Nash Social Welfare Overview

• Nash Social Welfare is defined as the geometric mean of agents’ utilities and serves as a fairness measure in allocations across diverse settings.
• It is maximized using techniques like convex programming and stable polynomials, yielding various approximation guarantees based on utility and valuation classes.
• NSW extends to market mechanisms and online allocation, influencing strategic behavior, load balancing, and reinforcement learning with strong fairness implications.

Searching arXiv for recent and foundational papers on Nash Social Welfare to support the article. Nash social welfare (NSW) is the geometric mean of agents’ utilities. In the standard allocation setting with utilities $u_1,\dots,u_n$, it is

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

and, because the $n$th root is monotone, maximizing NSW is equivalent to maximizing the Nash product $\prod_i u_i$ or, after taking logs, maximizing $\sum_i \log u_i$ (Anari et al., 2016, Gokhale et al., 2024). Across fair division, market design, matching, online allocation, load balancing, and multi-stakeholder reinforcement learning, NSW is used as a balance between fairness and efficiency; in market settings it is maximized by solution concepts such as Nash bargaining and the competitive equilibrium with equal incomes (Brânzei et al., 2016, Mandal et al., 2022).

1. Formal objective and canonical variants

In one-sided fair division, the basic problem is to allocate goods to agents so as to maximize the geometric mean of the agents’ valuations. For indivisible goods with additive valuations, if agent $i$ receives bundle $S_i$, then

$$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$

and the same objective is written in allocation-matrix form as $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$ under per-item feasibility constraints (Anari et al., 2016, Li et al., 2021). In divisible-goods and Fisher-market formulations, the logarithmic form $\sum_i \log u_i(x)$ is especially natural (Banerjee et al., 2020).

The literature also uses weighted NSW. In Fisher markets with unequal budgets $\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$0, the weighted objective is

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

and the Fisher market equilibrium with unequal budgets maximizes this weighted NSW objective (Brânzei et al., 2016).

Two-sided models replace bundles of goods by many-to-one matchings. In the worker–firm model, if $\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$2 is a matching, then

$\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$3

where workers and firms both contribute utilities to the same geometric mean (Jain et al., 2023). Capacitated variants preserve the same objective while adding side constraints such as $\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$4 and $\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$5 (Gokhale et al., 2024).

A cost-minimization analogue also appears. In selfish and online load balancing, the objective is to minimize the geometric mean of clients’ costs; for unweighted games this is

$\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$6

and weighted versions exponentiate each client’s cost by its weight (Bilò et al., 2020). This usage preserves the multiplicative aggregation but reverses the optimization direction.

An axiomatic account is available in socially fair reinforcement learning. There, Nash welfare is

$\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$7

equivalently $\mathrm{NSW}=\left(\prod_{i=1}^n u_i\right)^{1/n},$8, and it is the unique objective, up to monotone transformation, satisfying Pareto Optimality, Independence of Irrelevant Alternatives with Neutrality, Anonymity, and Continuity (Mandal et al., 2022).

2. Convex, market, and stable-polynomial formulations

A central algorithmic line formulates NSW maximization through convex programming and stable polynomials. For additive valuations over indivisible items, a convex relaxation introduced in “Nash Social Welfare, Matrix Permanent, and Stable Polynomials” (Anari et al., 2016) optimizes

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

subject to fractional assignment constraints. After solving the relaxation, each item is independently assigned to agent $n$0 with probability $n$1. The expected product of utilities is exactly the sum of square-free coefficients of the polynomial

$n$2

and a generalized Gurvits inequality lower-bounds this sum by an $n$3 factor times the relaxation value. Taking $n$4th roots yields a randomized polynomial-time $n$5-approximation for NSW (Anari et al., 2016).

The same paper’s main technical contribution is an extension of Gurvits’s lower bound from the coefficient of the fully multilinear monomial of a degree-$n$6 homogeneous stable polynomial to the sum of square-free coefficients of an arbitrary degree-$n$7 homogeneous real stable polynomial with nonnegative coefficients (Anari et al., 2016). This creates a direct bridge from NSW rounding to real stability, weighted matchings, and permanent-style counting.

For separable, piecewise-linear concave utilities, “Nash Social Welfare for Indivisible Items under Separable, Piecewise-Linear Concave Utilities” (Anari et al., 2016) gives two constant-factor algorithms. The first is a polynomial-time factor-$n$8 approximation based on a new spending-restricted utility allocation market, and the second is a randomized stable-polynomial method achieving an $n$9-approximation in expectation (Anari et al., 2016). The market construction generalizes the spending-restricted market of Cole–Gkatzelis by pricing utility rather than raw quantity.

Budget-additive valuations lead to capped Fisher markets. “Satiation in Fisher Markets and Approximation of Nash Social Welfare” (Garg et al., 2017) studies linear Fisher markets with both buyers’ utility caps and sellers’ earning caps. In these markets, money clearing is sufficient for equilibrium existence, the equilibrium set can be non-convex, approximate equilibrium admits an FPTAS, the exact equilibrium problem lies in $\prod_i u_i$0, and a rounding framework gives a $\prod_i u_i$1-approximation for NSW with budget-additive valuations (Garg et al., 2017). The same paper also improves the approximation hardness for additive valuations to $\prod_i u_i$2 (Garg et al., 2017).

3. Valuation classes and the approximation landscape

Beyond additive preferences, the NSW literature is organized by valuation class, oracle model, and whether the guarantee is on an allocation or only on the optimal value.

Setting	Guarantee	Source
Additive valuations	$\prod_i u_i$3-approximation	(Anari et al., 2016)
SPLC utilities	factor $\prod_i u_i$4; $\prod_i u_i$5 in expectation	(Anari et al., 2016)
Budget-additive valuations	$\prod_i u_i$6-approximation	(Garg et al., 2017)
Submodular valuations	deterministic $\prod_i u_i$7-approximation	(Garg et al., 2022)
Coverage / matroid-rank classes	$\prod_i u_i$8-approximation of the optimal value	(Li et al., 2021)
Subadditive valuations with demand queries	polynomial-time constant-factor approximation	(Dobzinski et al., 2023)
XOS with demand and XOS oracles	$\prod_i u_i$9-approximation	(Barman et al., 2021)
One-sided capacities with submodular valuations	$\sum_i \log u_i$0-approximation	(Gokhale et al., 2024)

For submodular valuations, “Approximating Nash Social Welfare by Matching and Local Search” (Garg et al., 2022) gives a simple deterministic $\sum_i \log u_i$1-approximation in the symmetric case, plus an asymmetric guarantee of $\sum_i \log u_i$2 and a sharper appendix bound via $\sum_i \log u_i$3 (Garg et al., 2022). The same work shows that $\sum_i \log u_i$4-EFX can be attained simultaneously with a constant-factor approximation, yielding an $\sum_i \log u_i$5-approximation that is also $\sum_i \log u_i$6-EFX (Garg et al., 2022).

For several submodular subclasses, “Estimating the Nash Social Welfare for coverage and other submodular valuations” (Li et al., 2021) gives a $\sum_i \log u_i$7-approximation of the optimal NSW value for weighted matroid rank functions, the cone generated by matroid rank functions including coverage valuations, bipartite matching with a matroid constraint, and related extensions (Li et al., 2021). A key limitation is explicit: the result approximates the optimal value of the relaxation, but the paper does not know how to find an allocation of corresponding value in polynomial time in the general cases treated there (Li et al., 2021).

For subadditive valuations, “A constant factor approximation for Nash social welfare with subadditive valuations” (Dobzinski et al., 2023) gives the first polynomial-time constant-factor approximation under demand queries. Its template reduces NSW optimization to solving an Eisenberg–Gale-type configuration relaxation and invoking a utilitarian social-welfare rounding black box (Dobzinski et al., 2023). For XOS valuations, stronger oracle access changes the picture: “Sublinear Approximation Algorithm for Nash Social Welfare with XOS Valuations” (Barman et al., 2021) gives the first sublinear approximation under demand and XOS oracles, namely $\sum_i \log u_i$8, but also proves that any $\sum_i \log u_i$9-approximation requires exponentially many demand and XOS queries (Barman et al., 2021).

Structured subclasses admit sharper results. For identical additive valuations, “Greedy Algorithms for Maximizing Nash Social Welfare” (Barman et al., 2018) gives a simple $i$0-approximation, despite NP-hardness persisting in this setting, and exact polynomial-time algorithms for binary valuations and for valuations of the form $i$1 with concave $i$2 (Barman et al., 2018). “An Additive Approximation Scheme for the Nash Social Welfare Maximization with Identical Additive Valuations” (Inoue et al., 2022) strengthens the approximation notion: for any $i$3, it returns an allocation with

$i$4

in time $i$5 (Inoue et al., 2022). For 2-value additive valuations, “Nash Social Welfare for 2-value Instances” (Akrami et al., 2021) gives an exact polynomial-time algorithm for integrally 2-valued instances (Akrami et al., 2021).

Binary-marginal complement-free valuations exhibit a sharp frontier. “Approximating Nash Social Welfare under Binary XOS and Binary Subadditive Valuations” (Barman et al., 2021) gives a polynomial-time $i$6-approximation for binary XOS valuations in the value-oracle model, proves APX-hardness even for identical binary XOS valuations, and shows that under binary subadditive valuations exponentially many value queries are necessary to obtain even a sub-linear approximation (Barman et al., 2021).

Coverage problems form another specialized line. “Nash Welfare Guarantees for Fair and Efficient Coverage” (Barman et al., 2022) studies smoothed coverage valuations $i$7 and gives a polynomial-time $i$8-approximation assuming black-box access to an FPTAS for weighted feasibility over the underlying combinatorial constraints; it also proves APX-hardness and shows that without the $i$9 smoothing there is no nontrivial multiplicative approximation unless $S_i$0 (Barman et al., 2022).

4. Strategic implementation and market mechanisms

A distinct literature studies how well market mechanisms implement NSW when agents are strategic. In Fisher markets, truthful equilibrium maximizes weighted NSW, but strategic behavior can degrade the outcome. “Nash Social Welfare Approximation for Strategic Agents” (Brânzei et al., 2016) analyzes the price of anarchy of the Fisher market mechanism and the Trading Post mechanism.

For additive valuations, the Fisher market mechanism has price of anarchy at most $S_i$1 and no better than $S_i$2 (Brânzei et al., 2016). For Leontief valuations, its price of anarchy is $S_i$3, and the bound is tight (Brânzei et al., 2016). This contrast isolates a basic structural distinction between substitutes and complements.

Trading Post behaves differently. It is an indirect mechanism in which agents directly bid their budgets across goods and receive proportional shares. Under additive valuations it also has price of anarchy at most $S_i$4 and at least $S_i$5, but the upper bound extends to all concave, non-decreasing valuations (Brânzei et al., 2016). For Leontief utilities, a parameterized variant $S_i$6 with a small entrance fee achieves price of anarchy $S_i$7 for every $S_i$8 provided

$S_i$9

(Brânzei et al., 2016). The same paper shows that all Nash equilibria of Trading Post are pure in the regimes analyzed there and that equilibria satisfy proportionality, with approximate proportionality for $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$0 (Brânzei et al., 2016).

These results place NSW at the center of implementation and price-of-anarchy analysis rather than only offline optimization. They also show that mechanism design conclusions can depend sharply on valuation geometry: Fisher is constant-factor good for perfect substitutes and linearly bad for perfect complements, whereas Trading Post retains a constant or near-optimal bound in the cases studied in (Brânzei et al., 2016).

5. Online, load-balancing, and learning formulations

In online divisible-goods allocation, NSW becomes an adversarial online optimization problem. “Online Nash Social Welfare Maximization with Predictions” (Banerjee et al., 2020) considers $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$1 agents and $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$2 rounds, with one divisible item arriving per round and adversarially chosen values revealed on arrival. Without predictions, no online algorithm can beat the trivial linear competitive ratio; with predictions of each agent’s monopolist utility $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$3, the Set-Aside Greedy algorithm achieves a competitive ratio

$$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$4

when predictions are perfectly accurate, and its performance degrades smoothly with multiplicative prediction errors (Banerjee et al., 2020). The paper also proves that no online algorithm, even with exact monopolist-utility predictions, can achieve $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$5 or $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$6 competitive ratio for any constant $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$7 (Banerjee et al., 2020).

Load balancing uses the geometric mean of costs rather than utilities. “Nash Social Welfare in Selfish and Online Load Balancing” (Bilò et al., 2020) proves tight bounds on the price of anarchy of pure Nash equilibria and on the competitive ratio of the greedy online algorithm under broad classes of latency functions. For polynomial latencies of maximum degree $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$8, the exact weighted and unweighted price of anarchy is $$\operatorname{NSW}(S_1,\ldots,S_n)=\left(\prod_{i=1}^{n} v_i(S_i)\right)^{1/n},$$9, the non-atomic price of anarchy is $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$0, and the greedy online strategy has competitive ratio $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$1; moreover, $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$2 is a lower bound for every online algorithm, so greedy is optimal in this class (Bilò et al., 2020).

In socially fair reinforcement learning, “Socially Fair Reinforcement Learning” (Mandal et al., 2022) studies an episodic MDP with multiple reward functions and compares minimum welfare, generalized Gini welfare, and Nash welfare. For NSW, it gives a regret upper bound

$\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$3

and a lower bound

$\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$4

showing an intrinsic exponential dependence on the number of agents $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$5 (Mandal et al., 2022). A separate line on equilibrium selection in specification-guided multi-agent RL ranks candidate equilibria by the arithmetic mean of satisfaction probabilities,

$\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$6

which is explicitly not the geometric-mean NSW objective (Jothimurugan et al., 2022). This distinction matters because “social welfare” terminology in learning papers is not always synonymous with Nash welfare.

6. Two-sided preferences, capacities, and fairness consequences

Two-sided NSW under preferences is substantially harder than one-sided allocation. “Maximizing Nash Social Welfare under Two-Sided Preferences” (Jain et al., 2023) studies many-to-one worker–firm matchings and proves that computing a Nash-optimal matching is NP-complete even when every firm has capacity $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$7, all valuations lie in $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$8, and every agent positively values at most three other agents (Jain et al., 2023). If all firms have unit capacity, however, the problem reduces to maximum-weight matching with edge weight $\nu(x)=\left(\prod_i v_i(x_i)\right)^{1/n}$9 and is polynomial-time solvable (Jain et al., 2023). The same work also gives a $\sum_i \log u_i(x)$0-approximation for positive valuations, a QPTAS for a constant number of firms, fixed-parameter algorithms in the number of workers, and polynomial-time algorithms for restricted domains such as symmetric binary valuations and several bounded-degree cases (Jain et al., 2023).

Capacities in one-sided and two-sided models change the approximation landscape again. “Approximating One-Sided and Two-Sided Nash Social Welfare With Capacities” (Gokhale et al., 2024) gives, for any $\sum_i \log u_i(x)$1, a $\sum_i \log u_i(x)$2-approximation for one-sided NSW with submodular valuations and capacities, a $\sum_i \log u_i(x)$3-approximation for two-sided NSW with subadditive valuations, and an $\sum_i \log u_i(x)$4-approximation for weighted two-sided additive NSW under capacity constraints (Gokhale et al., 2024). The paper also proves hardness-of-approximation results and emphasizes a computational separation between Nash welfare and utilitarian welfare (Gokhale et al., 2024).

Maximizers of NSW are also studied through their fairness properties under constraints. In the budget-feasible allocation problem with item costs and agents’ budgets, “Budget-feasible Maximum Nash Social Welfare Allocation is Almost Envy-free” (Wu et al., 2020) shows that every budget-feasible Max-NSW allocation is Pareto optimal and $\sum_i \log u_i(x)$5-EF1, and that the factor $\sum_i \log u_i(x)$6 is tight (Wu et al., 2020). When items are small relative to budgets, the guarantee improves to

$\sum_i \log u_i(x)$7

where $\sum_i \log u_i(x)$8, and the asymptotic limit $\sum_i \log u_i(x)$9 is also tight (Wu et al., 2020). This clarifies a common misconception: NSW maximization continues to imply strong approximate fairness under budgets, but not the exact EF1 guarantee known in the unconstrained setting.

Taken together, these results depict NSW as a unifying but highly model-sensitive objective. Its logarithmic and multiplicative structure enables convex relaxations, market-equilibrium interpretations, and axiomatic characterizations, yet its algorithmic behavior depends acutely on valuation class, oracle access, strategic behavior, side constraints, and whether preferences are one-sided or two-sided (Anari et al., 2016, Brânzei et al., 2016, Dobzinski et al., 2023, Jain et al., 2023).