Maximum Nash Welfare Rule

• Maximum Nash Welfare Rule is an allocation approach that maximizes the geometric mean of agents’ utilities, ensuring fairness by balancing efficiency and equity.
• Its logarithmic formulation transforms multiplicative objectives into separable sums, facilitating analysis in diverse settings such as indivisible goods and market equilibria.
• The rule extends to weighted, budget-constrained, and two-sided models, supporting EF1 fairness guarantees and competitive equilibrium approximations in theory and practice.

Searching arXiv for recent and foundational papers on Maximum Nash Welfare to ground the article and confirm citations. The Maximum Nash Welfare Rule, often abbreviated MNW, is an allocation rule that selects an allocation maximizing the Nash social welfare: the product, or equivalently the geometric mean, of the agents’ utilities. In indivisible-goods settings with additive valuations, divisible-resource settings with concave utilities, and several weighted or constrained variants, it functions as a canonical fairness–efficiency aggregator because it is scale-invariant and because its logarithmic form converts multiplicative welfare into a separable sum of log-utilities. The rule appears in fair division, market design, online allocation, matching, and coalition formation, but its exact structural properties depend sharply on the valuation domain, feasibility constraints, and whether utilities are one-sided or two-sided (Suksompong, 2022, Garg et al., 2024).

1. Formal objective and rule variants

In the standard indivisible-goods model, an allocation $A=(A_1,\dots,A_n)$ assigns disjoint bundles of goods to agents with utilities $u_i(A_i)$. The Nash social welfare is

$NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$

and maximizing $NSW(A)$ is equivalent to maximizing $\prod_i u_i(A_i)$, or, when all utilities are positive, $\sum_i \log u_i(A_i)$ (Caragiannis et al., 2019, Suksompong, 2022). In divisible settings with budgets or weights, the weighted form is

$NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$

equivalently $\sum_i w_i \log u_i(x_i)$, with the Eisenberg–Gale program as the canonical convex formulation when utilities are concave (Garg et al., 2024).

Several domain-specific conventions modify the basic objective without altering its conceptual role. Under binary additive valuations, zero utilities are handled by first maximizing the number of agents with positive utility and then maximizing the product over positive utilities; this avoids the $\log 0$ singularity and yields the version of deterministic MNW used in the binary-strategyproofness literature (Halpern et al., 2020, Brandl et al., 11 Jul 2026). Under weighted entitlements, the maximum weighted Nash welfare rule maximizes $\prod_i u_i(A_i)^{w_i}$, again with a two-stage definition in the binary domain to handle zeros (Suksompong et al., 2022). In budget-feasible allocation, costs constrain feasibility but do not enter utility, so MNW maximizes $u_i(A_i)$0 subject to per-agent budget constraints on bundle cost (Wu et al., 2020).

Viewed as a welfarist rule, MNW is the additive welfarist rule generated by $u_i(A_i)$1. This representation is central to several characterization theorems: the logarithm is not merely a convenient transform, but the unique welfare function within broad classes of additive or general welfarist rules that is compatible with EF1 guarantees under the relevant assumptions (Suksompong, 2022, Yuen et al., 2023). Across these formulations, scale invariance is preserved: multiplying one agent’s utilities by a positive constant does not change the argmax of the MNW objective (Caragiannis et al., 2019, Garg et al., 2024).

2. Indivisible goods: fairness, efficiency, and axiomatic position

For additive valuations over indivisible goods, MNW is a Pareto-optimal rule, and in the standard no-budget setting it is also envy-free up to one good (EF1). The cited characterization papers treat this EF1 property as the decisive fairness fact: among additive welfarist rules, MNW is the unique rule that guarantees EF1, and this uniqueness already holds for two agents (Suksompong, 2022, Yuen et al., 2023). A 2024 strengthening shows that the same uniqueness persists even on identical-good instances, two-value instances, and normalized instances with at least three agents, provided the welfare function is strictly increasing and continuous on $u_i(A_i)$2 (Celine et al., 2024).

In the binary additive domain, MNW has an unusually strong profile. Deterministic MNW with fixed lexicographic tie-breaking is EF1, Pareto optimal, group-strategyproof, and polynomial-time computable (Halpern et al., 2020). The same paper shows that fractional MNW coincides with fractional leximin, is ex ante envy-free and ex ante Pareto optimal, is ex ante group-strategyproof, and can be implemented as a lottery over deterministic MNW allocations, yielding ex post EF1 (Halpern et al., 2020). A later characterization sharpens the normative picture further: under binary valuations, the common rule induced by MNW, leximin, and all additive welfarist rules with a strictly concave objective is the only rule satisfying EF1, strategyproofness, neutrality, minimal completeness, and invariance under disapproving unassigned goods; for two agents, an alternative characterization replaces the last axiom with non-redundancy and resource-monotonicity (Brandl et al., 11 Jul 2026).

Weighted entitlements preserve part of this structure in the binary case. The maximum weighted Nash welfare rule is resource-monotone, population-monotone, group-strategyproof, and polynomial-time implementable under binary valuations, provided a specific tie-breaking convention is used (Suksompong et al., 2022). The broader matroid-rank generalization goes beyond logarithmic welfare: for matroid-rank valuations, weighted additive welfarist rules with concave $u_i(A_i)$3, including MWNW as the case $u_i(A_i)$4, are resource-monotone, population-monotone, and group-strategyproof, while weight-monotonicity holds for any strictly increasing $u_i(A_i)$5 (Suksompong et al., 2023). This suggests that the distinctive position of MNW in additive settings is partly a property of the valuation domain rather than of the product objective alone.

3. Constrained indivisible allocation: budgets, donation, and approximate fairness

Budget constraints alter the fairness frontier. In the budget-feasible indivisible-goods model, each item has a cost, each agent has a budget, and an agent’s envy toward another bundle is evaluated only over subsets she can afford. In this model, an MNW allocation remains Pareto optimal, but exact EF1 can fail. The sharp guarantee is that any budget-feasible MNW allocation is $u_i(A_i)$6-EF1, and the factor $u_i(A_i)$7 is tight (Wu et al., 2020). If

$u_i(A_i)$8

then any MNW allocation is $u_i(A_i)$9-EF1; as $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$0, the guarantee converges to $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$1, and $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$2 is itself the limiting barrier in this model (Wu et al., 2020). The proofs use contradiction-by-NSW-improvement, partitioning arguments, and, in the large-budget regime, a density-based removal lemma together with a heavy–light decomposition (Wu et al., 2020).

A different constrained variant is donation. For additive valuations, there always exists an EFX allocation of a subset of goods whose Nash welfare is at least $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$3 times the optimum Nash welfare on the full set, hence at least $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$4 of optimum, and this bound is tight (Caragiannis et al., 2019). The construction starts from an MNW allocation, never reallocates items across agents, donates selected goods to charity, and returns an EFX allocation $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$5 with $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$6 for every agent $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$7, where $NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$8 is the initial MNW allocation (Caragiannis et al., 2019). Under a large-market assumption,

$NSW(A)=\left(\prod_{i=1}^n u_i(A_i)\right)^{1/n},$9

the guarantee improves to

$NSW(A)$0

which tends to optimality as $NSW(A)$1 (Caragiannis et al., 2019). If the starting point is only a $NSW(A)$2-approximate MNW allocation, the same framework still yields an EFX allocation within a factor $NSW(A)$3 of optimal Nash welfare (Caragiannis et al., 2019).

Taken together, these results show that under knapsack-style feasibility or stronger fairness desiderata such as EFX, MNW typically ceases to deliver exact envy guarantees but remains a robust anchor for approximation. A plausible implication is that the Nash product is unusually stable under controlled loss of feasibility or controlled disposal of goods, even when exact envy-freeness becomes unattainable.

4. Divisible goods, market equilibria, and incentive interpretations

For divisible goods with concave utilities, maximizing weighted Nash welfare is a convex program. If agent $NSW(A)$4 has budget $NSW(A)$5 and utility $NSW(A)$6, the Eisenberg–Gale formulation is

$NSW(A)$7

When utilities are homogeneous concave of degree one, this program yields a competitive equilibrium, and the dual variables are equilibrium prices (Garg et al., 2024). For non-homogeneous concave utilities, the equivalence breaks: maximizing Nash welfare remains convex optimization, but computing competitive equilibrium becomes PPAD-hard already for separable piecewise linear concave utilities (Garg et al., 2024).

The 2024 CE-approximation analysis introduces Gale demand, Gale prices, and Gale-substitutes. In full generality, any MNW allocation together with any dual Gale prices forms a 2-demand-approximate competitive equilibrium: each agent gets at least half the utility available from best response at those prices, and equal budgets imply 2-envy-freeness (Garg et al., 2024). For $NSW(A)$8-Gale-substitutes, the guarantee strengthens: relative to any competitive equilibrium $NSW(A)$9, any MNW allocation $\prod_i u_i(A_i)$0 satisfies

$\prod_i u_i(A_i)$1

Generalized network utilities, which include SPLC and Leontief-free utilities, are Gale-substitutes, and all separable concave utilities are $\prod_i u_i(A_i)$2-Gale-substitutes (Garg et al., 2024). Conversely, every competitive equilibrium achieves at least $\prod_i u_i(A_i)$3 of the maximum Nash welfare, and this factor is tight (Garg et al., 2024).

The divisible-resource literature also studies MNW as a mechanism. In cake cutting and homogeneous divisible items, the MNW mechanism has incentive ratio exactly $\prod_i u_i(A_i)$4; for cake cutting this remains true even without free disposal (Bei et al., 2023). The same paper contrasts MNW with the Partial Allocation mechanism, whose incentive ratio in cake cutting lies between $\prod_i u_i(A_i)$5 and $\prod_i u_i(A_i)$6, and develops interpolation mechanisms trading off incentive ratio $\prod_i u_i(A_i)$7 against an $\prod_i u_i(A_i)$8-approximation to Nash welfare (Bei et al., 2023). In linear-divisible allocation with diversity constraints, MNW is distinguished by robustness: if one agent adds a convex diversity constraint, then for the unconstrained agents it guarantees constant-factor no negative externality, with $\prod_i u_i(A_i)$9 in the one-agent case, and for the constraining agent it guarantees monotonicity with factor $\sum_i \log u_i(A_i)$0 (Shen et al., 2021). For Leontief utilities, linear-price equilibria coincide with MNW allocations, but price curves support a strictly larger set of outcomes: strictly increasing price curves support exactly the group-domination-free allocations, and in bandwidth allocation any CES-welfare-maximizing allocation can be supported by price curves (Goel et al., 2018).

5. Computation, approximability, and tractable fragments

In the classical additive indivisible-goods model, exact MNW computation is hard. Maximizing Nash welfare with additive valuations is APX-hard, and the problem remains NP-hard even for identical additive valuations (Wu et al., 2020, Garg et al., 2021). This hardness persists under severe restrictions: many-to-one two-sided matching with capacities is NP-hard even when every firm has capacity $\sum_i \log u_i(A_i)$1, all valuations lie in $\sum_i \log u_i(A_i)$2, and each agent has degree at most $\sum_i \log u_i(A_i)$3 in the positive-edge graph (Jain et al., 2023). In general ASHGs, approximating Nash welfare within a factor of $\sum_i \log u_i(A_i)$4 is NP-hard (Pagano et al., 18 May 2026).

Despite this, several nontrivial fragments are tractable. For asymmetric agents with identical additive valuations, there is a PTAS for MNW; for identical agents with constant $\sum_i \log u_i(A_i)$5-ary additive valuations, there is an exact polynomial-time algorithm; for 2-valuable monotone valuations, there is a strongly polynomial-time exact algorithm; and for a constant number of asymmetric additive agents, there is an FPTAS (Garg et al., 2021). In the binary additive setting, a greedy algorithm computes an exact Nash-optimal allocation in polynomial time, while for identical additive valuations a simple greedy algorithm achieves a $\sum_i \log u_i(A_i)$6-approximation (Barman et al., 2018). For 2-value additive valuations with $\sum_i \log u_i(A_i)$7, there is a polynomial-time exact algorithm based on heavy-item multi-matchings, augmenting paths, and parity control among “small” agents (Akrami et al., 2021). Under binary XOS valuations, there is a polynomial-time 288-approximation using only value queries, and the same allocation also provides constant-factor approximations to social welfare and GMMS; by contrast, for binary subadditive valuations, exponentially many value queries are required to obtain even a sub-linear approximation (Barman et al., 2021).

For divisible goods, the computational picture is substantially friendlier. The Eisenberg–Gale program computes MNW efficiently for any concave utilities (Garg et al., 2024). In online divisible allocation without predictions, logarithmic competitive ratios are achievable under bounded-ratio assumptions: Half-and-Half is $\sum_i \log u_i(A_i)$8-competitive for $\sum_i \log u_i(A_i)$9-balanced instances, Myopic Greedy is $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$0-competitive for binary $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$1-impartial instances, and a rounded-value variant gives $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$2 for general valuations (Huang et al., 2022). This suggests a broad computational dichotomy: indivisibility creates hard combinatorial structure, whereas divisibility typically preserves convexity and tractable duality.

6. Generalizations beyond one-sided allocation

The MNW objective has been extended beyond one-sided fair division into coalition formation and two-sided matching. In additively separable hedonic games, the Nash welfare of an individually rational partition $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$3 is

$NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$4

restricted to IR partitions because utilities may otherwise be negative (Pagano et al., 18 May 2026). The maximum Nash welfare rule in this setting is scale-invariant and guarantees contractual Nash stability in symmetric ASHGs for any finite-factor approximation of the optimum, provided the optimum Nash welfare is positive (Pagano et al., 18 May 2026). Exact optimization is NP-hard even for symmetric aversion-to-enemies games, but there are approximation algorithms with ratio $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$5 for AEGs and $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$6 for appreciation-of-friends games, and the tractability boundary is sharp when the number or size of coalitions is bounded (Pagano et al., 18 May 2026).

In two-sided many-to-one matching, workers and firms both have utilities, firms have capacities, and the Nash social welfare is the geometric mean over all worker and firm utilities under the matching (Jain et al., 2023). Here the objective is genuinely bilateral rather than a projection of one-sided fair division. Exact optimization is polynomial-time when all capacities are $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$7, via maximum-weight bipartite matching with edge weights $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$8, but NP-hard already at capacity $NW(x)=\prod_{i\in N} u_i(x_i)^{w_i},$9 under very restrictive valuations (Jain et al., 2023). The model nevertheless admits a submodular-reduction approximation, a QPTAS for a constant number of firms, and fixed-parameter algorithms in the number of workers (Jain et al., 2023).

These generalizations indicate that the Nash product is portable across disparate allocation structures, but the induced normative content changes with the feasible set. In one-sided indivisible allocation, the signature theorem is EF1. In divisible markets, it is the Eisenberg–Gale–CE connection and its approximate generalizations. In hedonic games, it is contractual stability. In two-sided matching, it becomes a bilateral fairness–efficiency trade-off under capacities. A plausible synthesis is that the Maximum Nash Welfare Rule is best understood not as a single theorem, but as a unifying objective whose fairness meaning is domain-specific and whose algorithmic behavior is governed by whether the underlying feasible region is convex, matroidal, matching-based, or knapsack-constrained.