[Paper Review] Markovianity of the invariant distribution of probabilistic cellular automata on the line

Jérôme Casse, Jean‐François Marckert|arXiv (Cornell University)|Jan 21, 2014

Cellular Automata and Applications29 references7 citations

TL;DR

This paper provides a complete characterization of transition matrices for probabilistic cellular automata (PCA) on the integer line ℤ and the periodic line ℤ/nℤ with a two-neighborhood rule, identifying conditions under which the invariant distribution is Markovian. It extends prior results from the binary case (κ=1) to general finite alphabets (κ≥2), establishing that the invariant measure is Markovian if and only if a certain matrix commutativity condition involving the Perron-Frobenius eigenvectors of derived matrices holds.

ABSTRACT

We revisit the problem of finding the conditions under which synchronous probabilistic cellular automata indexed by the line $\\mathbb{Z}$, or the periodic line $\\cyl{n}$, depending on 2 neighbours, admit as invariant distribution the law of a space-indexed Markov chain. Our advances concerns PCA defined on a finite alphabet, where most of existing results concern size 2 alphabet. A part of the paper is also devoted to the comparison of different structures ($\\mathbb{Z}$, $\\cyl{n}$, and also some structures constituted with two consecutive lines of the space time diagram) with respect to the property to possess a Markovian invariant distribution.

Motivation & Objective

• To determine the conditions under which the invariant distribution of a synchronous probabilistic cellular automaton (PCA) on ℤ or ℤ/nℤ is Markovian with respect to the spatial index.
• To extend existing results—previously limited to the binary case (κ=1)—to PCA with finite alphabets of size κ≥2.
• To compare the structural requirements for Markovian invariance across different underlying graphs: ℤ, ℤ/nℤ, and horizontal zigzag structures in the space-time diagram.
• To provide a complete and explicit criterion for the existence of a Markovian invariant distribution in terms of the transition matrix T and associated Perron-Frobenius eigenvectors.

Proposed method

• Define the PCA as a Markov chain on configurations in E_κ^ℤ, with local transitions depending on two neighbors via a transition matrix T.
• Introduce derived matrices D^η and U^η based on the transition probabilities T, and analyze their spectral properties using Perron-Frobenius theory.
• Establish a necessary and sufficient condition for the existence of a Markovian invariant distribution by requiring that D^γU^γ = U^γD^γ, where γ is the normalized Perron-Frobenius left eigenvector of a derived matrix X.
• Use matrix commutativity and eigenvector normalization to derive a system of equations that uniquely determines the invariant measure as γ, the Perron-Frobenius left eigenvector of X.
• Adapt the proof to the periodic case ℤ/nℤ by verifying that the same matrix commutativity condition, along with positivity and irreducibility assumptions, ensures the existence of a unique invariant Markov measure.
• Verify that the condition D^γU^γ = U^γD^γ is sufficient for κ=1 but not in general for κ≥2, requiring additional analysis of diagonal entries and eigenvector alignment.

Experimental results

Research questions

• RQ1 Under what conditions on the transition matrix T does a PCA on ℤ with a finite alphabet E_κ admit a Markovian invariant distribution?
• RQ2 How does the requirement for a Markovian invariant distribution differ between the infinite line ℤ and the finite periodic line ℤ/nℤ?
• RQ3 What structural properties of the transition matrix T ensure that the invariant measure is a Markov chain in space, and how do these depend on the alphabet size κ?
• RQ4 Can the characterization of Markovian invariance be extended from the binary case (κ=1) to general finite alphabets (κ≥2), and what new mathematical conditions arise?
• RQ5 How do different underlying graphs—such as the space-time diagram’s horizontal zigzag—compare in their ability to support Markovian invariant distributions?

Key findings

• The invariant distribution of a PCA on ℤ or ℤ/nℤ is Markovian if and only if the matrices D^γ and U^γ commute, where γ is the normalized Perron-Frobenius left eigenvector of the matrix X defined by X_{a,b} = T_{a,a|0} ν_a / T_{a,b|0} γ_a.
• For κ=1, the condition D^γU^γ = U^γD^γ reduces to checking diagonal equality, which simplifies the characterization, but for κ≥2, full matrix commutativity is required.
• The unique invariant measure is given by the normalized Perron-Frobenius left eigenvector γ of the matrix X, which is derived from the transition matrix T and the Perron-Frobenius eigenvector ν of another matrix Y.
• The proof establishes that any solution η to the system of equations derived from the detailed balance condition must be equal to γ, implying uniqueness of the invariant Markov measure.
• The condition D^γU^γ = U^γD^γ is necessary and sufficient for the existence of a Markovian invariant distribution in the general case κ≥2, extending prior results limited to κ=1.
• The paper confirms that the structure of the underlying graph (e.g., ℤ, ℤ/nℤ, or zigzag in space-time) significantly affects the feasibility of Markovian invariance, with distinct conditions arising in each case.