[Paper Review] The zeta(2) limit in the random assignment problem

David Aldous|ArXiv.org|Oct 6, 2000

Graph theory and applications4 citations

TL;DR

This paper rigorously proves the long-conjectured limit of the expected minimum assignment cost in the random assignment problem as $\zeta(2) = \pi^2/6$, using a probabilistic analysis of optimal matching on an infinite random tree (the PWIT). It establishes the limiting distribution of edge costs and their rank-orders, and introduces the asymptotic essential uniqueness (AEU) property, showing that almost-optimal matchings deviate from the optimal solution on only a vanishing fraction of edges.

ABSTRACT

The random assignment (or bipartite matching) problem studies the random total cost A_n of the optimal assignment of each of n jobs to each of n machines, where the costs of the n^2 possible job-machine matches has exponential (mean 1) distribution. Mezard - Parisi (1987) used the replica method from statistical physics to argue non-rigorously that EA_n converges to zeta(2) = pi^2/6. Aldous (1992) identified the limit as the optimal solution of a matching problem on an infinite tree. Continuing that approach, we construct the optimal matching on the infinite tree. This yields a rigorous proof of the zeta(2) limit and of the conjectured limit distribution of edge-costs and their rank-orders in the optimal matching.

Motivation & Objective

• To provide a rigorous proof of the Mézard-Parisi conjecture that the expected minimum assignment cost converges to $\zeta(2) = \pi^2/6$ in the random assignment problem.
• To derive the limiting distribution of edge costs in the optimal assignment, confirming a non-rigorous prediction from statistical physics.
• To establish the asymptotic essential uniqueness (AEU) property, showing that almost-optimal matchings coincide with the optimal matching on most edges.
• To validate the conjectured limit distribution of rank-orders of edges in the optimal matching, confirming $q_n(k) \to 2^{-k}$.
• To formalize the connection between the random assignment problem and the probabilistic structure of the PWIT (Poisson Weighted Infinite Tree), enabling exact asymptotic analysis.

Proposed method

• Construct the optimal matching on the Poisson Weighted Infinite Tree (PWIT), a continuous limit object for the random assignment problem.
• Use distributional identities and recursive equations to characterize the limiting distribution of edge costs in the optimal matching.
• Define a matching or almost-matching on a rooted subtree $\mathbf{T}^+$ and derive a fixed-point equation for the random variable $X_\phi$ representing the likelihood of a complete matching.
• Establish the key distributional identity $X \stackrel{d}{=} \left(\sum_{i=1}^\infty e^{-\lambda \xi_i} X_i\right)^{-1}$, which governs the asymptotic behavior of edge inclusion probabilities.
• Prove the AEU property by showing that any matching deviating on a positive fraction of edges incurs a strictly larger expected cost, with a gap bounded below by $\varepsilon(\delta) > 0$.
• Use the PWIT framework to analyze the Gibbs measure and extend the method to the $\lambda$-dependent case, though full rigor for this extension remains open.

Experimental results

Research questions

• RQ1 Does the expected minimum assignment cost in the random assignment problem converge to $\zeta(2) = \pi^2/6$ as $n \to \infty$?
• RQ2 What is the limiting distribution of the cost of the edge incident to a fixed vertex in the optimal matching?
• RQ3 What is the limiting distribution of the rank-order of the matched edge among the $n$ edges incident to a fixed vertex?
• RQ4 Is the optimal matching essentially unique in the asymptotic sense, in that almost-optimal matchings differ on only a vanishing fraction of edges?
• RQ5 Can the replica method's non-rigorous predictions for the random assignment problem be formally justified using probabilistic analysis of the PWIT?

Key findings

• The expected minimum assignment cost converges to $\pi^2/6$, rigorously proving the Mézard-Parisi conjecture.
• The limiting distribution of the cost of the matched edge at a fixed vertex has density $h(x) = \frac{e^{-x}(e^{-x} - 1 + x)}{(1 - e^{-x})^2}$ for $x \geq 0$.
• The probability that the matched edge is the $k$-th smallest among the $n$ edges incident to a vertex converges to $2^{-k}$ as $n \to \infty$.
• The asymptotic essential uniqueness (AEU) property holds: any matching differing on a positive fraction $\delta$ of edges has expected cost exceeding $\pi^2/6 + \varepsilon(\delta)$ for some $\varepsilon(\delta) > 0$.
• The optimal matching on the PWIT is constructed explicitly, providing a rigorous foundation for the infinite-tree limit used in earlier work.
• The analysis confirms that the distribution of edge costs and their rank-orders in the optimal matching are asymptotically independent of the specific distribution of $c(i,j)$, depending only on the density at zero.