible set without retraining and without full metric access. In that construction, the $f$32 individuals are treated as basis vectors in $f$33, and the optimization objective becomes squared deviation from the original predictions subject to multifairness constraints. This shows that bounded metric access is compatible not only with end-to-end training but also with downstream fairness correction (Kim et al., 2018).

The design parameters encode explicit fairness–utility trade-offs. Choosing $f$34 to be more expressive strengthens fairness but increases metric sample complexity only logarithmically; runtime can be improved through agnostic learners for $f$35. Smaller $f$36 strengthens fairness but increases both iteration complexity and sample complexity as $f$37. The requirement that $f$38 be $f$39-large means that the metric sampling distribution $f$40 must cover the subpopulations of interest. The paper notes that this can be arranged by the metric-designing authority, but it also implies that poorly chosen $f$41 or poorly aligned $f$42 can weaken protections (Kim et al., 2018).

The limitations are equally explicit. Multifairness protects only efficiently identifiable comparisons in $f$43; harmful subpopulations absent from $f$44 may remain unprotected. Naive violation search is linear in $f$45 per iteration, and although the agnostic-learning reduction mitigates this, it depends on the availability of efficient agnostic learners. The paper is theoretical rather than empirical, so practical construction of $f$46, empirical validation across domains, and effective selection of $f$47 remain open directions. The post-processing guarantees are population-specific, and out-of-sample fairness and utility guarantees under multifairness are identified as an interesting extension (Kim et al., 2018).

6. Other formalizations of bounded work

Outside metric fair classification, the bounded-work idea is instantiated in several non-equivalent ways. In discrete fair division, one line of work measures the amount of local corrective action needed to eliminate envy. “Swap bounded envy” requires that if agent $f$48 envies agent $f$49, then a single swap of one item between the two bundles can remove the envy:

$f$50

A related relaxation allows a bounded number of shared items or sharings; for additive utilities, proportionality, envy-freeness, and equitability always admit allocations with at most $f$51 sharings, while consensus admits allocations with at most $f$52 sharings (Echenique et al., 12 Aug 2025, Bismuth et al., 2019).

In capacity-constrained decision systems, bounded work appears as bounded price of fairness. One formulation studies prediction with a capacity constraint $f$53 and utility $f$54, and shows that equal opportunity and max-min fairness can have unbounded price of fairness under information gaps, while normalized proportional fairness and an “Achievable equal opportunity” variant have bounded price of fairness. For two equal-sized groups, normalized proportional fairness satisfies

$f$55

and for $f$56 groups, Achievable EO guarantees at least a $f$57 fraction of the unconstrained optimum (Bachmat et al., 26 Feb 2026). In a different information-theoretic formulation, bounded work is identified with a representation-rate budget and a leakage budget: design $f$58 to maximize $f$59 subject to

$f$60

Here $f$61 is interpreted as bounded representational work and $f$62 as bounded statistical-parity leakage (Zamani et al., 18 Aug 2025).

Temporal and scheduling settings use yet another meaning. In periodic assignment, fairness requires that each worker performs the same work over time; the resulting price of fairness is at most one extra worker, and a fair solution exists with

$f$63

In FIFO financial markets, unbounded temporal fairness is replaced by $f$64-fairness:

$f$65

and an exchange is $f$66-bounded temporally fair if races are $f$67-fair with probability at least $f$68. In concurrent and parameterized systems, bounded fairness means that no process waits more than $f$69 steps between moves, and Prompt-LTL\X uses the prompt-eventually operator to express such bounds. In perpetual online allocation, fairness is tracked by deficits and a potential-minimizing online rule guarantees prefix-wise slack that grows on the order of $f$70, with a matching lower bound showing that such growth is unavoidable in general (Lieshout et al., 6 Jul 2025, Mavroudis, 2019, Jacobs et al., 2019, Kahana et al., 19 May 2026).

Taken together, these formulations do not define a single universal fairness criterion. They define a recurring methodological pattern: exact fairness is either enforced or approximated under an explicit bound on the work available to the mechanism, whether that work is metric access, local repair, item splitting, representational rate, timing skew, staffing overhead, scheduling delay, or cumulative deficit. The 2018 metric-multifairness framework remains the most direct machine-learning realization of this pattern, because it turns bounded informational access into a formal fairness notion with explicit statistical, algorithmic, and lower-bound guarantees (Kim et al., 2018).