[Paper Review] Merging costs for the additive Marcus-Lushnikov process, and Union-Find algorithms

Philippe Chassaing, Régine Marchand|arXiv (Cornell University)|Jun 5, 2004

Stochastic processes and statistical mechanics21 references3 citations

TL;DR

This paper analyzes the asymptotic behavior of merging costs in the additive Marcus–Lushnikov process, where particle clusters of size $x$ and $y$ merge at a rate proportional to $x+y$, with costs depending on cluster sizes and random factors. Using embeddings into Brownian excursion and parking processes, it establishes weak convergence of normalized cost processes to a limiting Brownian excursion, providing precise distributional limits and concentration results for cumulative merging costs across various regimes.

ABSTRACT

Starting with a monodisperse configuration with $n$ size-1 particles, an additive Marcus-Lushnikov process evolves until it reaches its final state (a unique particle with mass $n$). At each of the $n-1$ steps of its evolution, a merging cost is incurred, that depends on the sizes of the two particles involved, and on an independent random factor. This paper deals with the asymptotic behaviour of the cumulated costs up to the $k$th clustering, under various regimes for $(n,k)$, with applications to the study of Union--Find algorithms.

Motivation & Objective

• To understand the asymptotic distribution of cumulative merging costs in the additive Marcus–Lushnikov process, a stochastic coalescence model with kernel $K(x,y) = x+y$.
• To extend Knuth and Schönhage's results on Union-Find algorithms by providing distributional limits and concentration properties for cost functionals beyond just expectations.
• To establish a connection between the additive Marcus–Lushnikov process and the random parking scheme, enabling the use of known limit theorems for parking processes.
• To derive weak convergence of normalized cost processes to a Brownian excursion, offering a refined probabilistic description of merging cost trajectories.

Proposed method

• Embed the additive Marcus–Lushnikov process into a random parking scheme, where each merging step corresponds to a car attempting to park in a uniformly random slot.
• Use the fact that the size of the smaller cluster at each merge step corresponds to the number of failed parking attempts in a related parking process.
• Define normalized cost processes $D_n(eta)$ and $W_n(eta)$, representing partial sums of merging costs up to time $n - etainom{n}{1/2}$, and compare them via $L^2$-norm bounds.
• Leverage known weak convergence results for the normalized parking process to show that the normalized cost process converges weakly to a Brownian excursion.
• Apply the Skorohod representation theorem to upgrade convergence in distribution to almost sure uniform convergence on compact intervals.
• Use moment bounds and coupling arguments to control the $L^2$-distance between the cost process and its Brownian excursion limit, proving tightness and convergence.

Experimental results

Research questions

• RQ1 How do the cumulative merging costs in the additive Marcus–Lushnikov process behave asymptotically as $n \to \infty$?
• RQ2 Can the distribution of the total merging cost be described using a limiting Brownian excursion process?
• RQ3 What is the rate of convergence of the normalized cost process to its limiting distribution?
• RQ4 How does the embedding into the parking process allow for stronger convergence results than previous moment-based analyses?
• RQ5 Can the convergence be strengthened from convergence in distribution to almost sure uniform convergence?

Key findings

• The normalized merging cost process $D_n(eta)$ converges weakly to a Brownian excursion $W(\beta)$ as $n \to \infty$, with convergence in distribution of finite-dimensional distributions.
• Almost sure uniform convergence of $D_n(\beta)$ to $W(\beta)$ holds on $[0, \infty)$, under a suitable coupling, due to the almost sure continuity of the limit process.
• The $L^2$-distance between the normalized cost process and the Brownian excursion limit is $o(1)$, implying tightness and strong convergence.
• For the Quick-Find-Weighted algorithm, the total cost $C^{QFW}_{n,m}$ is asymptotically distributed as the integral of a Brownian excursion, with explicit concentration bounds.
• The results extend beyond expectations to full distributional limits, providing a non-asymptotic approximation for merging cost trajectories.
• The connection to parking functions allows the use of deep limit theorems from stochastic processes, enabling a refined analysis of cost functionals.