Erdos-Szekeres-type statements: Ramsey function and decidability in dimension 1

TL;DR

This paper extends the Erdos-Szekeres lemma to semialgebraic predicates, establishing bounds on Ramsey functions and providing an algorithm to decide if a predicate set is Erdos-Szekeres, enabling automated theorem proving in one dimension.

Contribution

It generalizes the classical Erdos-Szekeres lemma to arbitrary semialgebraic predicates and proves bounds on Ramsey functions, along with an algorithm for automatic verification.

Findings

Ramsey functions are at most doubly exponential in size.

There exists an algorithm to decide if a predicate set is Erdos-Szekeres.

Erdos-Szekeres properties can be proved automatically in one dimension.

Abstract

A classical and widely used lemma of Erdos and Szekeres asserts that for every n there exists N such that every N-term sequence a of real numbers contains an n-term increasing subsequence or an n-term nondecreasing subsequence; quantitatively, the smallest N with this property equals (n-1)^2+1. In the setting of the present paper, we express this lemma by saying that the set of predicates Phi={x_1<x_2,x_1\ge x_2}$ is Erdos-Szekeres with Ramsey function ES_Phi(n)=(n-1)^2+1. In general, we consider an arbitrary finite set Phi={Phi_1,...,Phi_m} of semialgebraic predicates, meaning that each Phi_j=Phi_j(x_1,...,x_k) is a Boolean combination of polynomial equations and inequalities in some number k of real variables. We define Phi to be Erdos-Szekeres if for every n there exists N such that each N-term sequence a of real numbers has an n-term subsequence b such that at least one of the Phi_j holds everywhere on b, which means that Phi_j(b_{i_1},...,b_{i_k}) holds for every choice of indices i_1,i_2,...,i_k, 1<=i_1<i_2<... <i_k<= n. We write ES_Phi(n) for the smallest N with the above property. We prove two main results. First, the Ramsey functions in this setting are at most doubly exponential (and sometimes they are indeed doubly exponential): for every Phi that is Erd\H{o}s--Szekeres, there is a constant C such that ES_Phi(n) < exp(exp(Cn)). Second, there is an algorithm that, given Phi, decides whether it is Erdos-Szekeres; thus, one-dimensional Erdos-Szekeres-style theorems can in principle be proved automatically.