[Paper Review] On uniformly generating Latin squares

Masood Aryapoor, E. S. Mahmoodian|arXiv (Cornell University)|May 2, 2010

graph theory and CDMA systems7 references3 citations

TL;DR

This paper presents a concise proof that the Markov chain used by Jacobson and Matthews for uniformly generating Latin squares is connected, using Latin bitrades—specifically intercalates and their algebraic properties. By showing that any two Latin squares can be transformed into each other via a sequence of ±1-moves (based on intercalates), the authors establish the connectivity of the state space, which is essential for the uniform ergodicity of the Markov chain, thus validating the method’s theoretical foundation for nearly uniform sampling of Latin squares.

ABSTRACT

By simulating an ergodic Markov chain whose stationary distribution is uniform over the space of nxn Latin squares, Mark T. Jacobson and Peter Matthews [4], have discussed elegant methods by which they generate Latin squares with a uniform distribution (approximately). The central issue is the construction of "moves" that connect the squares. Most of their lengthy paper is to prove that the associated graph is indeed connected. We give a short proof of this fact by using the concepts of Latin bitrades.

Motivation & Objective

• To provide a concise, alternative proof of the connectivity of the Markov chain used in uniform Latin square generation.
• To demonstrate that the space of Latin squares is connected under ±1-moves based on intercalates, using Latin bitrade theory.
• To validate the theoretical foundation of the Jacobson-Matthews algorithm for nearly uniform sampling of Latin squares.
• To show that proper moves (two- and three-rowed bitrades) suffice to connect all Latin squares, ensuring the Markov chain's irreducibility.

Proposed method

• The authors use the concept of Latin bitrades, particularly intercalates (bitrades of volume 4), to model local transformations between Latin squares.
• They define ±1-moves as the addition of an intercalate to a Latin square, which may result in an improper Latin square (with a -1 coefficient in one cell).
• The proof proceeds by induction on the number of rows, showing that any two Latin squares can be made identical row by row through a sequence of ±1-moves.
• The authors use Lemma 1 to correct improper squares back to proper ones using at most (n−1)/2 moves per row.
• Lemma 3 is used to swap entries in a row without altering the first k−1 rows, enabling incremental alignment of rows between two squares.
• The total number of ±1-moves required to transform any Latin square into another is bounded by 2(n−1)³, establishing a finite diameter for the state graph.

Experimental results

Research questions

• RQ1 Can the connectivity of the Markov chain for Latin square generation be proven more concisely than in Jacobson and Matthews’ original work?
• RQ2 Is the use of improper Latin squares and ±1-moves sufficient to connect all proper Latin squares via a finite sequence of moves?
• RQ3 Can Latin bitrades, especially intercalates, be used to construct a minimal and effective set of moves for generating all Latin squares?
• RQ4 What is the maximum number of moves required to transform any two Latin squares into one another using ±1-moves?

Key findings

• The diameter of the graph formed by all n×n Latin squares under ±1-moves is at most 2(n−1)³.
• Any two Latin squares can be transformed into each other using at most 2(n−1)³ ±1-moves, proving the graph is connected.
• The use of intercalates as building blocks for moves ensures that the Markov chain is irreducible, a necessary condition for uniform stationary distribution.
• The proof shows that proper moves (two- and three-rowed bitrades) are sufficient to connect all Latin squares, confirming the validity of the Jacobson-Matthews algorithm.
• The method provides a significantly shorter proof of connectivity compared to the original 1996 paper by Jacobson and Matthews.
• The result confirms that the Markov chain based on improper tables with one -1-cell is irreducible and thus suitable for nearly uniform sampling of Latin squares.