New Algorithm Challenges Claim of Efficient Nucleolus Computation

A recent claim of a breakthrough in computing a fundamental solution concept in cooperative game theory has been called into question by new research. The nucleolus, a fairness concept used to allocate payoffs among players in games where cooperation is possible, has long been computationally challenging to determine efficiently. While a 2025 paper by Maggiorano et al. proposed a strongly polynomial-time combinatorial algorithm for convex games, a new assessment argues this is flawed and fails to correctly compute the nucleolus. This debate underscores the persistent difficulty in finding practical algorithms for this important solution, which has applications in economics, political science, and resource allocation.

The core issue centers on the Davis/Maschler reduced game property (RGP), a mathematical axiom used to characterize solution concepts. The assessment paper contends that Maggiorano et al. incorrectly applied this property, assuming it uniquely identifies the nucleolus when it actually applies to multiple solutions including the core and pre-kernel. This creates a selection problem: their algorithm cannot guarantee it singles out the nucleolus rather than another solution sharing the same property. The paper demonstrates that their might at best verify a pre-computed nucleolus candidate, but cannot compute it from scratch as claimed. This flaw means the proposed algorithm does not achieve its stated goal of providing an efficient computation for convex games.

Ologically, the assessment compares several approaches to nucleolus computation. The ellipsoid by Faigle et al. (2001) offers polynomial-time computation for games where the pre-kernel and core intersect at a single point, but suffers from numerical instability with floating-point arithmetic. Maggiorano et al.'s approach uses submodular function minimization instead of the ellipsoid , avoiding some numerical issues but inheriting the RGP application error. In contrast, the paper highlights a Fenchel-Moreau conjugation-based from Meinhardt (2013) that computes the pre-kernel—which coincides with the nucleolus for certain game classes like convex games—by solving a sequence of linear equations with O(n³) runtime complexity. This avoids nucleolus catchers like the least-core entirely, navigating directly to balanced surplus conditions.

Show significant differences in practical applicability. Maggiorano et al.'s , even if corrected for verification purposes, has an estimated runtime complexity of O(n⁷·log²n·EO + n⁸·logᴼ⁽¹⁾ n), where EO is evaluation oracle time. This makes it impractical for today's double-precision computers, only outperforming exponential-time algorithms beyond 44-51 players. The Fenchel-Moreau , however, typically requires just n+1 iteration steps with O(n³) complexity per step, making it practical for games with around 10 players or more. The paper provides a step-by-step manual computation for a four-person convex game example, showing convergence to the nucleolus in three iterations. This demonstrates 's efficiency and accessibility compared to more cumbersome approaches.

These have important for both theory and practice. The nucleolus is used in real-world scenarios like cost allocation, network design, and political power distribution, where efficient computation is crucial. The failure of the proposed algorithm means researchers and practitioners must still rely on s with limitations or specialized approaches like the Fenchel-Moreau for specific game classes. The assessment also highlights the value of replication , showing that the nucleolus can remain stable under parameter variations in related games, which supports robust bargaining outcomes. This stability is demonstrated through a table of 10 related games that all share the same nucleolus as the original convex game example.

However, limitations remain. The Fenchel-Moreau applies primarily to game classes where the pre-kernel is single-valued and coincides with the nucleolus, such as convex games, almost-convex games, or veto-rich games. For general games, computing the nucleolus is NP-hard, and no polynomial-time algorithm exists without strong assumptions. The assessment notes that even the Fenchel-Moreau relies on an assumption that maximum surpluses can be computed efficiently, though this is less critical in practice. Additionally, the ellipsoid 's numerical instability and Maggiorano et al.'s flawed approach underscore s in developing universally practical algorithms. The paper concludes that while progress has been made, the quest for a robust, general-purpose polynomial-time algorithm for the nucleolus continues, with current s offering trade-offs between efficiency, applicability, and reliability.