[Paper Review] Random matrices and determinantal processes

Kurt Johansson|ArXiv.org|Oct 10, 2005

Random Matrices and Applications40 references227 citations

TL;DR

This paper establishes a deep connection between random matrix theory and determinantal point processes by analyzing exactly solvable models—random domino tilings of the Aztec diamond and the corner growth model—showing that their limiting distributions arise from determinantal processes via non-intersecting path representations. The key contribution is a rigorous mapping of these models to determinantal processes through recursive weight transformations and path non-intersection constraints, proving the weight-preserving and invertible nature of the correspondence via recursive reconstruction of path weights.

ABSTRACT

We survey recent results on determinantal processes, random growth, random tilings and their relation to random matrix theory.

Motivation & Objective

• To establish a rigorous connection between random matrix statistics and determinantal point processes in exactly solvable models.
• To analyze the statistical mechanics of random domino tilings of the Aztec diamond and the corner growth model as examples of such processes.
• To demonstrate that the limiting distributions in these models are naturally described by determinantal point processes.
• To develop a weight-preserving, invertible mapping between lattice path configurations and determinantal structures via recursive reconstruction of path weights.

Proposed method

• Modeling random tilings and growth processes as non-intersecting lattice paths in a two-dimensional grid.
• Defining a recursive weight system using functions $ G^{(k)}(i,j) $ that track path positions and ensure non-intersection through max/min inequalities.
• Introducing a transformation from $ W^{(k)}(\lambda) $ to $ W^{(k-1)}(\lambda) $ via path weight updates, preserving total weight across steps.
• Using the Laplace functional and moment generating functions to characterize point processes via correlation functions.
• Establishing invertibility of the path-to-weight mapping by reconstructing $ G^{(k-1)} $ values from $ G^{(k)} $ and path data along partition boundaries.
• Proving that the weight of $ W^{(k-1)}(\lambda) $ equals the product of the weight of $ W^{(k)}(\lambda) $ and the weight of path $ \pi_k $, ensuring weight preservation.

Experimental results

Research questions

• RQ1 How do random matrix eigenvalue distributions emerge as universal limits in statistical mechanical models like tiling and growth processes?
• RQ2 What is the precise mathematical mechanism that maps non-intersecting lattice paths to determinantal point processes?
• RQ3 How can the weight of a path configuration be preserved and reconstructed recursively across levels in a hierarchical path system?
• RQ4 What conditions ensure the non-intersection of paths in the transformed lattice model, and how are they encoded in the $ G^{(k)} $ functions?
• RQ5 To what extent is the mapping from path configurations to weight matrices invertible, and how is this invertibility proven?

Key findings

• The mapping from non-intersecting path configurations to determinantal point processes is weight-preserving, as shown by the identity $ \text{weight}(W^{(k-1)}(\lambda)) = \text{weight}(W^{(k)}(\lambda)) \cdot \text{weight}(\pi_k) $.
• The non-intersection of paths $ \pi_k $ and $ \pi_{k+1} $ is guaranteed by the inequality $ G^{(k)}(i,j) \geq \max(G^{(k+1)}(i,j+1), G^{(k+1)}(i+1,j)) $, which ensures vertical and horizontal separation.
• The path $ \pi_k $ becomes horizontal for $ k > \min(K,L) $, as $ G^{(k-1)}(i,j) = 0 $ when $ i $ or $ j < k $, leading to constant height paths.
• The reconstruction of $ W^{(k-1)}(\lambda) $ from $ W^{(k)}(\lambda) $ and path $ \pi_k $ is invertible, achieved by recursively computing $ G^{(k-1)}(m-1,n-1) $ from known values at $ (m-1,n) $, $ (m,n-1) $, and $ (m,n) $.
• The entire system is invertible: given the sequence of paths $ (\pi_1, \pi_2, \dots) $, the original weight matrix $ W(\lambda) $ can be reconstructed step-by-step from $ W^{(k)}(\lambda) $ down to $ W^{(0)}(\lambda) $.
• The process is fully characterized by recursive relations: $ w^{(k-1)}(i,j) = G^{(k-1)}(i,j) - \max(G^{(k-1)}(i-1,j), G^{(k-1)}(i,j-1)) - w^{(k)}(i,j) $, with $ G^{(k-1)}(i-1,j-1) $ recoverable from $ w^{(k)}(i,j) $ and the max/min expressions.