Monotonic Fairness: Principles & Applications

• Monotonic fairness is a fairness-by-design principle that ensures outcomes improve or deteriorate consistently with changes in qualification features, fairness slack, or entitlements.
• It encompasses varied constructions such as resentment-free classification, slack-consistent learning paths, and monotone aggregation, demonstrating its versatility across recommendation, credit scoring, and resource allocation.
• Practical implementations include sign-constrained neural networks, fairness-conditioned adapters, and monotonic neural additive models that balance fairness with performance.

Monotonic fairness is a family of fairness requirements in which predictions, fairness violations, or acceptable allocations are required to move in a consistent direction when a legitimate feature, a fairness-control parameter, an entitlement, or an auditing signal changes. Across the literature, the term covers at least five distinct constructions: resentment-free monotone classifiers over qualification features; slack-consistent fair learning paths; user-wise monotonic fairness control in recommendation; monotone aggregation of individual-fairness feedback from multiple auditors; and entitlement monotonicity in allocation. This suggests that monotonic fairness is best understood as a structural fairness principle rather than a single metric (Cole et al., 2019, Nachum et al., 2019, Chen et al., 28 Jan 2026, Bechavod, 2024, Feige, 2023).

1. Qualification monotonicity and resentment-free classification

A foundational formulation defines monotonic fairness through two forms of individual resentment. Protected attribute resentment occurs when an individual would receive a better outcome if only the protected attribute changed: $$\exists~A' \in \mathcal{A} \quad \text{s.t.}\quad f(X_u, A_u) < f(X_u, A').$$ Non-protected attribute resentment occurs when an individual with objectively better qualifications receives a worse outcome than someone with objectively worse qualifications: $f(X_u, A_u) < f(X', A').$ A function $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$ is monotonically fair if no possible individual $(X,A)$ experiences class resentment or score resentment. The construction assumes a partition $X=(X^+,X^\circ)$, where $X^+$ contains “qualification dimensions” endowed with a partial order $\preceq_Q$, and requires

$$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$

For non-increasing qualification dimensions, the ordering is reversed accordingly (Cole et al., 2019).

In this formulation, class resentment is eliminated when the predictor does not use the protected attribute, $f(x,a)=g(x)$, and score resentment is eliminated by monotonicity in the qualification features. The same work combines monotonicity with demographic balance through a penalized objective

$$\mathcal{L}(\theta;\alpha,\mathcal{M}) = \alpha\,\mathcal{L}_F(\theta) + (1-\alpha)\,\mathcal{L}_P(\theta),$$

where $f(X_u, A_u) < f(X', A').$0 is a demographic-parity penalty based on the difference in average predictions between groups. Monotonicity is enforced in a feedforward network by sign constraints on weights, with an offset ELU transformation

$f(X_u, A_u) < f(X', A').$1

and with positive first-layer weights for monotone non-decreasing features and negative first-layer weights for monotone non-increasing features. Empirically, the resulting monotone neural network achieves demographically balanced classifiers with zero resentment by construction, while maintaining accuracy close to unconstrained baselines (Cole et al., 2019).

2. Monotonicity along fairness-control and fairness-slack paths

A second line of work studies monotonicity with respect to the allowed fairness slack. In constrained fair learning, slack-consistency requires that, for any individual $f(X_u, A_u) < f(X', A').$2, the predictions vary monotonically as the allowable violation $f(X_u, A_u) < f(X', A').$3 changes. Formally, for any $f(X_u, A_u) < f(X', A').$4,

$f(X_u, A_u) < f(X', A').$5

The paper shows that standard constrained optimization and post-processing methods for Demographic Parity and Equal Opportunity violate this property, even with exhaustive threshold search. It then proves that post-processing with a group-aware Bayes-optimal score and randomized thresholds yields slack-consistent solutions for DP and EOp (Nachum et al., 2019).

A third construction treats fairness itself as a controllable post-training dimension. In recommendation, Cofair introduces $f(X_u, A_u) < f(X', A').$6 discrete fairness levels $f(X_u, A_u) < f(X', A').$7, a shared representation layer, fairness-conditioned adapters, and a user-level regularizer that enforces monotonic fairness improvement across levels. For user $f(X_u, A_u) < f(X', A').$8, the per-user adversarial fairness loss is

$f(X_u, A_u) < f(X', A').$9

and the monotonicity penalty is

$$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$0

The paper states that minimizing this term ensures that when the fairness level increases from $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$1 to $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$2, the per-user adversarial fairness loss does not increase. It also proves that the adversarial fairness objective upper bounds the demographic parity difference

$$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$3

and reports that a single training run can expose an entire fairness–accuracy curve by selecting $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$4 at inference time. The experiments on MovieLens-1M and Lastfm-360K show that fairness–accuracy curves are most Pareto-efficient in 15 of 16 comparisons, and that the framework needs far fewer epochs, approximately $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$5 of the baselines, to produce multiple fairness levels (Chen et al., 28 Jan 2026).

3. Monotone aggregation in online individual fairness

In online individual fairness, monotonicity appears in the aggregation of fairness reports from multiple auditors. An auditor $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$6 identifies an $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$7-violation when

$$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$8

where $$f:\mathcal{X}\times\mathcal{A}\rightarrow\mathbb{R}$$9 is the auditor-specific similarity function. The aggregation rule is called monotone if its entrywise Boolean combiner $(X,A)$0 is monotone under coordinatewise dominance: $(X,A)$1 OR, thresholded majority, weighted thresholds with nonnegative weights, and veto rules are all monotone in this sense (Bechavod, 2024).

The key structural result is a pivot-auditor characterization: for any ordered pair of individuals and any monotone aggregation function, there exists an instance-specific single auditor whose report is exactly the aggregated report for that pair. This reduces multi-auditor auditing to single-auditor analysis. Using this reduction, the paper gives oracle-efficient algorithms with improved regret–fairness frontiers. In the full-information setting, the algorithm achieves

$(X,A)$2

for regret and number of fairness violations. In the partial-information setting with labels observed only for positively predicted individuals, it achieves

$(X,A)$3

The corresponding oracle complexity is reduced to $(X,A)$4 per round in full information and $(X,A)$5 in partial information (Bechavod, 2024).

4. Monotonic neural and additive models

In regulated machine learning, monotonic fairness is often operationalized as monotonicity with respect to risk-relevant features and, more strongly, pairwise monotonicity across comparable features. One transparent formulation distinguishes individual monotonicity, weak pairwise monotonicity, and strong pairwise monotonicity. For a differentiable model, individual monotonicity on feature $(X,A)$6 is certified by

$(X,A)$7

Weak pairwise monotonicity for a more important feature $(X,A)$8 over $(X,A)$9 is verified by

$X=(X^+,X^\circ)$0

while strong pairwise monotonicity requires the same inequality for all $X=(X^+,X^\circ)$1. In additive models, strong pairwise monotonicity can be incompatible with diminishing marginal effects, motivating grouped components rather than purely univariate ones (Chen et al., 2023).

This perspective underlies monotonic neural additive models and monotonic groves of neural additive models. In credit scoring, the MNAM objective is

$X=(X^+,X^\circ)$2

with derivative penalties for individual monotonicity and pairwise monotonicity. Empirically, MNAMs match black-box fully connected neural networks on Taiwan and Give Me Some Credit while correcting non-monotone behavior on delinquency features (Chen et al., 2022). A broader empirical study reports analogous corrections in criminology, education, health care, and finance, arguing that violations can produce “catastrophic consequences” and that MNAMs impose both individual and pairwise monotonicity without sacrificing accuracy (Chen et al., 2023).

Several architectural mechanisms provide stronger guarantees. Smooth min-max networks replace hard min and max by scaled log-sum-exp,

$X=(X^+,X^\circ)$3

yielding strictly positive softmax and softmin gradients while preserving monotonicity under nonnegative weights on constrained features (Igel, 2023). Certified monotonic neural networks formulate global monotonicity as a counterexample MILP,

$X=(X^+,X^\circ)$4

and certify monotonicity when the optimum is nonpositive (Liu et al., 2020). Robust and provably monotonic networks instead use a monotonic residual,

$X=(X^+,X^\circ)$5

with $X=(X^+,X^\circ)$6 constrained to be $X=(X^+,X^\circ)$7, so that $X=(X^+,X^\circ)$8 for $X=(X^+,X^\circ)$9 follows from the Lipschitz bound (Kitouni et al., 2021).

5. Resource allocation, optimization, and online division

In multi-agent resource allocation, monotonic fairness can be expressed as monotone tuning of a fairness surrogate. GIFF defines a local fairness gain $X^+$0, a counterfactual advantage correction, and a fairness-augmented utility

$X^+$1

With $X^+$2, the allocation maximizes $X^+$3, and the paper proves that for $X^+$4,

$X^+$5

For $X^+$6-fairness, the surrogate equals the realized fairness improvement exactly; for generalized Gini, negative variance, and max–min fairness, it is a certified lower bound (Kumar et al., 30 Oct 2025).

In fair monotone $X^+$7-submodular maximization, monotonic fairness means maintaining extendability to lower and upper quotas at every step. A partial assignment is extendable iff

$X^+$8

Under this invariant, Fair-Greedy achieves a $X^+$9-approximation in $\preceq_Q$0 value-oracle calls, and Fair-Threshold achieves a $\preceq_Q$1-approximation with $\preceq_Q$2 evaluations. The experiments on influence maximization with $\preceq_Q$3 topics and sensor placement with $\preceq_Q$4 types show that fairness constraints do not significantly undermine solution quality (Zhu et al., 2024).

A related network-intervention formulation defines a DP-aware objective for influence blocking while preserving an approximately monotonic submodular structure. The scalarized objective

$\preceq_Q$5

is $\preceq_Q$6-approximately monotonic submodular, and greedy optimization yields

$\preceq_Q$7

The resulting CELF-R algorithm supports efficient Pareto-front construction over fairness–effectiveness trade-offs (Fang et al., 30 Jan 2026).

Online fair division with monotone utilities over bundles exposes a different role for monotonicity. In that model, utilities satisfy $\preceq_Q$8 whenever $\preceq_Q$9, yet the paper proves that no non-wasteful online mechanism is strategy-proof, envy-free, EF1, EFX, or Pareto efficient in full generality. On restricted identical-utility domains, MinLike is ex ante envy-free and, with non-zero marginals, EF1 ex post, while MinUtil is EF1 for identical monotone utilities (Aleksandrov et al., 2020). In a network-systems interpretation, a flow-fair MAC similarly uses “monotonic fairness” to avoid non-monotonic saturation caused by node-based fairness; the proposed MAC provides monotonic saturation and fairness among flows rather than nodes (Cloud et al., 2011).

6. Entitlement monotonicity, paradoxes, and limits

A distinct literature studies monotonic fairness under changing entitlements. Let $$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$0 mean that agent $$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$1's entitlement rises while all others weakly fall. Global monotonicity requires a deterministic selection rule consistent with the fairness notion such that $$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$2 does not decrease when moving from $$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$3 to $$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$4. Individual monotonicity is a set-valued analogue requiring upper and lower monotonicity across all acceptable allocations. The strongest failure mode is the inversion paradox, under which every acceptable allocation after the entitlement increase gives agent $$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$5 strictly less value than every acceptable allocation before it. In the indivisible-additive setting, APS and Prop are upper monotone, while WMMS, WEF, MWNSW, and CE can all exhibit inversion. In the divisible homogeneous-good setting, APS and Prop are globally and individually monotone, CE is globally monotone and lower monotone, WMMS is lower monotone but not upper monotone, and MWNSW becomes globally and individually monotone when all valuations are concave (Feige, 2023).

These results delimit monotonicity rather than universalize it. In dynamic pricing and admission control, the fairness constraint itself can force a non-monotone optimal policy. The unfairness index is

$$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$6

with the hard requirement $$f(x,a) \le f(x',a) \quad\text{whenever}\quad x^\circ = x'^\circ,\; x \preceq_Q x'.$$7. The paper shows that under heterogeneous price sensitivities, monotonic surge pricing can disproportionately exclude price-elastic users, and that when fairness is imposed as a hard state constraint, the optimal revenue-maximizing pricing policy is generally non-monotonic in demand. Under heavy load, lowering price may be necessary to maintain equitable access (Chen et al., 18 Mar 2026).

Monotonic fairness is therefore not synonymous with demographic parity, equalized odds, or any other group-parity condition. One line of work states explicitly that monotonicity “does not guarantee group fairness metrics,” and another frames monotonicity as a fairness-by-design constraint that is complementary to rather than a substitute for statistical parity notions (Chen et al., 2023, Chen et al., 2022). This suggests a general conclusion: monotonic fairness is most useful when the domain supplies a credible order relation—on qualifications, entitlements, slack parameters, or fairness levels—and least informative when no such order is normatively stable.