Rawlsian (Egalitarian) Objective

• Rawlsian (egalitarian) objective is a framework that prioritizes outcomes for the least advantaged using maximin principles and lexicographic fairness.
• It underpins a range of applications from algorithmic fairness in machine learning to equitable social policy and institutional design.
• The approach balances ethical mandates with computational challenges, revealing trade-offs between fairness, efficiency, and feasibility.

Searching arXiv for recent and foundational papers on Rawlsian / egalitarian objectives. The Rawlsian (egalitarian) objective is a family of social, algorithmic, and optimization criteria that evaluate an outcome by its treatment of the least advantaged individual, group, or generation. In its canonical form, it is a maximin rule: choose the feasible outcome that maximizes the minimum utility, or equivalently minimizes the maximum disutility. In contemporary computational work, this idea appears in several technically distinct but normatively related forms: worst-case utility optimization under a veil of ignorance, lexicographic protection of lower tails, conditional-independence constraints encoding fair equality of opportunity, welfare functions with explicit inequality aversion, and dynamic rules that prioritize the worst-off over time (Cousins, 2024, Rigobon, 2023, Liu et al., 2021, Wu et al., 1 Mar 2025).

1. Normative basis and scope

A central Rawlsian claim in this literature is that fairness is not exhausted by equal treatment at a single decision point. One line of work states Rawls’s principle of fair equality of opportunity as the requirement that “those who are at the same level of talent and ability, and have the same willingness to use them, should have the same prospects of success regardless of their initial place in the social system” (Khan et al., 2022). Another extends the same fairness logic across time, stressing that “persons in different generations have duties and obligations to one another just as contemporaries do,” and treating the “problem of justice between generations” as central to ethical theory (Zandi, 2021).

This normative basis broadens the Rawlsian objective beyond a single scalar formula. In AI governance, Rawlsian egalitarianism is applied to the “basic structure of society,” understood as a “composite of sociotechnical systems,” so that AI systems must be publicly justifiable, compatible with basic liberties, and oriented toward substantively fair outcomes for the least advantaged (Gabriel, 2021). In algorithmic fairness, the same distinction appears as the contrast between formal equal opportunity at a contest and substantive equal opportunity over a lifetime: Rawlsian fairness is forward-looking and institutional, not merely procedural (Khan et al., 2022).

A recurring misconception addressed by these papers is that Rawlsian egalitarianism requires simple equality of outcomes. The papers do not define it that way. They instead define it by priority to the least advantaged, by equality of prospects for equally talented persons, or by lexicographic protection against very bad outcomes. This is why Rawlsian objectives can differ from envy-freeness, calibration, predictive parity, or stability while still being recognizably egalitarian (Khan et al., 2022, Demeulemeester et al., 2022, Nana et al., 4 Nov 2025).

2. Canonical mathematical forms

The most direct formalization is maximin welfare. In the adversarial-game formulation of the veil of ignorance, if a world induces sentiment vector $s \in \mathbb{R}^g$, the Rawlsian objective is

$$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$

with the dual malfare formulation

$$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$

This makes the veil of ignorance a maximin or robust-optimization problem (Cousins, 2024).

In supervised learning, the same idea appears as minimax loss: $$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$ interpreted as the machine-learning analogue of protecting the least advantaged observation, namely the one with the largest loss (Rigobon, 2023). That paper also studies the smooth interpolation

$$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$

with

$$\lim_{\lambda\to 0}L(\theta;\lambda)=\frac{1}{n}\sum_i \ell(f_\theta(x_i),y_i), \qquad \lim_{\lambda\to\infty}L(\theta;\lambda)=\max_i \ell(f_\theta(x_i),y_i).$$

This yields a utilitarian–Rawlsian continuum rather than a binary opposition (Rigobon, 2023).

A related welfare-theoretic approach evaluates benefit vectors $b=(b_1,\dots,b_n)$ through cardinal social welfare functions

$$W_\alpha(b_1,\dots,b_n)=\sum_{i=1}^n w_\alpha(b_i),$$

with the practically emphasized regime $0<\alpha<1$, where the function is inequality-averse and satisfies the Pigou–Dalton transfer principle. That paper notes that the limit $\alpha\to -\infty$ corresponds to leximin or Rawlsian max-min welfare (Heidari et al., 2018).

Several papers refine maximin lexicographically. In graph orientations, the strongest egalitarian objective is lexicographic minimization of the indegree sequence sorted in non-increasing order, so that one first minimizes the maximum indegree, then the number of vertices at that maximum, then the second-largest indegree, and so on (Borradaile et al., 2012). In random assignment, the Rawlsian rule lexicographically minimizes the cumulative probabilities of receiving the worst objects, rank block by rank block, starting from the least preferred object (Demeulemeester et al., 2022). These lexicographic variants are stricter than plain maximin and are designed to protect the lower tail of outcomes rather than only a single worst coordinate.

3. Opportunity, institutions, and time

Not all Rawlsian objectives are immediate maximin problems over contemporaneous utilities. In the fair-equality-of-opportunity literature, the Rawlsian target is a structural condition on prospects. RAWLSNET encodes this as the conditional-independence requirement

$$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$0

equivalently

$$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$1

where $$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$2 are justified variables, $$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$3 are sensitive variables, $$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$4 is a controllable pre-hiring decision, and $$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$5 is the downstream advantageous social position (Liu et al., 2021). The method edits the CPT of $$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$6 in a Bayesian Network so that Rawlsian FEO is satisfied when possible, and otherwise minimizes the squared difference between the two sides of the fairness equations (Liu et al., 2021).

This FEO perspective is explicitly distinguished from formal equal opportunity. Formal EO is described as fair treatment at the point of contest; Rawlsian FEO instead asks whether people had fair opportunities to develop talent and whether equally talented people have equal future prospects after the decision (Khan et al., 2022). A plausible implication is that some Rawlsian objectives are not best understood as outcome equalization at all, but as the design of institutions that shape future opportunity sets.

The intergenerational literature extends the Rawlsian objective along a temporal axis. Rawls’s “just savings principle” is presented as “an understanding between generations to carry their fair share of the burden of realizing and preserving a just society,” but the usual interpretation of “saving” is criticized as too narrow if it means merely withholding resources (Zandi, 2021). Following Clark Wolf’s analogy between environmental harms and financial debt, this line of work argues that “saving” should include investment on behalf of future generations, and then extends that idea further: resources should be “fully use[d]” to produce sustainable and resilient systems (Zandi, 2021).

The resulting thesis is not passive preservation but productive transformation. Sustainable systems are defined as systems that preserve original resources and transfer them to the next generation “along with new achievements and outcomes,” while resilience is defined as the capacity to overcome natural disasters and repair or reform itself back to its previous condition (Zandi, 2021). This suggests a Rawlsian objective in which intergenerational fairness is realized by durable, adaptive institutions rather than by static stockpiling.

4. Optimization in machine learning, reinforcement learning, and language generation

In fair machine learning, Rawlsian objectives are frequently recast as robust optimization. The robust-fairness framework based on a zero-sum game between a Dæmon and an Angel generalizes the full Rawlsian case by restricting the Angel’s feasible distributions $$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$7, yielding objectives of the form

$$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$8

Special cases recover utilitarianism, weighted utilitarian-maximin social welfare, generalized Gini objectives, and power-mean families, all interpreted as partial-ignorance or partial-robustness variants of the Rawlsian game (Cousins, 2024). The same paper reduces optimization to saddle-point problems and shows that standard maximin methods can optimize these objectives under mild convexity and compactness conditions (Cousins, 2024).

The supervised-learning relaxation based on $$\arg\max_{s\in S}\min_{i\in\{1,\dots,g\}} s_i \;=\; \arg\max_{s\in S} M_{-\infty}(s),$$9 emphasizes a different technical trade-off. Because the objective is differentiable, it is easier to optimize than the non-smooth max-loss objective; its gradient is a loss-weighted average of individual gradients, so increasing $$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$0 concentrates optimization pressure on high-loss points (Rigobon, 2023). The paper also proves that the allowable step size and convergence rate deteriorate as $$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$1 increases, making more Rawlsian behavior computationally more expensive (Rigobon, 2023).

A third line operationalizes Rawlsian welfare through direct reward shaping. RAWL-E defines a Rawlsian ethical utility using maximin,

$$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$2

then compares successive minimum experiences $$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$3 and $$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$4 to generate a self-directed sanction $$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$5, with shaped reward

$$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$6

The DQN therefore keeps its base goal-directed objective while adding an ethical bias toward actions that improve the welfare of the least advantaged agent (Woodgate et al., 2024). In the reported harvesting scenarios, RAWL-E yields higher minimum experience, lower Gini index, higher social welfare, and slightly greater robustness than the non-Rawlsian baseline (Woodgate et al., 2024).

Large-language-model consensus generation introduces yet another form. In a token-level MDP, the egalitarian welfare of a complete statement $$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$7 is defined as

$$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$8

and the deterministic consensus problem is

$$\arg\min_{s\in S}\max_{i\in\{1,\dots,g\}} s_i \;=\; \arg\min_{s\in S} M_{\infty}(s).$$9

Finite lookahead and beam-search procedures then optimize worst-case agent alignment token by token, and the reported Egalitarian Perplexity results show lower worst-case perplexity than Best-of-$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$0 and Habermas-based baselines (Blair et al., 15 Oct 2025).

5. Allocative, combinatorial, and social-choice instantiations

Matching and assignment provide especially clear Rawlsian formulations. In many-to-one matching with non-linear college utilities, classical stable matchings may fail to exist, so the objective is replaced by

$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$1

that is, maximize the utility of the worst-off college (Nana et al., 4 Nov 2025). The proposed deterministic and stochastic greedy algorithms maintain feasibility and monotonically improve the minimum college utility, though they do not guarantee global optimality (Nana et al., 4 Nov 2025).

In two-sided matching markets with learning, the Rawlsian welfare of a stable matching $$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$2 is

$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$3

and the objective is to choose the stable matching that maximizes this minimum (Hosseini et al., 2024). The paper develops an epoch Explore-Then-Commit algorithm and analyzes regret and sample complexity in terms of the cross-side minimum preference gap $$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$4 (Hosseini et al., 2024).

Online bipartite matching generalizes Rawlsian max–min fairness to both sides of the market, with group and individual versions. Offline-side group fairness is the minimum average expected utility received by any offline group, and online-side group fairness is the minimum average expected utility accumulated over arrivals by any online group (Esmaeili et al., 2022). LP benchmarks and randomized probing algorithms then provide simultaneous competitive ratios for operator profit and the two Rawlsian side constraints (Esmaeili et al., 2022).

Random assignment without money yields a distinct ordinal formulation. A Rawlsian assignment lexicographically minimizes the sorted cumulative probabilities that agents receive their worst objects. The paper proves that every problem has a unique Rawlsian assignment, that the Rawlsian rule is anonymous and sd-efficient, and that it is not sd-strategyproof (Demeulemeester et al., 2022). This formulation is not standard utility maximin but a lexicographic minimization of lower-tail ordinal risk.

Clustering exhibits both early exploratory and more formal welfare-centric versions. The exploratory k-means paper defines individual utility as

$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$5

group utility as

$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$6

and treats the Rawlsian point as the clustering where the least-advantaged sensitive group has maximal utility (Simoes et al., 2022). The later welfare-centric formulation makes this explicit by defining group disutility

$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$7

and the Rawlsian objective

$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$8

combining distance and proportional-violation terms in a single max-min welfare criterion (Zhang et al., 14 Aug 2025).

Judgment aggregation and graph orientation show that Rawlsian ideas also govern purely combinatorial objects. In judgment aggregation, the maximin rule

$$\min_{\theta\in\Theta}\max_i \ell\!\left(f_\theta(x_i),y_i\right),$$9

minimizes the maximum Hamming-distance dissatisfaction, while the distinct equity rule

$$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$0

minimizes disparity in dissatisfaction (Botan et al., 2021). In graph orientation, the egalitarian objective is to share total indegree “as equally as allowed by the topology of the graph,” either by minimizing maximum indegree or by lexicographically minimizing the sorted indegree sequence (Borradaile et al., 2012).

6. Trade-offs, dynamic effects, and limitations

A major theme is that Rawlsian and utilitarian objectives are often different but not uniformly opposed. The long-run welfare-dynamics model compares Rawlsian policies that help the worst-off, such as $$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$1 and $$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$2, with utilitarian policies such as $$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$3, $$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$4, and $$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$5 (Wu et al., 1 Mar 2025). Under the survival condition

$$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$6

the paper proves that the Rawlsian asymptotic average welfare growth rate weakly dominates the utilitarian one almost surely; under the complementary ruin condition, the comparison reverses (Wu et al., 1 Mar 2025). This directly challenges the conventional assumption that prioritizing the worst-off must reduce average welfare in the long run.

Other papers formalize endogenous movement between Rawlsian and utilitarian criteria. The self-referential “fan” welfare functions define welfare as the fixed point

$$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$7

so that the admissible set of weights depends on the welfare level itself (Echenique et al., 29 Jan 2026). In the triage-oriented monotone increasing fan, low welfare leads to a criterion closer to utilitarianism, while high welfare expands the weight set toward the simplex and approaches Rawlsian maximin (Echenique et al., 29 Jan 2026).

The literature is equally explicit about limitations. Some Rawlsian rules are computationally hard: lexicographic indegree minimization becomes NP-hard under acyclicity constraints in graph orientation (Borradaile et al., 2012); outcome determination for the MaxEq judgment-aggregation rule is $$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$8-complete (Botan et al., 2021). Some sacrifice standard desiderata: the Rawlsian assignment rule is not sd-strategyproof and not sd-envy-free (Demeulemeester et al., 2022); Rawlsian many-to-one matching abandons stability when stability may not exist (Nana et al., 4 Nov 2025). In fair ML, stronger Rawlsianity can increase algorithmic cost or complicate statistical consistency (Rigobon, 2023). In clustering, the choice between Rawlsian and utilitarian welfare, and the choice of $$L(\theta;\lambda)=\frac{1}{\lambda}\log\!\left(\frac{1}{n}\sum_i e^{\lambda \ell(f_\theta(x_i),y_i)}\right),$$9, are explicitly normative rather than merely technical (Zhang et al., 14 Aug 2025).

Taken together, these results characterize the Rawlsian objective not as a single metric but as a principled family of egalitarian orderings. Its canonical operation is to protect the worst-off; its stronger versions protect the lower tail lexicographically; its substantive versions govern opportunity structures and intergenerational duties; and its algorithmic versions translate those commitments into optimization programs, LPs, saddle-point problems, reward-shaping rules, and search procedures. Across domains, the defining question remains the same: which feasible design best secures the position, prospects, or welfare of those who would otherwise fare worst (Cousins, 2024, Khan et al., 2022, Zandi, 2021, Gabriel, 2021).